Datasets:

CShorten

/

CORD-19-prototype

like 0

Rift Valley Fever Virus NSs Protein Promotes Post-Transcriptional Downregulation of Protein Kinase PKR and Inhibits eIF2α Phosphorylation

Rift Valley fever virus (RVFV) (genus Phlebovirus, family Bunyaviridae) is a negative-stranded RNA virus with a tripartite genome. RVFV is transmitted by mosquitoes and causes fever and severe hemorrhagic illness among humans, and fever and high rates of abortions in livestock. A nonstructural RVFV NSs protein inhibits the transcription of host mRNAs, including interferon-β mRNA, and is a major virulence factor. The present study explored a novel function of the RVFV NSs protein by testing the replication of RVFV lacking the NSs gene in the presence of actinomycin D (ActD) or α-amanitin, both of which served as a surrogate of the host mRNA synthesis suppression function of the NSs. In the presence of the host-transcriptional inhibitors, the replication of RVFV lacking the NSs protein, but not that carrying NSs, induced double-stranded RNA-dependent protein kinase (PKR)–mediated eukaryotic initiation factor (eIF)2α phosphorylation, leading to the suppression of host and viral protein translation. RVFV NSs promoted post-transcriptional downregulation of PKR early in the course of the infection and suppressed the phosphorylated eIF2α accumulation. These data suggested that a combination of RVFV replication and NSs-induced host transcriptional suppression induces PKR-mediated eIF2α phosphorylation, while the NSs facilitates efficient viral translation by downregulating PKR and inhibiting PKR-mediated eIF2α phosphorylation. Thus, the two distinct functions of the NSs, i.e., the suppression of host transcription, including that of type I interferon mRNAs, and the downregulation of PKR, work together to prevent host innate antiviral functions, allowing efficient replication and survival of RVFV in infected mammalian hosts.

Rift Valley fever virus (RVFV) is a mosquito-borne zoonotic pathogen, which is distributed in sub-Saharan Africa [1] and has also caused large outbreaks in Madagascar [2] , Egypt [3, 4, 5] , Saudi Arabia [6] , and Yemen [6] . In endemic areas, RVFV naturally circulates among mosquitoes and ruminants, such as sheep, goat and cattle [7] . RVFV infection in adult ruminants causes febrile illness and a high rate of abortions, while some newborn animals less than 1-2 weeks of age develop an acute infection which results in higher mortality rates than those in adults [8] . Humans infected with RVFV usually develop an acute febrile myalgic syndrome; however, a small percentage of patients have a lethal illness that results in hepatic damage, hemorrhagic fever-like illness, encephalitis and/or retinal vasculitis [8] . Due to the exotic origin of the virus, the potential for the aerosol transmission [9, 10, 11, 12] and serious consequences for humans and livestocks, wild-type (wt) RVFV is classified as a Risk Group 3 pathogen, that needs to be handled in a high-containment facility, e.g., a biosafety level (BSL)-4 laboratory, whereas a highly attenuated MP-12 strain of RVFV, produced after 12 serial passages of wt RVFV ZH548 in MRC-5 cells in the presence of 5fluorouracil [13] , is a Risk Group 2 pathogen. RVFV, which belongs to the genus Phlebovirus, family Bunyaviridae, is a negative-stranded RNA virus carrying a single-stranded, tripartite RNA genome composed of S, M and L segments [14] . The S segment encodes N and NSs genes and uses an ambi-sense strategy to express the N and NSs proteins in infected cells; N mRNA encoding N protein is transcribed from the viral-sense (negative-sense) S segment, while NSs mRNA encoding NSs protein is transcribed from the antiviral-sense (positive-sense) S segment. Monocistronic M mRNA and L mRNA are transcribed from the viral-sense M and L segments, respectively. M mRNA has one large open-reading frame (ORF) which encodes the nonstructural NSm protein, a 78-kDa glycoprotein and two major viral structural glycoproteins, Gn and Gc [15, 16] . L mRNA encodes L protein, a viral RNA-dependent RNA polymerase. Both N and L proteins are required for viral RNA synthesis, while Gn and Gc function as envelope proteins [14] . NSm and NSs, both nonstructural proteins, are dispensable for viral replication in cell cultures [17, 18, 19, 20] , but are involved in viral pathogenesis [21, 22, 23] . RVFV NSs protein is not essential for virus replication in cell cultures [19, 20] , yet plays a critical role in viral virulence [21] . A naturally occurring RVFV mutant Clone 13, which lacks approximately 70% of the NSs gene [20] , is highly attenuated in mouse, and, when reassortant viruses between wt RVFV and Clone 13 were characterized, the NSs was revealed as a major determinant of viral virulence in the mouse model [21] . The NSs localizes in the nucleus and cytoplasm in both RVFV-infected cells and NSs-expressing cells; further, the nuclear NSs, but not the cytoplasmic NSs, forms a unique filamentous structure [24] . The NSs suppresses the transcription of host mRNAs by interacting with the p44 subunit of TFIIH, an essential transcriptional factor for cellular RNA polymerase II [25] . Furthermore, the RVFV NSs binds to Sin3A-Associated Protein 30 (SAP30), which is important for maintaining the repressor complex containing histone deacetylase 3 on the interferon (IFN)-b promoter, and suppresses the IFN-b promoter activation early in infection [26] . Accordingly, the NSs protein in the nucleus of infected cells most probably exerts these host transcriptional-suppressive activities, including that of IFN production inhibition, and contributes to viral virulence [21] . In contrast, the biological function of cytoplasmic NSs is largely unknown. RVFV NSs expression promotes RVFV minigenome RNA synthesis driven by N and L protein in expression studies [27] . Because RNA synthesis of bunyaviruses occurs in the cytoplasm, NSs protein in cytoplasm may promote the minigenome RNA synthesis by unknown mechanism. To establish that RVFV NSs exhibits a novel function, we hypothesized that, when combined, RVFV replication and NSsinduced host transcription suppression likely induces a cellular environment that is unsuitable for viral replication. Thus, to secure efficient RVFV replication, the NSs protein, in turn, alters this putative, virally unfriendly cellular environment to one that supports efficient viral replication. To test this possibility, we examined the replication of RVFV lacking the NSs gene, a mutant which was generated by employing a reverse genetics system [19] , in the presence of a host transcription inhibitor, e.g., actinomycin D (ActD) or a-amanitin; these drugs were selected because they mimicked the host transcriptional suppressive activities of the NSs. Consistent with our supposition, RVFV lacking the NSs, but not RVFV, failed to replicate efficiently in ActD-treated cells. We noted that double-stranded RNA (dsRNA)-dependent protein kinase (PKR)-mediated eukaryotic initiation factor (eIF)2a phosphorylation suppressed the translation of RVFV lacking NSs in the presence of ActD. Further studies uncovered that RVFV NSs promoted PKR downregulation as early as 4 hours post-infection, and prevented eIF2a phosphorylation, which secured efficient viral translation. We speculate that this novel function of RVFV is important for counteracting the antiviral activities of PKR and allowing efficient virus replication and survival in infected hosts. To explore a novel biological function of RVFV NSs protein, in addition to its host transcriptional shutoff activity, we investigated the effect of ActD, a host transcriptional inhibitor, on the replication of MP-12 lacking the NSs gene in IFN-incompetent VeroE6 cells. VeroE6 cells were mock-infected or infected with MP-12 or rMP12-rLuc, which expressed Renilla luciferase (rLuc) in place of the NSs ( Figure 1A ) at a multiplicity of infection (moi) of 3. After 1 h virus adsorption, cells were incubated in the absence or presence of 5 mg/ml of ActD. Supernatants were harvested at 16 hours post-infection (h.p.i.), and virus titers were measured by plaque assay. ActD treatment had little effect on MP-12 titers, yet it significantly reduced the titer of rMP12-rLuc ( Figure 1B) , which suggested that the RVFV NSs was important for an efficient virus replication in the presence of ActD. To understand how the NSs protein exerted an efficient viral replication in the presence of ActD, we analyzed the status of host and viral translation. VeroE6 cells were mock-infected or independently infected with MP-12, rMP12-rLuc and rMP12-C13type, the latter of which lacks approximately 70% of the NSs ORF ( Figure 1A) , at an moi of 3. Cells were radiolabelled with [ 35 S] Methionine/Cysteine from 15 to 16 h.p.i. Cell extracts were prepared at 16 h.p.i., and the samples were applied to SDS-PAGE ( Figure 1C , top panel). In the absence of ActD, the host protein synthesis of rMP12-rLuc-infected cells ( Figure 1C , lane 5) and rMP12-C13type-infected cells ( Figure 1C , lane 7) was more efficient than that of MP-12-infected cells ( Figure 1C , lane 3). It is possible that NSs-mediated transcriptional suppression that occurred only in MP-12-infected cells, but not in cells infected with viruses lacking NSs, caused a reduction of host mRNA abundance, leading to the less efficient translation of host mRNAs in MP-12-infected cells. Efficient N protein synthesis occurred in cells infected with all three viruses in the absence of ActD, a finding which was consistent with our previous studies [19] , while for some unknown reasons the accumulation of N proteins of rMP12-rLuc and rMP12-C13type was slightly higher than that of MP-12. ActD treatment resulted in a strong inhibition of both host proteins and N protein synthesis in rMP12-rLuc-infected cells ( Figure 1C , lane 6) and rMP12-C13type-infected cells ( Figure 1C , lane 8). In contrast, ActD treatment did not inhibit N protein translation in MP-12-infected cells; rather, it moderately inhibited host protein synthesis in MP-12-infected cells ( Figure 1C, lane 4) . A similar ActD-induced, moderate host protein synthesis inhibi- The mosquito-borne bunyavirus Rift Valley fever virus (RVFV) devastates both humans and domestic animals; it causes abortions in ruminants and complications such as hemorrhage, encephalitis, or retinal vasculitis in humans. A major RVFV virulence factor, NSs, disables host cell mRNA synthesis. Here we describe our new evidence that showed NSs working in a second way; in addition to inhibiting host cell transcription, NSs kept translation active in infected cells. It is well-established that activated protein kinase PKR phosphorylates a translation factor, eIF2a, and then this phosphorylated eIF2a suppresses translation. We found that NSs decreased PKR abundance and prevented eIF2a phosphorylation in infected cells, allowing efficient viral translation and replication. In contrast, when cells were infected with an RVFV mutant lacking NSs in the presence of transcriptional inhibitors that mimic the transcription inhibition function of NSs, the PKR reduction did not occur and phoshorylated eIF2a was accumulated, resulting in the inhibition of virus gene expression and replication. Thus, NSs functions in two ways to help RVFV replicate in mammalian hosts: its newly identified PKR downregulation function secures efficient viral translation, and its host transcription inhibition function suppresses the expression of host innate antiviral functions. tion also occurred in mock-infected cells ( Figure 1C , lane 2). Western blot analysis of cell extracts at 16 h.p.i. clearly showed that ActD treatment strongly inhibited N protein accumulation in rMP12-rLuc-infected cells and rMP12-C13type-infected cells, but not in MP-12-infected cells ( Figure 1C , bottom panels). In summary, ActD treatment had little effect on MP-12 replication, whereas it strongly inhibited the expression of both host and viral proteins in the cells infected with MP-12 lacking the NSs gene, which resulted in poor virus production. To further confirm that the NSs exerted an efficient viral replication in the cells that underwent ActD-induced host transcriptional suppression, 293 cells, which showed higher RNA transfection efficiencies than did VeroE6 cells (data not shown), were infected with rMP12-rLuc at an moi of 2. After virus adsorption, infected cells were mock-transfected or independently transfected with in vitro-synthesized RNA transcripts encoding chloramphenicol acetyltransferase (CAT), MP-12 NSs, or wt RVFV ZH501 NSs. Then, the cells were mock-treated or treated with ActD. Analysis of rLuc activities at 16 h.p.i. demonstrated that NSs expression had little effect on rLuc activities in the absence of ActD (Figure 2A ). In the presence of ActD, rLuc activities were clearly higher in the cells transfected with MP-12 NSs RNA transcripts or ZH501 NSs RNA transcripts than in those transfected with CAT RNA transcripts or in the mocktransfected samples (Figure 2A ). In NSs RNA-transfected cells, similar levels of rMP-12-rLuc titers were observed in both the ActD-treated and mock-treated samples, whereas the rMP12-rLuc virus titers in mock-transfected cells and CAT RNA-transfected cells were significantly lower in the presence of ActD compared to the ActD-untreated cells ( Figure 2B ). Western blot analysis showed that NSs proteins were indeed expressed in the cells transfected with RNA transcripts encoding MP-12 NSs or wt RVFV ZH501 NSs ( Figure 2C , lanes 3, 4, 7, and 8). Also NSs protein expression increased the accumulation of rMP12-rLuc N protein in the presence of ActD ( Figure 2C, lanes 7 and 8) . These data demonstrated that NSs exerted efficient rMP12-rLuc replication in the presence of ActD. We noted that ActD treatment modestly reduced the accumulation of the CAT protein ( Figure 2C, lane 6) . Probably large amounts of CAT RNA transcripts that were introduced into the cells partly overcame the translational suppressive effects that were induced by the combination of rMP-12-rLuc replication and ActD treatment. To test the possibility that ActD treatment alone suppressed translation and RVFV NSs counteracted it, 293 cells were mocktreated or treated with 5 mg/ml of ActD. We examined the resulting polysome profiles at 16 h post-ActD treatment ( Figure S1A ). Because ActD treatment at 5 mg/ml inhibits the transcription mediated by RNA polymerases I, II and III [28] , we expected a reduction in the abundance of cellular mRNAs, tRNAs, and ribosomal RNAs, leading to reduced abundances of polysomes. ActD treatment indeed resulted in a reduced abundance of polysomes, whereas it did not substantially alter the polysome pattern, a finding which suggested to us that ActD treatment did not abolish translational activities. To test the translational competence of the cells treated with transcriptional inhibitors, we transfected 293 cells with in vitro-synthesized RNA transcripts encoding rLuc gene. Cells were mock-treated or treated with ActD or a-amanitin, an RNA polymerase II inhibitor [29] . The rLuc activities at 16 h post-transfection were slightly increased in the cells treated with ActD or a-amanitin ( Figure S1B ), demonstrating active host translation activities in the presence of either ActD or a-amanitin. These data led to the suggestion that by combining the replication of RVFV lacking the NSs and treatment of ActD, a cellular condition that is unfavorable for translation could be induced and that NSs expression somehow altered the cellular environment from one that was translationally inactive to one translationally active. To establish that NSs exerts an efficient viral translation in the presence of a transcription inhibitor, we tested whether coinfection of MP-12 and rMP12-rLuc increases the translation of rLuc mRNA of rMP12-rLuc in the presence of ActD. VeroE6 cells were mock-infected or co-infected with rMP12-rLuc and MP-12; rMP12-rLuc was infected at an moi. of 2, while MP-12 was infected at moi. of 0.1 or 1, as indicated in Figure 3 . The cell extracts were harvested at 16 h.p.i., and the rLuc activities ( Figure 3A ) and accumulation of viral RNAs ( Figure 3B ) were examined. As expected, ActD-treatment reduced the rLuc activities in cells infected with rMP12-rLuc alone ( Figure 3A ). Coinfection of MP-12 resulted in the reduction of both rLuc activities ( Figure 3A ) and the amounts of rLuc mRNA ( Figure 3B , lane 4) in the absence of ActD. In the presence of ActD, MP-12 co-infection also reduced the rLuc mRNA abundance ( Figure 3B , lane 9), whereas it increased rLuc activities ( Figure 3A ). The results suggested that NSs protein expressed from MP-12 S-segment promoted an efficient translation of the rLuc mRNA of rMP12-rLuc in the presence of ActD. We also noted that MP-12 coinfection did not reduce the abundance of the viral-sense rMP-12-rLuc S segment in the presence of ActD, and yet it caused the reduction of rLuc mRNA abundance ( Figure 3B , lane 9), which we believe implies that NSs expression and ActD treatment generated a cellular environment that was more favorable for rMP-12-rLuc RNA replication than for transcription. The dsRNAs generated during viral replication activate PKR, which in turn phosphorylates eIF2a [30] . Also 59-triphosphated single-stranded RNAs activate PKR [31] . The eIF2a is a component of eIF2, which binds to GTP and Met-tRNA to deliver the Met-tRNA to the start codon in capped mRNA, forming a 43S pre-initiation complex [32] . Upon the binding of the 60S ribosomal subunits to the 43S preinitiation complex, eIF2-GDP is released from the ribosome and undergoes a GTP exchange reaction catalyzed by binding with eIF2B, and the resultant eIF2-GTP is recycled for the next round of translation initiation. The phosphorylated eIF2a binds to eIF2B with a high affinity and prevents eIF2B to be used for the subsequent eIF2-GDP-to-eIF2-GTP exchange reaction, leading to the suppression of translation initiation. Hence, the phosphorylation status of eIF2a plays a critical role in translational control [32] . We suspected that rMP12-rLuc replication in the presence of ActD may generate dsRNAs and/or 59-triphosphated singlestranded RNAs, which activate PKR, resulting in the phosphorylation of eIF2a. When we analyzed the level of eIF2a phosphorylation in VeroE6 cells infected with rMP12-rLuc in the absence of transcription inhibitor, we found low levels of eIF2a phosphorylation from 8 to 24 h.p.i. and an efficient accumulation of N protein at 8 h.p.i. and onward ( Figure 4A -4C). In contrast, when we treated rMP12-rLuc-infected cells with ActD or aamanitin, either compound induced an efficient accumulation of phosphorylated eIF2a from 8 to 16 h.p.i., with a concomitant poor N protein accumulation ( Figure 4A -4C) and virus replication suppression ( Figure 4D ). The mechanism of reduction in the abundance of the phosphorylated eIF2a at 24 h.p.i. in the presence of ActD or a-amanitin ( Figure 4 , lanes 13 and 19) is unknown. When we treated rMP12-rLuc-infected cells with different concentrations of ActD or a-amanitin, we found that the inhibition of phosphorylated eIF2a accumulation was dependent on the concentrations used ( Figure S2 ). Also a strong correlation was seen between an increase in eIF2a phosphorylation and the decrease in both N protein accumulation and infectious virus production ( Figure S2 ). In contrast, no significant accumulation of phosphorylated eIF2a occurred in MP-12infected VeroE6 cells in the presence of ActD ( Figure 4A -4C). As expected, ActD treatment had little effect on MP-12 replication ( Figure 4D ). These data strongly suggested that the presence of highly phosphorylated eIF2a levels suppressed the translation of viral mRNAs, leading to inefficient virus production, and that the NSs protein somehow suppressed eIF2a phosphorylation, and facilitated efficient viral mRNA translation under the conditions of host transcriptional shutoff. Because activated caspases 3, 7 and 8 can cleave PKR, releasing the biologically active C-terminus kinase domain from the Nterminus inhibitory domain, resulting in eIF2a phosphorylation [33] , we subsequently tested whether the accumulation of phosphorylated eIF2a following the combined activity of viral replication and ActD-treatment was due to induction of apoptosis [34] . VeroE6 cells, either infected with rMP12-rLuc or mockinfected, received the pan-caspase-inhibitor, benzyloxycarbonyl-Val-Ala-DL-Asp(OMe) fluoromethylketone (Z-VADfmk), in the presence of ActD or a-amanitin ( Figure S3 ). Judging from the resulting inhibition of cleaved caspase-3 accumulation in these cells, the Z-VADfmk treatment indeed inhibited apoptosis, whereas it had little effect on the eIF2a phosphorylation status and infectious virus yield. These data suggested to us that the accumulation of phosphorylated eIF2a in cells supporting rMP12-rLuc replication in the presence of ActD or a-amanitin was caspase-independent. Four different kinases, including PKR, PKR-like ER kinase (PERK), heme-regulated inhibitor and the general control, nondepressible-2 (GCN2), are known to phosphorylate eIF2a [35] . To determine the role of PKR in the accumulation of phosphorylated eIF2a in rMP12-rLuc-infected cells treated with ActD or aamanitin, we generated a recombinant MP-12, rMP12-PKRDE7, which expresses a dominant-negative form of PKR, PKRDE7 [36] carrying the N-terminal Flag epitope tag, in place of the NSs ( Figure 5A ). If the replication of MP-12 lacking the NSs gene in cells subjected to host transcriptional suppression activates PKR, which in turn phosphorylates eIF2a, then the virally-encoded PKRDE7 in rMP12-PKRDE7-infected cells would interfere with the PKR function, resulting in the inhibition of PKR-mediated eIF2a phosphorylation and thereby leading to efficient viral translation and virus production. VeroE6 cells were mock-infected or independently infected with rMP12-rLuc, rMP12-PKRDE7 and MP-12 at an moi of 3. After the removal of the inocula, cells were treated with ActD or a-amanitin, and cell extracts were harvested at 16 h.p.i. As expected, rMP12-rLuc replication in the presence of ActD or a-amanitin induced eIF2a phosphorylation, resulted in reduced virus replication ( Figure 5B -5D). In contrast, efficient N protein accumulation and efficient virus replication, with no significant accumulation of phosphorylated eIF2a occurred in both MP-12-infected cells and rMP12-PKRDE7infected cells in the presence of ActD-or a-amanitin ( Figure 5B -5D). This finding strongly suggested that PKR is important for eIF2a phosphorylation in cells infected with the RVFV lacking the NSs in the presence of transcription inhibitors. To further confirm these data, viral protein accumulation and replication were analyzed in wt mouse embryonic fibroblast (MEF) cells and in Pkr 0/0 MEF cells lacking a functional PKR expression [37] . MP-12 efficiently replicated in both wt MEF and Pkr 0/0 MEF cells, and ActD treatment had little effect on N protein accumulation and virus replication ( Figure 5E , top and middle panels). rMP12-rLuc replication was not as efficient as MP-12 in wt MEF cells in the absence of ActD for an as yet unidentified reason ( Figure 5E , middle panel). In ActD-treated wt MEF cells, both rMP-12-C13type and rMP12-rLuc failed to efficiently accumulate N proteins, and rMP-12-rLuc replicated poorly, whereas rMP-12-rLuc underwent efficient N protein accumulation and viral replication in ActD-treated Pkr 0/0 MEF cells ( Figure 5E , top and middle panels). Furthermore, accumulation of phosphorylated eIF2a did not occur in Pkr 0/0 MEF cells that were infected with rMP12-rLuc in the presence of transcriptional inhibitors To know how the NSs suppressed the eIF2a phosphorylation activity of the PKR function, a dsRNA-binding assay was performed to test the possibility that the NSs binds to dsRNA, sequesters dsRNA from PKR, and interferes with the dsRNAmediated PKR activation. 293 cells were mock-infected or infected with rMP12-NSs-Flag carrying Flag-tagged NSs, rMP12-rLuc-Flag carrying Flag-tagged rLuc ( Figure 6A ) or rMP12-PKRDE7 ( Figure 5A ). In a separate experiment, 293 cells were transfected with in vitro-synthesized RNA transcripts encoding NSs. Lysates were prepared at 16 h.p.i. or 16 h post-transfection, and incubated with poly I:C beads (dsRNA) or poly (C) beads (ssRNA). Then the dsRNA-bound complexes were analyzed by a Western blot in which we used an anti-Flag antibody or anti-NSs antibody ( Figure 6B ). As expected, dsRNA bound to PKRDE7 [36] , whereas it poorly bound to the NSs from rMP12-NSs-Flaginfected cells and that from the NSs-expressing cells ( Figure 6B ), which suggested that NSs did not suppress PKR activation by its binding to dsRNA. Because activated PKR undergoes a structural alteration and autophosphorylation [30] , we tested whether NSs prevented PKR autophosphorylation. 293 cells were infected with rMP12-NSs-Flag, rMP12-rLuc-Flag or rMP12-PKRDE7, and mock-treated or immediately treated with ActD. Cell lysates were harvested at 16 h.p.i., and PKR was immunoprecipitated by anti-human PKR antibody, and the PKR bound to the protein A beads was subjected to an immunoprecipitation (IP)-kinase assay by using [c-32 P]ATP. We used 293 cells for this assay, because anti-human PKR antibody efficiently immunoprecipitated human PKR in 293 cells, but not non-human primate PKR in VeroE6 cells (data not shown). PKR is induced by IFN-a/b treatment [38] and the abundance of PKR could affect the results of the IP-kinase assay. We suspected that replication of MP-12 mutants lacking NSs may induce IFN-b production, leading to PKR induction [38] , whereas ActD treatment prevented the IFN-b production [39] . Indeed, IFN-b mRNA accumulation occurred at 8 h.p.i. in rMP-12-rLucinfected cells in the absence of ActD, but not in the presence of ActD ( Figure 6C ), a finding which suggested that ActD treatment inhibited the transcriptional induction of PKR which was induced by the type I IFN in infected 293 cells. We noted an efficient accumulation of rMP12-rLuc N mRNA at 8 h.p.i. in the presence of ActD ( Figure 6C ). Because rMP-12-rLuc replication in the presence of ActD did not induce an accumulation of phosphorylated eIF2a early in the course of the infection ( Figure 4A ), this efficient rMP12-rLuc replication probably occurred prior to 8 h.p.i. in the presence of ActD. As shown in Figure 6D , immunoprecipitated PKR from rMP12-rLuc-Flag-infected cells was phosphorylated both in the presence and absence of ActD, which led us to suggest that rMP12-rLuc-Flag replication activated PKR. In contrast, PKR phosphorylation did not occur in mockinfected cells or cells infected with rMP12-NSs-Flag or rMP12-PKRDE7 ( Figure 6D ). Most probably, a dominant-negative PKRDE7 suppressed PKR activation and prevented PKR phosphorylation [36] . To determine why we failed to detect the presence of PKR autophosporylation in rMP12-NSs-Flag-infected cells, we examined the amounts of the immunoprecipitated PKR in these samples ( Figure 6D, bottom) . Strikingly, substantial reductions in the abundance of cytoplasmic PKR occurred only in the cells supporting rMP12-NSs-Flag replication both in the presence and absence of ActD. We subsequently tested the possibility that the NSs downregulated PKR expression or sequestered PKR into the nuclear compartment, leading to the suppression of eIF2a phosphorylation. 293 cells were mock-infected or infected with rMP12-NSs-Flag, rMP12-rLuc-Flag or rMP12-PKRDE7, treated with ActD, and cytoplasmic and nuclear fractions of cell extracts were prepared at 16 h.p.i. Western blot analyses showed the presence of RVFV NSs both in the cytoplasmic and nuclear fractions, while rLuc and PKRDE7 signals were observed only in the cytoplasmic fraction. A substantial reduction in PKR abundance occurred in both the cytoplasmic and nuclear fractions of cells infected with rMP-12-NSs-Flag, but not in mock-infected cells and in those infected with rMP12-rLuc-Flag or rMP12-PKRDE7 ( Figure 7A ), demonstrating that NSs induced the PKR downregulation in the infected cells. The PKR downregulation also occurred in MP-12infected VeroE6 cells and in MRC-5 cells that were infected with the wt ZH501 strain of RVFV (data not shown). To determine whether NSs expression alone induces PKR downregulation, 293 cells were mock-infected or infected with rMP12-rLuc, immediately transfected with in vitro-synthesized RNA transcripts encoding rLuc or NSs, and treated with ActD ( Figure 7B, left panel) . Cells were harvested at 16 h posttransfection. The clear reduction in the PKR abundance occurred in mock-infected cells expressing NSs, but not in those expressing rLuc ( Figure 7B, left panel) , demonstrating that NSs protein alone exerted the PKR downregulation. We subsequently examined the requirement of the ActD treatment for the PKR downregulation. 293 cells were transfected with in vitro-synthesized RNA transcripts encoding NSs or rLuc in the absence of ActD. Analysis of cell extracts harvested at 8 h posttransfection showed the PKR downregulation in the NSsexpressing cells ( Figure 7B, right panel) , demonstrating that the NSs-mediated PKR downregulation occurred in the absence of ActD. To further understand the mechanism of the NSs-mediated PKR downregulation, we examined whether NSs expression promoted degradation of PKR mRNA. 293 cells were mocktransfected or transfected with in vitro-synthesized RNA transcripts encoding NSs or rLuc. Then the cells were mock-treated or treated with ActD. Total RNAs were harvested at 8 h posttransfection and the expression levels of PKR mRNA were examined by quantitative real-time reverse transcription polymerase chain reaction (RT-PCR) analysis ( Figure 7C ). The relative expression levels of PKR mRNA were significantly increased in cells that were transfected with rLuc RNA transcripts or NSs RNA transcripts both in the absence and presence of ActD, which demonstrated to us that NSs expression did not promote the degradation of PKR mRNA. Efficient ActD-mediated suppression of IFN-b mRNA accumulation in rMP-12-rLuc-infected 293 cells ( Figure 6C ) led us to suggest that unexpected increases in the abundance of PKR mRNA in the RNA-transfected 293 cells in the presence of ActD were IFN-b-independent. We suspect that the RNA transcripts that were taken up by the cells induced robust PKR mRNA synthesis prior to ActD-or NSs-mediated general transcription suppression. Since NSs expression did not decrease the abundance of PKR mRNA ( Figure 7C ), we next tested whether putative NSsmediated translational inhibition leads to a reduction in the abundance of PKR. To this end, 293 cells were mock-infected, infected with MP-12, or transfected with in vitro-synthesized RNA transcripts encoding NSs. Then cells were incubated with 100 mg/ ml of puromycin to completely shut off cellular translation or puromycin untreated, and cell extracts were harvested at 16 h.p.i. or post-transfection. As expected, puromycin treatment completely abolished the synthesis of N and NSs proteins in MP-12-infected cells ( Figure 7D, lane 4) and that of NSs in cells transfected with the RNA transcripts encoding NSs ( Figure 7D, lane 6) . We found that treatment of 293 cells with puromycin for 16 h decreased the abundance of PKR only slightly ( Figure 7D, lane 2) . Accordingly, it is highly unlikely that putative NSs-induced translational We next performed pulse-chase experiments to know whether NSs expression promoted PKR degradation. Because immunoprecipitation experiments using various anti-PKR antibodies failed to convincingly demonstrate a radiolabelled endogenous PKR signal from extracts of 293 cells (data not shown), we examined the effect of NSs expression on the stability of an expressed mutant PKR, PKRK296R, lacking kinase activity [40] and carrying a Nterminal myc tag; expression of wild-type PKR was not used due to its strong host translation suppression effects (data not shown). 293 cells were mock-transfected or transfected with pcDNA3.1-Myc-PKRK296R, a plasmid encoding PKRK296R under cytomegalovirus promoter and radiolabelled with [ 35 S] methionine/cycteine between 14 and 16 h post-DNA transfection. After pulse-radiolabelling cell extracts were prepared from some samples. In other samples, cells were transfected with in vitrosynthesized RNA transcripts encoding rLuc or NSs, and cell extracts were harvested at 8 h post-RNA transfection. Then the cell extracts were subjected to radioimmunoprecipitation analysis using anti-myc antibody, which immunoprecipitated the pulseradiolabelled, myc-tagged PKR ( Figure 7E , lane 2). After 8 h chase, the amount of the radiolabelled myc-tagged PKR was clearly reduced in the cells expressing NSs ( Figure 7E, lane 4) , but not those expressing rLuc ( Figure 7E, lane 3) . Instead, the amount of radiolabelled myc-tagged PKR was increased slightly in rLucexpressing cells, presumably due to radiolabelling of continuously synthesized myc-tagged PKR by residual [ 35 S] methionine/ cysteine. The complete shutoff of cellular translation by puromycin for 16 h did not result in the loss of endogenous PKR ( Figure 7D ), whereas as early as 8 h post-transfection of RNA transcripts encoding NSs, the abundance of PKR decreased substantially ( Figure 7B ) without reducing the abundance of PKR mRNA ( Figure 7C ). Furthermore, NSs expression reduced the abundance of myc-tagged PKR that had been radiolabelled prior to NSs expression ( Figure 7E ). These data strongly suggested that NSs induced the downregulation of PKR at a post-transcriptional level and pointed to a possibility that the NSs promoted PKR degradation. We next tested whether RVFV NSs downregulated PKR by promoting PKR degradation through a ubiquitin-proteasome pathway. 293 cells were infected with rMP12-rLuc or MP-12, and were immediately treated with proteasome inhibitor MG132 or lactacystin. Cells were harvested at 8 h.p.i. and analyzed in Western blotting. As expected, rMP12-rLuc replication did not induce PKR downregulation ( Figure 7F, lane 1) , while MP-12 replication induced the reduction of PKR abundance ( Figure 7F, lanes 4) . It was evident that the treatment of MP-12-infected cells with those proteasome inhibitors suppressed NSs-induced PKR downregulation ( Figure 7F, lanes 5 and 6) , suggesting that NSs promoted PKR downregulation by the degradation through the proteasome pathway. Phorbol 12-myristate 13-acetate (PMA), a potent activator of protein kinase C (PKC), induces PKR degradation [41] . Because the general PKC inhibitor, GÖ 6983, suppresses PMA-mediated PKR degradation [41] , we examined whether the NSs promoted PKR downregulation through PKC by treating MP-12-infected cells with GÖ 6983. Treatment of GÖ 6983 did not inhibit the NSs-mediated PKR downregulation ( Figure 7F, lane 8) , suggesting that PKC was not involved in NSs-induced PKR downregulation. To know how quickly PKR downregulation occurred in MP-12infected cells, whole-cell extracts were prepared at 2, 4, 6 and 8 h.p.i. from MP-12-infected 293 cells and the amounts of NSs and PKR were determined ( Figure 7G ). In the absence of proteasome inhibitor, a substantial reduction of PKR abundance occurred as early as 4 h.p.i., where the abundance of NSs ( Figure 7G , lane 2) was not as great as that at 8 h.p.i. (Figure 7G , lane 4). In the presence of MG132, the abundance of PKR decreased only slightly, and NSs accumulation was also somewhat less efficient. A similar trend for less efficient NSs accumulation in the presence of proteasome inhibitors was also shown in the data for Figure 7F . Our observation of less efficient NSs accumulation in MG132-treated cells was consistent with the report that MG132 induces eIF2a phosphorylation through GCN2 activation and translational suppression [42] . The abundance of NSs in MG132treated cells at 8 h.p.i. and that in untreated cells at 4 h.p.i. was similar, and yet PKR abundance in the former cells was clearly higher than that in the latter cells. Taken together, these data strongly suggested that RVFV NSs promoted PKR degradation through the proteasome-dependent pathway. The present study explored a novel function of the RVFV NSs protein by testing the replication of RVFV lacking the NSs gene initially in type I IFN-incompetent VeroE6 cells [43, 44] in the presence of ActD or a-amanitin, which served as a surrogate of the host mRNA synthesis suppression function of the NSs. The NSs protein was essential for efficient virus replication in the presence of ActD or a-amanitin. We found that the replication of RVFV lacking the NSs gene in the presence of a transcription inhibitor induced an accumulation of phosphorylated eIF2a, The accumulation of phosphorylated eIF2a in the presence of transcriptional inhibitors, did not occur in VeroE6 cells that were infected with RVFV expressing a dominant-negative form of PKR (PKRDE7) and in MEF cells lacking functional PKR ( Figure 5 ). These findings suggested to us that PKR played a major role in increasing phosphorylated eIF2a; however, it is currently unclear whether other kinases, such as PERK or GCN2, have a possible role in eIF2a phosphorylation in rMP-12-rLuc-infected cells in the presence of transcriptional inhibitors. The endogenous PKRmediated eIF2a phosphorylation suppressed translation of the RVFV lacking the NSs gene, resulting in poor virus replication. Our data further suggested that the NSs promoted PKR degradation most probably through a proteasome-dependent pathway and prevented eIF2a phosphorylation, leading to efficient viral translation. (8 h) . Radiolabelled myc-tagged PKR was immunoprecipitated by anti-myc monoclonal antibody, analyzed by SDS-PAGE with 7.5% poryacrylamide gel, and visualized by autoradiography. (F) 293 cells were infected with rMP12-rLuc or MP-12 at an moi of 3, and, then, treated with 10 mM of MG132 (MG) or 50 mM of lactacystin (LA) or they were mock-treated (2) . Whole-cell lysates were collected at 8 h.p.i., and the abundance of PKR, NSs and b-actin were examined by Western blot analysis. (G) 293 cells were mock-infected (Mock) or infected with MP-12 (MP-12) at an moi of 3, and treated with MG132 (MG132) at 10 mM or they were untreated (no drug used). Whole-cell lysates were collected at 2, 4, 6, and 8 h.p.i. Anti-PKR antibody, anti-NSs antibody, and anti-b-actin antibody were used to detect PKR, NSs, and b-actin, respectively (F and G). Data are representative of two to three independent experiments (A-G). doi:10.1371/journal.ppat.1000287.g007 The past studies [25, 26, 45] and the data shown in this study illustrate that two distinct biological activities of the RVFV NSs protein worked together to secure efficient RVFV replication. Namely, nuclear NSs protein inhibits transcription of host mRNAs, including the IFN mRNAs; this activity is critical for efficient RVFV replication in IFN-competent systems [25, 45] . However, a combination of RVFV replication and NSs-mediated host mRNA transcriptional suppression possibly induces PKR activation and subsequent eIF2a phosphorylation as a combination of RVFV replication and ActD or a-amanitin treatment induced it. The NSs protein, in turn, promotes PKR downregulation as early as 4 h.p.i., and prevents eIF2a phosphorylation to secure the translation of viral mRNAs and efficient virus replication. We think that both of these two NSs functions are tightly related and protect efficient viral replication by suppressing host antiviral responses. In RVFV-infected cells, the NSs establishes general transcriptional suppression at a later stage of infection (after 8 h.p.i.) [25] , while NSs also suppresses specific IFN-b mRNA transcription at early stages of infection (about 3 h.p.i.) by maintaining repressor complex including SAP30 on IFN-b promoter [26, 45] . PKR downregulation early in infection is probably important for maintaining efficient viral translation in combination with the suppression of host IFN responses. We suspect that the NSs-mediated PKR downregulation activity is important for RVFV replication and survival in infected mammalian hosts. RVFV-infection in rhesus monkeys showed that type I IFN is detectable around 1 day post-infection in both clinically ill surviving monkeys and lethally infected monkeys, and one dead monkey even kept high titer of IFN (120 to 480 U/ml) from 1 day post-infection [46] . In fact, the best correlation with outcome was early detection of IFN and not necessarily the height of the response. NSs suppresses IFN-b mRNA transcription early in the course of infection in cultured cells [26, 45] . Our present data suggest that the early downregulation of PKR in RVFVinfected cells might contribute to the inhibition of type I IFN induction, because PKR is known to serve as a pathogenrecognition receptor [47] . Furthermore, some of the host pathogen-recognition receptors, e.g., toll-like receptors 3 and 7, in uninfected cells located near the RVFV-infected cells, recognize virus-specific signals, e.g., viral RNAs in the virus particles or viral dsRNAs in the cell debris from infected cells, leading to type I IFN production. Because the transcriptional promoter of the PKR gene contains an IFN-stimulated response element, and IFN stimulation induces PKR mRNA transcription [38] , PKR abundance is likely increased in many uninfected cells of infected animals; RVFV needs to replicate in such cells to survive in infected hosts. The reduction in PKR abundance occurred as early as 4 h.p.i. ( Figure 7G ) and the NSs protein is produced very early in RVFV infection [48] . Immediate NSs synthesis and subsequent NSsinduced PKR downregulation in the cells, some of which have increased PKR abundance, could rapidly disarm the PKRmediated, antiviral functions and contribute to RVFV replication and survival in animal hosts. Possible mechanisms of the accumulation of phosphorylated eIF2a in ActD-treated, rMP12-rLucinfected cells ActD treatment of uninfected cells resulted in a significant reduction of the polysome fraction, but it did not abolish the translational activities ( Figure S1 ). rMP12-rLuc replication in the cells treated with ActD or a-amanitin promoted the accumulation of the phosphorylated eIF2a, resulting in translational suppression of viral proteins (Figure 4) . The increased amounts of phosphorylated eIF2a clearly correlated with the concentration of ActD or a-amanitin in rMP12-rLuc-infected cells ( Figure S2 ). In contrast, rMP12-rLuc replication in the absence of transcriptional suppressor induced only low levels of eIF2a phosphorylation (Figure 4 ). Yet, the phosphorylation of PKR occurred in cells supporting the replication of RVFV lacking NSs both in the absence and presence of ActD ( Figure 6D ). Thus, a significant accumulation of phosphorylated eIF2a occurred only in cells in which replication of RVFV lacking NSs was combined with the treatment with transcriptional inhibitors. Several different mechanisms are conceivable for the accumulation of phosphorylated eIF2a in rMP12-rLuc-infected in cells treated with ActD or a-amanitin (Figure 4 ). One possible mechanism relates to the eIF2a dephosphorylation step. Phosphorylation of eIF2a at Serine 51 induces a rapid synthesis of activating transcription factor (ATF)-4 mRNA, which can be translated in the presence of phosphorylated eIF2a [49] . Expressed ATF4 then induces GADD34 mRNA transcription [50] and GADD34 protein interacts with type 1 protein serine/ threonine phosphatase, PP1, and this complex dephosphorylates eIF2a to resume the cellular translation [51] . rMP12-rLuc replication in transcriptionally active cells probably induced PKR activation and eIF2a phosphorylation, the latter of which then induced GADD34 upregulation and subsequent eIF2a dephosphorylation allowing efficient viral translation. rMP12-rLuc replication in cells treated with ActD induced PKR activation, the extent of which was similar to that of infected, ActD-untreated cells ( Figure 6D ), whereas the transcription inhibitors would prevent GADD34 upregulation and subsequent eIF2a dephosphorylation, causing an accumulation of phosphorylated eIF2a, which inhibited viral translation. Other possible mechanisms relate to the failure to suppress the PKR function. Cells infected with adenovirus, human immunodeficiency virus-1, or herpes simplex virus undergo a dramatic increase in the abundance of Alu RNA, which carries a repetitive element of ,300-nt in length that is transcribed by RNA polymerase III [52, 53, 54, 55] . Alu RNA forms a stable complex with PKR, and antagonizes the PKR activation [56] . If rMP12-rLuc replication induces Alu RNA accumulation to prevent PKR activation, then the drug-induced transcriptional suppression would inhibit Alu RNA accumulation, preventing Alu RNA-mediated PKR inactivation. Another possibility is that rMP12-rLuc replication transcriptionally induces host mRNAs, and some of their gene products have PKR inhibition activities. An example of this possibility is that influenza virus replication induces P58 IPK , an inhibitor of PKR [57] . ActD or a-amanitin treatment suppresses the expression of the putative PKR inhibitor, resulting in an accumulation of phosphorylated eIF2a. Alternatively, transcriptional suppression may result in reduced amounts of ribosomal proteins, such as L18, which binds to PKR and inhibits PKR activation [58, 59] . These possibilities are not mutually exclusive, and the combination of these possibilities may contribute to the accumulation of phosphorylated eIF2a in the cells supporting replication of RVFV lacking the NSs in the presence of ActD or aamanitin. Although many viruses suppress PKR function using various strategies [32] , poliovirus is known to promote PKR degradation in infected cells [60, 61] ; poliovirus RNA and viral protein components are required for this activity [61] , and both eIF2a and PKR are highly activated in poliovirus-infected cells before PKR downregulation occurs [60] . In contrast to poliovirus, only RVFV NSs protein was probably required for PKR downregu-lation since the downregulation of PKR occurred as early as 4 h.p.i. in RVFV-infected cells ( Figure 7G ), whereas phosphorylated eIF2a accumulation occurred around 8 h.p.i. in rMP-12-rLuc-infected cells in the presence of ActD (Figure 4 ). Furthermore, NSs downregulated PKRK296R, a non-phosphorylatable mutant PKR ( Figure 7E ). These data suggest that the phosphorylation of PKR was not essential for the NSs-mediated PKR downregulation. A substantial reduction in the abundance of PKR occurred in both the cytoplasmic and nuclear fractions of MP-12-infected cells, but not in the mock-infected cells and in those infected with MP12 lacking NSs (Figure 7) . Expression of NSs alone resulted in a reduction in the amount of PKR ( Figure 7B ). These data established that NSs protein induced PKR downregulation. The NSs-induced PKR downregulation occurred as early as 4 h.p.i. of MP-12 ( Figure 7G ). We also demonstrated the presence of PKR mRNA in NSs-expressing cells ( Figure 7C ) and that ActD treatment had little effect on the NSs-induced PKR downregulation ( Figure 7B ). We failed to detect PKR at 16 h.p.i. of MP-12infected 293 cells, and yet we easily detected the PKR after a 16 hlong incubation of 293 cells with puromycin ( Figure 7D ). Accordingly, the putative NSs-induced translational inhibition was not a main reason for the reduction in the abundance of PKR. Furthermore, pulse-chase experiments showed that NSs expression reduced the abundance of radiolabelled myc-tagged PKR ( Figure 7E ). All of these data strongly suggested that the NSs induced PKR downregulation at post-transcriptional levels. It was reported that treatment of a macrophage cell line with IFN-c induced PKR degradation [62] . Because NSs induced PKR degradation in ActD-treated cells and ActD treatment completely inhibited the IFN-c mRNA accumulation ( Figure 6C ), it is highly unlikely that IFN-c was involved in the NSs-induced PKR downregulation. PKC activation also potentially induces PKR downregulation [41] . Treatment of MP-12-infected cells with a PKC inhibitor, GÖ 6983, did not inhibit NSs-mediated PKR downregulation ( Figure 7D ), which implied that PKC was not involved in it. Experiments using a proteasome inhibitor suggested that the NSs promoted PKR degradation through a proteasome pathway ( Figure 7F ). MG132 can induce GCN activation and translational suppression [42] . In fact, treatment of MP-12infected cells with MG132 and lactacystin moderately reduced the accumulation of NSs ( Figure 7F and 7G) . Accordingly, a possibility still exists that the MG132-induced moderate translational inhibition affected the NSs-induced PKR degradation in MG132-treated cells. For example, if the NSs-induced PKR downregulation requires an unstable host protein, then the MG132-induced moderate translational suppression would prevent accumulation of this putative host protein, resulting in inhibition of the NSs-induced PKR downregulation. Obviously further studies are required to elucidate the mechanism of PKR downregulation mediated by RVFV NSs. Vero E6 cells, wild-type mouse embryonic fibroblast (MEF) cells and Pkr 0/0 MEF cells [37] were maintained in Dulbecco's modified minimum essential medium (DMEM) (Invitrogen) containing 10% fetal bovine serum. BHK/T7-9 cells [63] , which stably express T7 RNA polymerase, were grown in MEM-alpha (Invitrogen) containing 10% fetal bovine serum (FBS). Penicillin (100 U/ml) and streptomycin (100 mg/ml) (Invitrogen) were added to the media. BHK/T7-9 cells were selected in medium containing 600 mg/ml hygromycin (Cellgro). An RVFV vaccine candidate MP-12 [13] and recombinant MP-12 [19] were used for the experiments, and infectivity was assessed by a plaque assay in Vero E6 cells. Cells were treated with transcriptional inhibitors, ActD (Sigma) (5 mg/ml) or a-amanitin (Sigma) (50 mg/ml) immediately after infection or transfection. To induce the inhibition of proteasome function, cells were immediately treated with MG132 (Sigma) at 10 mM or lactacystin (Sigma) at 50 mM after infection or transfection. To suppress PKC activity, cells were treated with a general PKC inhibitor, GÖ 6983 (Calbiochem) at 100 nM immediately after infection with rMP12-rLuc or MP-12. Cells were treated with puromycin (Cellgro) at 100 mg/ml immediately after infection or transfection to inhibit cellular translation. Standard molecular biological techniques, including a PCRbased mutagenesis method, were used for plasmid constructions. PCR fragments encoding PKRDE7 ORF with an N-terminal Flag-tag sequence were cloned between the Hpa I and Spe I sites of pProT7-S(+) plasmid [19] , designated as pProT7-S(+)-PKRDE7. PCR fragments encoding the rLuc ORF or NSs with a C-terminal Flag-tag sequence were cloned between the Hpa I and Spe I sites of the pProT7-S(+) plasmid, designated as pProT7-S(+)-rLuc-Flag or pProT7-S(+)-NSs-Flag, respectively. All of the constructs were confirmed to have the expected sequences. The PCR product of the entire human PKR ORF carrying a point mutation at K296R and the N-terminal myc tag was cloned between KpnI and XhoI of pcDNA3.1myc-His (Invitrogen), resulted in pcDNA3.1-Myc-PKR296R. A recombinant MP-12 carrying the PKRDE7 ORF, rLuc-Flag or the NSs-Flag in the place of the NSs ORF was recovered as described previously [19] . Briefly, subconfluent monolayers of BHK/T7-9 cells were co-transfected with an S-genome RNA expression plasmid, such as pProT7-S(+)-PKRDE7, pProT7-S(+)-rLuc-Flag or pProT7-S(+)-NSs-Flag, and a mixture of pPro-T7-M(+), pPro-T7-L(+), pT7-IRES-vN, pCAGGS-vG, and pT7-IRES-vL using TransIT-LT1 (Mirus Bio Corporation). The culture medium was replaced with fresh medium 24 h later. At 5 days post-transfection, the culture supernatants were collected, clarified and then inoculated into VeroE6 cells. The supernatant of infected VeroE6 cells at 3 days post-infection was used for the experiment. RVFV MP-12 or ZH501 NSs ORF, or CAT ORF were cloned downstream of the T7 promoter between the Kpn I and Xho I sites of the pcDNA3.1-myc-HisA (Invitrogen) plasmid. For rLuc RNA transcripts, the rLuc ORF in pRL-SV40 plasmid (Promega Corporation) was inserted downstream of the T7 promoter. Capped and polyadenylated RNA transcripts were synthesized in vitro by using mMESSAGE mMACHINE T7 Ultra (Ambion) according to the manufacturer's instructions [64] . One microgram of in vitrosynthesized RNA transcripts was transfected into 293 cells in a 12well plate with a TransIT-mRNA Transfection kit (Mirus Bio Corporation) according to the manufacturer's instructions. VeroE6 cells in 6-well plate were infected with a series of diluted virus samples in 400 ml. After 1 h adsorption at 37uC, we removed the inocula and added 2 ml of MEM containing 0.6% noble agar (Difco Laboratories), 5% FBS and 5% tryptose phosphate broth. The cells were incubated at 37uC for 3 days. Then, 2 ml of MEM containing 0.6% agar, 100 mg of neutral red (N2889, Sigma), 5% FBS and 5% tryptose phosphate broth were added into the wells, and incubated for 16 h. The virus titers were determined in triplicate. Cells were lysed in sample buffer and boiled for 10 min. Equal amounts of samples were subjected to sodium dodecyl sulfatepolyacrylamide gel electrophoresis (SDS-PAGE). Proteins were electroblotted onto polyvinylidene difluoride membranes (immobilon P; Millipore). Western blot was performed as described previously [27] . The following primary antibodies were used: anti-RVFV [27] ; anti-NSs [48] The IP-kinase assay was performed as described previously [65] . Briefly, 293 cells were mock-infected or infected with recombinant MP-12 at an moi of 3. Cells were dissolved in lysis buffer containing 10 mM Tris-HCl pH 7.6, 50 mM KCl, 2 mM Magnesium acetate, 10 mM 2-mercaptoethanol, 1% Triton X-100, 1 mM EDTA, phosphatase inhibitor cocktail (Sigma) and proteasome inhibitor cocktail (Roche). After centrifugation at 10,0006g for 5 min, a cytoplasmic fraction was collected. Cytoplasmic lysates were subjected to immunoprecipitation with anti-PKR antibody (Santa Cruz Biotech, K-17). Protein A beads that bound to the immunoprecipitated PKR were washed twice with a buffer-A containing 20 mM Tris-HCl, pH 7.6; 50 mM KCl; 400 mM NaCl; 5 mM 2-mercaptoethanol; 1% Triton X-100; 1 mM EDTA; phosphatase inhibitor cocktail (Sigma); proteasome inhibitor cocktail (Roche); and 20% glycerol. The beads were further washed twice with buffer-B containing 20 mM Tris-HCl, pH 7.6; 100 mM KCl; 5 mM 2-mercaptoethanol; 1% Triton X-100; 0.1 mM EDTA; proteasome inhibitor cocktail (Roche); and 20% glycerol. Then, washed beads were resuspended in 26 kinase buffer containing 30 mM Hepes-KOH, pH 7.4; 2 mM dithiothreitol; 2 mM MgCl 2 ; proteasome inhibitor cocktail (Roche); and 10 mCi of [c-32 P] ATP (MP Biomedicals). The suspension (20 ml) was incubated at 30uC for 20 min, and then 26 SDS sample buffer was added to terminate the reaction. The samples were separated on 10% SDS-PAGE gel and visualized on an autoradiograph. A portion of the samples were also used for Western blot analysis by employing anti-PKR monoclonal antibody (BD biosciences) to show the abundance of immunoprecipitated PKR. For the radiolabelling of host and viral proteins in infected cells, VeroE6 cells were incubated at 14.5 h post-infection for 30 min at 37uC with medium made up with MEM lacking methionine, cystine, and L-glutamine (M2289, Sigma); 1% dialyzed FBS (Invitrogen); 20 mM L-glutamine; penicillin (100 U/ml) and streptomycin (100 mg/ml). Then, Trans[ 35 S]label metabolic reagent (MP biomedicals) was directly added into the medium (100 mCi/ml). After 1 h labelling, cells were washed with PBS once and lysed in sample buffer. Equal amounts of samples were subjected to SDS-PAGE in 10% polyacrylamide gel. The gel was dried and exposed to X-ray film (KODAK BioMax XAR) overnight at 280uC. For the radiolabelling of PKR, 293 cells were mock-treated or transfected with pcDNA3.1-Myc-PKRK296R, and labelled with Trans[ 35 S]label metabolic reagent between 14 and 16 h post-DNA transfection. In some samples, cell extracts were prepared with chase using a lysis buffer. In other samples, cells were transfected with in vitro-synthesized RNA transcripts encoding rLuc or MP-12 NSs. Then, cells were washed, and incubated for 8 h in the presence of 2 mM methionine/ cysteine. At 8 h post-RNA transfection, cell extracts prepared using lysis buffer, were employed for immunoprecipitation with anti-Myc antibody (Santa Cruz: sc-40), as described in an IPkinase assay. Immunoprecipitated samples were separated on a 7.5% poryacrylamide gel and visualized on an autoradiograph. dsRNA-binding assay 293 cells were mock-infected or infected with rMP12-rLuc-Flag, rMP12-NSs-Flag or rMP12-PKRDE7 at an moi of 3. Alternatively, 293 cells were transfected with in vitro-synthesized RNA transcripts encoding MP-12 NSs. The cytoplasmic lysate was harvested at 16 h.p.i. or 16 h post-transfection, and incubated with poly C beads or poly I:C beads on ice for 45 minutes. After washing beads with buffer-A for 4 times, the beads were mixed with 26 sample buffer, and bound proteins were analyzed with anti-Flag antibody (Sigma) on a Western blot. The luciferase assay was performed on a Renilla Luciferase Assay System (E2810, Promega Corporation) according to the manufacturer's instructions. Total RNA was harvested by Trizol reagent (Invitrogen), and Northern blot was performed as described previously [27, 64] . Briefly, total RNA was denatured and separated on 1.2% denaturing agarose-formaldehyde gels and transferred onto a nylon membrane (Nylon Membrane, positively charged, Roche). Northern blot analysis was performed by using strand-specific RNA probes for detecting IFN-b mRNA [64] , GAPDH mRNA [64] , rLuc mRNA [27] or RVFV antiviral-sense S-segment / N-mRNA [48] . 293 cells in 6-well plates were transfected with 2 mg of in vitrosynthesized RNA transcripts encoding rLuc or MP-12 NSs by TransIT mRNA transfection kit (Mirus). Cells were mock-treated or treated with ActD (5 mg/ml) immediately after RNA transfection. Total RNA were extracted by using an RNeasy Mini kit (Qiagen) at 16 h post-transfection. For each sample, we used 500 ng of RNA to synthesize 1 st strand cDNA by High-Capacity cDNA Reverse Transcription Kits (Applied Biosystems). Real-Time PCR was performed at the Real-Time PCR core facility, Sealy Center for Cancer Cell Biology, UTMB. We used an Applied Biosystems made-to-order 206 assay mix of primers and TaqMan MGB probes (FAM dye-labled) for our target gene PKR (Applied Biosystems: assay ID#: Hs01091592_m1) and predeveloped an 18S rRNA (FAM-dye labelled probe) TaqMan assay reagent (Applied Biosystems: 4352930E) for endogenous control. Separate tubes (singleplex) real-time PCR was performed with 40 ng cDNA for both target gene and endogenous control by using a Taqman Gene expression master mix (Applied Biosystems: 4370074). The cycling parameters for real-time PCR were: UNG activation at 50uC for 2 min, AmpliTaq activation at 95uC for 10 min, denaturation at 95uC for 15 sec, and annealing/extension at 60uC for 1 min (repeat 40 times) on ABI7000. Duplicate C T values were analyzed by the comparative C T (DD C T ) method, as described by the manufacturer (Applied Biosystems). The amount of target (2 2DDCT ) was obtained by normalized to endogenous reference (18S rRNA) and relative to a calibrator (one of the experimental samples). Figure S1 Effects of ActD treatment on host translation activities. (A) 293 cells cultured in 10-cm dishes were treated with 5 mg/ml of ActD (ActD) or were left untreated (Mock). Cells were harvested at 16 h post-ActD-treatment in 900 ml of Lysis buffer (50 mM Tris-HCl, pH 7.5; 5 mM MgCl2; 100 mM KCl; 1% Triton X-100; 100 mg/ml cycloheximide; and 0.5 mg/ml heparin) on ice for 5 min. The cytoplasmic lysates were collected after the removal of nucleus by centrifugation at 10,0006g for 5 min. The lysates were loaded onto a 10-50% linear sucrose gradient containing 50 mM Tris-HCl, pH 7.5; 5 mM MgCl2; 100 mM KCl; 0.5 mM dithiothreitol; 100 mg/ml cycloheximide; and 0.5 mg/ml heparin, and centrifuged at 38,000 rpm for 3 h at 4uC using a Beckman SW41 rotor. The gradients were pumped by syringe pump (Brandel) and analyzed by a density gradient fractionator (Brandel) connected to an ISCO UA-6 (ISCO Inc.) at the absorbance of 254 nm according to the manufacturer's instructions. The data were representative of two independent experiments. (B) 293 cells were transfected with in vitrosynthesized rLuc RNA transcripts and mock-treated (no drug) or immediately treated with 5 mg/ml of ActD (ActD) or 50 mg/ml of a-amanitin (Amanitin). Luciferase activities were measured at 16 h post-transfection. The data shown in the graphs (mean+/ 2standard deviation) were obtained from three independent experiments with p values by using Student's t-test (*: p,0.05). Found at: doi:10.1371/journal.ppat.1000287.s001 (4.02 MB TIF) Figure S2 Effects of different concentrations of ActD or aamanitin on eIF2a phosphorylation and N protein accumulation in rMP-12-infected cells. VeroE6 cells were infected with rMP12-rLuc at an moi of 3, and then treated with different concentrations of ActD or a-amanitin. Cell extracts and culture supernatants were harvested at 16 h.p.i. Note that ActD suppresses about 80% of RNA polymerase I activity at 0.04 mg/ml, and about 80% of RNA polymerase II and III at 4.0 mg/ml [28] , while a-amanitin suppresses nearly 100% of RNA polymerase II and about 50% of RNA polymerase III at 50 mg/ml [64] . (A) Western blot analysis of N protein, phosphorylated eIF2a, total eIF2a and a-actin in each cell extract. (B) Top panels represent the relative abundance of phosphorylated eIF2a and total eIF2a. The relative abundance of phosphorylated eIF2a and total eIF2a in the untreated cells represents 100%. The middle panels and the bottom panels represent the abundance of N protein and the virus titers, respectively. The results were obtained from three independent experiments with p values by using Student's t-test (*: p,0.05). Found at: doi:10.1371/journal.ppat.1000287.s002 (4.33 MB TIF) Figure S3 Effects of pan-caspase inhibitor Z-VADfmk on PKRmediated eIF2a phosphorylation in infected cells. VeroE6 cells were mock-infected (Mock) infected with rMP12-rLuc (rMP12-rLuc) and then treated with 5 mg/ml ActD (Act) or 50 mg/ml of aamanitin (Ama) or mock-treated (M) in the presence or absence of 100 mM of Z-VADfmk. Cells and culture supernatants were harvested at 16 h.p.i. (A) Western blot analysis of eIF2a, phosphorylated eIF2a, cleaved caspase 3 using anti-Cleaved Caspase 3, Asp175, antibody (Cell Signaling Tech. #9661), and a-actin in the absence (Top) and presence (Bottom) of Z-VADfmk. The data are representative of three independent experiments. (B) The relative abundance of phosphorylated eIF2a and total eIF2a. The relative abundance of phosphorylated eIF2a and total eIF2a in the mock-infected, mock-treated cells represents 100%. The average and standard deviation from three independent experiments were shown (*: p,0.05 compared to mock-treated cells). (C) Virus titers of rMP12-rLuc from three independent experiments were shown (*: p,0.05 compared to mock-treated cells). Found at: doi:10.1371/journal.ppat.1000287.s003 (2.29 MB TIF)

Influenza Virus (H5N1) in Live Bird Markets and Food Markets, Thailand

A surveillance program for influenza A viruses (H5N1) was conducted in live bird and food markets in central Thailand during July 2006–August 2007. Twelve subtype H5N1 viruses were isolated. The subtype H5N1 viruses circulating in the markets were genetically related to those that circulated in Thailand during 2004–2005.

Infl uenza virus (H5N1) was identifi ed in 12 of 930 samples tested. In November 2006, a total of 5 samples with infl uenza virus (H5N1) were isolated from 1 healthy chicken and 4 visceral organs obtained from 1 live bird market (chicken) and 3 different food markets (moor hen, water cock, and quail). In December 2006, a total of 5 samples with infl uenza virus (H5N1) were isolated from 5 visceral organs (quail, water cock) from 1 food market. In January 2007, a total of 2 samples with infl uenza virus (H5N1) were isolated from 2 healthy ducks obtained from 1 live bird market. In the study, 7 isolates were sequenced for whole genome analysis, and the remaining 5 To analyze the isolates, nucleotide sequences were compared with those of infl uenza subtype H5N1 viruses in Thailand, People's Republic of China, Vietnam, Indonesia, Lao, Myanmar, and Cambodia. The sequences were aligned by using the DNASTAR program (3) to elucidate and compare the genetic changes. Phylogenetic analysis was conducted by applying the PAUP program (4) with the neighbor-joining algorithm and using branch swapping and bootstrap analysis with 1,000 replicates. In the course of the 14-month surveillance program, we isolated infl uenza virus (H5N1) from 12 samples from live birds and from bird meats obtained from the markets. Bird meats were the source of 9 virus-containing samples (5 quail, 2 moor hens, and 2 water cocks), which indicates a risk for infl uenza virus (H5N1) contamination in bird meats, especially quail. In addition, 3 highly pathogenic avian infl uenza viruses were isolated from healthy live poultry (1 chicken and 2 ducks). However, the samples that contained infl uenza virus subtype H5N1 were detected only during the 3-month winter season (November-January). A possible explanation for virus contamination in live bird and food markets may be animal movement from outbreak areas to the markets as well as an attempt to sell infected (dead or dying) birds, especially quail, as bird meat. In addition, most animals or meats in the markets came from backyard farms, where they were in unavoidably close contact with wild birds. Phylogenetic analysis of the virus HA and NA genes indicated that all 12 subtype H5N1 isolates were part of the Vietnam and Thailand lineage (clade 1). The viruses were closely related to those investigated in Thailand (2004) (2005) as well as to other subtype H5N1 isolates in clade 1. In contrast, they differed from infl uenza subtype H5N1 viruses of the south China and Indonesia lineages (clade 2) (Figure 2) . In this study, we did not discern any Thailand isolates closely related to the south China lineage, as previously established in Lao and Cambodia (5) . Phylogenetic analysis of 6 remaining genes showed them to be also closely related to the Vietnam and Thailand isolates. Analysis of the deduced amino acid sequences of the HA and NA proteins indicated that the viruses had characteristics of highly pathogenic avian infl uenza. The HA cleavage site consists of multiple basic amino acids RE-RRRKKR (in 1 isolate, CU-329, REKRRKKR). All infl uenza subtype viruses harbor Glu-222 and Gly-224 at the receptor binding site, indicating preferential binding to the avian receptor SA-α-2, 3-Gal. In addition, the virus sequences contain 7 glycosylation sites as previously identifi ed in most isolates from Thailand (6) . A glycosylation site adjacent to receptor binding sites may help increase virus infectivity in host cells (7) . In some isolates, polymorphisms of amino acids related to antigenic properties of the viruses at position V86A, L138Q, and K140N were observed. All 12 subtype H5N1 viruses had a 20-aa deletion in the NA protein, typical for the NA stalk region of recent subtype H5N1 isolates (2003) (2004) (2005) (2006) (2007) (8, 9) . None of the subtype H5N1 isolates had any amino acids indicating oseltamivir resistance at the crucial positions 119 (E), 275 Emerging Infectious Diseases • www.cdc.gov/eid • Vol. 14, No. 11, November 2008 (H), 293 (R), and 295 (N) of the NA protein. In summary, the 12 viruses isolated from this study were similar to the viruses from other sources in Thailand, which indicates that the viruses are endemic to Thailand, are circulating in the country, and can be found in any exposed species. The route of infl uenza virus (H5N1) introduction into the markets remains to be established. We suspect that this contamination might have occurred as a consequence of animal movement from outbreak areas or from viruscontaminated cages, trucks, and equipment. Unfortunately, the original sources of animals in the markets could not be identifi ed because birds from different sources were housed in 1 or several cages. Fortunately, no human infection was found during 2007-2008 in those provinces where the viruses were isolated. It has been known that live bird and wet markets play a major role in facilitating emergence or reemergence of infl uenza and some other respiratory diseases (10) (11) (12) . Moni-toring of live bird and food markets as an early warning system should be implemented in Asian countries where such markets are still commonplace, and routine surveillance of these markets should be conducted year-round. In addition, raw bird meats should be handled with caution, and consumption of raw bird meats should be avoided. Increased public awareness about the risks for infl uenza virus (H5N1) in association with live bird and food markets will help prevent and control subtype H5N1 infection in humans. Figure 2 . Phylogenetic analysis of the hemagglutinin (A) and neuraminidase genes (B) of infl uenza virus (H5N1) isolates. Phylogenetic trees were generated by using the PAUP computer program (4) and applying the neighbor-joining algorithm with branch swapping and bootstrap analysis with 1,000 replicates. The trees were rooted to A/goose/China/Guangdong/1/96 (H5N1).

Antimicrobial and antioxidant activities of Cortex Magnoliae Officinalis and some other medicinal plants commonly used in South-East Asia

BACKGROUND: Eight medicinal plants were tested for their antimicrobial and antioxidant activities. Different extraction methods were also tested for their effects on the bioactivities of the medicinal plants. METHODS: Eight plants, namely Herba Polygonis Hydropiperis (Laliaocao), Folium Murraya Koenigii (Jialiye), Rhizoma Arachis Hypogea (Huashenggen), Herba Houttuyniae (Yuxingcao), Epipremnum pinnatum (Pashulong), Rhizoma Typhonium Flagelliforme (Laoshuyu), Cortex Magnoliae Officinalis (Houpo) and Rhizoma Imperatae (Baimaogen) were investigated for their potential antimicrobial and antioxidant properties. RESULTS: Extracts of Cortex Magnoliae Officinalis had the strongest activities against M. Smegmatis, C. albicans, B. subtilis and S. aureus. Boiled extracts of Cortex Magnoliae Officinalis, Folium Murraya Koenigii, Herba Polygonis Hydropiperis and Herba Houttuyniae demonstrated greater antioxidant activities than other tested medicinal plants. CONCLUSION: Among the eight tested medicinal plants, Cortex Magnoliae Officinalis showed the highest antimicrobial and antioxidant activities. Different methods of extraction yield different spectra of bioactivities.

Some medicinal plants used in traditional Chinese medicine are effective in treating various ailments caused by bacterial and oxidative stress. As new drug-resistant bacteria strains emerge, especially methicillin-resistant Staphylococcus aureus and vancomycin-resistant enterococci, new drugs or adjuvants have been actively searched in medici-nal plants [1] [2] [3] . New antioxidants such as plant phenolics [4] [5] [6] [7] are sought for general health maintenance, antiaging and chemoprevention. Eight medicinal plants, namely Herba Polygonis Hydropiperis (Laliaocao), Folium Murraya Koenigii (Jialiye), Rhizoma Arachis Hypogea (Huashenggen), Herba Houttuyniae (Yux-ingcao), Epipremnum pinnatum (Pashulong), Rhizoma Typhonium Flagelliforme (Laoshuyu), Cortex Magnoliae Officinalis (Houpo) and Rhizoma Imperatae (Baimaogen) were tested for their potential antimicrobial and antioxidant properties. They have been long been used in treating of various infectious diseases, e.g. skin/wound infections, fever, cough and digestive ailments ( Table 1 , ). The traditional method for Chinese medicine preparation is to boil the medicinal plants in water for 20 minutes to one hour. The present study aims to test the effectiveness of traditional herb preparation methods for antimicrobial and antioxidant treatments. The rationales behind the selection of these eight plants are as follows. (1) They are commonly used in Asia. (2) They have long been used as medicinal plants. ( Reducing halitosis [11] , antioxidant [12] , antimicrobial [13] , antifungal [14] , antihyperglycemic and antihyperlipidemic properties [ [17] . Prevention of urinary infection, modulation of neutrophils and monocytes, inhibition of respiratory bacteria [18, 19] . Anti-inflammatory activity [20] . (Table 1) . magnolol, honokiol (99.9%) and quercetin were purchased from Sigma Aldrich (USA). Absolute ethanol (99.9%, Far East Distiller, Singapore) was diluted with water to produce 80% (v/v) solution of ethanol for extraction. De-ionized water was used for extraction (by boiling and maceration), reconstitution and dilution where appropriate. Methanol (analytical grade, Tedia, USA) was used for reconstitution and dilution in the DPPH assay. Four strains of bacteria and one strain of yeast were used for antimicrobial tests. The test bacteria included Grampositive Staphylococcus aureus (ATCC 6538P) and Bacillus subtilis (ATCC 6633), Gram-negative Pseudomonas aeruginosa (ATCC 9027) and acid-fast Mycobacterium smegmatis (ATCC 14468). Candida albicans (ATCC 2091) was used as a representative of yeast. All microorganisms were purchased in the form of inoculation loops from Oxoid (UK). Nutrient broth with agar and Sabouraud dextrose agar (Acumedia, USA) were used for the cultivation of bacteria and yeast respectively. Mueller Hinton agar (France) was used in antimicrobial screening. Standard antibiotic discs (diameter 6 mm) used in this study were: methicillin 5 μg, tetracycline 30 μg, carbenicillin 100 μg and streptomycin 10 μg. In our preliminary studies, these antibiotics were found to be active against Staphylococcus aureus, Bacillus subtilis, Pseudomonas aeruginosa and Mycobacterium smegmatis respectively. All standard antibiotic discs were purchased from Oxoid (UK). Disc containing chlorhexidine which was active against Candida albicans, were prepared by loading dry sterile filter paper discs (Whatman No. 54, diameter 5.5 mm) with chlorhexidine solution to give a total weight of approximately 100 μg of chlorhexidine per disc. The impregnated discs were dried overnight at 40°C and stored (less than five days) in a desiccator until use. The fresh plants were kept in a refrigerator for no longer than three days prior to extraction. Cortex Magnoliae Officinalis was dried in a cool, dark room (room temperature 19°C, relative humidity 60%) and subsequently stored in a drum with silica gel desiccants until use. Before extraction, the plants were cut into 1 cm pieces with pruning scissors, except Rhizoma Imperatae and Cortex Magnoliae Officinalis which were milled into fine powder using a pulverizer mill (Christy & Norris, UK). Triplicate preparations of each sample were carried out. Boiling, maceration and blending Two and a half grams of Folium Murraya Koenigii, Typhonium flagelliforme aerial parts and 5 g of the other plant materials, were each extracted with 200 ml of water or ethanol. Three extraction methods were employed: (1) boiling in water for 1 hour, (2) maceration for 24 hours in water or (3) 80% (v/v) ethanol at room temperature. Herba Houttuyniae was extracted using an additional extraction method that involved boiling in water for 20 minutes [36] . Additional extraction experiments were carried out on aqueous plant extracts that showed promising antimicrobial activities. Boiling time was limited to 20 minutes to minimize heat exposure. Blending-maceration was used as a non-heat extraction method with cell rupture mechanism. Blending was performed with a laboratory blender (Waring Commercial, USA) for one minute, followed by a pause and then blending for an additional minute. Maceration in de-ionized water for one hour was performed. Coarse particles were removed using Whatman No. 1 filter paper (Whatman International, UK) before evaporation. Fresh juices of Herba Houttuyniae, Epipremnum pinnatum stem and Typhonium flagelliforme aerial parts and rhizomes were prepared in a mortar, wrapped in linen cloth and squeezed for the juices. Coarse particles were removed using Whatman No. 1 filter paper before evaporation. The plant extracts were evaporated to dryness under reduced pressure at 40°C for ethanol extracts and 60°C for water extracts and fresh juices in a rotary evaporator (Model N1000, Eyela, Japan). The solid content of the extract was weighed. The dried extracts were stored in a freezer at -20°C. The crude and dried extracts were characterized by their odor, appearance and texture. The weights of the dried extracts were also determined. Preparation of extract-and standard-loaded discs Filter paper discs (Grade 54, diameter 5.5 mm, Whatman International, UK) were autoclaved at 121°C for 20 minutes and oven-dried at 40°C overnight. Plant extracts were diluted with the same extraction solvent to 50 μg/μl. Each diluted solution (2 μl, equivalent to 100 μg of the dried extract) was loaded on a sterile filter paper disc. All impregnated discs were dried in sterile glass Petri dishes placed in an oven at 40°C overnight. The discs were then allowed to condition to room temperature before use in the antimicrobial test. Solutions in methanol (5 μg/μl) were prepared for magnolol and honokiol respectively and a 1:1 solution of the two compounds (2.5 μg/μl) was made. 2 μl of the honokiol, magnolol or 1:1 solutions were loaded onto paper discs which were then left to airdry. These standard-loaded discs were freshly prepared before the antimicrobial screening experiments. The antimicrobial activities of the extracts were determined by the Kirby-Bauer agar diffusion method according to NCCLS standards [37, 38] . Sterilized molten agar (20 ml) was dispensed to each sterile disposable Petri dish (diameter 9 cm) and allowed to solidify. Mueller Hinton agar was used for bacteria and Sabouraud dextrose agar for yeast. Microbial suspension (200 μl) containing approximately 3 × 10 6 CFU was spread evenly onto the surface of the solidified medium. The plates were allowed to dry for 15 minutes before the test discs were placed at equidistance from each other. Each plate consisted of one standard antibiotic disc and three other discs impregnated with various extracts. After standing for 30 minutes, the Petri dishes were incubated in an inverted position at 37°C for 18 to 24 hours for bacteria and 24°C for 48 to 72 hours for yeasts. The diameters of the zone of inhibition (ZIH), defined by the clear area devoid of growth, was measured twice. The antimicrobial activities were determined by the ratio of the ZIH diameters of the extracts to that of the standard anti-biotic in the same Petri dish, whereby a higher ratio indicates a more potent extract. Antioxidant activities of the extracts were determined with 2,2-diphenyl-1-picryl-hydrazyl (DPPH) assay [39] . The free radical, DPPH, served as the model oxidizing agent to be reduced by the antioxidant present in the extracts. The amount of dried extract subject to DPPH assay was 100 μg, the same amount used for antimicrobial screening. The dried extract was dissolved in 1.56 ml of methanol and mixed with 40 μl of 2 mM DPPH dissolved in methanol to make up a total volume of 1.6 ml in each polyethylene microfuge tubes. The final solution was allowed to react in dim light for 15 minutes. It was then centrifuged (4000 rpm; 1165 × g, Kubota 2100 Centrifuge, Japan) for five minutes. The absorbance of the supernatant was measured at 517 nm with a UV spectrophotometer (Genesys 10 UV, ThermoSprectronic, USA). The tests were carried out in triplicates. The DPPH radical scavenging activity was calculated with the following formula: Where A 0 is the absorbance of the control solution containing only DPPH after incubation; A 1 is the absorbance in the presence of plant extract in DPPH solution after incubation; and A s is the absorbance of sample extract solution without DPPH for baseline correction arising from unequal color of the sample solutions (optical blank for A 1 ). Data are expressed as mean ± standard deviation (SD) of triplicates. Two-way ANOVA was used to analyze the effect of different plant materials and extraction methods on the extraction yields and DPPH radical scavenging activity while one-way ANOVA was performed to determine the effect of streptomycin, honokiol, magnolol and honokiol-magnolol combination on M. smegmatis. Both tests employed Bonferroni post hoc analysis. Student's ttest was used to compare antimicrobial activity of the extracts against the standard antibiotic. All statistical analyses were conducted with SPSS software (v.12, SPSS, USA) at a significance level of 0.05. The extraction yields obtained from different extraction methods were analyzed with two-way ANOVA and Bonferroni post hoc analysis. Among the 11 experimental groups, Rhizoma Imperatae produced the highest yields (P = 0.001) regardless of extraction methods, followed by Cortex Magnoliae Officinalis (Figure 1 ). These two dry herbs were processed through comminution producing fine powder prior to extraction. The reduced particle size decreases the internal mass resistance for compounds to traverse through the plant matrix and increases the specific surface area for extraction. The extraction yields obtained from boiling were higher than those from other extraction methods. Boiling Herba Houttuyniae aerial parts in water for 20 minutes or one hour produced comparable yields (P = 1.000). For Herba Polygonis Hydropiperis, Folium Murraya Koenigii and Cortex Magnoliae Officinalis, a shorter boiling time of 20 minutes was shown to be comparable to a boiling time of 60 minutes (P = 0.061, 0.053 and 0.798 respectively). While results from blending/maceration varied, this Solid content of extracts obtained by different methods* Figure 1 Solid content of extracts obtained by different methods*. *Error bars represent standard deviation (n = 3). method was as efficient as the boiling method in terms of solid yields (P = 0.261) of Folium murraya koenigii. The color, texture and odor of the plant extracts were characterized (Additional file 1). The ethanolic extracts were better than corresponding aqueous extracts in retaining the natural fragrances of the plants. This may be due to the preservative ability of ethanol (i.e. reducing breakdown of organic compounds by microorganisms), its enhanced extraction capability (i.e. more fragrant components extracted) or a combination of both. Extracts obtained by boiling generally appeared darker and more turbid than those obtained by maceration. The solid content by boiling was higher than that by maceration (Figure 1 ). Boiling is more likely to damage the plant cell membrane and cell wall which act as natural filters to keep larger extraneous compounds within the cell. Among all the extracts studied, the 100 μg of the ethanolic extract of Cortex Magnoliae Officinalis loaded on the filter paper disc demonstrated the most robust antimicrobial activities against S. aureus, B. subtilis, M.smegmatis and C. albicans, equivalent to at least 50% of the activities of the standard antibiotics. Among the test organisms, it was most active against M.smegmatis, 20% more than the standard antibiotic, streptomycin 10 μg (Student's t-test, P = 0.001) ( Table 2 ). The boiled extract of Cortex Magnoliae Officinalis had comparable antimicrobial activities to those of streptomycin 10 μg (Student's t-test, P = 0.279). These data suggest that Cortex Magnoliae Officinalis may be a potential agent to treat infections caused by M. smegmatis and Mycobacterium tuberculosis [40] . It was reported that magnolol and honokiol exhibited antibacterial activities against methicillin-resistant S. aureus and vancomycinresistant enterococci [33] , Propionibacterium sp [32] and periodontal pathogens [34] . Therefore, disk diffusion test was carried out on magnolol and honokiol individually and in combination ( Table 2 ). The one way ANOVA on the four treatment groups namely streptomycin, honokiol, magnolol and combination of magnolol and honokiol (1:1) demonstrated a significant difference between groups (P = 0.001). Bonferroni post-hoc test showed that honokiol and magnolol had comparable activities (P = 1.000) against M. smegmatis, accounting for 83.58 ± 3.06% (P = 0.015) and 82.09 ± 6.51% (P = 0.006) of those of Streptomycin 10 μg respectively. In terms of antibacterial activities, the combination of magnolol and honokiol (1:1) was comparable to the reference antibiotic (P = 1.000) but higher than either magnolol (P = 0.007) or honokiol (P = 0.017) alone. These results suggest a new discovery of synergism between magnolol and honokiol. Ethanolic extract of Folium Murraya Koenigii and boiled extract of Herba Polygonis Hydropiperis showed 80% and 50% of the activities of streptomycin 10 μg against M. smegmatis respectively. These extracts also exhibited antimicrobial activities against S. aureus and B. subtilis. Additionally, the boiled extract of Herba Polygonis Hydropiperis was active against C. albicans. Boiling was essential for the active principles to be removed from the laksa plant, as blended and water macerated extracts showed little antimicrobial activities ( Table 3 ). The duration of the boiling process also affected the antimicrobial activities of laksa plant, whereby herbs boiled for 20 minutes were more active against S. aureus and M. smegmatis. The aerial parts of Herba Houttuyniae and rodent tuber were only active against B. subtilis and S. aureus respectively. The leaves and Rhizoma Arachis Hypogea, Rhizoma Imperatae, Rhizoma Typhonium Flagelliforme, and the leaves and stems of Epipremnum pinnatum did not show any antimicrobial activities. An extract with a high yield, however, does not necessarily have high antimicrobial activities. For example, Rhizoma Imperatae whose yields topped all extraction methods, did not show any antimicrobial activities ( Figure 1 and Table 3 ). The fresh juices of Herba Houttuyniae aerial parts, Epipremnum pinnatum stems and Rhizoma Typhonium Flagelliforme were tested for their folkloric use to treat wounds and various skin ailments (Table 1 ). All these fresh juices displayed some activities (less than 30% of the activity of methicillin 5 μg) against S. aureus. However, they were inactive against the rest of the test organisms. While the yields of fresh juices were lower than those of other extraction methods, antibacterial activities against S. aureus implied reduced degradation of the bioactive principles. Among all extracts, only the fresh juices of Rhizoma Typhonium Flagelliforme and Epipremnum pinnatum leaves and stems possessed antimicrobial activities ( None of the extracts, however, inhibited Ps. aeruginosa. Both S. aureus and B. subtilis are Gram-positive, while Ps. aeruginosa is Gram-negative and has an outer lipid membrane [41] . The results suggest that the antimicrobial compounds in the extracts were unable to penetrate this lipid membrane to exert their effects inside a cell. This speculation will require further experiments to confirm. The antioxidant activities of the dried extracts and fresh juices are presented in Figure 2 . All tested plants possessed some DPPH radical scavenging activities to a certain extent. While Cortex Magnoliae Officinalis, stems and leaves of dragon tail, laksa aerial parts, Herba Houttuyniae aerial parts and curry leaves showed high activities, rodent tuber rhizomes and aerial parts showed low activities. The high antioxidant activities of the boiled and ethanolic extracts of the leafy materials were probably due to the extracted tannins and photosynthetic pigments. Cortex Magnoliae Officinalis is a rich source for antioxidative compounds, such as biphenols, polyphenols and tannins [42, 43] . Lo et al. found that the antioxidant effects of magnolol and honokiol isolated from Cortex Magnoliae Officinalis were 1000 times higher than those of alphatocopherol [44] . Earlier studies confirmed that several naturally occurring dietary phytochemicals, such as isothiocyanates, curcumin and Epigallocatechin-3-gallate, possessed cancer preventive properties [45, 46] . Boiled extracts showed greater antioxidant activities than those of other extraction methods (P = 0.001). Antioxidant compounds in leafy materials are generally located in conduit structures called the apoplast and symplast [47] [48] [49] . Maceration alone is not sufficient to extract these compounds from the structures. The application of heat, in the boiling process, facilitates cell rupture and leaching, Antioxidant activities of extracts tested by DPPH assay* Figure 2 Antioxidant activities of extracts tested by DPPH assay*. *Error bars represent standard deviations (n = 3). thereby improving the mass transfer of these compounds from the storage organs into the boiling water. Ethanol may partially solubilize the membranes of the plant cells and storage organs, helping leach the chemicals away. However, maceration in 80% ethanol took over 24 hours and exposed the extracts to oxidative and hydrolytic degradation. This may explain the relatively low antioxidant activities of some ethanolic extracts. The extracts of Cortex Magnoliae Officinalis, Herba Houttuyniae aerial parts and Folium Murraya Koenigii (ethanolic extract) had similar high DDPH radical scavenging activities (>85%) but markedly different antimicrobial properties ( Figure 2 and Table 3 ). The results suggest that the active components for antimicrobial and antioxidant activities do not share common biochemical pathways. The present study discovered that (a) the ethanolic extract of Cortex Magnoliae Officinalis had 20% greater antimicrobial activities against M. smegmatis than streptomycin; (b) the boiled extract of Cortex Magnoliae Officinalis demonstrated comparable activities to streptomycin (c) the synergism of magnonol and honokiol had comparable effects to those of streptomycin; (d) the aerial parts of rodent tuber had antimicrobial activities against S. aureus. Among the tested 107 extracts, Cortex Magnoliae Officinalis had (1) potent antimicrobial activities against S. aureus, B. subtilis, M. smegmatis and C. albicans and (2) highest antioxidant activities in DPPH assay regardless extraction methods. Cortex Magnoliae Officinalis is likely a potential medicinal plant resource for developing effective antimicrobials and antioxidants.

Results From a Hypothesis Generating Case-Control Study: Herpes Family Viruses and Schizophrenia Among Military Personnel

Background: Herpes family viruses can cause central nervous system inflammatory changes that can present with symptoms indistinguishable from schizophrenia and therefore are of interest in schizophrenia research. Most existing studies of herpes viruses have used small populations and postdiagnosis specimens. As part of a larger research program, we conducted a hypothesis-generating case-control study of selected herpes virus antibodies among individuals discharged from the US military with schizophrenia and pre- and postdiagnosis sera. Methods: Cases (n = 180) were servicemembers hospitalized and discharged from military service with schizophrenia. Controls, 3:1 matched on several factors, were members not discharged. The military routinely collects and stores members' serum specimens. We used microplate enzyme immunoassay to measure immunoglobulin G (IgG) antibody levels to 6 herpes viruses in pre- and postdiagnosis specimens. Conditional logistic regression was used, and the measure of association was the hazard ratio (HR). Results: Overall, we found a significant association between human herpes virus type 6 and schizophrenia, with an HR of 1.17 (95% confidence interval [CI] = 1.04, 1.32). Women and blacks had significant negative associations with herpes simplex virus type 2 and cytomegalovirus; among blacks, there was a significant positive association with herpes simplex virus type 1. Among men, there was a HHV-6 temporal effect with an HR of 1.41 (95% CI = 1.02, 1.96) for sera drawn 6–12 months before diagnosis. Discussion: Findings from previous studies of herpes family viruses and schizophrenia have been inconsistent. Our study is based on a larger population than most previous studies and used serum specimens collected before onset of illness. This study adds to the body of knowledge and provides testable hypotheses for follow-on studies.

Recent research has focused on infectious agents as potential players in the etiologic pathway of chronic diseases, [1] [2] [3] including psychiatric illnesses such as schizophrenia. [4] [5] [6] [7] [8] [9] [10] Due to their potential neurotropism and latency, viral organisms in particular are considered possible agents in many chronic central nervous system (CNS) disorders. 2, [11] [12] [13] Encephalitis and other conditions leading to CNS inflammatory changes often present with symptoms that are difficult to distinguish from new-onset schizophrenia. As ) are of prime interest in schizophrenia research. 14 Although laboratory-based research into herpes family viruses as possible etiologic agents for schizophrenia goes back decades, [15] [16] [17] [18] [19] [20] [21] [22] ascertaining the nature of a possible etiologic association between infection and schizophrenia is highly challenging. There have been few consistent findings between studies, which could be due to many factors, including the heterogeneity of schizophrenia itself, the use of different immunologic assays across the studies and over time as technology changes, focusing attention on a variety of different infections, and the possibility that maternal infection occurring at the right time during pregnancy may be enough to increase the risk of psychosis in offspring. 23 There is more information regarding risks associated with maternal, neonatal, or childhood infection and schizophrenia than for adult infection, with a number of studies reporting increased risk of schizophrenia among persons exposed in utero to a number of infections [24] [25] [26] or born after epidemic. 27 Little is known about potential mechanisms of action for herpes family virus infections and risk of schizophrenia. Studies of maternal infection provide some evidence that modulation of immune response 28 and adverse effects on in utero maturation of critical brain structural and functional componenets 24 may correlate with increased risk of schizophrenia in offspring. From animal models, maternal infection can alter interleukin 6, interleukin 1b, or tumor necrosis factor a (TNF-a) in amniotic fluid or placenta and TNF-a in the fetal brain. 29 There is evidence that cytokines have an important function in the development of fetal neurons and glial cells, and abnormal levels of these (proinflammatory) cytokines may contribute to abnormal brain development, [30] [31] [32] [33] [34] at least in animals. Findings from epidemiologic studies of herpes family viruses and schizophrenia among adults are mixed. In their 1995 review article, Yolken and Torrey 14 identified numerous published studies assessing viral (not just herpes family viruses) antibodies in serum of schizophrenia cases. Several of the studies were not interpretable because they lacked control groups or were evaluating changes in antibody titers over time. Of 21 interpretable studies, 3 reported positive significant associations with HSV-1 (or HSV unspecified), 35-37 1 study reported an association with EBV, 18 and 1 found a negative association with CMV. 38 The Military New-Onset Psychosis Project hypothesisgenerating studies afford unique opportunities to correct for some of the weaknesses identified in other studies of herpes viruses and schizophrenia. Because the US military routinely collects and stores serum samples and medical data from all active duty personnel, we are able to study large numbers of cases who have prediagnostic serum available and document the prevalence of infection prior to onset of disease. Because military members provide a serum sample at accession and about every 2 years thereafter while they remain on active duty, more than one prediagnostic sample will be available for many individuals. Sera from these routine samples are stored in the Department of Defense Serum Repository (DoDSR). 39 Informed consent is not routinely obtained when the serum is drawn; however, the DoDSR is maintained for surveillance and research purposes. Information obtained from this study may lead to more effective means of preventing, identifying, and treating new-onset schizophrenia in this population. This project was reviewed and approved by the appropriate human protection committees at the authors' institutions. We assayed the serum of 180 individuals diagnosed with schizophrenia who had been hospitalized in a military facility with a mental health diagnosis and subsequently medically discharged or retired from the military between 1992 and 2001 and the serum of 532 controls with no evidence of any mental illness. Diagnosis date for cases was estimated as the date of first mental health hospitalization, used as a proxy for onset. Controls were selected 3:1 to cases and matched on the following variables: date of birth (6 1 year), the corresponding case's accession date 6 6 months, sex, race (black, white, other), branch of military service, and the number of serum specimens available. We attempted to obtain the same number of specimens for cases and controls: the first available specimen (usually collected during the accession medical examination), a specimen collected in the 3-to 24-month period prior to the matched case's first mental health hospitalization, and the first available after the matched case's hospitalization. Additional details on study subject selection and inclusion, serum selection and shipping, and sources and types of ancillary data have been published. 40 Sera were assayed for 6 herpes family viruses: HSV-1, HSV-2, EBV, CMV, VZV, and HHV-6. IgG antibody to human herpes viruses were measured using microplate enzyme immunoassay. Reagents were obtained from the following sources: HSV-1 and HSV-2 from Focus Diagnostics, Cypress, CA; HHV-6 from Advanced Biotechnologies, Columbia, MD; VZV and EBV nuclear antigen from IBL Laboratories, Minneapolis, MN; and CMV from Viro-Immun Labor Diagnostika, Oberursel, Germany. Enzyme immunoassay consists of binding serum to solid-phase antigen and subsequent reactions with enzyme-labeled antihuman IgG and enzyme substrate. The amount of color generated by the enzyme substrate reaction was measured in optical density (OD) units with a microplate colorimeter. This method of analysis was selected because it allows for high throughput measurement of antibodies using a common platform and requiring only small amounts of sample. Samples were run on plates under code in matched groups in which case and control status was not identified. All matched case-control samples were tested on the same plates, but over the course of the study, samples were assayed on over 21 different plates per agent. To control for potential systematic error introduced by plate-toplate variation and to ensure that observed differences in OD are due to differential expression (cases or controls) and not experimental artifacts, data were normalized using the robust median normalization method which combines the within-plate and between-plates variance 41 using the following equation, where R ijk is the raw OD of ith subject's jth blood sample, M k is the median of all the control samples in plate k, V k is the variance of R ijk of control samples in plate k, and V b is the variance of all R ijk between plates. Therefore, S ijk is the scaled score for the ith subject's jth blood sample on plate k. To investigate the relationship between subject status (case or control) and antibody level in a matched study design, conditional logistic regression is used. Failure to account for matching in analysis may lead to biased results, usually toward the null. The conditional analysis, which has a higher (less negative) log likelihood, suggests a better ''fit'' for this data. 42 A guiding concern in regression modeling is that the relationship between the independent and dependent variables (the latter assumed to be continuous) should be as inherently ''linear'' as possible; hence, OD was analyzed as a continuous, rather than dichotomous (positive vs negative) variable. This approach also preserves the power to detect a difference between cases and controls, particularly because infection with most of these herpes viruses is ubiquitous and differences cannot be detected based on prevalence data. For this analysis, we chose the proportional hazards (PH) model for conditional logistic regression. 42 The PH model is similar to regular conditional logistic regression but modified to allow for multiple and variable numbers of controls per case and specimens per subject. Dummy ''survival time'' to diagnosis was generated so that all samples for a given case had the same event time with corresponding controls censored at a later time. Proportional hazard regression was then used to study associations, reported as hazard ratios (HRs), between ''survival time'' and the risk factors. Because the PH assumption might not be true for all the data, but may be true among specific demographic subgroups, we performed stratified analysis on several of the matched variables. First, we assessed the overall antibody effect for each agent separately. The analysis was stratified by the matched factors with case or control status as the outcome. Variables analyzed included antibody level, the time from serum collection to the case's diagnosis, and time in service for both cases and controls. Logistic models were developed to assess the antibody effect for modeling the agents separately as well as simultaneously. To study the homogeneity of the agent effect across demographic levels and time, interaction terms were also evaluated. Where an interaction with demographic factor was observed, separate models were developed to explore the possibility of effect modification by those factors. For homogeneity of the agent effect across the time, we studied the interaction with time to diagnosis. Time to diagnosis was categorized as follows: greater than 2 years, 1-2 years, 0.5-1 year, less than 0.5 year before diagnosis, and after diagnosis. The interaction with time to diagnosis shows the temporal effect of antibody level and describes the consistency of risk across time periods. The interactions of the agents with demographic factors show the heterogeneity of the agent effect by those factors. Because we used scaled values to represent measured antibody level, there is no recognized unit with which to describe increasing or decreasing levels. In this case, we chose the SD as the unit of measure. All results are reported as HRs for each increase of 1 SD of antibody level. This method also allows comparison between the effects of different antibody agents on schizophrenia. The SD of antibody level for the 6 agents ranged from 0.10 for HHV-6 to 0.59 for HSV-2. A total of 180 cases (contributing 404 serum specimens) and 532 controls (with 1180 specimens) were included in the study population. Eight cases could only be matched to 2 controls. Table 1 shows the distribution of cases and controls by demographic factors. Overall, about 83% were males, 49% were whites, 44% were blacks, over 57% were younger than 25 years, 10% were older than 35 years, about 12% were Hispanic, and over 56% were in the army. Approximately 35% of cases had greater than 3 years of military service. We found that antibody levels for all 6 agents were only weakly correlated (data not shown), and therefore, No significant effect modification was noted for sex (male vs female) or age (<25 vs !25) for any of the infectious agents (P values > .10). A significant interaction effect with race (black vs white) was found for HSV-1 (P < .01) and for CMV (P = .03), suggesting that significant differences exist in these antibody effects between black and white cases. As seen in Hazard ratio expressed per 1 SD increase of each agent antibody level. Our hypothesis-generating study found a statistically significant positive HR between HHV-6 and schizophrenia among men and between HSV-1 and schizophrenia among blacks discharged or retired from the military with a diagnosis of schizophrenia and a history of mental health hospitalization. A negative association with HSV-2 and CMV was noted among women and blacks. Blacks dominated the results for women. These findings should be interpreted with caution, however, because they are driven by a small number of subjects (n = 80 for black cases, n = 30 for female cases, n = 22 for black females) and may be the result of type I error. Further analysis is warranted with a larger sample size. No significant associations were observed for HSV-2, EBV, or VZV among men or women. No significant association was found among whites for any agent. Our subanalysis of HHV-6 IgG levels by time period for males around diagnosis is obviously limited by the sample size. We conducted this analysis to replicate the analytic modeling in our previous, related work on Toxoplasma gondii IgG level and risk of schizophrenia. 40 The P values of .04 in this hypothesis-generating study warrant further evaluation in the hypothesis-testing phase of our research. More recently, studies of antibody levels in serum and cerebrospinal fluid demonstrate mixed findings. One analysis of untreated subjects with recent-onset schizophrenia found increased IgG antibody levels to CMV, decreased levels of antibodies to HHV-6 and VZV, and no differences in antibody level to HSV-1 and -2 and EBV. 9 Several other studies of cerebrospinal fluid yielded conflicting results with some reporting increased CMV antibody titers 38, [43] [44] [45] while others demonstrate no association. [46] [47] [48] Increased levels of HSV-1 antibody were demonstrated in one group of schizophrenic patients compared with normal controls, and cases with higher levels of antibody also demonstrated decreased gray matter in 2 areas of the brain. 49 Another study noted that deficit schizophrenics were more likely to have antibodies to CMV than were nondeficit patients. 50 A recent review of the literature regarding CMV and schizophrenia identified a number of studies reporting more frequent infection or higher levels of antibody in serum or cerebrospinal fluid. 51 The authors noted that studies conducted in 1992 were all null but that the serum assays utilized had been complement fixation or other less sensitive methods. They note 3 unpublished studies (M. J. Schwarz and N. Mueller, S. Bachmann; J. Schrö der; and R. H. Yolken, unpublished data) in which patients with schizophrenia were more likely to have antibodies to CMV, or had higher levels of antibodies, than did the control subjects. One of these studies (F. B. Dickerson, C. Stallings, A. Origoni, J. J. Boronow, R. H. Yolken, unpublished data) was of 415 outpatients with schizophrenia and 164 matched controls, in which the odds ratio for CMV positivity was 2.1. The authors note that patients who were seropositive were more likely to be female, black, older, and less educated. Leweke et al 9 found that CMV IgG antibody levels, but not HSV-1, HSV-2, EBV, HHV-6, or VZV, were higher among patients with schizophrenia. 9 Given the limited amount of research reported and the discordant findings among the existing articles regarding herpes viruses and schizophrenia, interpretation of our findings is challenging. Recent work has implicated HHV-6 in acute 52 and chronic 53,54 neurologic diseases. We note the negative association with HSV-2 and CMV among women and blacks and the positive association with HSV-1 among blacks. Although speculative, and limited by sample size, there is a potential for underlying genetic differences that could explain some portion of the racial differences. There are a number of factors that could potentially account for the discrepancies observed between the various reports above and the present study. These include but are not limited to differences in diagnostic criteria for schizophrenia, all cases were adult onset, different time frames of serum collection related to illness onset, differences in laboratory assays, and multiple vs single serum specimens. In addition, our sample was drawn from the military population that differs from the general US population in several important ways. The male to female ratio in the military is much higher than in the general population, making it difficult to achieve adequate power when analyzing females separately. Comprehensive medical screening prior to entry into the military creates a healthy worker effect in the population and skews our sample toward individuals with later onset of schizophrenia. Also, our cases are a convenience sample of individuals with schizophrenia in the military. A small degree of bias introduced by any of these factors could account for the significance and direction of our findings. This study has 2 important strengths. First, we used cases that were diagnosed and discharged from the military after a careful evaluation process. 40 A record review conducted on a sample of study subjects demonstrated a high level of concordance between the diagnoses documented in the military records and those assigned by 2 psychiatrist reviewers. 55 Also, because military applicants receive an extensive administrative and medical examination prior to accession, are directly supervised while on active duty, and have access to health care, we assumed that cases were not psychotic on accession and that the onset of psychosis would generally result in a mental health hospitalization for an expedited psychiatric evaluation. This assumption was validated by the record review. 55 Therefore, we are confident that diagnostic misclassification is not a major source of error in our findings. In addition, the current study obtained multiple (between 1 and 3) specimens for most subjects both prior to and after onset of illness in the matched case. The second specimen, collected in the 3-to 24-month period prior to first mental health hospitalization was chosen as preonset of psychosis. The ability to analyze longitudinal specimens may be important if illness alters behaviors in a way that could impact antibody levels or if medical treatment for illness changes antibody responses. Although this hypothesis-generating study does not resolve the issue of diverse findings between studies, we feel that our study has advantages over other studies with our high degree of diagnostic validity, adequate numbers of appropriate controls, and multiple serum specimens. It is clear that additional studies are needed to clarify the many remaining questions, particularly regarding HHV-6, CMV, and HSV-1. As part of our broad research program, we will be conducting a much larger casecontrol study with approximately 1600 cases and 2100 controls. Although herpes family viruses will not be the primary focus of this follow-on study, we intend to further explore the associations identified in this hypothesisgenerating study.

Science into policy: preparing for pandemic influenza

Authoratative government pandemic preparedness requires an evidence-based approach. The scientific advisory process that has informed the current UK pandemic preparedness plans is described. The final endorsed scientific papers are now publicly available.

Public expectations of effective government interventions in health crises are high in developed countries. Authoritative action and provision of information to the public can help in avoiding public disquiet or panic and in mitigating the societal risks of a pandemic, complementing the direct health effects of any interventions. Conversely, disagreement over the scientific evidence base, particularly where considerable uncertainties and gaps in information exist, can open the way to debate based primarily on established beliefs and prejudices. In the face of a future event such as an influenza pandemic, the timing and precise nature of which is unknown, robust preparation will be strengthened by an agreed scientific understanding of the risks and the options for response. The UK Government has followed an extensive process to review and confirm an agreed summary of the international evidence base. This underpins policy development on countermeasures within its pandemic influenza preparedness programme and can be of use to other countries developing pandemic preparedness plans as well. Under the auspices of the UK Scientific Advisory Group on Pandemic Influenza, five scientific papers dealing with the main clinical countermeasures (antivirals, pre-pandemic and pandemic specific vaccines, antibiotics and facemasks) and the risk of a pandemic originating from an H5N1 virus were developed. These papers were reviewed and revised by additional national and international expert scientific reviewers and subsequently at a colloquium, convened by the Secretary of State for Health, of scientific experts. Revised papers were then submitted to the Scientific Advisory Group for final endorsement as reflecting an accurate and comprehensive summary of the state of knowledge in June 2007. The final endorsed papers have now been made publicly available as a resource to all. 1 Papers reviewing the scientific evidence base in the following areas are available at: http://www.advisorybodies.doh.gov.uk/ spi/evidence.htm (i) The use of antiviral drugs in a pandemic; (ii) pre-pandemic and pandemic specific influenza vaccines; (iii) the use of antibiotics for pandemic influenza; (iv) the use of face masks during a pandemic; and (v) the risk of a pandemic originating from H5N1. This widely agreed scientific state of the art offers a firm foundation for complex and potentially expensive policy and procurement decisions on pandemic countermeasures. Within the UK, the papers have already informed the recently published framework 2 and policy statement. 3 They will continue to inform policy decisions across Government. The scientific knowledge in this field is continually evolving and improving, and the UK will therefore continue to review and refine its assessment of the evidence base.

Smallpox and Season: Reanalysis of Historical Data

Seasonal variation in smallpox transmission is one of the most pressing ecological questions and is relevant to bioterrorism preparedness. The present study reanalyzed 7 historical datasets which recorded monthly cases or deaths. In addition to time series analyses of reported data, an estimation and spectral analysis of the effective reproduction number at calendar time t, R(t), were made. Meteorological variables were extracted from a report in India from 1890–1921 and compared with smallpox mortality as well as R(t). Annual cycles of smallpox transmission were clearly shown not only in monthly reports but also in the estimates of R(t). Even short-term epidemic data clearly exhibited an annual peak every January. Both mortality and R(t) revealed significant negative association (P < .01) and correlation (P < .01), respectively, with humidity. These findings suggest that smallpox transmission greatly varies with season and is most likely enhanced by dry weather.

Smallpox is the only disease to have been eradicated worldwide [1] . Despite the success story of vaccination and other public health interventions, the number of susceptible individuals has grown to date following cessation of routine vaccination, and the threat of bioterrorist attack has led to debates on countermeasures in such an event [2] . Various mathematical studies have been conducted as part of a preparedness program, including large-scale simulation of a bioterrorist attack and the public health countermeasures against it [3] [4] [5] [6] . Theoretical studies on the spread of smallpox include not only simulations but also quantitative analysis of historical data [7] [8] [9] [10] . A statistical modeling study suggests that a small outbreak could be contained only implementing contact tracing and isolation [11] . Moreover, those who underwent vaccination in the past are believed to be still protected against severe and fatal manifestations of smallpox even today [7, 12] . Studies on smallpox control have progressed in parallel with the development of epidemiological and statistical methods, and because of the eradication before maturation of biostatistics, many questions have remained in regards to the details of the epidemiology. Seasonal variation in smallpox transmission is one of the most pressing ecological questions playing a key role in determining the transmission dynamics, should a future outbreak occur following the deliberate release [1, 4] . For example, clarification of the seasonal preference of variola virus is crucial for identifying and forecasting the disease risk using ecological data [13] . Although seasonal occurrence of smallpox was documented early on among directly transmitted infectious diseases [14, 15] , and whereas the disease is believed to be one of the "winter diseases" in industrialized countries, even the presence of seasonality has not been investigated in detail. The best available evidence stems from a series of studies by Sir Leonard Rogers (1868 Rogers ( -1962 [16] , who conducted epidemiologic surveys of smallpox outbreaks in India over a long period of time [17] [18] [19] . He also conducted a similar survey in England and Wales [20] . By analyzing the monthly mortality data from the late 19th to the early 20th century in these countries, Rogers argued that the smallpox epidemic in India is relatively uniform (i.e., not apparently cyclical) compared to that in England and Wales [17, 19, 21] . Further, he descriptively and implicitly suggested that there is a negative correlation between humidity and smallpox mortality, but there was little association between smallpox and rainfall [17, 18] . This effort was followed by Russell and Sundararajan [22] who supported the notion that a dry environment offers favorable conditions for smallpox transmission. These consistent findings have also been reported during the Smallpox Eradication Program (SEP), where a peak of smallpox incidence occurred from December to January in the Northern hemisphere (e.g., Indonesia and Bangladesh) and from May to June in the Southern hemisphere (e.g., Brazil) [23] [24] [25] [26] . However, the observed data during the SEP were greatly modified by intensive immunizations, and perhaps because of this, epidemics in other locations were not suggested to be seasonal [26] [27] [28] , leaving this issue yet to be clarified. Despite the rigorous efforts before the global eradication, later progress on this issue was unfortunately subtle. Upham once revisited Rogers's dataset from India, anthropologically discussing potential reasons why the American Southwest was less infested by smallpox [29] . A time series technique was applied to historical data in Finland and England [30] [31] [32] , showing that periodicity is mainly regulated by the susceptible fraction of a population in question [33] . However, despite the analyses on the impact of vaccination and migration on periodicity, seasonal patterns of transmission have not been explicitly studied, mainly because of a lack of data precision. In a historical study examining smallpox in England from the 16th to 17th centuries, the time referred to as the "little ice age," it has been documented that longterm climatic changes did little to the smallpox transmission [34] , but this conclusion was drawn without quantitatively and explicitly analyzing the data. Instead, the quality of time series data and its impact on seasonality were discussed in relation to social backgrounds of smallpox control [35, 36] , but again no rigorous statistical analyses were made using observed data. Accordingly, several lingering questions remain. Is smallpox transmission really seasonal? If so, is the seasonality associated with humidity? Clarification of these points will not only enhance our understanding of the pattern of smallpox transmission, but also will be crucial for identifying the seasonal preference of variola virus with some implications for bioterrorism preparedness plans. The present study is aimed at examining the presence of seasonality and clarifying the relationships between smallpox and climate. We reanalyzed various historical datasets, suggesting a new method for the analysis of time series. Records. Seven temporal distributions of smallpox at different times and locations were extracted from historical literature. This literature review was based on references collected by tracking all the references given in the relevant publications and repeating this task until we could find no further references; the details are given elsewhere [37, 38] . Figure 1 shows the time series data by location with a monthly reporting interval. Chronologically, epidemic records for The Hague (1755-1773), Berlin (1758-1774), Zurich (1870-1887), the entire Netherlands (1870-1873), Northwest Frontier province in India (1890-1921), Shanghai (1900) (1901) (1902) (1903) (1904) (1905) (1906) (1907) (1908) (1909) (1910) (1911) (1912) (1913) , and Bombay (1902) (1903) (1904) (1905) (1906) (1907) provide monthly data of smallpox with time and were used for further analysis [17, [39] [40] [41] [42] [43] [44] [45] . The first two records contain data before the introduction of vaccination. Except for Zurich, which documents the monthly number of cases, the remaining datasets record only monthly deaths. Death data are given as the absolute number of deaths, except where indicated. With regard to the magnitude of the epidemics, the annual averages of the disaster size were 10.1 deaths (The Hague), 32.9 deaths (Berlin), 9.9 cases (Zurich), 428.6 deaths (the entire Netherlands), 5.28 deaths per 100 000 (Northwest Frontier province in India), 21.5 deaths (Shanghai), and 2.45 deaths per 100 000 (Bombay). By examining another historical record of the smallpox epidemic in Tokyo, it was found that the mean (and the standard deviation) and the median (25-75% quartile) time from onset to death were 29.1 (13.8) and 26.0 (19.0-37.0) days, respectively [46] . Thus, it is reasonable to assume that the relative frequency of death with time represents that of onset accompanied by approximately a 1 month delay. Moreover, the infection may have happened approximately half a month before the onset [9] . Meteorological variables with time were given only in Rogers's observations [17] , which contained the monthly rainfall (inch) and the absolute humidity. First, the presence of seasonality was examined for all 7 datasets using spectral density analysis. Spectral analysis is based on the idea of a theoretical power-spectrum, which partitions the total variance of the series among sinusoidal components [47] . In other words, spectral density decomposes a time series function into a sum of sines and cosines. The density plot (i.e., correlogram) was graphically plotted to determine if a sharp peak at a period of 12 months exists, corresponding to an annual cycle (i.e., seasonality). Second, seasonality that was evaluated using the effective reproduction number, R(t), defined as the actual average number of secondary cases per primary case at calendar time t. R(t) shows a time-dependent variation with a decline in susceptible individuals (intrinsic factors) and with the implementation of control measures (extrinsic factors). If R(t) < 1, it suggests that the epidemic is in decline (vice verca, if R(t) > 1). This approach was employed to clearly show the seasonal patterns of transmission and to partly address the issue of dependence among cases, that is, statistically, the observation of an infected individual is not independent of other infected individuals, since the disease is transmitted directly from human to human. The following approximation was made to derive estimates of R(t). Supposing that the number of new infections at calendar time t is I(t), the transmission dynamics are described by the renewal equation using R(t) [48, 49] : 1898 1900 1902 1904 1906 1908 1910 1912 1914 1916 1918 1920 Year where ω(τ) is the probability density function of the generation time. The right-hand side of (1) represents secondary transmissions at calendar time t, which are determined by the number of those who were infected at time t − τ, I(t − τ), and the magnitude of secondary transmissions at time t, R(t). Since the data in the present study were recorded only monthly, the equation has to be simplified to comply with discrete points of time data. From the beginning of the history of mathematical modeling of smallpox in the late 19th century [50] , cases tended to be modeled by generation, the idea of which is applied as follows. Given the number of cases in generation i, I i , the expected number of cases in generation i + 1, E(I i+1 ) is given by where R i is the effective reproduction number of generation i [51] . That is, the reproduction number is simply given by ratio of successive generations of cases, which was implicitly understood in history by a pioneering epidemiologist, Clare Oswald Stallybrass (1881 Stallybrass ( -1951 who applied the theory to analyze the seasonality of various infectious diseases [52, 53] . Since the mean generation time of smallpox is approximately 15 days (i.e., half a month) [50, 54] , monthly data contains exactly two generations. Let us consider three successive generations, i, i+1, and i+2. Given the reproduction numbers R i and R i+1 , we get Considering that the generations i and i + 1 are grouped together and reported in the same month j, the reproduction number cannot be estimated by generation i. Instead, by assuming that the reproduction numbers in the successive generations are identical, that is, R i = R i+1 (:= R j ), (3) can be rearranged as The expected number of cases in the next generation i + 3 is given by product of I i+2 and the reproduction number in the next month j + 1, R j+1 , that is, Given (4) and (5), the number of cases in month j + 1, C j+1 (:= I i+2 + I i+3 ) is given using C j (:= I i + I i+1 ), that is, We assume that the expected values are sufficient to characterize Poisson distributions. This assumption indicates that the conditional distribution of the reported number of cases in month j + 1, C j+1 , givenC j is given by Thus, for the observation of cases (or deaths with a 1 month lag) from month 0 to N, the likelihood of estimating R j is given by By minimizing the negative logarithm of (8), the maximum likelihood estimates of the monthly-approximated reproduction numbers, R j were obtained. Modeling. Third, to identify the characteristic factors of seasonal variation in smallpox transmissions, the relationships between meteorological variables (i.e., rainfall and humidity) and incidence (mortality) as well as the effective reproduction number were examined. To examine the influence of seasonal variables on the temporal trend of smallpox, we employed one of the generalized linear models with the construction of a Poisson regression model incorporating monthly and yearly terms [55] : where E(C j ) is the expected number of cases (deaths) in month j, α is a constant, and β i are the regression coefficients for year or month. The relationship was investigated using both univariate and multivariate models. In the multivariate model, the year of occurrence was controlled for, but the sine and cosine of the month were not included due to colinearity with rainfall. The mortality rate ratios (MRR) for the occurrence of smallpox death were used to evaluate the impact of each meteorological variable on smallpox. With regard to the relationship between meteorological variables and R(t), multiple linear regression analysis was employed to determine factors contributing to R(t). Because of the obvious cyclical nature of the observed data yielding an autocorrelation in the linear regression analysis (Durbin-Watson = 0.23), the monthly periodic terms (as shown in (9)) were added to the list of independent variables. The level of statistical significance was set at α = 0.05. All statistical data were analyzed using the statistical software JMP version 7.0 (SAS Institute Inc., Cary, NC, USA). Density. The spectral densities are shown in Figure 2 which can be reasonably interpreted by comparatively examining the temporal distributions ( Figure 1 ). With regard to the data collected from The Hague and Berlin, the observations of which were made before the introduction of vaccinations, periodic epidemics (i.e., super-annual cycles) are apparent where the interepidemic period ranges from 3 to 5 years (see Figures 1(a) and 1(b) ). However, the annual cycle is not seen, and thus, the spectral densities do not show a clear seasonal pattern (Figures 2(a) and 2(b) ). On the contrary, the time series data in Zurich and Shanghai clearly revealed a peak at 12 months (Figures 2(c) and 2(f)). The entire Netherlands data covers a relatively short period of time compared to the other datasets (Figure 1(d) ) with unclear seasonal and periodic frequencies in the spectral diagram (Figure 2(d) ). Although a small peak is seen at 12 months for the data in the Northwest Frontier province in India (Figure 2(e) ), the density plot exhibits a multimodal pattern, reflecting an irregular temporal distribution (Figure 1(e) ). In the Bombay data, the annual cycle is most clearly highlighted in the temporal distribution (Figure 1(g) ), which is also reflected in the spectral density (Figure 2(g) ). Figure 3 plots estimates of the effective reproduction number as a function of calendar time. The vertical broken lines represent January in every year, while a horizontal dashed line is a reference value yielding R(t) = 1, that is, the threshold condition of an epidemic. R(t) tends to increase during the winter season for three early records (Figures 3(a), 3(b), and 3(c) ), but the annual cycles are not seen. However, the shortterm epidemic data for the entire Netherlands clearly shows that three peaks of R(t) coincide in every January with estimates above unity (Figure 3(d) ). A similar pattern is observed in Shanghai and Bombay (Figures 3(f) and 3(g) ). Figure 4 shows the spectral density plots of R(t) for the entire Netherlands and Northwest Frontier province in India. Although spectral densities of death and mortality (Figures 2(d) and 2(e)) did not exhibit a clear annual cycle, the obvious peak at 12 months is seen for both datasets in terms of R(t) (Figures 4(a) and 4(b) ). That is, seasonal patterns of smallpox transmission were reasonably shown with the use of R(t) even for the short-and long-term time series. Table 1 shows the output of univariate and multivariate models for explaining smallpox mortality in India using meteorological variables. In both models, rainfall was not significantly associated with smallpox mortality. However, significant negative association was found for humidity (adjusted MRR = 0.387 (95% confidence interval (CI): 0.311, 0.481), P < .01). Table 2 summarizes the relationship between the effective reproduction number and meteorological variables using a multiple linear regression model. On a whole, the model showed a weak predictive ability. However, humidity was again identified as an explanatory variable which significantly reduces the effective reproduction number (P < .01). No significant correlation was found between R(t) and rainfall. The present study reanalyzed historical records of smallpox to examine the presence of seasonality and to partly clarify the characteristic factors. Although 18th century data did not show an apparent annual cycle, the remaining records reasonably showed seasonal variations either in the monthly observation or the reproduction number. In particular, even the short-term epidemic data for the entire Netherlands clearly revealed peaks of transmission every January. Although several important meteorological variables were missing (e.g., temperature and atmospheric pressure), Rogers's observation permitted investigations of a few variables as underlying factors characterizing the seasonality. Analyzing the meteorological data in India, both smallpox mortality and the reproduction number yielded significant negative association and correlation with humidity. Rainfall did not appear to be a useful predictor of seasonality. One important message drawn from this exercise is that smallpox transmission is confirmed as seasonal and this is most likely associated with dry weather. This finding is consistent with implicit suggestions which have accumulated in the historical literature [1, 17, 19] . Whereas the data from The Hague and Berlin did not offer the relevant interpretations, their periodic peaks were also observed during the winter seasons. Assuming that these records captured mainly the large periodic outbreaks alone, it is plausible that the old data were accompanied by underreporting during less intensive years, and thus, did not precisely contain subtle seasonal fluctuations. Given that the seasonal force of infection was obvious even in the shortterm epidemic data from the entire Netherlands, not only endemic but also epidemic smallpox would greatly vary with the season and most likely would be enhanced by dry weather. Historically, virologists attempted to attribute the annual cycle to the seasonal preference of the variola virus [56] [57] [58] . To date, it is known that the variola virus could survive in an infective state under different conditions of temperature and humidity [56, 57] . However, as temperature and humidity rise above 30 • C and 55%, respectively, the virus is known to immediately lose infectivity [57] . Such a virological explanation supports the epidemiologic findings from this present study and reasonably explains the seasonal preference of the virus as a factor behind the seasonality of outbreaks. The above-mentioned point implies that we cannot ignore the seasonality even in the event of a shortterm reintroduction of variola virus due to bioterrorist attack. A technical development in analyzing seasonal data is also notable. Since the observation of an infected individual is not independent of other infectious individuals, direct application of the generalized linear model has not been appropriate to date. One approach to resolve this issue is to employ a Bayesian method, explicitly dealing with dependence in a Poisson regression model [59] , which is, Interdisciplinary Perspectives on Infectious Diseases however, computationally complicated for nonspecialists. As an alternative, the present study suggested the use of R(t). R(t) reasonably reflects time-dependent changes in the susceptible fraction of the population in question and other various factors determining the transmission (including seasonality) [60, 61] . In particular, it should be noted that R(t) does not reflect onset or death but can represent an infection event with time, proving its potential as a marker to model seasonal and periodic transmission cycles. In addition, quantitative assessment of R(t) is theoretically important, because the amplitude of seasonal forces of infection characterizes the length of the interepidemic period [33, 62, 63] . A continued super-annual cycle mathematically requires seasonally varying forces of infection, which determines the season-specific threshold condition [64] and evolutionary dynamics of a disease [65, 66] . To the best of our knowledge, the present study is the first to suggest a quantitative method to reasonably extract the amplitude using the notation of R(t) and extending the previous efforts of Stallybrass [53] . Most infectious diseases show characteristic seasonal variations in incidence. The present study confirms that the transmission of smallpox does vary with season. However, compared to the seasonal ecology of insects in vector-borne diseases, seasonal factors for directly transmitted diseases are far more complex, and thus, questions remain as to what exactly are the factors behind the seasonality of smallpox. At least, experimental evidence supports the role of dry weather in the dynamics of influenza [67, 68] ; a recent study found that low (dry) relative humidity in the range of 20 to 30% produced the spread of the influenza virus faster than at relative humidity in higher percentages. In fact, at a humidity of 80% or above, the study found no spread of the flu [68] . Since there are also various social factors which vary with the season, the seasonal preference of pathogens cannot be captured without sufficiently highlighting the time-varying human contact patterns, and thus, further analyses (e.g., reanalysis of small-scale outbreaks where we can adjust the contact frequency) are needed. We hope that the present study enhances the similar reanalysis of historical data, triggering an interest in investigating the relationship between the transmission of directly transmitted infectious diseases and climatic changes. Seven historical datasets of smallpox were reanalyzed to examine the presence of seasonality and to identify the characteristic factors. Annual cycles were clearly shown not only in the monthly reports but also in the estimates of the effective reproduction number. Even for a short-term epidemic, the transmission of smallpox would most likely be enhanced by dry weather.

Chinese journals: a guide for epidemiologists

Chinese journals in epidemiology, preventive medicine and public health contain much that is of potential international interest. However, few non-Chinese speakers are acquainted with this literature. This article therefore provides an overview of the contemporary scene in Chinese biomedical journal publication, Chinese bibliographic databases and Chinese journals in epidemiology, preventive medicine and public health. The challenge of switching to English as the medium of publication, the development of publishing bibliometric data from Chinese databases, the prospect of an Open Access publication model in China, the issue of language bias in literature reviews and the quality of Chinese journals are discussed. Epidemiologists are encouraged to search the Chinese bibliographic databases for Chinese journal articles.

The Chinese have had a long history in infectious disease control, and records of epidemics can be traced back two millennia [1] . Since the introduction of modern medicine by missionary doctors in the 19 th century [2] , modern epidemiological studies have been conducted in China, first by Western doctors, and then gradually superseded by their Chinese colleagues in the 1930s [3] . Since the 1950s, huge reductions in the incidence of infectious diseases like measles and schistosomiasis have been achieved through national vaccination programmes and environmental intervention programmes [1, 4] . The adoption of the Open Door Policy in 1978 marked the beginning of remarkable social and economic development unprecedented in China's modern history. However, rapid industrialization and urbanization are accompanied by many social problems, from the increasing rich-poor, urban-rural, coastal-interior disparity to heavy environmental pollution. Changes in disease profile with the increasing burden of non-communicable diseases as a result of an aging population with a successful one-child policy posed new challenges in the 21 st century [1] . The SARS epidemic in 2003 exposed how a lack of transparency and delayed dissemination of information on the part of the Chinese government made an epidemic of then unknown aetiology a global problem [3] . Epidemiologists from the non-Chinese world may wonder what resources of scientific knowledge and epidemiological information China (whose health research serves a fifth of the world's population) may offer us. In 1994, the British Medical Journal published an editorial recommending to its readers the Chinese medical journals [5] . However, 13 years have gone by, and the Chinese medical and scientific literature is still largely terra incognita outside China [6] . Recent enthusiasm among Westerners in learning the Chinese language [7, 8] may rekindle their interest in this untapped resource. As Beijing prepares for the Olympics in 2008 celebrating China's arrival in the modern world, perhaps an update of the development of Chinese biomedical journals may whet the reader's appetite. This paper is intended to serve as a guide. This article will first provide a general overview to Chinese biomedical journals. Next, Chinese bibliographic databases will be described, using Wan Fang and iLib as examples. Chinese journals in epidemiology and public health will then be discussed, followed by a comprehensive examination of issues arising from switching the publication language to English, the effect on impact factors and Open Access. Lastly, the problems of language bias and quality of articles will be discussed. Three appendices are included. Appendix 1 provides additional information on bibliographic indexing of Chinese biomedical journals. Appendix 2 illustrates the historical background to the choice of language of publication using three journals as examples. Appendix 3 is a review of a survey of English language biomedical journals of China previously published in a Chinese journal. For the purpose of this study, Chinese journals and databases discussed here are confined to that of mainland China, excluding Hong Kong, Macau and Taiwan. For a more in-depth study of the research potential of Chinese biomedical bibliographic databases, illustrated by the example of schistosomiasis research, please refer to the paper in this thematic issue by Liu et al. [9] . Today there are more than 5000 academic periodicals published in mainland China, and around a thousand of these are related to biomedicine and health. Seventy-four journals from mainland China were indexed in 2006 Journal Citation Reports ® Science Edition (JCR) published by Thomson Scientific, of which 12 were biomedical journals and two were multi-disciplinary science journals that publish biomedical articles. Of these 14 journals, only one was published in Chinese, while the rest were in English. Eighty-two mainland Chinese journals are indexed for MEDLINE [10, 11] , among which, 62 publish articles in Chinese, 16 in English, one in either English or German, and three in either Chinese or English. Only six of the MEDLINE-indexed mainland Chinese journals receive impact factors from JCR. All six publish articles in English (Table 1) . Altogether, 146 mainland Chinese journals that cover subjects such as general science, biology, medicine, veterinary science, agriculture and forestry, are indexed in the PubMed journal database (some of these are indexed in MEDLINE). Of these 146 journals, 110 publish articles in Chinese, 24 in English and seven in either Chinese or English (with one in Chinese or Latin and one with missing language data). For a detailed discussion, please refer to Appendix 1. Full texts of more than five thousand Chinese journals are now available online. There are six mainland Chinese bibliographic databases through which Chinese language biomedical journal articles can be searched and located and of which two provide English interfaces: (a) Chinese Biomedical Literature Database (CBM) [12] , (b) Chinese Medical Current Content (CMCC) [13] , Users of traditional Chinese characters can use Yahoo! Taiwan Academia Search [20] whose mainland Chinese journal article entries are provided by iLib. In addition, Google Scholar [21, 22] , as a multi-lingual bibliographic database, also facilitates searches in the Chinese language (Table 2) As a recent paper [23] has given a detailed description and analyses of five of the Chinese bibliographic databases, the following discussion is restricted to three of them: Google scholar as related to searches in Chinese has not yet been covered by any academic paper in English and the same is true of iLib, which is not covered by [23] ; Wan Fang database, which is freely available through terminals in the British Library, will be used as an example to illustrate the wealth of biomedical journals available to us through the internet. Google Scholar provides a convenient starting point for searching Chinese articles, of which the bibliographic data is mainly provided by VIP information, Wan Fang database and iLib (all accessed on 21 st February, 2007). For Chinese speakers, Google Scholar also provides a Chinese interface [22] . There are two apparent advantages (especially for non-Chinese speakers) of searching for Chinese articles in Google Scholar. Firstly, Google Scholar (Chinese interface) provides 'pinyin search', i.e. using a standardised Romanised form of Chinese, known as pinyin in Chinese [24] . For example, if I type 'bing du' in the Google Scholar English interface, I will obtain journal articles with authors of the family name Bing Du. However, if I use the Chinese In additional to these functions, Google Scholar also provides links to institutional libraries and the British Library, citation records, links to related articles, and it groups different entries of the same article together. For a more structured search, the Advanced Scholar Search is needed, of which a Chinese interface is also available [25] . As of 13 th February 2008, the Chinese links in Google Scholar provided by VIP information are linked to the PDF full text which requires subscription to VIP information. If the user is not covered by subscription, the link will be redirected to the webpage on which the title, author, abstract and keywords (all in Chinese) are displayed. The full text can then be purchased individually. Chinese links in Google Scholar provided by the Wan Fang database and iLib will directly lead to the Chinese abstract page. From there a link is provided to the full text PDF file which requires payment or subscription. Although a previous study performed in 2005 found an English language bias in Google Scholar [26] , the search engine has evolved so quickly that a new study of its article coverage is definitely worthwhile. Both Wan Fang database and iLib are run by Wanfang Data, an affiliate of the Chinese Ministry of Science & Technology (cf. [27] ). While Wan Fang provides access to databases of journal articles, conference proceedings, degree theses, patents, national and industrial standards and even listed companies in China, iLib is essentially a subset of Wan Fang and is restricted to journal articles only. The Wan Fang database maintains two portals, one in Chinese [17] and one in English [18]. Cross-searches of different databases (e.g. journal articles and conference proceedings) using simplified Chinese in the domestic portal and English in the international one, are available. The advantage of iLib over Wan Fang for journal article searches is that the interface of iLib is more user-friendly and, unlike Wan Fang, there are links to the author, the journal issue, the journal, the references cited in the paper and some related papers in the iLib database, similar to the AbstractPlus format of PubMed. Like PubMed, the Wan Fang databases or iLib can be searched for free. However, only Chinese abstracts are available for free in HTML format. Although many Chinese journals provide English abstracts to their articles nowadays, these English abstracts are not uploaded onto the public domain by Wan Fang or iLib. To access the English abstract online, one has to download the PDF full text which requires subscription. The only exceptions are those indexed by PubMed, through which they are freely available. A difference in the search mechanism is that in Wan Fang, one has to choose whether to search the English Online Journals category or the China Online Journals (Chinese language journals) in the first place, while in iLib, there is no separation of the journals by language. Thus, if one types 'influenza' in iLib, one will find articles published in Chinese language journals (as the English titles of the Chinese articles are actually being searched) as well as in English language journals. Subscription or payment for full text of mainland Chinese journal articles For individual users, there are various methods of payment. However, most (if not all) of these methods apply only to users in mainland China. While VIP information accepts VISA card online payment, Wan Fang and iLib do not accept any credit cards; they accept only bank cards issued in mainland China or payment through a mainland Chinese mobile phone company, remittance via post offices or banks, or some 'pay-asyou-download cards', which provides you with a password to top-up your download credit online, using your personal Wan Fang or iLib account. To the knowledge of the author, as of 14 th February 2007, the British Library has subscriptions to full text (PDF files) of all academic journals (both English language journals and Chinese language journals) available in the Wan Fang database (around 5700 periodicals). Readers have access to these journals through the computer terminals in the library. Below I describe in more detail what is available in the Wan Fang database. There are 141 titles under the category of English China Online Journals. According to Wan Fang categories, eight are on agriculture, 58 on fundamental science, 24 on health and medical science, 48 on science & technology and three on social science, as of 14 August 2007 [28] . Table 3 lists 24 English language journals on health and medical science available in Wan Fang. A full list in alphabetical order is available in [29] . Chinese language journals in the Wan Fang database Under the category of China Online Journals, there are more than 5600 titles (5638 as of March 2007). When subdivided into five categories, over a thousand titles are found to be related to health, medicine and biology (1056 as of June 2007) [30] . Currently, it is the Chinese national standard that scientific periodicals published in mainland China in the Chinese language should contain English abstracts for every original research article and English titles for selected important articles (e.g. editorials, reviews, forums and short research articles, depending on the judgement of the editorial board) [31] . The English table of contents is available online, free of charge, through the Wan Fang database. However, the English abstracts are only available in the full text PDF file from the Wan Apart from impact factors published by Thomson Scientific in JCR, VIP Information also publishes bibliometric data of some of the journals indexed in its database [40] . A handful of journals listed in Table 4 have their VIP impact factor and immediacy index available, which can contribute towards evaluation of their quality. Two tiers: national and provincial Guangxi Zhuang Autonomous Region in southern China now has an international board of editors [31] . The development of the internet has prompted a drastic change in the ecology of academic publication worldwide and Chinese journals are no exception. Some observers may note that the purpose of publishing provincial journals may be to present epidemiological findings mainly of local use and serve as a local publication outlet. However, as all of these journals are now available online, the original raison d'être of provincial journals to foster the exchange of research output on a provincial level may diminish. A doctor from Sichuan can now easily download a paper published in the Shanghai Journal of Preventive Medicine, while a scientist from Guangzhou (Canton) can easily publish his/her paper in the Zhejiang Journal of Preventive Medicine. One can imagine fierce competition for good research papers among these journals in the near future and through the invisible hand of the market, some journals may prosper and attain international status while others may wither and die. An interesting exception to the two tiers of national and provincial journals is the Journal of Preventive Medicine of Chinese People's Liberation Army, in which research articles related to public health issues in a military context, from hygiene in training camps to the temperature inside tanks, are published. Apart from those, there are also articles on civilian public health issues written by scientists in the military academy. Table 5 lists 23 journals related to tropical medicine, including journals in parasitology, HIV and tuberculosis. Table 6 lists 23 journals on non-communicable diseases, medical statistics, school health, occupational health, port/ frontier health and quarantine, evidence-based medicine and reproductive health and family planning. All but one are published in Chinese. The exception, the Journal of Reproduction and Contraception is published in English with a sister publication, Reproduction and Contraception, published in Chinese [42] . Some of these journals have been listed as Chinese core journals in 2004: three in parasitology (Table 5) , one in medical statistics, one in school health, five in Emerging Themes in Epidemiology 2008, 5:20 http://www.ete-online.com/content/5/1/20 occupational health and two in reproductive health and family planning (Table 6 ). Specialist national journals like the Chinese Journal of Parasitology and Parasitic Diseases, of the Chinese Preventive Medicine Association (CPMA) series [43] , publish papers of high academic standard in their respective specialties. Articles of epidemiological relevance may also be found in medical university journals. It is common among mainland Chinese universities to publish university journals that contain predominantly their own research outputs. These journals number around 2000 in total, of which nearly half belong to natural sciences [44] . University journals published by medical universities or medical faculties of comprehensive universities cover the whole spectrum of medical specialities. Some are indexed in MEDLINE, like the Journal of Peking University (Health Sciences) (cf. Table 1 ). The performances of 41 of these medical university journals have been analysed recently [36] and they varied greatly. Reform proposals have been suggested [44] . [48] . On the contrary, in mainland China, with its huge population and considerable number of scientists and medical professionals, the internal market for biomedical journals is substantial enough to sustain a sizable number of Chinese language journals. Thus I suggest that the size of the prospective market (as a result of the linguistic and political divide) plays a significant role in shaping the language trend of the world's journal publication. Through the international language of scientific communication, English language journals provide a platform for Chinese (and foreign) scientists with a broad international readership. Hopefully, some of these journals will manage to receive their impact factors from JCR. However, by switching to English and internationalising their scope (e.g. by dropping the word 'Chinese' from their titles), they face severe competition from their counterparts in North America and Europe. Nevertheless, there are a few successes so far, like Cell Research and the World Journal of Gastroenterology: WJG [49] that are now indexed in Science Citation Index Expanded and MEDLINE. As the Chinese share of the world's scientific output increases and as Chinese scientists become more fluent in English, more English language biomedical journals published in China will receive their limelight in the international arena. For further discussion, please refer to Appendix 3. The increasing trend of using impact factors published by Thomson Scientific as an indicator in academic evaluation in universities and research institutes has received much criticism from non-English-speakers of the developing world [50] . One of the criticisms against it is the alleged language bias of the Thomson Scientific database coverage towards journals published in English and in the industrialised world [51, 52] . The 'what-if' scenarios of inclusion of non-Science Citation Index (SCI)-indexed journals upon the impact factors of SCIindexed journals have been studied and the 'hypothetical' impact factors of the non-SCI-indexed journals calculated [52, 53] . In order to better evaluate the performances of Latin American journals, SciELO publishes bibliometric indices of its own, similar to that of Thomson Scientific, using data from its database which reflect more the regional context [53] . Brazilians can now evaluate their journals using the SciELO impact factor, rather than relying solely on that published in JCR [54] . Should the Chinese do the same? At the moment VIP Information publishes bibliometric indices using data from its own database [40] (cf. Table 4 ). These data should be used in our evaluation of the quality of Chinese journals, especially in our fields of epidemiology and public health, as hardly any of these are indexed in Thomson Scientific database. Currently only a sub-set of Chinese journals receive their impact factors from VIP Information. Hopefully, in the future, more journals will receive their bibliometric data, perhaps not only from VIP Information alone, but pooling data from the other Chinese databases as well. In the long run, I envision an international collaboration between Thomson Scientific, SciELO, the Chinese databases and other bibliographic databases to provide authors and editors alike with a more accurate and comprehensive bibliometric data of journal performance by collating data across the various databases. Open Access (OA) online publishing in China falls into two categories: non-peer-reviewed and peer-reviewed [55] . The former provides an online interface for authors to publish their papers directly online, without peerreview or other form of quality control. Examples include Qiji.cn [56] , the Chinese Preprint Service System [57] and Sciencepaper Online [58] . The latter transfers paper-based peer-reviewed journals onto the web for free access (usually in PDF format). The Alliance of open access journals (OAJs) [59] , sponsored by the Society of China University Journals in Natural Sciences, provides access to a number of OA journals, predominantly Chinese university journals. The international Directory of Open Access Journals (DOAJ) [60] also provides links to the websites of individual Chinese OA journals, including the Chinese Medical Journal [33] . Emerging Themes in Epidemiology 2008, 5:20 http://www.ete-online.com/content/5/1/20 However, to date the proportion of mainland Chinese journals adopting the OA publishing model is small. One possible reason is that Chinese bibliographic databases, unlike their Western counterparts, provide subscribed readers with PDF full text on behalf of the journals at an affordable rate -CNY three yuans (equivalent to USD 39 cents, as of 17 August 2007 [61] ) per paper. Thus, the cost of setting up and maintaining an individual website for a journal may seem to be a potential financial disincentive. Furthermore, most OA journals, like that of BioMed Central [62] and Public Library of Science [63] , adopt an authorpay model. In the mainland Chinese context where research funding is inadequate, more often than not, authors are less willing to pay for publication in OA journals. While OA journals in the West can grant waivers to authors from low income countries because their overhead costs are met by membership fees and article processing charges paid by universities and authors from the West, for most Chinese journals this will be difficult as most of their authors come from mainland China. However, there are reasons to believe that many mainland Chinese authors welcome the development of OA publishing [64] . Given the current limited accessibility of full text Chinese journal articles from outside China, OA journals may prove to be an option for rapid scientific communication between authors and readers from within China and without. One may ask why bother with Chinese journal articles after all. Apart from those who do field work in China, what important epidemiological information does the Chinese literature offer us? Avoid language bias Perhaps one important application is to avoid language bias in our literature reviews [65] . Back in 1995, Grégoire et al. [66] found that among the 36 consecutive metaanalyses that they analysed, one would produce a different conclusion had it not excluded studies based on linguistic reasons. Comparing English and German journals, Egger et al. [67] found that randomised controlled trials (RCTs) were more likely to publish in English language journals if they gave statistically significant results. This led to the worry that language bias could be introduced to reviews and meta-analyses restricted to data published in English, leading to distorted results. However, subsequent studies [68] [69] [70] found little evidence supporting this assertion. Pham et al. found that language bias led to an under-estimation of the protective effect of intervention in RCTs in complementary and alternative medicine (CAM) systematic reviews but not in that of conventional medicine [70] . Regarding the quality of reports, trials and systematic reviews published in English and those in languages other than English (LOE), are similar [71] [72] [73] . Inclusion of studies published in LOE in systematic reviews and meta-analyses is "likely to increase precision and may reduce systematic errors" [72] , but financial budget and time constraints should also be taken into account [70] . The quality of articles published in Chinese medical journals has led to debates in Western academia. The conclusion of a recent systematic review on the clinical effectiveness of treatment with hyperbaric oxygen for neonatal hypoxic-ischaemic encephalophathy that the "Chinese medical literature may be a rich source of evidence to inform clinical practice and other systematic reviews" [73] was disputed. In an online rapid response, Peter C. Gotzsche ("No double standards in research, please" dated 26 th August 2006) argued that Liu et al. had provided no evidence for their statement. The standard adopted by Cochrane and CONSORT by which the Chinese trials identified in [73] are judged to be of poor quality, are not "Western" as declared by Liu et al. since they are adopted internationally, including by the Chinese Cochrane Centre. Gotzsche also cited two reviews [74, 75] to argue that "Chinese trials are far more positive, on average, than trials performed in other countries". In another study, Wang and Zhang found that by 1995, the "frequency of using statistical tests in Chinese medical journals appears comparable to that in other parts of the world", but "the lack or inappropriate use of statistics remains a problem" [76] . In spite of this scepticism, the present author agrees with Smith that Chinese medical journals are "a treasure house of medical science available for explorers" [5] provided that we evaluate the evidence published therein with no double standard. There are examples of reviews that cover Chinese journals and evaluate the evidence available, e.g. in a recent review on the effectiveness of hand-washing in preventing SARS, among the ten case-control studies identified, four were published in Chinese journals [77] . Chinese journals are a mine of epidemiological information that is yet to be explored by the outside world. Thanks to the development of the internet and bibliographic databases, they can now be explored with relative ease. It has been suggested that in order to be comprehensive, we should apply LILACS in our literature search to cover Spanish and Portuguese articles in our systematic Emerging Themes in Epidemiology 2008, 5:20 http://www.ete-online.com/content/5/1/20 reviews [78, 79] . Perhaps it is time to add to our list the Chinese databases and also include Chinese papers. In 1990, Gastel and Weng [80] published a detailed overview of Chinese medical journals written for Western readers. At that time, the number of medical journals published in China was estimated to be 500, rising to 700 only four years later [5] . In 2007, around 1000 titles related to biomedicine and health, from more than five thousand academic periodicals, were published in mainland China. To see this in a bigger picture, let us take Journal Citation Reports ® (JCR) and MEDLINE as bench marks. Chinese journals (including those published by foreign publishers and thus registered the country of publication of its publisher). There were a few records whose country of publication data were missing or mistaken and were corrected for in the analysis. According to a recent study [83] , from 1932 to 2005, there were 1093 journals (537 titles current in 2005) from greater China, including Hong Kong (n = 7, 0.6%) and Taiwan (n = 58, 5.3%) indexed in Chemical Abstracts, among which 51 (4.7%) belonged to biological sciences and 216 were health-related (yiyaoweisheng), i.e. category Q and R according to the Chinese Library Classification [84] . English language journals made up of 9.7% (n = 106) of the total. The majority of the journals indexed were established in or after 1980 (69.2%). The first journal being indexed was Chinese Medical Journal (see Appendix 2) . Up to 14 th October 2005, a total of 693610 articles had been indexed. A full list of the indexed journals with their ISSN, indexed years and number of indexed articles can be found at [85] . [94] . Here, we observe the colonial footprint of the Japanese in the early history of Journal of the Formosan Medical Association. The end of Japanese rule brought an end to the use of Japanese among Taiwanese doctors. The decision to publish in Chinese coincided with the tide of decolonisation. The decision to split the journal into two, one in English for the publication of research output and one in Chinese for the continuing medical education at the grass-roots level, is a classic example of the dilemma between serving local needs and communicating research outputs to the world. [98] . The above section illustrated how Chinese doctors adapt to changes in the socio-political arena by their choices of language of publication of their flagship general medicine journals. The Chinese Medical Journal, as the only English medical journal published in mainland China for many years, provided a showcase of Chinese medical achievement to the world and a channel of communication between Chinese doctors and their colleagues aboard. The Journal of the Formosan Medical Association demonstrated the process of colonisation, decolonisation and internationalisation in its switch of language from Japanese to Chinese and then to English. The Hong Kong Medical Journal is an example of how a colonial legacy has left a language heritage that fosters internationalisation in this globalising world. The choice of language of publication by biomedical journals is often a consequence of many different sociopolitical factors. [101] . The survey covered 31 journals and gave a good summary of their basic information, bibliometrics, and details of their management, editorial board, publication and distribution. As this survey was published in Chinese with no English abstract available, and is therefore, less accessible to the average English-speaking readers, a review highlighting its major findings that are relevant to our present study will be of benefit to interested readers and is provided below. While 19 of these 31 English language journals reported an adequate supply of submitted manuscripts, nine reported that theirs were inadequate (three did not reply to this question). Major reasons for this inadequacy were (1) dearth of scientific research output leading to dearth of manuscripts; (2) huge amount of high-quality manuscripts being drained to foreign journals; and (3) the limited capacity of writing in English on the part of some authors. The annual number of manuscripts received varied greatly (Table 7) . Eighteen journals received manuscripts from outside China, ranging from a few manuscripts per year to 30% of its total number of manuscripts received. The majority of these submissions came from other developing countries. One of the consequences of low supply of high-quality manuscripts is that the frequency of publication of journals in China is low: 19 of the 31 English language journals are quarterly or semi-annual (Table 8 ) [101] . Another area that awaits improvement, according to Yu et al. [101] , is the unequal distribution and nonspecialisation of journals in China. Nearly half of the 31 English language journals surveyed are general medical journals, while there are no English language journals from China that are specialised in fields like epidemiology and preventive medicine. As it has been suggested [44] , general journals should merge to raise their profile while others should specialise to avoid overlap in disciplines. Among these 31 journals, the only full-time Editor-in-Chief is that of the World Journal of Gastroenterology: WJG. All the others work part-time for the journals. Yu et al. [101] argued that this was very disadvantageous to the development of these journals. The number of full-time editors varied across the 31 journals (Table 8 ), while 13 journals had additional part-time editors varying from one to six. Yu et al. [101] commented that in China, there are very few editors who have high academic qualifications in biomedical sciences and at the same time are proficient in the English language. More training is needed. In order to attract talent and prevent further brain-drain, Yu et al. [101] suggested that scientific editors in China should receive the same pay and benefits as scientific researchers to remove the impression that editors are second-class scientific professionals. Twenty-four of the 31 journals received funding from the government; 20 had page-charges; six received remuneration from advertisements; six received sponsorship from the National Fund for Natural Sciences; and eight received sponsorship from other sources. Seven journals ran a deficit balance; 14 achieved breakeven and two made profits (no reply from eight journals). Given these disturbing facts, Yu et al. [101] suggested that while the Chinese government should increase its financial investment in these journals, the editorial boards should also learn how to manage the journals more efficiently. International peer-review has been archived by thirteen of the 31 journals. Peer reviewers were drawn mainly from the West and Japan. Non-Chinese editors are found in eight journals. Apart from two Germans, they all come from English-speaking countries, and their number is limited to one per journal, with one exception which has three non-Chinese editors. All 31 journals studied are now online, of which eleven have their own websites. To different extents, they all manage their editorial process of submission, peer-review and re-submission electronically [101] . According to Yu et al. [101] , the crux of the problem of journals in mainland China is that their model of operations remains that of planned economy, rendering them unfit to compete in today's Chinese market economy. As of 2006, eighteen of the 31 journals were distributed internationally, mainly through the agency of international publishing groups, like Elsevier, Nature, Springer and Blackwell. Through collaboration with these publishing groups, Chinese journals can benefit in terms of efficiency, economy of scales and share of the international market. Not only does this illustrate the feasibility of international collaboration, but it also provides Chinese publishers a model of development into a commercially viable publishing group of scientific periodicals. However, only seven of these 18 journals achieved an international circulation of more than 100 copies. This reflects the difficulty of breaking through into the international market. Publishing in English is correlated to higher international visibility [6] . Using data of 2003, Yu et al. [101] showed that English language biomedical journals of China were more likely to be indexed in international databases than their Chinese language counterparts (Table 9 ). However, compared to the international English language journals, their impact was rather low (as indicated by their low impact factor in JCR). Interestingly, by moving towards an international readership, these English language journals of China fared not so well in China either (

The Complete Genome and Proteome of Laribacter hongkongensis Reveal Potential Mechanisms for Adaptations to Different Temperatures and Habitats

Laribacter hongkongensis is a newly discovered Gram-negative bacillus of the Neisseriaceae family associated with freshwater fish–borne gastroenteritis and traveler's diarrhea. The complete genome sequence of L. hongkongensis HLHK9, recovered from an immunocompetent patient with severe gastroenteritis, consists of a 3,169-kb chromosome with G+C content of 62.35%. Genome analysis reveals different mechanisms potentially important for its adaptation to diverse habitats of human and freshwater fish intestines and freshwater environments. The gene contents support its phenotypic properties and suggest that amino acids and fatty acids can be used as carbon sources. The extensive variety of transporters, including multidrug efflux and heavy metal transporters as well as genes involved in chemotaxis, may enable L. hongkongensis to survive in different environmental niches. Genes encoding urease, bile salts efflux pump, adhesin, catalase, superoxide dismutase, and other putative virulence factors—such as hemolysins, RTX toxins, patatin-like proteins, phospholipase A1, and collagenases—are present. Proteomes of L. hongkongensis HLHK9 cultured at 37°C (human body temperature) and 20°C (freshwater habitat temperature) showed differential gene expression, including two homologous copies of argB, argB-20, and argB-37, which encode two isoenzymes of N-acetyl-L-glutamate kinase (NAGK)—NAGK-20 and NAGK-37—in the arginine biosynthesis pathway. NAGK-20 showed higher expression at 20°C, whereas NAGK-37 showed higher expression at 37°C. NAGK-20 also had a lower optimal temperature for enzymatic activities and was inhibited by arginine probably as negative-feedback control. Similar duplicated copies of argB are also observed in bacteria from hot springs such as Thermus thermophilus, Deinococcus geothermalis, Deinococcus radiodurans, and Roseiflexus castenholzii, suggesting that similar mechanisms for temperature adaptation may be employed by other bacteria. Genome and proteome analysis of L. hongkongensis revealed novel mechanisms for adaptations to survival at different temperatures and habitats.

Laribacter hongkongensis is a recently discovered, Gram-negative, facultative anaerobic, motile, seagull or S-shaped, asaccharolytic, urease-positive bacillus that belongs to the Neisseriaceae family of bproteobacteria [1] . It was first isolated from the blood and thoracic empyema of an alcoholic liver cirrhosis patient in Hong Kong [2] . In a prospective study, L. hongkongensis was shown to be associated with community acquired gastroenteritis and traveler's diarrhea [3, 4] . L. hongkongensis is likely to be globally distributed, as travel histories from patients suggested its presence in at least four continents: Asia, Europe, Africa and Central America [4] [5] [6] . L. hongkongensis has been found in up to 60% of the intestines of commonly consumed freshwater fish, such as grass carp and bighead carp [4, 7, 8] . It has also been isolated from drinking water reservoirs in Hong Kong [9] . Pulsed-field gel electrophoresis and multilocus sequence typing showed that the fish and patient isolates fell into separate clusters, suggesting that some clones could be more virulent or adapted to human [8, 10] . These data strongly suggest that this bacterium is a potential diarrheal pathogen that warrants further investigations. Compared to other families such as Enterobacteriaceae, Vibrionaceae, Streptococcaceae, genomes of bacteria in the Neisseriaceae family have been relatively under-studied. Within this family, Neisseria meningitidis, Neisseria gonorrhoeae and Chromobacterium violaceum are the only species with completely sequenced genomes [11] [12] [13] . In view of its potential clinical importance, distinct phylogenetic position, interesting phenotypic characteristics and the availability of genetic manipulation systems [14] [15] [16] [17] , we sequenced and annotated the complete genome of a strain (HLHK9) of L. hongkongensis recovered from a 36-year old previously healthy Chinese patient with profuse diarrhea, vomiting and abdominal pain [4] . Proteomes of L. hongkongensis growing at 37uC (body temperature of human) and 20uC (average temperature of freshwater habitat in fall and winter) [9] were also compared. The complete genome of L. hongkongensis is a single circular chromosome of 3,169,329 bp with a G+C content of 62.35% ( Figure 1 ). In terms of genome size and number of predicted coding sequences (CDSs), rRNA operons and tRNA genes (Table 1) , L. hongkongensis falls into a position intermediate between C. violaceum and the pathogenic Neisseria species. A similar intermediate status was also observed when the CDSs were classified into Cluster of Orthologous Groups (COG) functional categories, except for genes of RNA processing and modification (COG A), cell cycle control, mitosis and meiosis (COG D), replication, recombination and repair (COG L) and extracellular structures (COG W), of which all four bacteria have similar number of genes ( Figure 2 ). This is in line with the life cycles and growth requirements of the bacteria. C. violaceum is a highly versatile, facultative anaerobic, soil-and water-borne free-living bacterium and therefore requires the largest genome size and gene number. The pathogenic Neisseria species are strictly aerobic bacteria with human as the only host and therefore require the smallest genome size and gene number. L. hongkongensis is a facultative anaerobic bacterium that can survive in human, freshwater fish and 0-2% NaCl but not in marine fish or $3% NaCl and therefore requires an intermediate genome size and gene number. The L. hongkongensis genome lacks a complete set of enzymes for glycolysis, with orthologues of glucokinase, 6-phosphofructokinase and pyruvate kinase being absent (Table S1 ). This is compatible with its asaccharolytic phenotype and is consistent with other asaccharolytic bacteria, such as Campylobacter jejuni, Bordetella pertussis, Bordetella parapertussis and Bordetella bronchiseptica, in that glucokinase and 6-phosphofructokinase are also absent from their genomes [18, 19] . On the other hand, the L. hongkongensis genome encodes the complete sets of enzymes for gluconeogenesis, the pentose phosphate pathway and the glyoxylate cycle (Table S1) . Similar to C. jejuni, the L. hongkongensis genome encodes a number of extracellular proteases and amino acid transporters. These amino acids can be used as carbon source for the bacterium. The genome encodes enzymes for biosynthesis of the 21 genetically encoded amino acids and for biosynthesis and b-oxidation of saturated fatty acids (Tables S2 and S3 ). The L. hongkongensis genome encodes a variety of dehydrogenases (LHK_00527-00540, LHK_01219-01224, LHK_02418-02421, LHK_00801-00803, LHK_01861, LHK_02912-02913 and LHK_00934) that enable it to utilize a variety of substrates as electron donors, such as NADH, succinate, formate, proline, acyl-CoA and D-amino acids. The presence of three terminal cytochrome oxidases may allow L. hongkongensis to carry out respiration using oxygen as the electron acceptor under both aerobic conditions [type aa 3 oxidase (LHK_00169-00170, LHK_00173)] and conditions with reduced oxygen tension [type cbb 3 (LHK_00995-00996, LHK_00998) and type bd (LHK_02252-02253) oxidases]. The genome also encodes a number of reductases [fumarate reductase (LHK_02340-02342), nitrate reductase (LHK_02079-02085), dimethylsulfoxide (DMSO) reductase (LHK_02496-02498) and tetrathionate reductase (LHK_01476-01478)], which may help carry out respiration with alternative electron acceptors to oxygen (fumarate, nitrate, DMSO and tetrathionate) under anaerobic conditions. This is supported by the enhanced growth of L. hongkongensis under anaerobic conditions in the presence of nitrate (data not shown). Further studies are required to confirm if the bacterium can utilize other potential electron acceptors. There were 441 transport-related proteins (13.6% of all CDSs) in the L. hongkongensis genome, comprising an extensive variety of transporters, which may reflect its ability to adapt to the freshwater fish and human intestines, and freshwater environments. According to the Transporter Classification Database (TCDB) (http:// www.tcdb.org/), all seven major categories of transporters are present in L. hongkongensis. Primary active transporters (class 3 transporters) were the most abundant class of transporters, accounting for 43.3% (191 CDSs) of all annotated CDSs related to transport, among which 104 belong to the ATP-binding cassette (ABC) transporter superfamily and 41 were oxidoreduction-driven transporters. Electrochemical potential-driven transporters (class 2 transporters) were the second most abundant class of transporters, accounting for 27.9% (123 CDSs) of all annotated CDSs related to transport, most of which (117 CDSs) are various kinds of porters including major facilitator superfamily (MFS) (19 CDSs), resistance-nodulation-cell division (RND) superfamily (22 CDSs), amino acid-polyamine-organocation family (8 CDSs), dicarboxylate/amino acid:cation symporter (DAACS) family (5 CDSs) and monovalent cation:proton antiporter-2 family (3 CDSs), and various heavy metal transporters which may be involved in detoxification and resistance against environmental hazards. Three different types of class 2 transporters, belonging to the DAACS, tripartite ATP-independent periplasmic transporter and Laribacter hongkongensis is a recently discovered bacterium associated with gastroenteritis and traveler's diarrhea. Freshwater fish is the reservoir of L. hongkongensis. In order to achieve a rapid understanding on the mechanisms by which the bacterium adapts to different habitats and its potential virulence factors, we sequenced the complete genome of L. hongkongensis, compared its gene contents with other bacteria, and compared its gene expression at 37uC (human body temperature) and 20uC (freshwater habitat temperature). We found that the gene contents of L. hongkongensis enable it to adapt to its diverse habitats of human and freshwater fish intestines and freshwater environments. Genes encoding proteins responsible for survival in the intestinal environments, adhesion to intestinal cells, evasion from host immune systems, and putative virulence factors similar to those observed in other pathogens are present. We also observed, in gene expression studies, that L. hongkongensis may be using different pathways for arginine synthesis regulated at different temperatures. Phylogenetic analysis suggested that such mechanisms for temperature adaptation may also be used in bacteria found in extreme temperatures. C 4 -dicarboxylate uptake C family, are likely involved in the transport of malate, which can be used as the sole carbon source for L. hongkongensis in minimal medium [unpublished data]. The remaining class 2 transporters were ion-gradient-driven energizers belonging to the TonB family (6 CDSs). The third most abundant class of transporters was the channels and pores (class 1), with 39 CDSs including 12 a-type channels, 26 b-barrel porins. Among the 12 a-type channels, four were mechanosensitive channels and G+C content (10-kb window with 100-b step); circles 3 to 7, red, light purple, orange, aqua and teal bars show BLAST hits to Neisseria gonorrhoeae FA 1090, Neisseria gonorrhoeae MC58, Neisseria gonorrhoeae FAM18, Neisseria gonorrhoeae Z2491 and Chromobacterium violaceum ATCC 12472, respectively; circle 8, green arcs show location of eight putative prophages; circles 9 and 12, colors reflect Cluster of Orthologous Groups of coding sequences (CDSs). Maroon, translation, ribosomal structure and biogenesis; navy, transcription; purple, DNA replication, recombination and repair; light brown, cell division and chromosome partitioning; aqua, posttranslational modification, protein turnover, chaperones; teal, cell envelope biogenesis, outer membrane; blue, cell motility and secretion; orange, inorganic ion transport and metabolism; light purple, signal transduction mechanisms; olive, energy production and conversion; lime, carbohydrate transport and metabolism; green, amino acid transport and metabolism; fuchsia, nucleotide transport and metabolism; light pink, coenzyme metabolism; red, lipid metabolism; yellow, secondary metabolites biosynthesis, transport and catabolism; gray, general function prediction only; silver, function unknown; circles 10 and 11, dark blue, dark red and dark purple indicate CDSs, tRNA and rRNA on the 2 and + strands, respectively. doi:10.1371/journal.pgen.1000416.g001 which are important for mediating resistance to mechanophysical changes. The remaining transporters belong to four other classes, namely group translocators (class 4, 9 CDSs), transport electron carriers (class 5, 16 CDSs), accessory factors involved in transport (class 8, 9 CDSs) and incompletely characterized transport system (class 9, 54 CDSs). In line with their asaccharolytic nature, the genomes of L. hongkongensis and C. jejuni do not contain genes that encode a complete phosphotransferase system. The five families of multidrug efflux transporters, including MFS (6 CDSs), RND (8 CDSs), small multidrug resistance family (2 CDSs), multidrug and toxic compound extrusion family (2 CDSs) and ABC transporter superfamily (5 CDSs), were all present in L. hongkongensis, which may reflect its ability to withstand toxic substances in different habitats [20] . 20 CDSs were related to iron metabolism, including hemin transporters, ABC transporters of the metal type and ferrous iron, iron-storage proteins and the Fur protein responsible for iron uptake regulation. In contrast to C. violaceum which produces siderophores for iron acquisition, but similar to the pathogenic Neisseria species, proteins related to siderophore formation are not found in L. hongkongensis genome. In addition to a TonB-dependent siderophore receptor (LHK_00497), a set of genes (LHK_01190, LHK_01193, LHK_01427-1428) related to the transport of hemin were present, suggesting that L. hongkongensis is able to utilize exogenous siderophores or host proteins for iron acquisition, which may be important for survival in different environments and hosts. Except the first strain of L. hongkongensis isolated from the blood and empyema pus of a patient which represented a non-motile variant, all L. hongkongensis strains, whether from human diarrheal stool, fish intestine or environmental water, are motile with polar flagella. The ability to sense and respond to environmental signals is important for survival in changing ecological niches. A total of 47 CDSs are related to chemotaxis, of which 27 encode methyl-accepting chemotaxis proteins (MCPs) and 20 encode chemosensory transducer proteins. While most MCPs are scattered throughout the genome, the transducer proteins are mostly arranged in three gene clusters ( Figure S1 ). At least 38 genes, in six gene clusters, are involved in the biosynthesis of flagella ( Figure S2 ). Enteric bacteria use several quorum-sensing mechanisms, including the LuxR-I, LuxS/AI-2, and AI-3/epinephrine/norepinephrine systems, to recognize the host environment and communicate across species. Unlike the genomes of C. violaceum and the pathogenic Neisseria species which encode genes involved in LuxR-I and LuxS/AI-2 systems respectively, the L. hongkongensis genome does not encode genes of these 2 systems. Instead, the AI-3/epinephrine/norepinephrine system, which is involved in interkingdom cross-signaling and regulation of virulence gene transcription and motility, best characterized in enterohemorrhagic E. coli [21, 22] , is likely the predominant quorum-sensing mechanism used by L. hongkongensis. Several human enteric commensals or pathogens, including E. coli, Shigella, and Salmonella, produce AI-3 [23] . A two-component system, QseB/C, of which QseC is the sensor kinase and QseB the response regulator, has been found to be involved in sensing AI-3 from bacteria and epinephrine/ norepinephrine from host, and activation of the flagellar regulon transcription [21] . While the biosynthetic pathway of AI-3 has not been discovered, two sets of genes, LHK_00329/LHK_00328 and LHK_01812/LHK_01813, homologous to QseB/QseC were identified in the L. hongkongensis genome, suggesting that the bacterium may regulate its motility upon recognition of its host environment. The presence of two sets of QseB/QseC, one most similar to those of C. violaceum and the other most homologous to Azoarcus sp. strain BH72, is intriguing, as the latter is the only bacterium, with complete genome sequence available, that possesses two copies of such genes. Before reaching the human intestine, L. hongkongensis has to pass through the highly acidic environment of the stomach. In the L. hongkongensis genome, a cluster of genes, spanning a 12-kb region, related to acid resistance, is present. Similar to Helicobacter pylori, the L. hongkongensis genome contains a complete urease gene cluster (LHK_01035-LHK_01037, LHK_01040-LHK_01044), in line with the bacterium's urease activity. Phylogenetically, all 8 genes in the urease cassette are most closely related to the corresponding homologues in Brucella species (a-proteobacteria), Yersinia species (c-proteobacteria) and Photorhabdus luminescens (c-proteobacteria), instead of those in other members of b-proteobacteria, indicating that L. hongkongensis has probably acquired the genes through horizontal gene transfer after its evolution into a distinct species ( Figure S3 ). Upstream and downstream to the urease cassette, adi (LHK_01034) and hdeA (LHK_01046) were found respectively. Their activities will raise the cytoplasmic pH and prevents proteins in the periplasmic space from aggregation during acid shock respectively [24, 25] . In addition to the acid resistance gene cluster, the L. hongkongensis genome contains two arc gene clusters [arcA (LHK_02729 and LHK_02734), arcB (LHK_02728 and LHK_02733), arcC (LHK_02727 and LHK_02732) and arcD (LHK_02730 and LHK_02731)] of the arginine deiminase pathway which converts L-arginine to carbon dioxide, ATP, and ammonia. The production of ammonia increases the pH of the local environment [26, 27] . Similar to other pathogenic bacteria of the gastrointestinal tract, the genome of L. hongkongensis encodes genes for bile resistance. These include three complete copies of acrAB (LHK_01425-01426, LHK_02129-02130 and LHK_02929-02930), encoding the best studied efflux pump for bile salts, and two pairs of genes (LHK_01373-01374 and LHK_03132-03133) that encode putative efflux pumps homologous to that encoded by emrAB in E. coli [28] . Furthermore, five genes [tolQ (LHK_00053), tolR (LHK_03174), tolA (LHK_03173), tolB (LHK_03172) and pal (LHK_03171)] that encode the Tol proteins, important in maintaining the integrity of the outer membrane and for bile resistance, are also present [29] . In the L. hongkongensis genome, a putative adhesin (LHK_01901) for colonization of the intestinal mucosa, most closely related to the adhesins of diffusely adherent E. coli (DAEC) and enterotoxigenic E. coli (ETEC), encoded by aidA and tibA respectively, was observed ( Figure S4 ) [30, 31] . aidA and tibA encode proteins of the autotransporter family, type V protein secretion system of Gramnegative bacteria. All the three domains (an N-terminal signal sequence, a passenger domain and a translocation domain) present in proteins of this family are found in the putative adhesin in L. hongkongensis. Moreover, a putative heptosyltransferase (LHK_01902), with 52% amino acid identity to the TibC heptosyltransferase of ETEC, responsible for addition of heptose to the passenger domain, was present upstream to the putative adhesin gene in the L. hongkongensis genome ( Figure S4 ). In addition to host cell adhesion, the passenger domains of autotransporters may also confer various virulence functions, including autoaggregation, invasion, biofilm formation and cytotoxicity. The L. hongkongensis genome encodes a putative superoxide dismutase (LHK_01716) and catalases (LHK_01264, LHK_01300 and LHK_02436), which may play a role in resistance to superoxide radicals and hydrogen peroxide generated by neutrophils. The same set of genes that encode enzymes for synthesis of lipid A (endotoxin), the two Kdo units and the heptose units of lipopolysaccharide (LPS) are present in the genomes of L. hongkongensis, C. violaceum, N. meningitidis, N. gonorrhoeae and E. coli. Moreover, 9 genes [rfbA (LHK_02995), rfbB (LHK_02997), rfbC (LHK_02994), rfbD (LHK_02996), wbmF (LHK_02799), wbmG (LHK_02800), wbmH (LHK_02801), wbmI (LHK_02790) and wbmK (LHK_02792)] that encode putative enzymes for biosyn-thesis of the polysaccharide side chains are present in the L. hongkongensis genome. In addition to genes for synthesizing LPS, a number of CDSs that encode putative cytotoxins are present, including cytotoxins that act on the cell surface [hemolysins (LHK_00956 and LHK_03166) and RTX toxins (LHK_02735 and LHK_02918)] and those that act intracellularly [patatin-like proteins (LHK_00116, LHK_01938, and LHK_03113)] [32, 33] . Furthermore, a number of CDSs that encode putative outer membrane phospholipase A1 (LHK_00790) and collagenases (LHK_00305-00306, LHK_00451, and LHK_02651) for possible bacterial invasion are present. To better understand how L. hongkongensis adapts to human body and freshwater habitat temperatures at the molecular level, the types and quantities of proteins expressed in L. hongkongensis HLHK9 cultured at 37uC and 20uC were compared. Since initial 2D gel electrophoresis analysis of L. hongkongensis HLHK9 proteins under a broad range of pI and molecular weight conditions revealed that the majority of the proteins reside on the weakly acidic to neutral portion, with a minority on the weak basic portion, consistent with the median pI value of 6.63 calculated for all putative proteins in the genome of L. hongkongensis HLHK9, we therefore focused on IPG strips of pH 4-7 and 7-10. Comparison of the 2D gel electrophoresis patterns from L. hongkongensis HLHK9 cells grown at 20uC and 37uC revealed 12 differentially expressed protein spots, with 7 being more highly expressed at 20uC than at 37uC and 5 being more highly expressed at 37uC than at 20uC (Table 2, Figure 3 ). The identified proteins were involved in various functions (Table 2 ). Of note, spot 8 [N-acetyl-L-glutamate kinase (NAGK)-37, encoded by argB-37] was up-regulated at 37uC, whereas spot 1 (NAGK-20, encoded by argB-20), was upregulated at 20uC (Figures 3, 4A and 4B ). These two homologous copies of argB encode two isoenzymes of NAGK [NAGK-20 (LHK_02829) and NAGK-37 (LHK_02337)], which catalyze the second step of the arginine biosynthesis pathway. The transcription levels of argB-20 and argB-37 at 20uC and 37uC were quantified by real time RT-PCR. Results showed that the mRNA level of argB-20 at 20uC was significantly higher that at 37uC and the mRNA level of argB-37 at 37uC was significantly higher that at 20uC ( Figure 4C and 4D), suggesting that their expressions, similar to most other bacterial genes, were controlled at the transcription level. When argB-20 and argB-37 were cloned, expressed and the corresponding proteins NAGK-20 and NAGK-37 purified for enzyme assays, their highest enzymatic activities were observed at 37-45uC and 45-50uC respectively ( Figure 4E) . Moreover, NAGK-20, but not NAGK-37, was inhibited by 0.25-10 mM of arginine ( Figure 4F ). L. hongkongensis probably regulates arginine biosynthesis at temperatures of different habitats using two pathways with two isoenzymes of NAGK. L. hongkongensis and wild type E. coli ATCC 25922, but not E. coli JW5553-1 (argB deletion mutant), grew in minimal medium without arginine, indicating that L. hongkongensis contains a functional arginine biosynthesis pathway. NAGK-20 is expressed at higher level at 20uC than 37uC, whereas NAGK-37 is expressed at higher level at 37uC than 20uC. Bacteria use either of two different pathways, linear and cyclic, for arginine biosynthesis. Similar to NAGK-20 of L. hongkongensis, NAGK of Pseudomonas aeruginosa and Thermotoga maritima, which employ the cyclic pathway, can be inhibited by arginine as the rate-limiting enzyme for negative feedback control [34] [35] [36] [37] . On the other hand, similar to NAGK-37 of L. hongkongensis, NAGK of E. coli, which employs the linear pathway, is not inhibited by arginine [35, 36] . We speculate that L. hongkongensis can use different pathways with the two NAGK isoenzymes with differential importance at different temperatures of different habitats. Phylogenetic analysis of NAGK-20 and NAGK-37 showed that they were more closely related to each other than to homologues in other bacteria ( Figure 5 ). The topology of the phylogenetic tree constructed using NAGK was similar to that constructed using 16S rRNA gene sequences (data not shown). This suggested that the evolution of argB genes in general paralleled the evolution of the corresponding bacteria, and argB gene duplication has probably occurred after the evolution of L. hongkongensis into a separate species. The requirement to adapt to different temperatures and habitats may have provided the driving force for subsequent evolution to 2 homologous proteins that serve in different environments. Notably, among all 465 bacterial species with complete genome sequences available, only Thermus thermophilus, Deinococcus geothermalis, Deinococcus radiodurans, Roseiflexus castenholzii and Roseiflexus sp. RS-1 possessed two copies of argB, whereas Anaeromyxobacter sp. Fw109-5 and Anaeromyxobacter dehalo- genans 2CP-C possessed one copy of argB and another fused with argJ ( Figure 5 ). The clustering of argB in two separate groups in these bacteria suggests that argB gene duplication has probably occurred in their ancestor, before the divergence into separate species. The prevalence of T. thermophilus, Deinococcus species and Roseiflexus species in hot springs suggested that this novel mechanism of temperature adaptation may also be important for survival at different temperatures in other bacteria. Further experiments on differential expression of the two isoenzymes at different temperatures in these bacteria will verify our speculations. Traditionally, complete genomes of bacteria with medical, biological, phylogenetic or industrial interests were sequenced only after profound phenotypic and genotypic characterization of the bacteria had been performed. With the advance in technology and bioinformatics tools, complete genome sequences of bacteria can be obtained with greater ease. In this study, we sequenced and analyzed the complete genome of L. hongkongensis, a newly discovered bacterium of emerging medical and phylogenetic interest, and performed differential proteomics and downstream characterization of important pathways. In addition, putative virulence factors and a putative novel mechanism of arginine biosynthesis regulation at different temperatures were discovered, further characterization of which will lead to better understanding of their contributions to the survival and virulence of L. hongkongensis, the Neisseriaceae family and other bacteria. A similar ''reverse genomics'' approach can be used for the study of other newly discovered important bacteria. The genome sequence of L. hongkongensis HLHK9 was determined with the whole-genome shotgun method. Three shotgun libraries were generated: one small-insert (2-4 kb) library and one medium-insert (5-6 kb) library in pcDNA2.1, and a largeinsert (35-45 kb) fosmid library in pCC2FOS. DNA sequencing was performed using dye-terminator chemistries on ABI3700 sequencers. Shotgun sequences were assembled with Phrap. Fosmid end sequences were mapped onto the assembly using BACCardI [38] for validation and support of gap closing. Sequences of all large repeat elements (rRNA operons and prophages) were confirmed by primer walking of fosmid clones. The nucleotide sequence for the complete genome sequence of L. hongkongensis HLHK9 was submitted to Genbank under accession number CP001154. Gene prediction was performed by Glimmer [39] version 3.02, and results post-processed using TICO [40] for improving predictions of translation initiation sites. Automated annotation of the finished sequence was performed by a modified version of AutoFACT [41] , supplemented by analysis by InterProScan [42] . Manual curation of annotation results was done with support from the software tool GenDB [43] . In addition, annotation of membrane transport proteins was done by performing BLAST search of all predicted genes against the curated TCDB [44] . Ribosomal RNA genes were annotated using the online RNAmmer service [45] . Putative prophage sequences were identified using Prophage Finder [46] . Frameshift errors were predicted using ProFED [47] . CRISPRs (Clustered Regularly Interspaced Short Palindromic Repeats) were searched by using PILER-CR [48] , CRISPRFinder [49] and CRT (CRISPR recognition tool) [50] . Single colony of L. hongkongensis HLHK9 was inoculated into brain heart infusion (BHI) medium for 16 h. The bacterial cultures were diluted 1:100 in BHI medium and growth was continued at 20uC for 20 h and 37uC for 6 h, respectively, with shaking to OD 600 of 0.6. After centrifugation at 6,5006g for 15 min, cells were lysed in a sample buffer containing 7 M urea, 2 M thiourea and 4% CHAPS. The crude cell homogenate was sonicated and centrifuged at 16,0006g for 20 min. Immobilized pH gradient (IPG) strips (Bio-Rad Laboratories) (17 cm) with pH 4-7 and 7-10 were hydrated overnight in rehydration buffer containing 7 M urea, 2 M thiourea, 4% CHAPS, 1% IPG buffer pH 4-7 (IPG strip of pH 4-7) and pH 6-11 (IPG strip of pH 7-10) (GE Healthcare) and 60 mM DTT with 60 mg of total protein. The first dimension, isoelectric focusing (IEF), was carried out in a Protean IEF cell electrophoresis unit (Bio-Rad Laboratories) for about 100,000 volt-hours. Protein separation in the second dimension was performed in 12% SDS-PAGE utilizing the Bio-Rad Protean II xi unit (Bio-Rad Laboratories). 2D gels were stained with silver and colloidal Coomassie blue G-250 respectively for qualitative and quantitative analysis, and scanned with ImageScanner (GE Healthcare). ImageMaster 2D Platinum 6.0 (GE Healthcare) was used for image analysis. For MALDI-TOF MS analysis, protein spots were manually excised from gels and subjected to in-situ digestion with trypsin, and peptides generated were analyzed using a 4800 Plus MALDI TOF/TOF Analyzer (Applied Biosystems). Proteins were identified by peptide mass fingerprinting using the MS-Fit software (http://prospector.ucsf. edu) and an in-house sequence database of L. hongkongensis HLHK9 proteins generated using the information obtained from the complete genome sequence and annotation. Only spots with at least two-fold difference in their spot volume between 20uC and 37uC and those uniquely detected at either temperature were subjected to protein identification by MALDI-TOF MS analysis. Three independent experiments for each growth condition were performed. Essentiality of Arginine for Growth of L. hongkongensis HLHK9 L. hongkongensis HLHK9 cells were grown in minimal medium M63 [51] supplemented with 20 mM L-malate as carbon source and 19 mM potassium nitrate as nitrogen source, and 1 mM each of vitamin B1 and vitamin B12. The pH of all media was adjusted to 7.0 with KOH. Essentiality of arginine for growth of L. hongkongensis HLHK9 was determined by transferring the bacterial cells to the modified M63 medium with or without 100 mM of Larginine. Escherichia coli ATCC 25922 and JW5553-1 (argB deletion mutant) [52] were used as positive and negative controls respectively. All cultures were incubated at 37uC with shaking for 5 days. Growth in each medium was determined by measuring absorbance spectrophotometrically at OD 600 . The experiment was performed in duplicate. mRNA levels of argB-20 and argB-37 in L. hongkongensis HLHK9 cells grown in 20uC and 37uC were compared. Total RNA was extracted from culture of L. hongkongensis HLHK9 (OD 600 of 0.6) grown in conditions described in proteomic analysis by using RNeasy kit (Qiagen) in combination with RNAprotect Bacteria Reagent (Qiagen) as described by the manufacturer. Genomic DNA was removed by DNase digestion using RNase-free DNase I (Roche). The total nucleic acid concentration and purity were estimated using A 260 /A 280 values measured by NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies). Bacteria were harvested from three independent replicate cultures. cDNA was synthesized by RT using random hexamers and SuperScript III kit (Invitrogen) as described previously [53, 54] (Table S4) . Reactions were first incubated at 50uC for 2 min, followed by 95uC for 10 min in duplicate wells. Reactions were then thermal-cycled in 40 cycles of 95uC for 15 s and 60uC for 1 min. Absolute standard curve method was used for determination of transcript level for each gene. Standard curves were made by using serial dilutions from plasmids containing the target sequences with known quantities. Housekeeping gene RNA polymerase beta subunit, rpoB, was used as an internal control. Triplicate assays using RNAs extracted in three independent experiments confirmed that transcript levels of rpoB were not significantly different (P.0.05) at 20uC compared with 37uC (data not shown). The transcript levels of argB-20 and argB-37 were then normalized to that of rpoB. Triplicate assays using RNAs extracted in three independent experiments were performed for each target gene. The phylogenetic relationships among NAGK-20 and NAGK-37 of L. hongkongensis HLHK9 and their homologues in other bacteria with complete genomes available were analyzed. Phylogenetic tree was constructed by the neighbor-joining method using Kimura's two-parameter correction with ClustalX 1.83. Three hundred and eleven positions were included in the analysis. Cloning and Purification of (His) 6 -Tagged Recombinant NAGK Proteins of L. hongkongensis HLHK9 Cloning and purification of (His) 6 -tagged recombinant NAGK proteins of L. hongkongensis HLHK9 was performed according to our previous publications, with modifications [53, 55] . To produce plasmids for protein purification, primers (59-GGAATTCCA-TATGCTGCTTGCAGACGCCC -39 and 59-GGAATTCCA-TATGTCAGGCTGCGCGGATCAT -39 for argB-20 and 59-GGAATTCCATATGGTTATTCAATCTGAAGT -39 and 59-GGAATTCCATATGTCAGAGCGTGGTACAGAT -39 for argB-37) were used to amplify the genes encoding NAGK-20 and NAGK-37, respectively, by PCR. The sequence coding for amino acid residues of the complete NAGK-20 and NAGK-37 was amplified and cloned, respectively, into the NdeI site of expression vector pET-28b(+) (Novagen) in frame and downstream of the series of six histidine residues. The two recombinant NAGK proteins were expressed and purified using the Ni 2+ -loaded HiTrap Chelating System according to the manufacturer's instructions (GE Healthcare). Purified NAGK-20 and NAGK-37 were assayed for N-acetyl-Lglutamate kinase activity using Haas and Leisinger's method [56] , with modifications. The reaction mixtures contained 400 mM NH 2 OH?HCl, 400 mM Tris?HCl, 40 mM N-acetyl-L-glutamate, 20 mM MgCl 2 , 10 mM ATP and 2 mg of enzyme in a final volume of 1.0 ml at pH 7.0. After incubation at 25uC, 30uC, 37uC, 45uC, 50uC, 55uC or 60uC for 30 min, the reaction was terminated by adding 1.0 ml of a stop solution containing 5% (w/ v) FeCl 3 ?6H 2 O, 8% (w/v) trichloroacetic acid and 0.3 M HCl. The absorbance of the hydroxamate?Fe 3+ complex was measured with a spectrophotometer at A 540 [57] . Inhibition of the kinase activities of NAGK-20 and NAGK-37 were examined with and without 0.25, 0.5, 0.75, 1, 2.5, 5, 10, and 20 mM of L-arginine and incubated at 37uC for 30 min. One unit of N-acetyl-Lglutamate kinase is defined as the amount of enzyme required to catalyze the formation of 1 mmol of product per min under the assay conditions used. Each assay was performed in duplicate. Results were presented as means and standard deviations of three independent experiments. Figure S1 Physical map of the chemotaxis-related genes in L. hongkongensis. While the three gene clusters contain the transducer proteins and some of the methyl-accepting proteins (MCPs), most MCPs are scattered outside the clusters. Genes in orange are coding for chemotaxis transducer proteins; genes in green are coding for MCPs; genes in grey are coding for hypothetical proteins. The numbers refer to the coding sequences in the L. hongkongensis genome.

Malignant mesothelioma

Malignant mesothelioma is a fatal asbestos-associated malignancy originating from the lining cells (mesothelium) of the pleural and peritoneal cavities, as well as the pericardium and the tunica vaginalis. The exact prevalence is unknown but it is estimated that mesotheliomas represent less than 1% of all cancers. Its incidence is increasing, with an expected peak in the next 10–20 years. Pleural malignant mesothelioma is the most common form of mesothelioma. Typical presenting features are those of chest pain and dyspnoea. Breathlessness due to a pleural effusion without chest pain is reported in about 30% of patients. A chest wall mass, weight loss, sweating, abdominal pain and ascites (due to peritoneal involvement) are less common presentations. Mesothelioma is directly attributable to occupational asbestos exposure with a history of exposure in over 90% of cases. There is also evidence that mesothelioma may result from both para-occupational exposure and non-occupational "environmental" exposure. Idiopathic or spontaneous mesothelioma can also occur in the absence of any exposure to asbestos, with a spontaneous rate in humans of around one per million. A combination of accurate exposure history, along with examination radiology and pathology are essential to make the diagnosis. Distinguishing malignant from benign pleural disease can be challenging. The most helpful CT findings suggesting malignant pleural disease are 1) a circumferential pleural rind, 2) nodular pleural thickening, 3) pleural thickening of > 1 cm and 4) mediastinal pleural involvement. Involvement of a multidisciplinary team is recommended to ensure prompt and appropriate management, using a framework of radiotherapy, chemotherapy, surgery and symptom palliation with end of life care. Compensation issues must also be considered. Life expectancy in malignant mesothelioma is poor, with a median survival of about one year following diagnosis.

Malignant mesothelioma is a cancer originating from the lining cells (mesothelium) of the pleural and peritoneal cavities, as well as the pericardium and the tunica vaginalis [1] . Its distribution may be uni-or multifocal or may involve the lining cells in a continuous manner. Before 1950, malignant mesothelioma was so rare that some pathologists even questioned its existence [2] . However, the increasing use of asbestos after the second world war led to the description of a causal relationship between asbestos exposure and the development of mesothelioma in 1960 [3] . Although its use was widely abandoned in the western world in the 1980s, the long latency period between exposure to asbestos and onset of mesothelioma, which can range from 15 to 60 years [4, 5] , meant that the mortality rates from mesothelioma have continued to rise. In the USA the annual deaths from mesothelioma peaked at 3060 in 2002 and have subsequently declined, however the incidence of mesothelioma will continue to rise in the UK until it reaches a peak in about the year 2020 [6] . According to data gathered in the US Surveillance Epidemiology and End Results programme for 1973-1992, there has been virtually a constant rate of mesothelioma in females but a consistently higher rate for males. In the 1960s a mesothelioma register was set up in the UK to systematically record the mortality rates from mesothelioma and to try to identify the incidence of tumour development without known occupational exposure. It is predicted that around 90,000 deaths will occur from mesothelioma by 2050, with 65,000 of these occurring from 2002 onwards [7] . Asbestos is a naturally occurring fibrous silicate, and the risk of developing mesothelioma depends on the exposure to different types of the asbestos mineral fibre. The main asbestos mineral groups are serpentine fibres, which are long and curly, or amphibole fibres which are straight and rod-like. This distinction is important as the serpentine fibre shape is more easily cleared from the respiratory tract. Epidemiologic data suggests that the amphibole, crocidolite, is associated with the highest risk of mesothelioma [8] and that the serpentine fibre, chrysotile has the lowest. The diagnosis of mesothelioma is directly attributable to occupational asbestos exposure; however there is evidence that mesothelioma may result from both paraoccupational exposure (e.g.: women having laundered their husband's overalls) and non-occupational "environmental" exposure [9] . Idiopathic or spontaneous mesothelioma can also occur in the absence of any exposure to asbestos in both animals [10] and humans [8] , and a recent review suggests a spontaneous mesothelioma rate in humans of around one per million [11] . Non-asbestos mineral fibres have also been shown to induce mesothelioma, such as erionite found in certain areas of Turkey [12] or tremolite in north-western Greece [13] , and whilst not specifically mined for commercial purposes, tremolite is often found as a contaminant of chrysotile asbestos and has been causally linked with an increased risk of mesothelioma [14] . The role of Simian virus (SV-40) in the pathogenesis of mesothelioma is more controversial; SV-40 was found to contaminate polio vaccines in the 1950s and 60s in the UK, and although it has been suggested that it is a causative factor in the development of mesothelioma [15] [16] [17] , recent studies have found no link [18] [19] [20] . Mesothelioma is primarily a disease of adults and usually presents in the fifth to seventh decades, and 70-80% of cases occur in men. Those diagnosed between the ages of 20 to 40 years usually have a history of childhood exposure [1] . Typical presenting features are those of chest pain, dyspnoea or both [11, 21] , and in one series up to a third of patients presented with breathlessness due to a pleural effusion without chest pain [22] . When it occurs, the chest pain tends to be dull or boring in nature, but a pleuritic-type pain can occur in the presence of pleural effusions. Involvement of the mediastinal structures is well recognized but hoarseness of the voice and superior vena caval obstruction only rarely causes major symptoms. Dysphagia can also occur but this is a late finding. Unlike bronchogenic carcinomas, presentation with haemoptysis, lymphadenopathy and metastatic symptoms are unusual. A chest wall mass, weight loss and sweating are less common presentations, as is peritoneal mesothelioma although involvement may be found in up to one third of cases at autopsy. Presentation of peritoneal mesothelioma is with non-specific symptoms including loss of appetite, nausea and vomiting, diarrhoea or constipation and occasionally ascites. Small bowel obstruction is usually a late feature, and overall the prognosis of peritoneal mesothelioma is worse than pleural mesothelioma with a mean survival time of about 7 months [11] . Physical examination is usually unremarkable except for signs of pleural effusion and pleural thickening due to tumour infiltration. Finger clubbing is more common than in benign asbestos related pleural disease, and can occur in up to 30% of cases [23] . The tumour originates mainly on the parietal pleura and spreads via the fissures, to encase the lung surfaces. Infiltration of the pericardium can result in signs of cardiac tamponade, and mesothelioma can grow along needle tracks and incisions. Blood tests can reveal an elevated erythrocyte sedimentation rate (ESR) [24] , and there have been isolated case reports of mesothelioma associated with autoimmune haemolytic anaemia [25] . The combination of accurate history, examination, radiology and the acquisition of pathology is essential in the diagnosis of mesothelioma. A careful history of asbestos exposure is essential, and the identification of at-risk occupations are strong markers of exposure. However, the delay between exposure and presentation may naturally preclude accurate recall of occupational exposure and working conditions which may have occurred up to 60 years previously. In those patients with a pleural effusion, sampling of the fluid for cytological examination is the first step in con-firming the diagnosis. Pleural fluid cytology is positive for malignant cells in about a third of cases [1] and if the clinical, radiological and cytological results support a diagnosis of mesothelioma then this can be accepted. However, it is uncommon for the definitive diagnosis to be made on pleural fluid cytology alone and pleural biopsy for tissue diagnosis is therefore recommended. A contrast enhanced computed tomogram (CT) scan is essential to both identify the extent of the disease, and help guide a percutaneous biopsy if the pleural fluid cytological analysis is not sufficient. Radiological imaging is essential for the diagnosis, staging and management of mesothelioma. X-ray, CT, magnetic resonance imaging (MRI) and positron emission tomography (PET) have all been used to evaluate the disease. Intravenous contrast-enhanced CT is the primary imaging modality for suspected pleural malignant disease. CT allows visualisation of the whole pleural surface and diaphragm and use of a 45-60 second scan delay enables the pleural surfaces to be studied whilst still allowing assessment of the mediastinal nodes [26] . A standard protocol should include the liver and adrenal glands, but in cases where there is a past history of abdominal or pelvic malignancy, the scan should also include the lower abdomen and pelvis [11] . Distinguishing malignant from benign pleural disease can be challenging. The most helpful CT findings suggesting malignant pleural disease are 1) a circumferential pleural rind, 2) nodular pleural thickening, 3) pleural thickening of > 1 cm and 4) mediastinal pleural involvement [27] . The specificities of these findings were 100%, 94%, 94% and 88% respectively. The sensitivities were 41%, 51%, 36% and 56% respectively. The presence of bilateral pleural calcification on CT is uncommon in malignant mesothelioma [27] . A significant reduction in thoracic volume seen on CT is more common, however, occurring in up to 73% of cases according to some series [28] . Whilst these features have a high positive predictive value, absence of these signs does not reliably exclude the diagnosis of pleural malignancy. MRI screening is not used routinely in the assessment of malignant mesothelioma, however in patients with potentially resectable disease, MRI can help to provide additional staging information over and above CT. Using gadolinium enhancement, MRI can improve the identification of tumour extension into the diaphragm or chest wall, allowing better assessment of the individual for surgical treatment. MRI also is the imaging modality of choice in those in whom intravenous iodinated contrast is contraindicated [29] . The standardized uptake value (SUV) in PET is a semiquantitative measure of the metabolic activity of a lesion and the SUV is significantly higher in mesothelioma than in other benign pleural diseases such as pleural plaques or inflammatory pleuritis [29, 30] , and one study found PET scanning to have a 96.8% sensitivity and an 88.5% specificity at distinguishing benign from malignant pleural disease [31] . PET scanning has also increased the accuracy in diagnosing mediastinal nodal metastases [30] and therefore the combination of metabolic and anatomical information provided by PET makes it useful in the staging and preoperative evaluation of mesothelioma. PET may also help as a guide to the optimal site for CT guided pleural biopsy, and there is evidence that changes in the fluorodeoxyglucose (FDG) uptake within the tumour might indicate response to treatment suggesting its role in the assessment of response to both chemotherapy and chemo-radiotherapy [32] . At least six different staging systems have been suggested for malignant mesothelioma, but none have been accurately shown to predict survival. Currently, a TNM staging system ( Table 1 ) similar to that used in non-small cell lung carcinoma has been proposed by the International Mesothelioma Interest Group (IMIG) [33] . Tumour response to treatment is an important surrogate for patient benefit. The World Health Organization (WHO) criteria for tumour response were most useful for measuring bi-dimensional lesions, whereas the irregular growth pattern of mesothelioma as a rind around the chest wall makes these criteria poorly applicable [34] . More recently, the Response Evaluation Criteria in Solid Tumours (RECIST) [35] uses a uni-dimensional measurement of tumour size to assess response, but this is based on the assumption that tumours are largely spherical, so again there are limitations to the applicability of this technique in the assessment of malignant mesothelioma. A modified RECIST criteria has now been published, however, with particular reference to malignant mesothelioma [36] . Assessment of response to treatment is now made by measuring uni-dimensional tumour thickness perpendicular to the chest wall in 2 sites at three different levels on CT. According to the WHO classification, malignant mesothelioma can be classified as epithelioid, sarcomatoid, or biphasic based on tissue obtained by biopsy. The patho-logical diagnosis is reached with the aid of immunohistochemistry to demonstrate the presence of mesothelial, epithelial, or true sarcomatous differentiation in the malignant cells [37] . The reported diagnostic yield from CT guided biopsy varies from 60% with a single attempt up to 85% with multiple attempts [38] . Ultimately, how-ever, the highest yields are obtained with open or thoracoscopic pleural biopsy. There is currently no individual immunohistochemical mesothelial marker that provides 100% specificity and high sensitivity for the diagnosis of malignant mesotheli- [37] and so a number of mesothelial and epithelial markers have been developed. Clearly, distinguishing a malignant pleural process from an inflammatory process is a priority and immunohistochemistry has demonstrated that mesothelioma frequently shows immunoreactivity for keratin, p53 and epithelial membrane antigen (EMA) which is unusual for benign pleural disease. Differentiating malignant mesothelioma from other malignancies, however, can be more difficult as mesotheliomas, especially those of the epithelioid type, can mimic several other tumours. Again, immunohistochemistry can help in the differentiation of mesothelioma and markers such as mesothelin, cytokeratin 5/6, calretinin, thrombomodulin and WT-1 have been used. The most specific and sensitive markers for mesothelioma are mesothelin (in epithelioid mesothelioma), calretinin and cytokeratin 5/ 6. However, according to a recent review article, mesothelin is positive in 27% of squamous cell carcinomas, and calretinin and cytokeratin 5/6 can be raised in both squamous cell carcinoma of the lung and adenocarcinoma of both the lung and other sites [37] . Clearly there can be difficulty in making the diagnosis, and so usually a panel of two positive and two negative markers will be sufficient to confirm the presence of malignant mesothelioma. Reliably diagnosing malignant mesothelioma early in the disease is notoriously difficult due to the variability in time to presentation from exposure in the disease. Recently there has been much interest in the use of serum markers for the diagnosis of the disease. The ideal serum marker would be one that is able to offer: 1) early diagnosis, 2) identification of all the subtypes, 3) differentiation of malignant mesothelioma from benign pleural disease and other metastatic pleural malignancies, and 4) be able to track response to therapy and predict survival. As yet no such ideal marker exists, but studies have suggested the use of osteopontin, mesothelin or megakaryocyte potentiating factor in this role. Osteopontin is a glycoprotein that is over-expressed in lung, breast, colorectal, gastric and ovarian carcinomas and in melanoma. Increased levels correlate with tumour progression, invasion and metastases. Although increased levels do not exclude other malignancies, recent data suggest that osteopontin has great potential use as a marker for mesothelioma [39] and has a positive predictive power for mesothelioma equivalent to Ca-125 for ovarian cancer. Studies have suggested a sensitivity and specificity for mesothelioma of 77% and 85% respectively [40] . Mesothelin is a membrane bound glycoprotein expressed by mesothelial cells and over expressed in malignancy, especially mesothelioma [41] . Soluble mesothelin related proteins (SMRP) are serum proteins thought to be released by alternative splicing of the mesothelin protein and thereby preventing adherence to cell membranes. Serum concentrations are increased in malignant mesothelioma, and therefore this protein is a potential serum marker for the disease [42] . However, although most patients with mesothelioma have raised serum levels of SMRP, giving a sensitivity of between 80-83% and specificity between 80-100% [43] [44] [45] , it is mostly associated with the epithelioid sub-type, leading to difficulties in identifying the other sub-types of mesothelioma, especially sarcomatoid, using this marker alone [46] . Nevertheless a commercially available FDA-approved assay, MESOMARK (Fujirebio Diagnostics Inc, Malvern, PA, USA), has been recently approved for use in the monitoring of disease in epithelioid and biphasic mesothelioma, and data seems to support its usefulness [47] . Megakaryocyte Potentiating Factor (MPF) is secreted by cells of several mesothelioma cell lines. In recent studies, MPF was elevated in 91% of 56 malignant mesothelioma patients compared with controls, and levels returned to normal after surgery in patients with peritoneal mesothelioma [46, 48] . This might make MPF useful in the monitoring of treatment response in mesothelioma. Because there can be difficulty reaching a diagnosis of mesothelioma despite radiology, cytology and biopsy, the general management of patients with mesothelioma should be under a multidisciplinary team including respiratory physicians, oncologists, radiologists, palliative care physicians and lung cancer specialist nurses. In addition, once the diagnosis has been reached, there should be close liaison with the patient's primary care physician, and both the primary care physician and family should be warned that, at least in the UK, a post mortem examination will usually be required after the death of a patient with mesothelioma. The general treatment strategy of mesothelioma should cover the following areas: • Management of pleural effusions Management of malignant pleural effusions begins with therapeutic thoracocentesis, which assesses the response of dyspnoea to fluid removal. If symptoms do not respond to thoracocentesis, alternative causes of dyspnoea should be sought such as pulmonary thromboembolic disease, or lymphangitis carcinomatosis. Early, successful management on pleural effusions with pleurodesis is essential for the palliation of symptoms and the prevention of a trapped lung. However with repeated thoracocentesis, the pleural fluid may undergo loculation making it difficult to drain subsequently and also the risk of pleural infection increases. If a conclusive diagnosis has been made, chemical pleurodesis can be performed via a small bore intercostal chest drain (9-14F), which has equivalent success rates to larger bore chest drains with the added benefit of patient comfort [49, 50] . The ideal sclerosing agent is sterile talc, with success rates between 70 and 96% [50] [51] [52] , although care must be taken to ensure the talc particles are of the optimal calibration to prevent the rare risk of adult respiratory distress syndrome (ARDS) [53] . If a firm diagnosis is yet to be made, and the patient is fit enough for surgery, then thoracoscopy is an extremely useful technique in the management of suspected malignant pleural effusions. This procedure allows visualisation of the pleural surface, enables histological sampling for diagnosis and allows complete drainage of the effusion followed by pleurodesis via talc poudrage. In patients with a trapped lung, or in whom pleurodesis has failed, a pleuro-peritoneal shunt can be considered. Whilst symptoms can improve in over 90% of patients, complications (including shunt occlusion and infection) occur in 15% and therefore their use is diminishing [54] . More recently, an ambulatory pleural drainage system (Pleur x Pleural Catheter, Denver Biomedical Inc, USA) has been developed. This system enables the patient to control the pleural effusion at home by means of a long-term drainage catheter and vacuum bottles. This might help patients with trapped lungs as it both avoids the need for surgical intervention, and palliates dyspnoea [55] . Palliative radiotherapy Radiotherapy can be used to control local tumour growth and occasionally causes regression of disease, but there is no evidence that radiotherapy alone affects survival [56] . It has, however, been shown to be helpful in the palliation of pain [57] , and approximately half the patients treated with palliative radiotherapy derive some benefit [58] . It also appears that short courses of radiotherapy (for example 20 Gy in 5 fractions) are as effective as longer courses such as 30 Gy in 10+ fractions [59] , although there is a total dose response effect [60] . Unfortunately, radiotherapy is rarely helpful in either the palliation of dyspnoea or the management of symptoms of mediastinal infiltration such as superior vena caval obstruction (SVCO), and alternative methods of treatment such as SVC stenting should be sought [11, 61] . Whilst hemithorax irradiation may provide symptomatic benefit, no studies have shown that it prolongs survival [61] . Mesothelioma can involve large areas of the pleural cavity and use of radical radiotherapy over large fields places a range of organs such as lungs, liver, spinal cord or heart at a significant risk of dose related damage. As a result, a recent Cochrane review found no randomised clinical trial evidence to support the use of radical radiotherapy alone (or in combination with other treatment modalities) in mesothelioma [62] , and this treatment option has been shown to confer a significant mortality, with rates as high as 17% in one series [59] . Failure to increase survival using single treatment modalities has led to a multimodality approach to treatment in mesothelioma. Combining debulking surgery with radiotherapy is the cornerstone of this approach and can both reduce systemic recurrence and influence the natural history of the disease [63, 64] . Typically, hemithorax radiotherapy is combined with extra pleural pneumonectomy (EPP -see later) and correctly staged patients with epithelioid mesothelioma can have median survival rates of 33 months [65] . Due to the large, irregular nature of malignant mesothelioma along with the close proximity of other organs, a more directed modality of irradiation which could help improve local tumour control whilst limiting exposure to surrounding organs has been developed, known as intensity modulated radiotherapy (IMRT) [66] . Despite the possibilities for improved tumour control, however, IMRT still carries a potential risk of fatal pulmonary toxicity in mesothelioma [67, 68] . Mesothelioma may seed malignant cells in procedure scars. Whilst pain from these metastases is rare, they can become uncomfortable, and evidence has suggested that radiotherapy to intervention sites can prevent this complication [69, 70] . Whilst this practice is recommended in current guidelines [11] , evidence is emerging that it should only be offered to those patients who are symptomatic from these subcutaneous tumours as the radiotherapy itself may have side effects [71, 72] . The role of surgery in the treatment of malignant mesothelioma is still uncertain. The three most common sur-gical procedures are surgical pleurodesis via video assisted thoracoscopic surgery (VATS), debulking surgery (also known as cytoreductive surgery or pleurectomy/decortication (P/D)) and extra pleural pneumonectomy (EPP). This comprises en-bloc resection of the parietal pleura, lung, pericardium and diaphragm and mediastinal nodes [56] . P/D allows the removal of the visceral, parietal and pericardial pleura as well as debulking the tumour and is therefore less demanding, with mortality rates < 5% [73] . P/D has limitations, however, as it does not remove the tumour completely, and the preservation of the ipsilateral lung makes postoperative radiotherapy difficult due to the risk of pulmonary side effects [74] . The impact of P/D on overall survival is contentious, with some evidence suggesting that P/D surgery via a VATS approach may provide a survival benefit in patients with advanced disease not suitable for EPP [75] , whilst other studies do not clearly distinguish the benefit of P/D over EPP [76] . EPP is a much more demanding procedure with morbidity rates as high as 60% [77] and mortality rates of 4-9% [11, 78, 79] . However, the pneumonectomy in EPP allows the use of high dose hemithoracic radiotherapy, and this combined treatment has been shown to reduce local recurrence and prolong survival in early disease [65] . As a result, as part of the trimodality approach to treatment with surgery chemotherapy and radiotherapy, EPP has arisen as the surgical treatment of choice, albeit without clear randomised controlled trial evidence [78] . In order to clarify this issue, a large randomized trial (Mesothelioma And Radical Surgery (MARS)) is underway [80] . In this trial, EPP is sandwiched between induction chemotherapy and radical radiotherapy. The control arm offers full active tri-modality therapy, although the surgery is limited to debulking surgery. In addition patients receive the same induction chemotherapy and are given radiotherapy to any drain or port sites. The use of chemotherapy in malignant mesothelioma aims to lengthen survival, improve quality of life and provide symptomatic relief. Currently, there is no single drug or combination therapy that could be considered as standard treatment for mesothelioma. A variety of single agent and combination regimens have been tried in clinical trials with response rates of between 0% and 45% [81] . A review of studies by Berghmans et al [82] revealed that cisplatin was the most effective single agent for mesothelioma, with carboplatin having similar activity, and in combination with doxorubicin provided a response rate superior to other regimens studied, although there was no clear benefit in terms of survival. Single agent vinorelbine and combination treatment MVP (mitomycin C, vinblas-tine and cisplatin) have also been shown to provide good symptom relief with acceptable toxicity [83] [84] [85] . Recently, there has been much interest in the use of the antifolate pemetrexed (Alimta; Eli Lilly). Pemetrexed exerts its effect by interrupting folate dependent metabolic synthesis of purines and pyrimidines (the building blocks of DNA and RNA), and a Cochrane review of the effectiveness of combination treatment with cisplatin/pemetrexed [81] suggests a survival benefit, albeit with increased toxicity which is reduced by the co-administration of vitamin B12 and folate supplements. Much of the evidence supporting the use of pemetrexed comes from a single study [86] in which 331 patients, fully supplemented with vitamin B12, demonstrated a median survival of 13.3 months with combination pemetrexed/cisplatin compared to 10.0 months with cisplatin alone (p = 0.05). This was associated with a significant improvement in quality of life and symptom relief compared to cisplatin alone [87] . Combination with carboplatin instead of cisplatin, however, may be an alternative regimen which provides similar efficacies and reduced side effects [88] . An alternative antifolate agent, raltitrexed, has also been shown to confer a survival benefit in combination with cisplatin, with median overall survival increasing from 8.8 months (CI 7.8-10.8) with cisplatin alone to 11.4 months (CI 10.1 to 15) with combination cisplatin-raltitrexed (p = 0.05) [89] . Although this study only demonstrated borderline significance, likely as a result of sample size, the results of these trials mean that combination cisplatin and an antifolate should be the reference regimen in mesothelioma. Recent advances now favour the use of neoadjuvant chemotherapy followed by EPP and radiation for malignant mesothelioma; Weder et al recently published their experience of a prospective trial of neoadjuvant chemotherapy consisting of cisplatin and gemcitabine followed by EPP, demonstrating a median survival of 23 months [90] . Similarly, Flores et al prospectively studied patients given induction gemcitabine and cisplatin followed by EPP. Median survival of all patients in the study was 19 months, but those patients who completed induction chemotherapy and subsequently underwent EPP had a median survival of 33.5 months [91] . As a result, all patients who are fit enough (ECOG performance status 0-2) should have the opportunity to discuss the merits of chemotherapy in mesothelioma with an oncologist [11] , in the knowledge that there are no published data which compares survival or symptom control in patients treated with chemotherapy or best supportive care only. The first such trial (MSO-1) [92] is complete, and the preliminary results were published in abstract form in 2007 [93] . These results suggest that the addition of chemotherapy to active symptom control with best supportive care did not infer a significant survival benefit in mesothelioma, however this trial was started before the emergence of pemetrexed. Given that malignant mesothelioma has a poor overall prognosis, referral of patients to specialist palliative care services will be appropriate at some time in the disease course. Most patients need palliation of symptoms early on in the course of the disease and recognition of this by the patient, family and primary care physician is essential in the management of the patient with mesothelioma. The use of a central lung cancer nurse specialist provides a means by which the patient can gain access to the delivery of care needed for this disease, through the following core elements [11] : • Communication The majority of patients who survive for more than 2 years have epithelioid histology and death from mesothelioma tends to be due to respiratory failure. Eligibility for compensation for mesothelioma may vary from country to country. In the UK for example, a diagnosis of mesothelioma allows compensation via the Industrial Injuries Disablement Benefit, War Pensions Scheme or through a Common Law claim from the firm/firms where the asbestos exposure occurred. Patients who cannot identify occupational exposure to asbestos are not entitled for compensation, however pathological confirmation of the diagnosis is not mandatory; establishing the diagnosis and causation on the balance of probability is sufficient, although an unequivocal diagnosis obviously removes any cause for debate. Earlier this year, a new legislation was made in the UK Child Maintenance and Other Payments Act whereby dependants of persons with mesothelioma became eligible to claim for lump-sum compensation after death http://www.opsi.gov.uk/acts/ acts2008a.

Antiviral resistance during pandemic influenza: implications for stockpiling and drug use

BACKGROUND: The anticipated extent of antiviral use during an influenza pandemic can have adverse consequences for the development of drug resistance and rationing of limited stockpiles. The strategic use of drugs is therefore a major public health concern in planning for effective pandemic responses. METHODS: We employed a mathematical model that includes both sensitive and resistant strains of a virus with pandemic potential, and applies antiviral drugs for treatment of clinical infections. Using estimated parameters in the published literature, the model was simulated for various sizes of stockpiles to evaluate the outcome of different antiviral strategies. RESULTS: We demonstrated that the emergence of highly transmissible resistant strains has no significant impact on the use of available stockpiles if treatment is maintained at low levels or the reproduction number of the sensitive strain is sufficiently high. However, moderate to high treatment levels can result in a more rapid depletion of stockpiles, leading to run-out, by promoting wide-spread drug resistance. We applied an antiviral strategy that delays the onset of aggressive treatment for a certain amount of time after the onset of the outbreak. Our results show that if high treatment levels are enforced too early during the outbreak, a second wave of infections can potentially occur with a substantially larger magnitude. However, a timely implementation of wide-scale treatment can prevent resistance spread in the population, and minimize the final size of the pandemic. CONCLUSION: Our results reveal that conservative treatment levels during the early stages of the outbreak, followed by a timely increase in the scale of drug-use, will offer an effective strategy to manage drug resistance in the population and avoid run-out. For a 1918-like strain, the findings suggest that pandemic plans should consider stockpiling antiviral drugs to cover at least 20% of the population.

Taking into account the above assumptions, the model can be expressed as the following system of deterministic differential equations: drug-sensitive infections drug-resistant infections (LTF) drug-resistant infections (HTF) where f = β(δ A A + I U + δ T I T ), g = δ r β(δ A A r + I r,U + I r,T ), and the prime ' ′ ' denotes the derivative of the numbers in compartments with respect to the time; β is the baseline transmission rate of the sensitive strain; p is the probability of developing clinical symptoms; δ A represents the relative transmissibility of asymptomatic infection; δ T is the relative transmissibility of treated individuals infected with the sensitive strain; δ r represents the relative transmissibility of the resistant strain (high fitness); d U and d T are disease-induced death rates of untreated and treated individuals infected with the sensitive strain, respectively; d U,r and d r,U are disease-induced death rates of individuals infected with mutants of LTF and HTF, respectively; µ A is the recovery rate of asymptomatic infection; µ U and µ T represent recovery rates of untreated and treated infected individuals, respectively; α is the rate at which treated individuals develop drug-resistance (rate of de novo resistant mutation); γ is the conversion rate of resistant strains with LTF to HTF; and q is the fraction of infected individuals which receives treatment (treatment level). The equations for removed subpopulations are given by where these compartments include both recovered and dead individuals following infection. While national pandemic plans consider stockpiling antiviral drugs, it is imperative to evaluate the impact of emergence of resistance on antiviral use, particularly for the scenario of limited supply in which run-out is likely to occur. In order to demonstrate the relationship between the level of treatment and drug stockpile, we appended the following equation to the model where M 0 is the initial size of the stockpile and M (t) represents the number of antiviral courses available at time t. We assumed that the supply is only depleted through treatment of indexed cases with a single course of oseltamivir during symptomatic infection, and no additional antiviral courses will be provided during the pandemic. We applied a previously established technique to calculate R c [1] , and considered infectious classes in the order X = (A, I U , I T , A r , I r,U , I r,T , I T ,r ) to simplify computations. The system (2)-(4) can be written as Taking the Fréchet derivatives F = DF and V = DV , and evaluating at the disease free equilibrium Then, by evaluating V −1 , the control reproduction number is defined by the spectral radius of F V −1 and given by R c = max{R s c , R r 0 }, where R s c and R r 0 are expressed in (1) and (2) of the main text, respectively. where S 0 and S ∞ are respectively the initial and final sizes of the susceptible population. Integrating , and qpS ′ + I ′ T , and substituting into (6), we can express the final size equation as where I U (0) is the initial number of infections at the onset of the outbreak. By integrating equation (3), we can also compute the total number of resistant cases resulted from the treatment of sensitive infections as To investigate the effect of parameter changes on the results presented by simulations in the main text using baseline values, we performed sensitivity analyses by considering a sampling approach that allows for the simultaneous variations of several key parameters, including the basic reproduction number R 0 , the rate of de novo resistant mutations α, the rate of conversion between resistant strains γ, the relative transmissibility of the resistant strain δ r , and the probability of developing clinical disease p. Using the Latin Hypercube Sampling technique [2] , we generated samples of size n = 1000 in which each parameter is treated as a random variable and assigned a probability function. In this technique, the parameters are uniformly distributed and sampled within their respective ranges. The reproduction number was uniformly sampled from the range [1.5, 2.5] (values found in references [3, 4] ), and the rate of de novo resistant emergence was sampled from the range [0.018, 0.072] (values taken from references [5, 6] ). The corresponding range for the conversion rate of resistant strains was computed using the constraint that the fraction of treated individuals hosting resistance, which undergoes compensatory mutations and subsequently generates resistant strains with high fitness, lies between 1/5000 and 1/500 [7, 8] . Furthermore, we considered a range of [0.5, 0.7] for the probability of developing clinical disease [9, 10] , and sampled the relative transmissibility of resistant strain with HTF from the range [0.8, 1] [8] . The same ranges of parameter values were also used in a previous study for sensitivity analyses [8] . To evaluate the effect of parameter changes on the required stockpile and the total number of infections, we ran the simulations for different treatment levels (between 0% and 80%) in the presence/absence of resistance. These simulations, illustrated in Figure 1 , correspond to Figure 2 in the main text, and show that a substantially larger stockpile is required when treatment exceeds a certain level in a constant treatment strategy. This is due to the wide spread of highly transmissible resistance under high pressure of antiviral drugs, which in turn leads to a larger number of total infections. We further ran simulations for the optimal time t * in an adaptive treatment strategy (using a sample of size n = 100), to determine the sensitivity of the results on the parameters variation in minimizing the total number of infections. The results of this sensitivity analysis are illustrated in Figure 2 , with different sizes of stockpile (8.5%, 12%, 25%), when initial treatment levels (0%, 25%) change to 80% at time t * . Regardless of the level of stockpiles, the results show that aggressive treatment should be implemented with shorter delay after the onset of the outbreak as the basic reproduction number increases. However, the implementation of intensive treatment requires a significantly longer delay for higher initial treatment levels as the reproduction number decreases.

Transmission Dynamics and Prospects for the Elimination of Canine Rabies

Rabies has been eliminated from domestic dog populations in Western Europe and North America, but continues to kill many thousands of people throughout Africa and Asia every year. A quantitative understanding of transmission dynamics in domestic dog populations provides critical information to assess whether global elimination of canine rabies is possible. We report extensive observations of individual rabid animals in Tanzania and generate a uniquely detailed analysis of transmission biology, which explains important epidemiological features, including the level of variation in epidemic trajectories. We found that the basic reproductive number for rabies, R(0), is very low in our study area in rural Africa (∼1.2) and throughout its historic global range (<2). This finding provides strong support for the feasibility of controlling endemic canine rabies by vaccination, even near wildlife areas with large wild carnivore populations. However, we show that rapid turnover of domestic dog populations has been a major obstacle to successful control in developing countries, thus regular pulse vaccinations will be required to maintain population-level immunity between campaigns. Nonetheless our analyses suggest that with sustained, international commitment, global elimination of rabies from domestic dog populations, the most dangerous vector to humans, is a realistic goal.

Rabies has been one of the most feared diseases throughout human history and has the highest human case-fatality proportion of any infectious disease [1, 2] . Every year over 7 million people receive post-exposure prophylaxis, and an estimated 55,000 people die from rabies [3] (more than yellow fever, dengue fever, or Japanese encephalitis [4] ). Over 99% of these deaths occur in developing countries where rabies is endemic in domestic dog populations [5] . However, the impacts of canine rabies are often overlooked, largely because human rabies deaths are now extremely rare in Western Europe and North America, where mass vaccination successfully eliminated the disease from domestic dog populations [6] . Increasing incidence of canine rabies in Africa and Asia has prompted concerns that similar strategies may not be effective in these areas [7, 8] . The critical question now is whether global elimination of domestic dog rabies is achievable. Keys to answering this question include: a quantitative understanding of the transmission dynamics of rabies in domestic dog populations, particularly the basic reproductive number, R 0 ; a quantitative understanding of domestic dog demography; and information about the practicality and effectiveness of various vaccination strategies. While recent data support the feasibility and practicality of domestic dog vaccination strategies [9] [10] [11] , there are very little quantitative data on rabies transmission dynamics [12] and the underlying demographic processes. Transmission is the most important process underlying infectious disease dynamics [13] , but it is also the least understood. Rates of transmission are usually inferred from population patterns of disease incidence, but populationlevel analyses do not capture between-individual variation in transmission resulting from differences in behaviour, genetics, immune status, and environmental and stochastic factors, which play an important role in determining disease dynamics [14, 15] . Contact tracing has been used to directly measure case-to-case transmission, and applications of the technique to emerging infections such as SARS have generated important insights into disease transmission and control in human populations [16, 17] , but transmission processes for diseases circulating in animal populations are much harder to study. Rabies is an acute viral encephalitis that is spread through the saliva of infected hosts [2] . Clinical manifestations vary, but the neurological phase often includes increased aggression and the tendency to bite and thereby transmit infection; rapid progression to death is inevitable [4] . These distinctive signs make transmission of rabies easier to track than that of most other diseases and provide an unusual opportunity to explore epidemiological patterns at the scale of the individual. Here, we present data on rabies transmission in two districts of rural Tanzania, Serengeti and Ngorongoro ( Figure 1 ). We were able to monitor the spread of infection using contact-tracing methods, which were feasible due to the discrete and memorable nature of transmission events. We recorded .3,000 potential transmission events between 2002 and 2006 and reconstructed case histories of over 1,000 suspect rabid animals that illustrate heterogeneity in several aspects of transmission, including the latency, movement patterns, and biting propensity of infected individuals. Although these districts border the Serengeti ecosystem, we have argued that domestic dogs are the sole maintenance population of rabies in this community: they make up over 90% of our observations of rabid animals, and the .70 isolates that have been sequenced (from 13 host species) are all consistent with the Africa 1b canid strain [18, 19] . This is one of the most extensive datasets on individual transmission events assembled in an animal population; it has potential to shed light on critical, but often elusive, details of infectious disease transmission. We also analyze data from rabies outbreaks around the world, which provide a global and historical context for the Tanzania dataset. Analyses of the contact-tracing data generated robust estimates of epidemiological parameters that have important implications for rabies control (Table 1 , Figures 2 and 3 , and Figure S1 ) and provide insight into how infectious disease transmission scales from individual behaviour to populationlevel dynamics. We estimated R 0 for rabies in Serengeti and Ngorongoro districts directly from infectious histories, from reconstructed epidemic trees based on the spatiotemporal proximity of cases, and from the exponential rate of increase in cases at the beginning of an epidemic. Biting behaviour of rabid dogs during the course of infectious periods was highly variable (mean bites per rabid dog ¼ 2.15, 95% confidence interval (CI) from fitting a negative binomial distribution: 1.95-2.37; variance ¼ 5.61, CI: 4.63-6.92; shape parameter k ¼ 1.33; CI: 1.23-1.42) ( Figure 3A ). The probability that an unvaccinated dog developed rabies after being bitten by an infectious animal was high (P rabiesjbite ¼ 0.49, CI: 0.45-0.52) ( Table 1) if the bitten dog was not vaccinated or killed immediately after exposure. Multiplying the average number of dogs bitten per rabid dog by the probability of developing rabies following exposure gave an R 0 estimate of 1.05 (CI: 0.96-1.14) ( Figure 3A and Table 1 ). These estimates should be regarded as lower bounds, because not all transmission events were observed (this calculation excludes rabid dogs that were killed before biting other animals or that disappeared and likely corresponded to unknown or unobserved rabid dogs in other areas; see Materials and Methods). Detailed data on the timing and location of transmission events and infections allowed us to estimate the spatial infection kernel and generation interval (distances and times between source cases and their resulting infections, respectively) ( Figure 2) and Canine rabies has been successfully eliminated from Western Europe and North America, but in the developing world, someone dies every ten minutes from this horrific disease, which is primarily spread by domestic dogs. A quantitative understanding of rabies transmission dynamics in domestic dog populations is crucial to determining whether global elimination can be achieved. The unique pathology of rabies allowed us to trace case-to-case transmission directly, during a rabies outbreak in northern Tanzania. From these unusual data, we generated a detailed analysis of rabies transmission biology and found evidence for surprisingly low levels of transmission. We also analysed outbreak data from around the world and found that the transmission of canine rabies has been inherently low throughout its global historic range, explaining the success of control efforts in developed countries. However, we show that when birth and death rates in domestic dog populations are high, such as in our study populations in Tanzania, it is more difficult to maintain population-level immunity in between vaccination campaigns. Nonetheless, we conclude that, although the level of vaccination coverage required is higher than would be predicted from naïve transmission models, global elimination of canine rabies can be achieved through appropriately designed, sustained domestic dog vaccination campaigns. probabilistically reconstruct transmission networks (Videos S1 and S2). Calculating the average number of secondary cases per rabid dog during the period of exponential epidemic growth (before vaccinations were implemented) from these reconstructions gave similar R 0 estimates of 1.1 in Serengeti district and 1.3 in Ngorongoro (CIs: 1.04-1.10 and 1.26-1.42, respectively) ( Table 1 ). The more traditional approach of estimating R 0 , by fitting a curve to incidence data over the same interval of exponential epidemic growth, also produced similar estimates of 1.2 in Serengeti and 1.1 in Ngorongoro (CIs: 1.12-1.41 and 0.94-1.32, respectively) ( Table 1 and Figure 3B ). This approach is robust to underreporting (Text S1 and Figure S2 ) but should likewise be considered a lower bound, because some local control measures were instituted (such as tying or killing). We also estimated R 0 from the intrinsic growth rate of outbreaks of domestic dog rabies elsewhere in the world ( Table 2) and obtained values between 1.05 and 1.85, which are consistent with our estimates from northwest Tanzania. For many diseases, R 0 is expected to increase with host density [12, 13, 20, 21] . Despite the domestic dog population density in Serengeti (9.38 dogs/km 2 ) being considerably higher than the dog population density in Ngorongoro (1.36 dogs/ km 2 , see Table 3 ), we were unable to detect significant differences in our estimated values of R 0 between the two districts. Nor did we find any conspicuous differences in R 0 estimated from the outbreaks listed in Table 2 , which represent a wide range of population densities. There may, in fact, be no relationship between R 0 and population density for canine rabies. On the other hand, a subtle relationship between dog density and transmission rates might be difficult to detect for a number of reasons. To investigate whether it would be possible to decipher systematic differences in R 0 across the range of values that we estimated, we simulated outbreaks using our epidemiological parameter estimates, but varied R 0 (from R 0 ¼ 1 to R 0 ¼ 2), whilst maintaining individual variance in biting behaviour (same shape parameter k, see Text S2). Although the mean estimates of R 0 from fitting to these simulated trajectories were accurate, they were surrounded by wide confidence intervals ( Figures S2 and S5 ), suggesting that if only a small number of epidemics were sampled, any underlying relationship might not be apparent. Several mass domestic dog vaccination campaigns were carried out in villages in the study districts during the 5-y period. We analysed the impacts of these interventions at the village level to capture the wide variation in achieved levels of vaccination coverage. We incorporated demographic processes (Table 3 gives demographic parameter estimates) and waning of vaccine-induced immunity (see Materials and Methods), because these affect the level of herd immunity within the population at any one time. There were no rabies outbreaks (defined as at least two cases not interrupted by an interval of more than one month) in villages when vaccination coverage exceeded .70%. Small outbreaks occurred in villages with lower coverage and the largest (and longest) outbreaks only occurred in villages with ,20% coverage. Observed outbreak sizes were within the range expected from the heterogeneity of biting behaviour and the coverage achieved by village-level vaccination campaigns ( Figure 4A ). The effective reproduction number R, which describes transmission once an epidemic is underway, declined during the course of the observed epidemics ( Figure 3C ). At the level of individuals, vaccination coverage reduced the number of secondary cases per rabid dog ( Figure 4B ). More than 300 vaccinated dogs were identified by contact tracing as having been bitten by rabid animals. Only ten of these animals showed any signs indicative of rabies, although in the absence of vaccination approximately 50% (P rabiesjbite ¼ 0.49) ( Table 1) of these would have been expected to succumb to the disease. Individual actions by dog owners such as tying or killing exposed or infectious animals also had an impact. By killing rabid dogs, villagers reduced the overall average infectious period by around 16% (3.7 d for rabid animals that died from the disease versus 3.1 d for all infected animals, including those that were killed). However, there were no consistent declines through time in the number of bites by rabid dogs ( Figure S3 ). Thus we consider vaccination to have been the overwhelming factor in curtailing the outbreaks ( Figure 4A ). From our estimates of R 0 , we calculate the deterministic critical vaccination threshold for rabies elimination in rural To calculate R 0 , we excluded dogs that were killed, tied, or those that disappeared before biting any other dogs. Variability in biting behaviour means that a small number of individuals disproportionately affect transmission and can potentially spark an epidemic, but since most individuals cause few, if any, infections, R 0 is low and most introductions quickly die out ( Figure 4C ). (B) Exponential epidemic growth in Serengeti (blue, R 0 ; 1.2) and Ngorongoro (red, R 0 ; 1.1) districts. The R 0 estimates from the epidemic trajectories were relatively insensitive to the period used for fitting the exponential curve. The inset shows the distribution of R 0 estimates based on fitting to different regions of the time series. (C) The effective reproductive number, R, (averaged over three-month intervals) for Serengeti (blue) and Ngorongoro (red) districts measured from reconstructed epidemic trees that incorporate prior knowledge on who infected whom. Dots indicate the number of secondary cases resulting from each primary case (inferred from the composite tree of most likely links, with random jitter to avoid superposition on the y-axis). R 0 estimated from these reconstructions (during the period of exponential epidemic growth) was ;1. The exponential growth rates of the epidemics were estimated by fitting exponential curves to monthly time series of rabies incidence and converted to estimates of R 0 using the serial interval distribution from the contact tracing data in Tanzania (see Materials and Methods). Estimates based on weekly data are shown in parentheses. The estimated period of exponential epidemic growth, the year of the epidemic onset, and a description of the epidemic setting (where available) are listed. For populations that were partially vaccinated, we corrected our R 0 estimates by dividing by the proportion of vaccinated animals at the onset of the outbreak. Our estimates show that R 0 for canine rabies is inherently low throughout its historic global range. doi:10.1371/journal.pbio.1000053.t002 We found no effect of age on the frequency of litters for female dogs older than 3 months. Domestic dog population growth was estimated from the demography data collected in Serengeti district (r dogs ) and confidence intervals generated from bootstrapping the data. Domestic dog population densities for 2004 were estimated from 2002 national census data (with projected human population growth rates of 2.6% and 3.8% per annum in Serengeti and Ngorongoro respectively) and human:dog ratios (generated from household questionnaires). Alternate estimates of domestic dog population growth rates were extrapolated for each district (r Serengeti and r Ngorongoro ) using these data. Overall domestic dog densities are presented despite considerable variation at the village level. doi:10.1371/journal.pbio.1000053.t003 Figure 4 . The Impact of Vaccination on Transmission (A) The size of village-level outbreaks (defined as at least two cases not separated by more than one month, isolated cases are assumed to be nonpersistent introductions) in Serengeti (blue, n ¼ 138) and Ngorongoro (red, n ¼ 20) districts plotted against village-specific vaccination coverage at the outbreak onset. Coverage was extrapolated from a demographic model initialized with village-specific dog population estimates and incorporating village-specific vaccination data. Gray shading and contours correspond to the probability of observing an outbreak of a particular size or less, generated from 10,000 stochastic simulations of rabies transmission for every initial vaccination coverage (contours were calculated conditional upon .1 secondary case occurring). The inset illustrates a village-level example of the susceptible reconstruction used to calculate instantaneous vaccination coverage plotted beside rabies cases in that village. (B) The distribution of secondary cases per infectious dog as inferred from reconstructed epidemic trees in Serengeti (blue) and Ngorongoro (red) districts, plotted against vaccination coverage in the village where the primary case occurred. Random jitter was added to prevent superposition on the y-axis. (C) Probability of an outbreak being seeded by an introduced case under different levels of vaccination coverage. Due to heterogeneity in the transmission process outbreaks rarely occur when coverage is maintained above P crit . However if infections are frequently imported from outside the vaccinated region, at least 40% coverage would need to be maintained to reduce the probability of subsequent outbreaks (of at least ten cases) to ,0.05. doi:10.1371/journal.pbio.1000053.g004 Tanzania to be only 20% (P crit ¼ 1 -1/R 0 ), and even in areas where R 0 is higher, P crit rises to just 40% (Table 2 ). Our observations and simulations ( Figure 4 ) demonstrate that small outbreaks occur by chance even when coverage exceeds P crit and should be expected more frequently when there is individual variation in transmission (Text S2). Herd immunity declines rapidly in the interval between vaccination campaigns because of births and deaths in the domestic dog population (Figure 4 , inset). To maintain herd immunity above P crit between campaigns, therefore, requires a larger proportion of the dog population, P target , to be vaccinated (P target ¼ e (mþdþr)T P crit , where r is the rate of dog population growth, d is the death rate, 1/m is the duration of vaccineinduced immunity, and T is the interval between campaigns (see Materials and Methods)). By incorporating demographic parameters (Table 3) , we estimate that annual campaigns should therefore aim to vaccinate 60% of the dog population to avoid coverage falling below P crit . The basic reproductive number, R 0 , is the average number of secondary infections produced by an infected individual in an otherwise fully susceptible population [20] . R 0 is the most important parameter in infectious disease epidemiology, and considerable effort has been devoted to its estimation and to understanding its implications for disease control [20, [22] [23] [24] [25] [26] , although it is important to note that some factors not incorporated in R 0 , e.g., host births as well as deaths, may also have important control implications. Depending upon the quality and quantity of data, a number of approaches can be used to estimate R 0 . Choosing the most appropriate method and assessing its accuracy can be difficult, given the associated assumptions and shortcomings [22] . Most methods do not account for variability in the pathogenesis and behaviour of infected animals; some methods make inferences from quantities that are confounded by (often unmeasured) responses to disease incidence (e.g., epidemic size or prevalence at equilibrium); and different methods are variously biased due to measurement and process error. Although our attempts to estimate R 0 are also imperfect, they do incorporate individual variation in behaviour and pathogenesis, explicitly address several common assumptions, and have been carefully checked for biases through extensive simulations. The overall consistency in the low values of R 0 that we estimated (;1.1 , R 0 , 2) is therefore reassuring and provides optimism for the feasibility of canine rabies control by vaccination. If R 0 increases with host density in this system, different threshold levels of vaccination coverage would be necessary to eliminate disease in different density populations [12, 20] . However, our data on individual variation in biting behaviour also illustrate that it would be difficult to detect statistical differences in the range of R 0 values that we estimated ( Figure S2 ). Thus in practice, when only a small number of epidemics are observed, individual variation in transmission may mask any underlying variation in R 0 driven by population density. So although we cannot decipher the relationship between population density and rabies transmission, the consistency of our individual-and population-level estimates from Tanzania and from a wide range of sites around the world allow us to estimate the threshold vaccination coverage necessary to eliminate the disease. Our estimates of R 0 predict that only relatively low levels of vaccination coverage are required to eliminate rabies (;20-45%), but there is considerable variation in empirically observed levels of coverage that have successfully controlled the disease; low levels of coverage (30-50%) have been successful in some circumstances [27] , although higher levels have also failed [28] . Our analyses suggest that these inconsistencies are, in large part, a consequence of host demography. When vaccinations are carried out in pulses, births and deaths within the host population will continuously reduce the level of herd immunity attained during campaigns (Figure 4, inset) . Turnover of domestic dogs in rural Tanzania is very high (Table 3) ; therefore, annual campaigns should aim to vaccinate 60% of the dog population to maintain vaccination coverage above P crit for the duration of the interval between campaigns. When successive campaigns have achieved this, rabies incidence has declined dramatically despite high endemic levels in adjacent areas [29] . Domestic dog population turnover therefore appears to have had a marked influence on rabies dynamics that explains the variable success of vaccination efforts. The empirically derived consensus that 70% coverage is sufficient for long-term rabies elimination [30, 31] was likely reached because it is effective as a target for annual campaigns in almost all demographic settings, including those with particularly high turnover such as those we describe from Tanzania. There are other potential explanations and caveats. The nutritional and health status of animals might affect the development of protective immunity in response to vaccination. However, more than 97% of dogs sampled from Serengeti district developed strong antibody titres (.0.5 IU/ ml) in response to vaccination [32] , suggesting that these factors do not impair the efficacy of dog vaccination in rural Tanzania. In addition, numerous practicalities-such as occasional failures in the cold chain, improper vaccination of animals, mistaken registrations, etc.-will all reduce the level of population immunity below the estimated vaccination coverage. Furthermore, our observations and simulations confirm that small outbreaks may occur simply by chance even when coverage exceeds P crit [33] , and these are particularly likely when there is individual variation in transmission (Figure 4 ). Higher levels of coverage are therefore necessary to reduce the chance of outbreaks with greater certainty; especially where the risk from imported infections is highest ( Figure 4C ). This could be a concern if canine rabies were to be eliminated from domestic dog populations but continued to circulate in sympatric wildlife; however, canine rabies was successfully eliminated in Western Europe and North America despite the presence of wildlife hosts capable of transmission. Thousands of people die every year from this horrific and preventable disease, because the control of canine rabies has been severely neglected in developing countries [2] . Inherent inter-annual periodicity of epidemics exacerbates the situation, with rabies only intermittently perceived as problematic [6] , as illustrated by the recent outbreak in China [34] . The problem of canine rabies has often been considered intractable in rural Africa, because of poor infrastructure, limited capacity, and the misperception that large popula-tions of wild carnivores are responsible for disease persistence. Our analyses show that global control of canine rabies is entirely feasible and that successful elimination of canine rabies in many parts of the world has likely been achieved precisely because R 0 is so low and institutional commitment to maintain high levels of vaccination coverage has been sustained [6] . Achieving vaccination coverage of 60% or more in dog populations in Africa is both logistically and economically feasible through annual vaccination campaigns [9] [10] [11] 29] . The resultant reduction in costs of human postexposure prophylaxis suggest that vaccination interventions targeted at domestic dog populations could translate into appreciable savings for the public health sector [3, 8, 29] . Furthermore, the inherently low R 0 and the tractability of rabies contact-tracing indicates that once endemic rabies is controlled, elimination could be achieved through active case detection in remnant foci of infection (much like the strategy used to eradicate smallpox [35] ); similar measures are proving effective in programmes to eliminate canine rabies in the Americas [36] . However, the most crucial step towards global elimination of canine rabies will be sustained commitment and coordinated efforts to maintain sufficient vaccination coverage in domestic dog populations. Study areas. We collected data from two districts in northwest Tanzania: Serengeti, inhabited by multi-ethnic, agro-pastoralist communities and high-density dog populations, and Ngorongoro, a multiple-use controlled wildlife area, inhabited by low-density pastoralist communities, predominantly Maasai, and lower-density dog populations (Figure 1 ). Attributes of the dog populations in these districts are presented in Table 3 . Wildlife populations also differ in the two districts, but domestic dogs are the focus of this study because they are the only maintenance population of rabies in the area [18] . Incidence data. Data on patients with animal-bite injuries from hospitals and dispensaries, case reports of rabid animals from livestock offices, and community-based surveillance activities were used as primary sources [18] . Visits were made to investigate incidents reported in 2002 to 2006 involving suspected rabid animals. Cases were mapped at the site of the incident (wherever possible) and villagers interviewed to evaluate the status of the biting animal, determine its case history, and identify its source of exposure and subsequent contacts (if known). The same procedure was exhaustively followed for all associated exposures/cases. Interviews were conducted with veterinary officers, local community leaders, and livestock field officers in attendance, resulting in an active reporting network. Cases were diagnosed on epidemiological and clinical criteria, adapting the ''six-step'' method through retrospective interviews with witnesses [37] . Rabies was suspected if an animal displayed clinical signs [37] and either (a) disappeared or died within 10 days, or (b) was killed, but had a history of a bite by another animal or was of unknown origin. Additional clinical criteria for wild carnivores (;10% of human exposures were caused by wild animals and ;10% of inferred transmission events involved rabid wildlife) included tameness, loss of fear of humans, diurnal activity (for nocturnal species), and unprovoked biting of objects and animals without feeding. When multiple incidents involving suspected rabid wildlife were reported on the same/consecutive days within neighbouring homesteads, we assumed a single animal was involved. Brain samples were collected and tested for confirmation wherever possible, but despite efforts to obtain diagnostic samples, most cases reported here were suspected rather than confirmed. Inadequate sample preservation such as storage at room temperature and long intervals between sample collection and testing (during which samples underwent repeated freeze-thaw cycles) probably caused specimens to deteriorate. Composite samples of each brain necessary to achieve the highest test reliability were also rarely available. Nevertheless, a high percentage of samples from suspected cases of rabies were confirmed by laboratory diagnosis (;75%) suggesting that use of epidemiological and clinical criteria is justified and reliable [18] . Researchers are encouraged to contact the authors regarding data availability. Vaccination data. Dog vaccination campaigns in Serengeti district in 2000 resulted in low and patchy vaccination coverage (35-40% estimated from post-vaccination household surveys). Annual campaigns conducted from 2003 onwards in a 10-km zone adjacent to the western border of Serengeti National Park achieved higher coverage levels of between 40 and 80%. In 2004, the Tanzanian government conducted vaccinations in villages in Serengeti district beyond the 10-km zone reaching 55% coverage across the remainder of the district, but in subsequent years, campaigns were less systematic and conducted in fewer villages. Vaccination in Ngorongoro was restricted to small-scale localised campaigns in the district town centre until 2004, whereupon widespread annual vaccinations were implemented with overall coverage exceeding 80% [9] . Data on the number of dogs vaccinated in each village and on each campaign date were collected from 2003 onwards. Parameter estimation. The incubation period and duration of infectiousness were estimated for rabies in domestic dogs from records of when individual dogs were bitten, developed clinical signs, and were killed or died. Gamma distributions were fitted to these data using maximum likelihood with interval censoring to account for cases where the relevant data were only approximately known ( Figure 2 and Table 1 ). To estimate the probability distribution of the generation interval, G(t), an incubation and an infectious period were drawn from their respective distributions, a ''time-to-bite'' deviate was drawn from a uniform distribution over the interval of the infectious period, and the two intervals were summed. There was a significant correlation between the length of the infectious and incubation periods, but significance was entirely due to a single data point; we therefore treated the distributions as independent. The spatial infection kernel K(d) was estimated by fitting a gamma distribution to the distances between known source cases and animals that they contacted. Many contacts occurred within the same, or neighbouring, homesteads. In these cases, the precise distance was not always recorded, but we assumed it was less than 100 m. We therefore replaced the probability of a contact within 100 m by the probability distribution averaged over the range 0-100 m. The basic reproductive number R 0 . (1) Direct estimates from infectious histories. Using maximum likelihood, we fitted a negative binomial distribution to data on biting behaviour of rabid dogs ( Figure 3A ). The probability of developing rabies following a bite (P rabiesjbite ) was estimated, excluding bitten animals that had previously been vaccinated, or that were either killed or vaccinated immediately after the bite, and binomial confidence intervals were calculated. R 0 was estimated as the probability P rabiesjbite multiplied by the average number of bites per rabid dog and confidence intervals were calculated using a resampling procedure. Dogs that were removed (killed or tied up) before causing secondary cases in other dogs (even if they bit people) were excluded from this calculation, as were suspect rabid dogs that either disappeared before biting other dogs or that were of unknown origin and were killed before being observed to bite other dogs ( Figure 3A ). We pooled data from both districts for this estimate because insufficient complete case-histories of rabid dogs (after excluding cases with interventions) were traced to accurately estimate R 0 for Ngorongoro (35 versus 477 in Serengeti). We also estimated R 0 directly from the distribution of secondary cases per rabid dog. Dogs that were bitten by rabid animals but did not develop rabies because of interventions (previous vaccination or being killed/vaccinated immediately after the bite) were multiplied by P rabiesjbite and added to observed secondary cases, giving an expected number of secondary cases per rabid dog in the absence of intervention and a similar estimate of R 0 (1.14, CI: 1.03-1.25) ( Figure S1 ). (2) Epidemic tree reconstruction. We used an algorithm for probabilistically constructing epidemic trees based on the location of cases in space and time [38] . For each suspected case (i), we chose a progenitor ( j) at random with probability p ij from all n cases preceding that case, where: Gðt ik ÞKðd ik Þ G is the distribution of generation times, t ij is the length of time (in days) between the occurrence of case i and its potential progenitor j (G(t) ¼ 0 for t , 0), K is the spatial infection kernel, and d ij is the distance (in km) between the locations of case i and its potential progenitor j (using the average probability when distance ,100 m, see above). Because the dates that some individuals were bitten or developed rabies were only approximately known, 1,000 bootstrapped datasets were generated with the dates drawn randomly from a uniform distribution over the window of uncertainty and a consensus tree of the most probable links was determined and used to generate secondary case distributions illustrated in Figure S1 . Because transmission of rabies from livestock is recorded extremely rarely, we did not allow livestock progenitors, which considerably improved the match between known and assigned links compared to an algorithm where all species could be assigned as progenitors. All detected cases in carnivores (including domestic cat and wildlife cases) were included in the tree reconstructions using the spatial infection kernel and generation interval parameters estimated for domestic dogs. The contribution of nondomestic dog carnivores to the overall epidemic was small, and estimates of within-and between-species transmission are described elsewhere [18] . When known links between primary and secondary cases were not retained in the trees, they were correctly assigned in more than 60% of cases in both districts, indicating that probabilistic reconstruction was effective. The average number of secondary cases putatively produced from each primary case was calculated from the bootstrapped trees. R 0 was estimated as the average number of infections caused per rabid dog that was infectious during the period of exponential epidemic growth. Determining the period of exponential growth is somewhat subjective; for consistency between methods, we used the interval that gave the median R 0 value for time series regression estimates (see below). The choice of interval caused more variance in R 0 estimates for this reconstruction technique than for other methods because it averages the heterogeneous behaviour of a small number of individual animals that spark an epidemic. Thus inclusion or exclusion of particularly infectious individuals has a large effect on R 0 . (3) Inference from the epidemic curve. A single infection will cause future cases distributed according to the probability distribution of the generation interval. Therefore the number of cases arising in any given interval is the result of those cases that occurred at times in the past whose secondary cases occur in this interval and is determined by the probability distribution of the generation interval. This intuitive description is formalized by the Euler-Lotka equation, adapted for an infection process [25] and an expression for R 0 can be obtained: GðsÞe Àrs We estimated the initial growth rate of the epidemic (r) by fitting an exponential curve to incidence data using a generalized linear model. We compared Akaike's Information Criterion values to determine the appropriate error structure (Poisson or negative binomial). The choice of which part of the epidemic curve the model should be fit to was subjective, therefore the model was fit to all possible sections of the epidemic curve (using a minimum of nine consecutive months) and the median, the 2.5th and the 97.5th percentile of the R 0 estimates are presented in Table 1 . Figure 3B (inset) shows that the estimate of R 0 was robust to the interval chosen for fitting the curve. We used the same method to estimate R 0 from data that we had compiled on outbreaks of canine rabies from elsewhere in the world. For these time series, we fitted exponential curves to the intervals between the first recorded case and the month (or week) with highest rabies incidence (Table 2 ) and converted the estimated growth rates to estimates of R 0 using the serial interval distribution data gathered by contact tracing in Tanzania. For partly vaccinated populations, we corrected our R 0 estimates by dividing by the fraction of dogs which were vaccinated prior to the outbreak [12] . For all the outbreaks considered, including those in Tanzania, some localized and individual control measures may have been instituted (such as tying up or killing infected animals), and therefore our R 0 estimates should be regarded as lower bounds. However simulations also revealed that for very low values of R 0 (,1.2), estimates from the epidemic trajectory can be slightly biased upwards ( Figure S2 ). This is probably because at very low levels of R 0 , most introductions do not initiate further cases and therefore a small number of individuals with higher than average biting behaviour are needed to trigger epidemics, thus biasing trajectories. The effective reproductive number R. The effective reproductive number R measures the average number of secondary cases per primary infection once an epidemic is underway. R changes through space and time depending upon the implementation of control measures, the depletion of susceptibles and the build-up of local correlations in the spatial distribution of infected and susceptible individuals. Numbers of secondary cases per rabid dog (inferred from the epidemic tree reconstructions) were calculated monthly and averaged across bootstrapped trees to give a time-varying estimate of R ( Figure 3C ). Although R declined through time in both districts, there was no apparent temporal trend in the biting behaviour of rabid dogs ( Figure S3 ), suggesting that domestic dog vaccination was the main factor responsible for reducing transmission. Domestic dog demography. To calculate vaccination coverage and the decline in herd immunity due to population turnover and waning of vaccine-induced immunity, it was necessary to estimate the size of the domestic dog populations (N) and their rates of growth (r dogs ). We projected human population sizes in both districts using 2002 national census data [39, 40] , and we calculated human:dog ratios from household questionnaires conducted in 1994, 2003, and 2008 in Serengeti district and in 1994 and 2004 in Ngorongoro district. We then estimated dog populations from the projected human population sizes and the human:dog ratios and calculated the rate of increase of the dog population in each district (r dogs ¼ log(N t /N 0 )/t) ( Table 3 ). An alternative estimate of the rate of domestic dog population growth was derived from demographic data collected using household questionnaires. The death rate of dogs (d) was calculated using a Cox proportional hazards model of survival from longitudinal data (n ¼ 802). When pups (dogs under 3 months of age) were excluded from the model, neither age nor sex significantly affected survival. The percapita birth rate (b) was assumed to be the product of the sex ratio (q), the average litter size (l), and frequency (/) and pup survival (s) (b ¼ ql/s). These demographic parameters were estimated from crosssectional data (309 litters) and the rate of increase was calculated (r dogs ¼ b -d). Pup survival was estimated from a subset of puppies that remained in the household, because of the unknown fate of puppies that were given away or sold. We suspect that mortality of female puppies is greater than males. However, obtaining reliable data to accurately estimate pup survival is difficult, and the result of assuming equal mortality rates is an estimate of r dogs that is more conservative with respect to vaccination coverage (i.e., results in lower population-level immunity). This estimate of r dogs (0.088 dogs/y) was similar to other estimates from the region [41, 42] and close to those calculated directly from population sizes (r Serengeti ¼ 0.090 dogs/ y, r Ngorongoro ¼ 0.102 dogs/y) ( Table 3) . A comparison of the stable age distribution (calculated from cross-sectional data assuming a roughly constant rate of population growth) was consistent with age distributions predicted from the estimated demographic parameters. Analysis of the impacts of intervention. To evaluate whether the predicted level of vaccination coverage required to control rabies (P crit ¼ 1 -1/R 0 ) was sufficient in practice [20] , we plotted the size of village-level outbreaks (an outbreak was defined as at least two cases not interrupted by an interval of more than one month) against vaccination coverage in that village at the time of the case that initiated the outbreak. Vaccination coverage was modeled by susceptible reconstruction using demographic parameters described above (we show the results from using the largest estimate of r dogs (0.10 dogs/y) because this gives the most conservative predictions of the impacts of vaccination, but results are very similar using the lower r dogs estimates). We assumed coverage was approximately 20% in January 2002 and that the duration of vaccine-induced immunity (1/m) was approximately 3 y (http://www.intervet.co.uk/Products_Public/Nobivac_Rabies/ 090_Product_Datasheet.asp). Numbers of vaccinated and susceptible animals within a village were adjusted according to the doses of vaccine used at village vaccination stations on each campaign date (sufficient vaccine was provided such that all animals brought to the station could be vaccinated). A time series of cases in a village and the associated susceptible reconstruction are shown in the inset of Figure 4A . To predict the expected size of outbreaks given the observed variability in transmission, we simulated outbreaks in a starting population of 500 dogs (similar to the domestic dog population size in an average village); this choice had little effect on our results. We used our parameter estimates (Table 1) to randomly assign secondary cases and corresponding generation intervals. Each realization was seeded by a single animal and the starting population was initialized with vaccination coverage generated from a binomial distribution. For comparison with the outbreak data we conditioned each realization upon .1 secondary case ( Figure 4A ). Demographic parameters were incorporated, and 10,000 runs were completed for each starting condition. We also calculated the probability of an outbreak of a particular size or larger being seeded by one infectious case to evaluate the coverage needed to prevent outbreaks with different degrees of certainty ( Figure 4C and Figure S4 ). If V and N denote numbers of vaccinated individuals and the total population size respectively, then vaccination coverage can be expressed as a proportion P ¼ V/N. The number of vaccinated dogs declines following a campaign as individuals die and as vaccineinduced immunity wanes (V t ¼ V 0 e -(dþm)t , where d is the death rate and 1/m is the duration of vaccine-induced immunity), whereas the total population grows at the rate of population increase (N t ¼ N 0 e rt ). To prevent sustained endemic transmission, vaccination coverage must be maintained above P crit (such that R is held below 1). From our estimates of demographic parameters and R 0 , we calculated the proportion of the population that needs to be vaccinated, P target , to prevent vaccination coverage falling below P crit during the interval, T, between campaigns: P target ¼ e (mþdþr)T P crit . This formulation for estimating the coverage needed to interrupt endemic transmission given turnover in the domestic dog population assumes that immunity from vaccination lasts an average of 1/m time units and declines exponentially. In reality, vaccine-induced immunity is likely to be closer to a fixed duration, and thus fewer dogs would be expected to lose immunity during the 1-y interval between campaigns than under the exponential model. This indicates that our estimate of P target may be slightly overestimated, although this is an important area for further investigation. Supporting Information Figure S1 . The distributions of R 0 estimates from fitting curves to simulated epidemic trajectories generated from biting behaviour described by a negative binomial distribution (black) with mean and variance equal to observed biting behaviour or by a poisson distribution (red) with the same mean. The range of R 0 estimates from simulations spans the range of estimates from compiled outbreak data from around the world ( collection and analysis, decision to publish, or preparation of the manuscript.

Detection of genetically modified organisms (GMOs) using isothermal amplification of target DNA sequences

BACKGROUND: The most common method of GMO detection is based upon the amplification of GMO-specific DNA amplicons using the polymerase chain reaction (PCR). Here we have applied the loop-mediated isothermal amplification (LAMP) method to amplify GMO-related DNA sequences, 'internal' commonly-used motifs for controlling transgene expression and event-specific (plant-transgene) junctions. RESULTS: We have tested the specificity and sensitivity of the technique for use in GMO studies. Results show that detection of 0.01% GMO in equivalent background DNA was possible and dilutions of template suggest that detection from single copies of the template may be possible using LAMP. CONCLUSION: This work shows that GMO detection can be carried out using LAMP for routine screening as well as for specific events detection. Moreover, the sensitivity and ability to amplify targets, even with a high background of DNA, here demonstrated, highlights the advantages of this isothermal amplification when applied for GMO detection.

The ability to detect the presence of GMO is pivotal for consumers to be able to exercise their lifestyle choice of whether to consume, or not, food containing GMOs. Though the detection and quantification of GMO proteins using immunoassay has been reported [1] , denaturation of the protein during processing makes the method unsuitable for GMO testing and quantification of food. The durability of DNA makes it a better substrate for testing and its amplification by PCR is the method of choice, as recommended by the EC (2004/787), for detection and quantification of GMO in samples. An alternative DNA amplification method was described by Notomi and coworkers [2] called 'loop mediated isothermal amplification' (LAMP). The LAMP assay relies on the design of a set of primers that generate stem looped (hairpin) structures during the early stage of DNA synthesis. The hairpin structures form because two of the primers used contain, at their 5' end, a reverse complement of a sequence that is present in the target further downstream of the initial binding site. Displacement primers help the formation of these hairpins at the ends of the DNA strands and once formed, these structures can be copied into a series of DNA fragments containing multiple units of the target sequence under isothermal conditions utilizing the displacement properties of Bst polymerase (see [3] ). Although LAMP was first described using a set of four primers, enhanced sensitivity was reported using an additional pair of loop primers [4] . As the reactions are performed at a single temperature, LAMP assays can be performed very quickly since there are no separate denaturation, annealing and extension steps, and as such, reactions do not require thermalcyclers. Here we assess the LAMP protocol for the detection of GMOs using primers that target event-specific sequences for transgenic MS8 and RF3 oilseed rape (Brassica napus L.) and generic GMO sequences such as the cauliflower mosaic virus 35S promoter (P-35S) and the promoter and terminator for the nopaline synthase gene (P-nos and T-nos, respectively) from Agrobacterium spp. The LAMP technique relies on the design of an interrelated set of primers. The orientation and positions are important for self-priming through stem-looped products that drive and perpetuate the reaction. The OSR events MS8 and RF3 arise from the insertion of two closely related transgenes from the plasmids, pTHW107 and pTHW118, respectively [5] . The former encodes the Barnase gene that give rise to male sterility, which is replaced in the latter by the Barstar gene which restores fertility: both also have the selectable marker bar which confers tolerance to the herbicide glufosinate ammonium. Though the left border of the RF3 insert has undergone rearrangement in the form of duplication and inversion [5] , the right borders of both events are relatively intact (our data [6] which agree with the two sequences in the database). Even though the insertions of the transgenes have different breakpoints from the plasmids, they are very close so it was possible to design assays for the RF3 and MS8 events utilizing a common set of primers within the transgenes ( Figure 1 ). When used in conjunction with primers for the plant sequence at the border of each event, they are able to detect each event ( Figure 2 ). Since they have 50% of primers in common, it was important to determine whether there was any cross reaction between the assays. Specificity of the two assays was tested using plasmid DNA for each event. No cross-reaction between the two targets was observed ( Figure 2 ). The sensitivity of the LAMP reaction was assessed in two ways: copy number detection and background in which 10 copies of the target could be detected. Copy number detection was measured by serial dilutions of known amounts of DNA containing the target sequences, either as genomic or plasmid DNA ( Figure 3 ). Reactions fail in both assays at DNA molecule number of less than 1 which is consistent with the stochastic probability of a target being present [7] . We note that sometimes non-specific amplification can also take place, especially where the target DNA is absent (see Figure 3 ) and there is low amounts of DNA present in the reaction (cf Figures 3a and 4) . These do not form the specific banding patterns representing the different multimeric LAMP products that are characteristic for each assay and thus can be easily distinguished on a gel. Alternative banding patterns have been observed, also for low template reactions; analyses of these products show that they are formed by interactions of the primers used [8] , to form LAMP equivalents of primer-dimers. Two factors seem to be important: in the absence of target, low background DNA may aid the formation of non-specific products that go on to be amplified; and freeze-thaw repetitions may induce damage to the primers to permit the formation of 'primer intermediates' which can then be amplified. Since LAMP is capable of non-specific amplification, techniques that rely on the detection of by-products of DNA synthesis, e.g., the use of magnesium pyrophosphate precipitation [9] or the use of SyBr Green dye [10] may not be able to distinguish between real and Right border sequences of MS8 and RF3 Figure 1 Right border sequences of MS8 and RF3. Sequences of the plant (above) and transgene (below) at the right border junctions for MS8 and RF3. Highlighted sequences are the targets of the LAMP primers. The plant sequences are those shown in Table 1 : dark blue bases highlighted are the outer displacement primers, yellow and green sequences are the 5' and 3' ends of the LAMP primers respectively, and light blue sequences depict the loop primers. The sensitivity of the LAMP assay and its suitability for practical GMO detection was tested using assays for commonly-used sequence motifs, the CaMV 35S promoter (P-35S) and the Agrobacterium promoter and terminator for the nopaline synthase gene (P-nos and T-nos, respectively). These sequences are commonly used in constructs used to create approved GM events (see [11] ). RoundUp Ready™ soya construct contains both P-35S and T-nos so provided convenient template for testing the assays. We used a sample where the copy number of the GM has been well characterized and thus control the number of template in each reaction. LAMP sensitivity was assessed by the detection of ten RoundUp Ready™ GMO targets in a background of 100 ng of genomic oilseed rape DNA (Figure 3 ). Since the C value of both species is approximately 1 pg [12] , this background DNA represents a GMO level of 0.01% for both Tnos and P-35S assays. OSR DNA was used because we did not have any soya DNA free from RoundUp Ready™. DNA extracted from our 0% CRM was shown to contain 0.002% GMO [13] . The use of 100 ng of this sample would be equivalent to adding 200 copies of the transgene sequence, considerably more than the experimental input of 10 copies. We believe the use of OSR DNA to be a valid substitute since none of the primer sequences for either assay will be present in non-transgenic soya or oilseed rape. We have tested the upper limits of DNA that LAMP reactions can tolerate and found that up to 200 ng DNA in a 20 μl reaction, positive detection is reproducible. Above this DNA level, reactions become unreliable (data not shown). We have found that denaturation of template was a prerequisite step prior to amplification, unlike results found by Nagamine and co-workers [14] . This can be explained by the fact that we do not use a base pair destabiliser, such as betaine, in the reaction buffer. Since we are detecting down to near single copies of templates, our results suggest no benefit to the sensitivity of the assay by their inclusion. The consistent amplification within all dilutions showed that LAMP is an 'all or nothing' reaction, with little of the tailing off effect that is often observed in PCRs with diluting templates. This makes it easy to identify positive reactions. Together with specificity and the speed at which reactions can be performed, LAMP is an excellent method for diagnostics [8, 10, 15] . The use of CaMV 35S promoter sequence in LAMP has previously been reported as a screening method [16] . Here we demonstrate the sensitivity and reliability of the LAMP method for GMO detection, both with generic and GMO-specific assays. The ability to perform reactions in a simple heated block or water bath without the need for thermal cycling makes testing using LAMP more accessible. That LAMP is able to detect very small amounts of target and do that even in high amounts of background DNA makes it ideal for GMO detection. GMO testing can be performed in steps: routine screening for the presence of GMOs using generic assays such as for 35S promoters and T-nos; and if required, identification of specific events can be performed using event-specific assays. Equally, direct screening using event-specific assays is also feasible. The levels of sensitivity are orders of magnitude below the permissible threshold for GMO in food and feed (EC regulation 1830/2003), ensuring the detection of the presence of GMOs at acceptable levels and the reliable detection of any presence of unauthorised GM events, for which at present there is no legally acceptable lower limit (according to EC regulations). Conventional oilseed rape (OSR) seed, variety 'Hearty' was a gift from Christine Lewis, NIAB. The sample was originally purchased from Monsanto UK Ltd (Cambridge, UK). DNA was extracted by grinding 2 g seed with 10 ml extraction buffer [0.5 M NaCl, 0.1 M EDTA pH 8 and 1% (w/v) SDS] in a mortar with a pestle. The sample was emulsified with 5 ml of chloroform:isoamyl alcohol (24:1) and poured into a 20 ml Falcon polypropylene tube. After centrifugation at 1000 g for 5 mins, the aqueous phase was transferred to a new tube and nucleic acids DNA from the oilseed rape MS8/RF3 was extracted from seedlings from a selfed MS8/RF3 plant [17] using DNeasy Plant DNA Extraction Kit (QIAGEN, Crawley, UK). The parent plant was genotyped to be MS8MS8/RF3rf3 using real-time PCR (data not shown). The sample was quantified using picogreen fluorescence (Molecular Probes Inc., Invitrogen). DNA containing RoundUp Ready™ soya was extracted from Soya Roundup Ready™ GMO Reference Material (Fluka Biochemika, Sigma-Aldrich, Dorset, UK) and the GMO concentration of the sample has been accurately determined by dilutions of template combined with statistical analysis [13] . The plasmid pGreenII 0049 was a gift from Mark Smedley and Wendy Harwood of the John Innes Centre. Details of the plasmid can be found at the website: http:// www.pgreen.ac.uk/JIT/pGreen0000_fr.htm Sensitivity assessment of LAMP Detection Figure 3 Sensitivity assessment of LAMP Detection. A. Sensitivity of LAMP using genomic target. DNA from MS8/RF3 sample (16 ng.μL -1 ) was serially diluted and amplified by LAMP, in triplicate, using primers to assay for the RF3 junction. The numbers in parenthesis represent the approximate copy numbers of the target assuming that the sample represents RF3 in a hemizygous state (determined using RT-PCR data not shown) for the transgene and using 1 pg as the genome size for oilseed rape. C is the no DNA control and M represents molecular size markers. The smear (*) shows an example of non-specific amplification. B. Sensitivity of LAMP using plasmid target. Serial dilutions of the plasmid pGreenII were amplified using primers for the Pnos target. Numbers represent the calculated copy numbers of the plasmid derived from the DNA value. C is the no DNA control and M represents molecular sized markers. Plasmids containing each event were constructed to test the specificity of the MS8 and RF3 assays separately. The junction at the right borders of the transgenes were amplified by PCR from the MS8/RF3 DNA sample using the displacement (outer) primers of the LAMP reactions, MS8-RF3 DisplR (B3c) separately with MS8 DisplF (F3) and RF3 DisplF (F3), to amplify the MS8 and RF3 junctions, respectively (see Figure 1 and Table 1 ). The fragments were cloned into pGEM-T vector (Promega, Southampton, UK) and transformed into DH5α. Clones containing the correct inserts were confirmed by sequencing. Target sequences for P-35S, P-nos and T-nos were chosen based upon common identity between different plasmids in the EMBL database containing the promoters and terminator. The sequences of the targets and positions of the primers are provided (see Additional file 1). Primers for each target segment have Tm's of 50-52°C (calculated using the 2 × AT, 4 × GC formula), except for the F2 and B2 regions (3' of the LAMP primers), where the Tm was 54-56°C. Primer sequences are listed in Table 1 . For LAMP reactions, primers were purchased from Sigmagenosys (Table 1) . Reactions were performed in 20 μl containing 1 × Bst pol buffer (NEB, Ipswich, UK) with 0.4 mM each dNTP with the appropriate primers listed in Table 1 : Displacement primers were each used at a concentration of 0.2 μM; Loop primers at 0.4 μM and LAMP primers at 0.8 μM in the reactions. Enough reaction reagents, without template and enzyme, were mixed together and split into two. Template (1 μl) was added to 9 μl of the mix and the samples denatured at 95°C for 2 mins and then cooled to 4°C. Bst pol was added to the remaining reaction mix to a concentration of 3.2 U.μl -1 , mixed thoroughly, and 10 μl was added to each reaction. Reactions were incubated at 55°C for 2 hours, followed by 80°C for 10 mins to inactivate the enzyme and stored at 4°C until analysed. Aliquots of the reactions (5 μl) were run on 1.5% (w/v) agarose gels, containing ethidium bromide

Multivalent HA DNA Vaccination Protects against Highly Pathogenic H5N1 Avian Influenza Infection in Chickens and Mice

BACKGROUND: Sustained outbreaks of highly pathogenic avian influenza (HPAI) H5N1 in avian species increase the risk of reassortment and adaptation to humans. The ability to contain its spread in chickens would reduce this threat and help maintain the capacity for egg-based vaccine production. While vaccines offer the potential to control avian disease, a major concern of current vaccines is their potency and inability to protect against evolving avian influenza viruses. METHODOLOGY / PRINCIPAL FINDINGS: The ability of DNA vaccines encoding hemagglutinin (HA) proteins from different HPAI H5N1 serotypes was evaluated for its ability to elicit neutralizing antibodies and to protect against homologous and heterologous HPAI H5N1 strain challenge in mice and chickens after DNA immunization by needle and syringe or with a pressure injection device. These vaccines elicited antibodies that neutralized multiple strains of HPAI H5N1 when given in combinations containing up to 10 HAs. The response was dose-dependent, and breadth was determined by the choice of the influenza virus HA in the vaccine. Monovalent and trivalent HA vaccines were tested first in mice and conferred protection against lethal H5N1 A/Vietnam/1203/2004 challenge 68 weeks after vaccination. In chickens, protection was observed against heterologous strains of HPAI H5N1 after vaccination with a trivalent H5 serotype DNA vaccine with doses as low as 5 µg DNA given twice either by intramuscular needle injection or with a needle-free device. CONCLUSIONS/SIGNIFICANCE: DNA vaccines offer a generic approach to influenza virus immunization applicable to multiple animal species. In addition, the ability to substitute plasmids encoding different strains enables rapid adaptation of the vaccine to newly evolving field isolates.

The highly pathogenic H5N1 influenza virus causes lethal multi-organ disease in poultry, resulting in significant economic losses and a public health concern in many parts of the world. The greatest threats posed by this virus are its ability to cause mortality in humans, its potential to compromise food supplies, and its possible economic impacts. Viral maintenance in poultry potentiates the risk of human-to-human transmission and the emergence of a pandemic strain through reassortment. An effective, safe poultry vaccine that elicits broadly protective immune responses to evolving flu strains would provide a countermeasure to reduce the likelihood of transmission of this virus from domestic birds to humans and simultaneously would protect commercial poultry operations and subsistence farmers. DNA vaccines have been shown to elicit robust immune responses in various animal species, from mice to nonhuman primates [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] . In human trials, these vaccines elicit cellular and humoral immune responses against various infectious agents, including influenza, SARS, SIV and HIV. In addition to their ability to elicit antibody responses, they also stimulate antigenspecific and sustained T cell responses [1] [2] [3] 6, 12, 13] . DNA vaccination has been used experimentally against various infectious agents in a variety of mammals, including cattle (against infectious bovine rhinotracheitis/bovine diarrhea virus, leptospirosis and mycobacteriosis) [14, 15] , pigs (against classical swine fever virus and mycoplasmosis) [16] , and horses (against West Nile virus and rabies) [17] . In addition, DNA vaccines have been tested against avian plasmodium infection in penguins [18] and against influenza and infectious bursal disease in chickens [7, 8, 19] , duck hepatitis B virus in ducks [6] , and avian metapneumovirus and Chlamydia psittaci in turkeys [20, 21] (reviewed in ref. [22] ). While they have been used in chickens to generate antisera to specific influenza viruses and confer protection against the low pathogenicity H5N2 strain [23] , there is only one previous report of a monovalent DNA vaccine effective against H5N1 (and that only against a matched H5N1 isolate) [24] ; no protection with multivalent DNA vaccines against heterologous strains has been reported. Development and characterization of a DNA vaccine modality for use in poultry offers a potential countermeasure against HPAI H5N1 avian influenza outbreaks. The virus can infect humans, typically from animal sources, including commercial and wild avian species, livestock, and possibly other non-domesticated animal species [25] [26] [27] . While there is marked diversity in the host range of type A influenza viruses, many experts have speculated that a pandemic strain of type A influenza could evolve in avian species or avian influenza viruses could contribute virulent genes to a pandemic strain through reassortment [28, 29] . Thus, there is reason to consider vaccination of poultry that would stimulate potent and broad protective immune responses [7, 30, 31] . In undertaking such efforts, it is important that there be a differentiation of infected from vaccinated animals [32] so that animals can be protected and permit monitoring of new infections using proven and sensitive methodologies. In this study, we used an automated high capacity needle-free injection device, Agro-JetH (Medical International Technology, Inc., Denver, CO) to explore the feasibility of DNA vaccination of poultry. After optimization of injection conditions, alternative multivalent DNA vaccine regimens were analyzed and compared for magnitude and breadth of neutralizing antibodies, as well as protective efficacy after challenge in mouse and chicken models of HPAI H5N1 infection. The findings suggest that it is possible to develop a multivalent DNA vaccine for poultry that can protect against multiple HPAI H5N1 strains and that could keep pace with the continued evolution of avian influenza viruses. Immunogenicity and neutralizing antibody specificity of alternative HA DNA vaccines in mice To evaluate the efficacy of multivalent DNA vaccines, initial studies were performed in mice. Expression vectors encoding HAs from ten phylogenetically diverse strains of influenza viruses [33] were generated by synthesis of cDNAs (see Materials and Methods) in plasmid expression vectors, pCMV/R or pCMV/R 8kB, which mediates high level expression and immunogenicity in vivo [34, 35, 36] . Animals were immunized with each expression vector intramuscularly (IM) at three week intervals, and the antisera were evaluated on day 14 after the third immunization for their ability to neutralize HPAI H5N1 pseudotyped lentiviral vectors as previously described [35, 36] . We have previously shown that lentiviral assay inhibition (LAI) yields similar results to microneutralization and HAI analyses with higher sensitivity in mice [35, 36] To determine whether immunization with multiple HAs simultaneously could expand the breadth of the neutralizing antibody response without significant loss of magnitude, a combination of 10 HA DNA vaccine immunogens was administered IM at proportionally lower concentration (1.5 mg per immunogen) into groups of 10 mice (see Materials and Methods). Remarkably, despite a log lower DNA concentration of each component, significant neutralizing antibody titers were generated to each of the 10 immunogens, with .80% neutralization against 6 out of 12 H5 HA pseudoviruses at dilutions of up to 1:400 ( Fig. 2A) . To evaluate whether similar breadth of immunity could be generated with fewer immunogens, two different combinations of 5 immunogens were selected, based on the phylogenetic diversity of HA among the avian influenza viruses [33] and the crossreactivity of the neutralizing antibody responses of select individual immunogens (Fig. 1 ). As expected, there were substantial differences in the breadth of neutralization between these two sets of 5 immunogen multivalent vaccines (Fig. 2 , B vs. C). In one set, while neutralization of homologous strains was comparable to the monovalent and the 10 immunogen multivalent immune response, fewer cross-reactive antibodies were detected, directed most prominently against A/Iraq Protection of DNA-vaccinated mice against challenge with heterologous H5N1 A/Vietnam/1203/2004 influenza virus Mice immunized as described above were challenged with a heterologous H5N1 virus 68 weeks after the final DNA vaccination. Animals were then challenged with 10 LD 50 of the highly pathogenic A/Vietnam/1203/2004 virus intranasally, and morbidity and mortality were monitored for 21 days after the viral challenge. The control animals, injected with the plasmid expression vector with no insert, died within 10 days of infection. Complete survival was observed in the groups immunized with the 10 component and set 2 of the 5 component multivalent DNA vaccines (Fig. 3) . Immunization with HA derived from the A/ Indonesia/05/2005 strain or set 1 of the 5 component multivalent DNA vaccine showed a survival rate approaching 90%. In contrast, animals injected with HA plasmid DNA derived from A/ Anhui/1/2005, which has diverged more from A/Vietnam/ 1203/2004, showed a lower percent survival (70%) after lethal viral challenge. Survival differences between groups were assessed using a log-rank test and the Gehan-Wilcoxon test on the survival curves for pairs of groups. A test was deemed significant if the pvalue was ,0.01. Mice injected IM with different HAs, A/ Indonesia/5/05, A/Anhui/1/05, 10HA, 5 HA (Set 1), or 5 HA (Set 2) showed a significant difference compared to control (all p values,0.001). Among the HA-immunized groups, there was no significant difference between any two groups (p.0.08 for all comparisons). For example, no significant difference was observed between the A/Anhui/1/05 group, which had the least survival among the HA immunized groups (7 out of 10), and other HA groups: A/Indonesia/5/05 (p = 0.377), 10 HA (p = 0.082), 5 HA (Set 1) (p = 0.101), or 5 HA (Set 2) (p = .411). Therefore, we cannot exclude the possibility that the 3 deaths in the A/Anhui/1/05 group may have been due to random chance. Since it is desirable to confer protective immunity in poultry and HA DNA vaccination was effective in mice, we next examined the breadth and potency of single or multiple HA plasmid immunization in chickens. The ability of chickens to generate specific antibodies was assessed with three strains that showed broad cross protection in mouse studies (A/Vietnam/1203/2004, A/Anhui/ 1/2005 and A/Indonesia/05/2005), administered individually or in combination, by different injection methods. In addition to needle injection, a needle-free repetitive injection device, Agro-JetH (Medical International Technology, Inc., Denver, CO), was analyzed. This device disperses the 0.1 to 5 ml injection doses into the dermal, subcutaneous, or intramuscular tissue depending upon the pressure adjustments, powered by a CO 2 gas pressure plunger [39] . The injection conditions were determined by histologic analysis of tissues that received injections of India ink; a pressure of 48 psi was chosen since it enabled consistent delivery into intradermal and subcutaneous tissues (Fig. S1 ). Immunization of chickens with the control plasmid (CMV/R) without an HA gene insert elicited minimal neutralizing antibody titers compared to HA-immunized animals 1 week after 3 DNA immunizations. Nearly all chickens immunized with either monovalent or multivalent HA DNA vaccines generated significant neutralization titers ( Fig. 4 and Table S1 ). In general, there was a progressive increase in the amount of neutralization after each successive DNA vaccination (data not shown) with maximal response at 1 week after the 3 rd DNA immunization, with highest and most consistent levels in the trivalent vaccine group delivered with the Agro-JetH device. Neutralization of the Indonesia HA strain was the most robust, with neutralization nearing 100% at titers greater than 1:3200. Both the monovalent and multivalent vaccines elicited robust homologous ( vaccine (Fig. 4 ). Even though one chicken (238) in the multivalent vaccine group produced almost the same degree of neutralization at each time point and was protected, it did not produce a high neutralizing antibody titer for reasons that were uncertain but possibly related to a non-specific inhibitor in the sera. To determine whether chickens immunized with single or multiple DNA vaccines were protected from a lethal challenge of a heterologous HPAI H5N1 virus, vaccinated chickens were In panels B and C, mice (n = 10) were immunized with 15 mg of plasmid (3 mg each) three times at 3 week intervals. Serum pools from the immunized animals were collected 14 days after the third immunization. The antisera were tested against the 12 indicated pseudotyped lentiviral vectors at varying dilutions. Error bars at each point indicate the standard deviation; each sample was evaluated in triplicate. In general, the immunized serum neutralized all tested pseudotyped lentiviruses at low dilutions while differences were often observed at high dilution. doi:10.1371/journal.pone.0002432.g002 inoculated with 20 LD 50 of highly pathogenic A/Vietnam/1203/ 2004 heterologous virus intranasally using standard methods [25, 40] and monitored for morbidity, mortality, viral shedding and serum antibodies. While all the control animals died within 2 days of infection, 100% survival was noted in the rest of the chickens (Fig. 5A ). The animals that were healthy, showing no signs of clinical disease or malaise, were euthanized on day 14. There was no evidence for viral shedding monitored via tracheal and cloacal swabs of infected chickens 2-14 days after challenge as determined by embryonal inoculation (data not shown: egg infectious dose 50 (EID 50 ) limit of detection ,100 virus particles). To compare the relative efficacy of DNA vaccines delivered IM by needle and syringe versus the needle-free Agro-JetH device injection, a dose-response study was performed with amounts of DNA vaccine ranging from 500 to 0.5 mg with two inoculations. In these experiments, the HA derived from A/chicken/Nigeria/641/ 2006 was substituted for A/Vietnam/1203/2004 since it represented a more contemporary isolate. The observed rate of protection was higher among the animals receiving 5 mg by Agro-Jet (8/8) than by IM injection (6/8) (Fig. 5, B vs. C). Both modes provided complete protection for all animals at doses higher than this, and 25% protection for the animals receiving 0.5 mg doses (Fig. 5B, C) . Survival differences between consecutive doses were assessed using a log-rank test on the survival curves for pairs of groups. A test was deemed significant if the p-value was ,.01, and marginally significant if the p-value was ,.05 but ..01. Chickens injected IM showed a marginally significant difference between 0.5 and 5 mg (p = .047). In the same group there was a significant difference between control and 5, 50 and 500 mg (p,.001 for all comparisons) and the difference between control and 0.5 mg was marginally significant (p = .016). Chickens that were injected using Agro-JetH showed a significant difference between 0.5 and 5 mg (p = .004) and between control and 5, 50, and 500 mg (p,.001 for all comparisons). There were no differences between control and 0.5 mg or between 5, 50, and 500 mg. Lastly, the survival differences between Agro-JetH and IM for each dose group were not significant. The neutralizing antibody response to homologous and heterologous HAs corresponded with protection and correlated with dose, with higher titers elicited by injection with Agro-JetH compared to needle (Table S2) . We assessed viable viral shedding after inoculation by chick embryo inoculation three days after virus challenge (Week 8). While we noted some embryonic lethality at the 0.5 mg dose, there was no embryonic lethality at 5, 50 or 500 mg groups (data not shown). Since the HPAI H5N1 virus first appeared ten years ago, this highly pathogenic avian influenza virus has shown increasing diversification and dissemination in Asia, Africa, and Europe [28, [41] [42] [43] [44] . In addition to its effects on human health by crossspecies transmission [28, 45, 46] and ability to compromise food sources, it poses a continuing threat to public health as it evolves and adapts in different species. The pandemic potential of this virus, especially as it relates to the poultry industry and for reservoir avian hosts, underscores the need for a vaccine that offers broad spectrum immunity and protection against lethal viral challenge. While the virus remains restricted in its ability to infect humans and undergo efficient human-to-human transmission [28, 47] , its persistence and spread in poultry increases the risk of the emergence of a pandemic strain. One approach to pandemic risk reduction is to limit the propagation of the virus in poultry and other relevant avian species. We have previously reported that DNA vaccines encoding HA can confer protection against a highly lethal human pandemic influenza virus, the 1918 H1N1 virus, in mice [36] . DNA vaccines offer several advantages, including the ability to express diverse antigens, tolerability in various hosts, ease of delivery, and stability for storage and distribution without the necessity of maintaining a cold chain; they have been shown to be safe and efficacious in a variety of animal models [2, 4, 12, 22, 48] . Because they do not contain other viral proteins used to screen for infection, they also address the need to differentiate vaccinated from infected animals. There is evidence that DNA vaccination elicits cell-mediated immunity against influenza HA in addition to inducing an antibody response [36] , an effect that could significantly contribute to protective immunity as viruses show genetic drift and reduced susceptibility to neutralization. Ideally, a highly effective influenza vaccine should not only be able to let the host develop a protective immune response against a matching live virus challenge but also elicit robust protective immune responses against a broad range of homologous and heterologous H5 influenza strains. A multivalent H5 vaccine containing diverse serotypes could expand the antigenic breadth sufficiently to provide protection against heterologous challenge and may preclude the emergence of vaccine-resistant strains that may arise due to evolutionary vaccine pressure on the virus. Due to the antigenic drift and shift of the influenza virus genome, it has been very difficult to predict the next dominant strain of an avian endemic outbreak. DNA vaccines can be synthesized in a relatively short period of time, and the targeted mutations can be tailored to specific viral serotypes. The mutations promote a focused and enhanced immune response [3, 49, 50] that may be particularly important in the event of an outbreak where specificity is the key to epidemic control. The use of modified codons ensures maximal expression in the host and eliminates the possibility of recombination with influenza viruses that might potentially generate new strains. A more broadly protective murine vaccine was developed here by including more HAs from varying strains in the multivalent vaccine (Figs. 2 and 3) . However, it is less practical to include large numbers of different HAs in one vaccine due to the cost and complexity of manufacturing such a vaccine. Therefore, we simplified the vaccine regimen based on cross-neutralization studies and phylogenetic relationships. A trivalent vaccine was subsequently identified for further studies. Due [51] . While three DNA immunizations were used initially to demonstrate protective immunity and have been used previously to elicit protection in mice [36] , we found that effective protective immunity could be induced with two DNA vaccinations and as little as 5 mg trivalent DNA immunization using the ID/SC route with the Agro-JetH device. In addition, based on the chick embryo inoculation data, we believe that there is effective neutralization of the virus and lack of infectious viral shedding in chicken vaccinated with as little as 5 mg of DNA. The device's capacity for rapid repetitive injection and the lower quantity and stability of DNA enhance the practicality and utility of this approach for vaccination of endangered species in captivity or administration to poultry or other animals. A/Vietnam/1203/2004 (H5N1) (A/VN/1203/04) was obtained from the repository at the Centers for Disease Control and Prevention (CDC), Atlanta, Georgia. The virus was propagated in 10-day old embryonated chicken eggs at 35uC and stored at 270uC until use. The virus was titrated by the Reed and Muench method to determine EID 50 [52] . GenBank ABD28180) were synthesized using human-preferred codons (GeneArt, Regensburg, Germany) [36] . HA cDNAs from diverse strains of influenza viruses were then inserted into plasmid expression vectors, pCMV/R or pCMV/R 8kB, which mediates high level expression and immunogenicity in vivo [34, 35, 36] . For initial trivalent immunizations in chickens, the A/Vietnam/1203/ 2004, A/Anhui/1/2005 and A/Indonesia/05/2005 strains were used and in the dose response study, the Vietnam strain was replaced with A/chicken/Nigeria/641/2006. The immunogens used in DNA vaccination contained a cleavage site mutation (PQRERRRKKRG to PQRETRG) as previously described [35, 36] . This mutation was generated by site-directed mutagenesis using a QuickChange kit (Stratagene, La Jolla, CA). DNA immunization of mice [6] [7] [8] week old female BALB/c mice were purchased from The Jackson Laboratory and maintained in the AAALAC-accredited Vaccine Research Center Animal Care Facility (Bethesda, MD) under specific pathogen-free conditions. All experiments were approved by the Vaccine Research Center Animal Care and Use Committee. The mice were immunized as previously described [5] . Briefly, mice (10 animals for all test groups, 20 animals for the The study was carried out in the AAALAC-accredited animal facility at the University of Maryland School of Medicine. Six groups of 8 one-day-old male and female SPAFAS White Leghorn Chickens, Gallus domesticus, were obtained from Charles River Laboratories (Connecticut). The animals were housed in brooder and grower cages (McMurray Hatcheries, Iowa). Feed (Teklad Japanese Quail Diet -3050, Harlan-Teklad, WI) and water were provided to the animals ad libitum. The study was performed in strict accordance with the ''Guide'' after approvals from the Animal Care and Use Committees of the Vaccine Research Center, NIH and the University of Maryland. DNA immunizations were performed as described at 0, 3 and 6 weeks. A total dose of 500 mg of one or a combination of the following DNA plasmids in a volume of 250 ml was administered to each animal: pCMV/ R, pCMV/R-HA Agro-JetH is a needle-free device used for mass delivery of vaccines and drugs in livestock and poultry. The device is semiautomatic and requires a small CO 2 tank or compressed air for low pressure delivery. Upon trigger activation, CO 2 disperses the injectate at a precise dose into the muscle, dermis or subcutaneous tissue depending on the setting that was standardized for our use. We used an effective volume of 0.1 ml in our injectate [39] . In this study we were able to effectively deliver 0.1 ml of injectate into the animal's dermis/subcutaneous tissue at a pressure of 48-55 psi. Sixty-eight weeks after the last immunization, female BALB/c mice were lightly anesthetized with Ketamine/Xylazine and inoculated intranasally with 10 LD 50 of A/Vietnam/1203/2004 virus diluted in phosphate-buffered saline in a 50 ul volume. Mice were monitored daily for morbidity and measured for weight loss and mortality for 21 days post infection. Any mouse that had lost more than 25% of its body weight was euthanized. All experiments involving the HPAI virus were conducted in an AAALAC accredited facility (BioQual Inc., Gaithersburg, MD) under BSL 3 conditions that included enhancements required by the USDA and the Select Agent Program. White Leghorn chickens were challenged one week after the last immunization with 20 lethal dose 50 (LD 50 ) of A/Vietnam/1203/04 (H5N1) influenza A virus, equivalent to 2610 4 EID 50 based on previous challenges [53] . Chickens were infected with 200 ml virus intranasally. Tracheal and cloacal swabs were collected days 3 and 5 post-challenge and stored in glass vials containing BHI medium (BBL TM Brain Heart Infusion, Becton Dickinson) at 280uC. Blood was collected 14 days post-challenge and serum was titered by microneutralization assay. Chickens were observed and scored daily for clinical signs of infection, morbidity and mortality. Chickens that survived the study were bled and humanely euthanized at day 14 post-challenge. Lungs, heart, intestine and kidney were collected and samples were stored in formalin for histopathology. Experiments were carried out under BSL3+ conditions with investigators wearing appropriate protective equipment and compliant with all Institutional Animal Care and Use Committee-approved protocols and under Animal Welfare Act regulations at the University of Maryland, College Park, Maryland. Representative tracheal and cloacal swabs were chosen to run an EID 50 assay for comparison and virus titers were determine by the method of Reed and Meunch [52] . Briefly, swabs were used to infect 10 day-old embryonated chicken eggs in 10-fold dilutions. Three eggs were inoculated per dilution and incubated for 48 hours before titration. Neutralizing antibodies were titrated from serum samples collected week 5 and 7 post-vaccination and day 14 post-challenge. The microneutralization assay was performed using a 96-well plate format. Serum was treated with receptor-destroying enzyme (Denka Seiken Co.) and treated at 37uC per the manufacturer's instructions. After an overnight incubation and subsequent inactivation samples were brought to a final dilution of 1:10 using PBS and each sample was serially diluted and virus, diluted to 100 TCID 50 , was added to each well. The plates were then incubated at 37uC, 5% CO 2 for 1-2 hours. Following incubation, supernatants were used to infect a second 96-well plate of MDCK cells. Microplates were incubated at 4uC for 15 minutes and then 37uC, 5% CO 2 for 45 minutes. Supernatants of serum and virus were then discarded and 200 ml of OptiMEM (containing 1X antibiotics/antimycotics, 1 mg/ml TPCK-trypsin) was added and incubated at 37uC, 5% CO 2 for 3 days. After 3 days, 50 ml of the supernatant from each well was transferred into a new 96-well microplate, and an HA assay was performed to calculate the antibody titers. Virus and cell controls were included in the assay. Two-fold dilutions of heat-inactivated sera were tested in a microneutralization assay as previously described [54] for the presence of antibodies that neutralized the infectivity of 100 TCID 50 (50% tissue culture infectious dose) of the A/Vietnam/ 1203/2004 H5N1 virus on MDCK cell monolayers by using two wells per dilution on a 96-well plate. The recombinant lentiviral vectors expressing a luciferase reporter gene were produced as previously described [35, 36] . For the neutralization assay, antisera from immunized animals were heat-inactivated at 55uC for 30 minutes and mixed with 50 ml of pseudovirus at various dilutions. The sera/virus mixture was then added to 293A cells in 96-well B&W TC Isoplates (Wallac, Turku, Finland; 12,000 cells/well). Two hours later, the plates were washed and fresh medium was added. Cells were lysed in mammalian cell lysis buffer (Promega, Madison, WI) 24 hrs after infection and luciferase activity was measured using the Luciferase Assay System (Promega, Madison, WI). The following strains were used for the production of pseudotyped viruses: for HA we used A/Thailand/1(KAN- The HA/HI titers were determined as previously described [54] . Briefly, HA titers were calculated using 50 ml of 0.5% chicken red blood cell suspension in PBS added to 50 ml of twofold dilutions of virus in PBS. This mix was incubated at room temperature for 30 minutes. The HA titers were calculated as the reciprocal value of the highest dilution that caused complete hemagglutination. HI titers were calculated by titrating 50 ml of antiserum treated with receptor-destroying enzyme and an equivalent amount of A/Vietnam/1203/2004 virus (four hemagglutinating doses) was added to each well. Wells were incubated at room temperature for 30 minutes and 50 ml of a 0.5% suspension of chicken red blood cells was added. HI titers were calculated after 30 minutes as the reciprocal of the serum dilution that inhibited hemagglutination. Table S1 Hemagglutination inhibition (HI), microneutralization titer (NT), and LAI of sera from individual chickens immunized with different vaccines. Sera from immunized animals were obtained at week 5 or 7, a week before or after the final boost, and neutralization was assessed by HI, microneutralization (NT) and LAI (shown as IC 50 ). Individual animal serum of each group is shown and was analyzed as described in the Materials and Methods section. Figure S1 Characterization of needle-free (Agro-JetH) DNA immunization in chickens. To evaluate the distribution of fluid into superficial or deep layers of subcutaneous tissues after delivery by AgroJetH, 4 or 7 week old chickens were injected with a solution containing India ink with this needle-free device at various pressures, ranging from 45 to 55 mm Hg. Three sites (thigh, wing and breast) were used, and biopsies were taken for routine hematoxylin and eosin staining. Representative sections of thigh injections are shown from 7 week old chickens and were similar at 4 weeks (data not shown). While the 48 mm Hg pressure deposited the injectate into the dermis/subcutaneous region (left), the higher pressure injections, 52 and 58 mm Hg, deposited the injectate into the subcutaneous and muscle layers (middle, right). 48 mm Hg consistently provided an optimal pressure to deposit the injectate into the dermis and subcutaneous tissue and was chosen for all AgroJetH immunizations. Found at: doi:10.1371/journal.pone.0002432.s003 (10.74 MB DOC)

Virus-Specific Read-Through Codon Preference Affects Infectivity of Chimeric Cucumber Green Mottle Mosaic Viruses Displaying a Dengue Virus Epitope

A Cucumber green mottle mosaic virus (CGMMV) was used to present a truncated dengue virus type 2 envelope (E) protein binding region from amino acids 379 to 423 (EB4). The EB4 gene was inserted at the terminal end of the CGMMV coat protein (CP) open reading frame (ORF). Read-through sequences of TMV or CGMMV, CAA-UAG-CAA-UUA, or AAA-UAG-CAA-UUA were, respectively, inserted in between the CP and the EB4 genes. The chimeric clones, pRT, pRG, and pCG+FSRTRE, were transcribed into full-length capped recombinant CGMMV transcripts. Only constructs with the wild-type CGMMV read-through sequence yielded infectious viruses following infection of host plant, muskmelon (Cucumis melo) leaves. The ratio of modified to unmodified CP for the read-through expression clone developed was also found to be approximately 1:1, higher than what has been previously reported. It was also observed that infectivity was not affected by differences in pI between the chimera and its wild counterpart. Analysis of recombinant viruses after 21-days-postinculation (dpi) revealed that deletions occurred resulting in partial reversions of the viral population to near wild type and suggesting that this would be the limiting harvest period for obtaining true to type recombinants with this construct.

The development of plant virus vectors as in planta expression systems for foreign genes provides an attractive alternative biotechnological approach for peptide expression [1] [2] [3] [4] [5] . This method has been exploited in vaccine production, where small foreign peptides are expressed as a fusion with the viral coat proteins. Essentially, an insertion site has to be determined in the virus genome so that the resulting product will be displayed on the surface of the virus particle which is then propagated in plants and consequently isolated and used as antigen presenting vehicles [5, 6] . Modifications that do not interfere with the normal functions of the particular virus are a prerequisite for this peptide fusion approach. One strategy suggests that foreign gene segments could be fused to the terminus of a viral gene in a way that permits the production of both the fusion product and the native viral protein, thus avoiding interference with normal gene functions. The success of this epitope presentation strategy depends on a detailed knowledge of virus structure at the atomic level, which is only available for a limited number of viruses. We have recently developed Cucumber green mottle mosaic virus (CGMMV) as a candidate for expressing antigenic peptides in plants [7] . CGMMV is a tobamovirus with a genome size of ∼6. 4 kb which has been well characterized both biologically [8, 9] and structurally [10, 11] . In this study, a truncated dengue virus type 2 envelope (E) protein binding region from amino acid 379 to 423 (EB4) was inserted into the end of the coat protein (CP) open reading frame (ORF) of a previously constructed CGMMV fulllength clone, pCGT7X [7] . The antigenic peptide was chosen based on a recent study that suggests its importance in enabling dengue virus to bind to specific host cell receptors (S. Abu Bakar personal communication). The present study explores the possibility of extrapolating the CGMMV antigenic epitope presentation system for developing diagnostics Figure 1 : Position of primers in constructed plasmid clones. Primers 4-8 were used to amplify the EB4 gene from pCANTAB 5E. Plasmid pCGT7X and amplified EB4 fragments were digested with HindIII and ligated together to obtain the respective plasmid clones as shown. HindIII is the insertion site of EB4 at the end of CGMMV CP. PCR amplification using primer pairs 1-9, 3-9, and 2-7 will yield amplified products of approximately 6.5 kb, 0.85 kb, and 2.2 kb, respectively. and potentially therapeutics against dengue. The study was also used to challenge the size limits of foreign gene insertion into the CGMMV vector as in the previous study the hepatitis B surface antigen (HBsAg) used was only 33 amino acids [7] . The 45 amino acid-long EB4 protein used in this study has been previously shown to react with the dengue-specific antibody 3H5-1 (S. Abu Bakar personal communication). Vector. The EB4 coding sequence was amplified from a pCANTAB 5E vector carrying the virus gene using 3 primer pair sets. The resulting PCR products were purified then digested overnight with HindIII restriction endonuclease, with the same treatment carried out on the full-length clone of CGMMV (pCGT7X, ∼9.0 kb) previously constructed [7] . The digested PCR products and the linearized pCGT7X were purified following 1% agarose gel electrophoresis, and then ligated to form pRT, pRG, both containing the TMV read-through sequence [12] and pCG+FSRTRE (containing the CGMMV read-through sequence) [13] , respectively ( Figure 1 ). The three primer sets used in the amplifications were as follows. Forward RT (5 -CCAAGCTTGCCAATAGCAATTAAT-CATAGGAGTAGAGC-3 ) and Reverse E (5 -CCAAGC-TTCTCCAAAATCCCAAGCTGT-3 ) for construction of clone pRT, Forward (RT) TGG (5 -AAGCTTGGCAATAGC-AATTAATCATAGGAGTAGAGCCG-3 ) and Reverse Q (5 -CCAAGCTTGTCCAAAATCCCAAGCTGTGT-3 ) for clone pRG, and Forward SRT (5 -CCAAGCTTCCAAATAGCA-ATTAATCATAGGAGTAGAGCCG-3 ) and Reverse E (5 -CCAAGCTTCTCCAAAATCCCAAGCTGT-3 ) for clone pCG+FSRTRE. The templates used in the in vitro transcription reactions were synthesized through long-distance PCR (LD-PCR) in 50 μL PCR cocktails containing 1X HF Buffer of Phusion DNA Polymerase (Finnzymes, Espoo, Finland) with 1.5 mM MgCl 2 (Finnzymes, Espoo, Finland), 0.2 mM dNTP mix, 0.5 μM forward primer, CGT7dG (5 -CCGAGCTCGTAATAC-GACTCACTATAGGTTTTA-3 ), 0.5 μM reverse primer, CGMMV 3 -UTR (5 -TGGGCCCCTACCCGGGGAAAA-GGGGGGAT-3 ), 10-20 ng of DNA template, and 1 U of Phusion DNA Polymerase (Finnzymes, Espoo, Finland). The reaction was set up in 0.2 mL tubes, and the thermal cycling was conducted with initial denaturation at 98 • C for 60 seconds, followed by 30 cycles of 98 • C denaturation for 10 seconds, annealing at 63 • C for 20 seconds and elongation at 72 • C for 1 minute and 50 seconds, and finally an extension step at 72 • C for 5 minutes. The amplified product was purified through phenol-chloroform extraction followed by ethanol precipitation. The pellet was dissolved in an appropriate volume of RNase-free distilled water to 1 μg/μL and stored at −20 • C till further use. The in vitro transcription was carried out using the (Ambion, Calif, USA) High Yield Capped T7 RNA Transcription Kit according to the manufacturer protocol. Aliquots of the in vitrosynthesized transcripts were denatured and electrophoresed alongside RNA markers showing its integrity and the expected transcript size of approximately 6.5 kb. Since no DNase I treatment was done, traces of DNA template of the transcription reactions were detected. 2.5. Inoculation with RNA Transcripts. One transcription reaction was used to inoculate 2 plantlets by gently rubbing the reaction mixture over carborundum-dusted first leaf and cotyledons of 10-day-old plantlets. Mock inoculation was done by gently rubbing distilled water onto carborundumdusted first leaves. The excess inoculum was rinsed off using distilled water from the leaf surfaces 60 minutes after inoculation. Total RNA was isolated from the new leaf of the inoculated and healthy plants using RNeasy Plant Mini Kit (QIAGEN). RT-PCR was performed using AccuPower RT/PCR PreMix (Bioneer, Daejeon, South Korea) with primers CGMMV 3 UTR (5 -TGGGCCCCTACCCGGGGAAAAGGGGGG-AT-3 ) paired with Pst I sense (5 -TAGGAAAAAACC-AGAAGATCTGCAGGAATTTTTCTC-3 ) or C5500F (5 -GTCGCTACAACTAACTCTATTATCAAAAAGGGTC-3 ). Reactions were carried out according to the manufacturer protocols. Infected plants will give a PCR-amplified product of approximately 2.2 kb (with PstI sense and CGMMV 3 UTR primers) or 0.85 kb (with C5500F and CGMMV 3 UTR primers). RT-PCR reactions were carried out for plants at 14, 21 , and 30 day-postinoculation (dpi). Plant virus isolation procedures used in this study were modified from [8] . Infected plants showing typical symptoms were harvested, weighed, and homogenized in ice-cold 0.1 M phosphate buffer (pH 7.0 containing 1% of β-mercaptoethanol) at 1 mL/g of plant material for 10 minutes using a mechanical blender. The homogenate was filtered through 2 layers of cheesecloth and then mixed with equal volume of chloroform:butanol (1:1). The mixture was then stirred for 1-2 hours at room temperature and then the organic phase was separated from the mixture through centrifugation at 8000 g for 15 minutes. The aqueous layer was transferred to a beaker, 100 mL of NaCl (4 g/L) and PEG6000 added, and the mixture stirred on ice for 1 hour. The precipitated virus was separated from the solution through centrifugation at 10 000 g for 30 minutes at 4 • C. The resulting pellet was reconstituted in 10 mL of 0.1 M phosphate buffer pH 7.0. Any undissolved material was cleared by centrifugation at 10 000 g for 30 minutes at 12 • C. Then 0.2 M EDTA (pH 7) (50 mL/L) was added to the supernatant and the mixture subjected to centrifugation at 110 000 g for 90 minutes at 4 • C. The supernatant was discarded, and the pellet was left to air dry. The virus pellet was then reconstituted in 100 μL of distilled water and stored at 4 • C until used. Analyses of sequences of the amplified products were carried using BioEdit Sequence Alignment Editor Software (version 6.0.5) (http://www.mbio .ncsu.edu/BioEdit/bioedit.html). The pI and charge values of the coat protein were calculated using the protein calculator developed by Chris Putnam of The Scripps Research Institute (http://www. scripps.edu/∼cdputnam/protcalc.html). In this study, the chimeric CGMMV vectors pRT, pRG, and pCG-FSRTRE were constructed by inserting the EB4 coding sequence to the end of the CGMMV CP ORF in plasmid pCGT7X. The maps of these constructed clones are shown in Figure 1 , which indicates their respective position of the primers during amplification and cloning procedures. Maps of plasmids pCGT7X are carrying the wild-type CGMMV, and pCANTAB 5E are carrying EB4 with their respective priming sites are also as indicated in Figure 1 . The genome size of wild-type CGMMV is approximately 6.4 kb (without the plasmid backbone), and the resulting chimeric CGMMV genome would be approximately 6.5 kb in size and contains the EB4 and readthrough sequences as well as inserted HindIII restriction recognition sites and additional nucleotides enabling inframe cloning. The pRT and pRG chimeric clones were constructed with the read-through sequence of TMV (CAA-UAG-CAA-UUA). This leaky sequence meets the minimal sequence requirement for effective read-through of the stop codon [12] and had been used successfully in previous reports [14] . The templates for in vitro transcription of these two clones were generated through LD-PCR (data not shown). The resulting amplified products (∼6.5 kb) consisted of a T7 promoter fused with the chimeric CGMMV genome carrying EB4. Transcripts produced from the two constructs were separately tested for infectivity by inoculating the host plants. After repeated attempts, both the pRT-and pRG-generated transcripts did not cause infection of the inoculated plants (Table 1 ). There was no evidence of virus genomic material in the inoculated plant tissues tested (data not shown). It is speculated that the read-through sequence of TMV may not be suitable for the CGMMV chimeric clones, hence contributing to the absence of infectious virus transcripts. To overcome this possibility, another chimeric clone (pCG+FSRTRE) was constructed using the read-through sequence AAA-UAG-CAA-UUA of the wild-type CGMMV itself ( Figure 1 ). The template for in vitro transcription, based on the pCG+FSRTRE clone, was generated through LD-PCR. The in vitro transcription products ( Figure 2) were used in inoculation studies. The new leaves of plants (Table 1 and Figure 3 ). Infection could be detected through RT-PCR when total RNA of new noninoculated leaf was used as template. Detection of the virus by RT-PCR ( Figure 4 ) suggests that the viruses had moved from the inoculated leaf to new leaves. This implies that the chimeric virus pCG+FSRTRE carrying the read-through sequence from the CGMMV genome (AAA-UAG-CAA-UUA) was infectious and that the virus particles were able to assemble. The plants, however, continued to grow without any further noticeable symptoms. Virus particles were extracted from the infected plants and an aliquot was electrophoretically separated on 15% SDS-PAGE ( Figure 5 ) resulting in two distinct suggesting that the virus population consisted of two species of coat proteins, the EB4-CP fusion (larger in size) and the wild-type CGMMV CP (smaller in size). The ratio of modified to unmodified CP was approximately 1:1. Apart from the usage of leaky UAG amber stop codons, it has been reported that pI:charge can affect the production of viable recombinant virus [15] . The pI of the epitope is thought to be an important factor as the hybrid coat protein pI:charge value can affect epitope presentation. It was also reported that TMV was more tolerant to positively charged epitopes on its surface. Thus, it was initially speculated that the failure in expression of the foreign peptide was possibly due to the pI:charge value of recombinant CGMMV CP which was different from the wild-type CGMMV CP pI:charge value ( Table 2) . Table 2 shows the isoelectric point (pI) and charge of the wild-type CGMMV CP, the read-through recombinant CGMMV CP, and the EB4 insert. The charge of the EB4 insert is positive and potentially suitable for expression on the surface of the CGMMV CP [2] . Hence, the inserted peptide is speculated to be expressed if the pI:charge value of modified virus CP resembles the pI:charge value of unmodified virus CP. Earlier transcripts (data not shown) generated from fusion clones without a read-through sequence, where their pI values deviated significantly (>6.0) from the wildtype CP (5.08), were not able to cause infection in inoculated plants leading to the suggestion initially that pI:charge value Journal of Biomedicine and Biotechnology 410 420 430 440 450 460 470 480 490 500 510 520 530 540 550 560 570 580 590 600 610 620 630 640 650 660 670 680 690 700 Sequence not determined Deleted sequence Gaps introduced for alignment Figure 6 : Sequence analyses of RT-PCR-amplified products from putative chimeric CGMMV RNA at different days postinoculation (dpi). The sequenced alignment shows that EB4 sequence was truncated and not complete after 21 dpi. The putative chimeric CGMMV produced did not express the EB4 and its genome resembled the wild-type CGMMV. Introduced read-through sequences and extra codons are underlined. Complete EB4 sequence is aligned accordingly with the other sequences. played an important role in virus particle assembly and infectivity. Thus, the pI value of the recombinant CP constructs was adjusted to more closely resemble the wild-type CP pI value by inserting the acidic amino acid (glutamate) to the 3 end of CP ( Table 2 ). The experiments (Table 1) , however, showed that although the pI was still higher than that of the wild type (Table 2) , the construct pCG+FSRTRE remained infectious. This implies that infectivity of the clones was not directly related to the deviation in pI value with the wild-type virus CGMMV CP. Sequencing was carried out on RT-PCR-amplified products of viral RNA extracted from putative chimeric virus particles at 30 days postinoculation (dpi) and total plant RNA isolated from inoculated plant materials at 14 and 21 dpi to confirm the expression of EB4. Sequence analysis was done using BioEdit Sequence Alignment Editor Software (version 6.0.5) (http://www.mbio.ncsu.edu/BioEdit/bioedit.html). The chimeric CP carried the complete EB4 sequence within its genome at 14 dpi ( Figure 6 ). However, the EB4 sequence appeared to be truncated at 21 dpi with an upstream portion of the EB4 sequence was not present. Only part of the downstream sequence of EB4 was detected together with the 3 untranslated region of the CGMMV. EB4 was totally undetectable at 30 dpi. Interestingly, the 5 end of introduced read-through sequence (position 490 to 501) was retained in the genome. The introduced HindIII recognition site at EB4 downstream from position 639 to 644 was found to have been deleted after 30 dpi. Even though only a single band was visible on RT-PCR screening, it was suspected that there could be products with similar sizes which could not be separated in normal agarose gel electrophoresis, therefore, the RT-PCR products from the chimeric CGMMV RNA at 30 dpi were cloned into pGEM-T Easy vector and subjected to sequence analyses. The results (Figure 7 ) confirmed the absence of the EB4 sequence, except for clone pR P3U4 (from nucleotide position of 702 to 746), where only part of the downstream sequence of the EB4 was detected together with the 3 untranslated region of the CGMMV. The 5 end of introduced read-through sequence of AAA-TAG-CAA-TTA (position 594 to 605) was retained within the genome for all sequenced clones (Figure 7) . The introduced HindIII site from position 742 to 747 was deleted for clones pR P3U1, pR P3U3, and pR P3U11 after 30 dpi. The chimeric CGMMV sequence analyses showed one common similarity (Figures 6 and 7) , that is, that part of the read-through amber stops codon sequence, and additional nucleotides "CC-AAA-TAG" were retained for all sequenced clones. This suggests that deletion had occurred within the host plant. The sequence analysis also showed that the EB4 was not fully expressed in the putative chimeric CGMMV. : Sequence alignments of cloned RT-PCR-amplified products from putative chimeric CGMMV RNA at 30 dpi. Plasmid clones sent for sequence analysis were pR P3U1, pR P3U3, pR P3U4, and pR P3U11. The sequence alignment shows that EB4 sequence was not within the insert of plasmids, except for a partial sequence for pR P3U4 clone. Only part of the downstream sequence of EB4 was detected together with 3 untranslated region of CGMMV for pR P3U4 clone, the rest were without EB4 sequence. Introduced read-through sequences and extra codons are underlined. Complete EB4 sequence is aligned accordingly with the other sequences. The positive results from the initial RT-PCR screening of transcript-inoculated plants at 14 dpi ( Figure 4 ) suggest the presence of EB4. The EB4 was, however, not detectable at 30 dpi, and only present in some (< 50% tested) plants at 21 dpi. These findings strongly suggest that deletion had occurred within the host plants after 14 dpi. The sequence analyses in this section are summarized in Table 3 . Plant virus vectors-based expression systems have been widely studied for their development as antigen presentation systems as well as for the production of pharmaceutically important peptides. The CGMMV has been previously shown to be suitable for expression of foreign peptide [7] . In this study, CGMMV vector was used to express a 45 amino acid EB4 gene. The integration of the EB4 gene into the end of CGMMV coat protein gene was done via a leaky UAG read-through sequence. Transcripts generated from chimeric clones of pRT and pRG carrying CAA-UAG-CAA-UUA read-through codon sequences were not infectious. This is possibly caused by the failure of self assembly [16] , and thus none of the inoculated plants was systemically infected. The assembly of CGMMV into virus particle has been shown to be essential for the viral movement through phloem [17] , hence another chimeric clone pCG+FSRTRE was constructed carrying read-through sequence (AAA-UAG-CAA-UUA) from the wild-type CGMMV genome. The clone containing this read-through signal was infectious and produced chimeric CGMMV (Table 1 and Figures 4 and 5) . It is, thus, postulated that there is a read-through signal preference between different species of tobamoviruses, in this case between TMV and CGMMV causing possibly viral particle assembly failure. It has also been shown that KGMMV, the tobamovirus, which has the closest genome similarity to CGMMV [13] also utilizes the same CGMMV read-through signal reaffirming the differences between the tobamoviruses. Additionally, unlike other plant virus vectors [15, 18] , this study reaffirms that with CGMMV pI deviation did not appear to be a factor affecting infectivity [7] . The chimeric (carrying the EB4) and putative wild-type CGMMV were shown to coexist in the virus population of the infected plants ( Figure 5 ). Previous reports show that the efficiencies of the leaky UAG codon varied from 0.5% to 5% so that the ratio of modified to unmodified CP would be between 1:200 and 1:20 [12, 19] . However, in this study, relatively high levels of chimeric coat protein was observed ( Figure 5 ) giving a ratio of modified to unmodified CP of approximately 1:1. It has been suggested that muskmelon host plant could be producing higher levels of translation nonsense suppressor tRNA making the application of the translation read-through signal favorable in this host [7] . Due to their relatively higher rate of mutation during replication, RNA viruses are evolving rapidly and this is the basis of their ubiquity and adaptability [20, 21] . In this study, it is shown that the EB4 gene sequence carried by the chimeric CGMMV was systemically removed during the infection process. The order of the removal of the transgene was speculated to be the 5 to 3 direction ( Figures 6 and 7) . This is further supported by the detection of two additional nucleotides together with the read-through sequence "CC-AAA-TAG" downstream from the CGMMV CP ORF. This report shows the temporal in-host truncation of the transgene from a chimeric virus in a natural host. Recent report has shown truncation occurring in transgenic plants expressing the same or similar transgenes as the chimeric virus [22] suggesting targeting by a resistance mechanism or competition with the parental virus as the mechanism involved. The exact mechanism of truncation of the transgene in our study is less clear as a previous study using the same vector and host with a different transgene did not exhibit the same instability [7] . The larger size of the EB4 peptide in comparison to the Hepatitis B epitope, however, suggests that the truncation mechanism or transgene recognition by the virus was size dependent. In summary, we have shown that CGMMV has a readthrough codon preference and that the read-through codon for TMV was shown to be not efficient, as the chimeric CGMMV transcripts utilizing this signal were not infectious. The reported limitation of low-modified coat protein yield of this type of read-through transient expression system appears to have been overcome as relatively equal yield of chimeric and wild-type CGMMV coat protein were produced. This report also provides a rational harvesting timeline for the chimeric virus making this system exploitable for implementation in a plantation scale in the future. It can be suggested that once host plants are infected with the chimeric virus carrying the inserted foreign peptides, the optimum harvesting time would be at around 14 dpi or not more than 20 dpi in order to obtain maximum yield of the full-length transgene. Growth of the infected plants for longer periods to obtain higher yields of the chimeric virus may induce unwanted transgene deletions. This and other factors described earlier should be relevant information for the further development of CGMMV or other plant viruses as vectors for medically important peptides such as for dengue (this study) and Hepatitis B [7] viral antigens.

Enhanced Hygiene Measures and Norovirus Transmission during an Outbreak

Control of norovirus outbreaks relies on enhanced hygiene measures, such as handwashing, surface cleaning, using disposable paper towels, and using separate toilets for sick and well persons. However, little is known about their effectiveness in limiting further spread of norovirus infections. We analyzed norovirus outbreaks in 7 camps at an international scouting jamboree in the Netherlands during 2004. Implementation of hygiene measures coincided with an 84.8% (95% predictive interval 81.2%–86.6%) reduction in reproduction number. This reduction was unexpectedly large but still below the reduction needed to contain a norovirus outbreak. Even more stringent control measures are required to break the chain of transmission of norovirus.

The estimated maximum likelihood estimates are α = 3.35 and β = 1.09, resulting in a peak generation time of 2.6, and a mean generation time of 3.6 days. Other positively skewed unimodal distributions such as the Weibull distributions did not produce a significantly better fit. As the generation time distribution might also be a realization of a mixture of several components, we fitted the data with a mixture of 2 or 3 gamma distributed components. This did not give a significantly better fit than a 1-component model (Technical Appendix 1 Table) . Let t = (t 1 ,…,t n ) be the vector of observed times of symptom onset of observed cases {1, …, n}. We assume that the elements of t are ordered such that t i ≤t j for all i <j. For subsets {i k , … i k+j } ⊆ {1,…,n} with all permutations of observations within this subset are equivalent. We chose 1 possible ordering arbitrarily. We now define a transmission matrix V = (v i,j ), whose elements represent the probability that the person with time of symptom onset t i was infected by the person with time of symptom onset t j , thus . Assuming that every case i > 2 was infected by another case in the set of observed cases, we get: for all i > 2. For i =1, the index case, we assume that: Furthermore, we assume that v i,j = 0 for all j < i. This assumption means that the ordering of times of infection is equivalent to the ordering of observed times of symptom onset, and more specifically, that persons cannot have infected themselves and cannot have infected persons with earlier time of symptom onset than their own. The matrix V is a lower triangular matrix and therefore does not contain cycles. To translate the transmission matrix V to reproduction number estimates, any The expected number of secondary cases produced by case j in these possible outbreaks based on transmission matrix V is: To translate this to an estimate of R for each day in the outbreak t; the mean R j of all cases with the same date of symptom onset is calculated, for all dates with observations: where m represents the label of the first case with symptom onset on day t, and q the total number of cases on day t. The likelihood that an observed time interval t i -t j represents a transmission event is determined as a product of the probability that i was infected by j and the probability that the time interval of symptom onset is t i -t j . That is, The likelihood of any case-patient j transmitting infection to case-patient i, becomes: Combining for all observed cases, the likelihood of a transmission matrix V becomes: for a given value of θ, and omitting the index case (i = 1) from the multiplication. Given the parameters for the generation time distribution θ, and all dates of symptom onset t, the parameters v i,j can be estimated. To estimate v i,j , the above likelihood function was evaluated in an adaptive rejection algorithm (Metropolis Hastings sampler) obtaining sets of V matrices with relative frequencies proportional to their likelihood (2) (3) (4) . To be reasonably certain of convergence and sufficient mixing, we have run 4 independent chains of 40,000 iterations and 3 independent chains with additional information about population structure and pathogen genotype and compared resulting estimates of reproduction numbers. Adding additional information is possible by setting implausible transmission probabilities in the transmission matrix V to 0. This may be considered a very strong prior assumption, but we have seen (Figure 4 in main text) that the resulting reproduction numbers are not strongly influenced by this radical assumption. In a true Bayesian approach, we might have applied different weights to pairs of cases within and between camps by multiplying a matrix containing these weights with the transmission matrix V. As described above, case-patients with a date of symptom onset on the same day are given an arbitrary order of infection within that day. Sampled transmission matrices represent all possible (noncyclic) patterns among cases, given the arbitrary order. Now any other possible pattern can be found by permutation of indexes among cases with the same date of symptom onset. Because these all have the same contribution to the likelihood such permutations do not change the likelihood: all permutations are equally likely. Such permutations also have the same reproduction numbers, only for different cases (indices). If we average over all such permutations with identical contributions, the resulting reproduction numbers do not change. The expected time course of the reproduction number R(t) is given by the following equation: Here, is the day of implementation of enhanced hygiene measures, G is the cumulative probability function of the generation time distribution, h t ρ is the relative reduction of the reproduction number due to implementation of hygiene measures and is the effective reproduction number without enhanced hygiene measures. u R

Polyomaviruses KI and WU in Immunocompromised Patients with Respiratory Disease

Polyomaviruses KI (KIPyV) and WU (WUPyV) were recently identified, mainly in respiratory specimens from children. Among 200 patients with respiratory disorders admitted to Saint Louis Hospital, Paris, France, KIPyV was detected in 8% and WUPyV in 1%. KIPyV was significantly more frequent among human stem cell transplant patients (17.8% vs. 5.1%; p = 0.01).

R ecently, 2 new, distinct polyomaviruses (PyVs), KI (KIPyV) and WU (WUPyV), were identifi ed in respiratory specimens, mainly from children <5 years of age with respiratory tract infections. The fi rst retrospective studies of respiratory specimens in Sweden and Australia showed a KIPyV prevalence of 1% and 2.5%, respectively (1, 2) . Studies conducted in Australia and the United States showed a WUPyV prevalence in respiratory specimens of 3% and 0.7%, respectively (3) . Further studies conducted in Canada and South Korea have shown similar frequencies (4, 5) . In this study, we examined the prevalence of KIPyV and WUPyV in immunocompromised patients with suspected respiratory tract infections. From January through June 2007, 265 respiratory samples were received in the laboratory of Saint Louis Hospital, Paris: 154 nasal aspirates (NA) and 111 bronchoalveolar lavage (BAL) specimens collected from 200 patients with suspected upper or lower respiratory tract infections. This hospital specializes in the management of immunocompromised patients. Respiratory samples were collected for the diagnosis of acute respiratory illness; 89% of samples were from immunocompromised patients. Their median age was 46 years (range 3.6-85.3 years). Given the observational nature of the study, French law did not require ethical approval or informed consent. The specimens were routinely tested for infl uenza A and B viruses, respiratory syncytial virus, and parainfl uenza viruses 1, 2, and 3 by immunofl uorescence assay (Imagen; DakoCytomation, Trappes, France). Specimens positive for KIPyV or WUPyV were tested for adenoviruses; human bocavirus; human rhinoviruses; human metapneumovirus; human coronaviruses OC43, 229E, NL63, HKU1; and human PyVs BK and JC by using PCR methods (6) (7) (8) (9) (10) (11) . Total nucleic acid was extracted from 200 μL of NA, BAL, or stool specimens by using the EasyMag System (bio-Mérieux, Marcy l'Etoile, France). KIPyV was detected with an in-house real-time PCR assay targeting the VP1 gene. The primers and hydrolysis probe were designed by using Primer Express 3.0 software (Applied Biosystems, Foster City, CA, USA). The fi nal reaction volume was 25 μL and contained 12.5 pmol of SLKI-VP1s (5′-GGAAATACAGCTGCTCAGGAT-3′) and SLKI-VP1as (5′-CTTTGATACTTGAACCGCTTTCCTT-3′), 6.25 pmol of corresponding probe SLKI-VP1PR (5′-6FAM-C GTGACCCCACCCCTCATTACTGGTC-TAMRA-3′), 12.5 μL of TaqMan Universal Master Mix (Applied Biosystems), and 5 μL of DNA extract. The reaction was run on a 7500 Real-Time PCR System (Applied Biosystems). The specifi city of positive specimens was confi rmed by using PCR and nested PCR with primers POLVP1-39F/POLVP1-363R and POLVP1-118F/POLVP1-324R, as described (1) . The PCR products were then sequenced and compared with the previously described sequences from Sweden and Australia (GenBank accession nos. EF127906, EF127907, EF127908, EF520287, EF520288, and EF520289). WUPyV was detected by PCR as described (3) . PCR products with the expected molecular weights were sequenced by using primers AG0044 and AG0045 and compared to published sequences (GenBank accession nos. EF444550, EF444551, EF444552, EF444553, and EF444554) (3). KIPyV was detected in 17 (6.5%) of the 265 respiratory samples and in 16 (8.0%) of the 200 patients. All cases were confi rmed by a nested PCR targeting another region of the VP1 gene. Twelve of the 17 PCR products were successfully sequenced, and all shared 100% homology with published sequences. WUPyV was detected in only 2 patients (1.0%). Genome sequencing showed 98% homology with reported WUPyV sequences. Six KIPyV-positive patients (37.5%) had co-infections with other respiratory viruses, and 2 of them (12.5%) had a pulmonary bacterial infection (online Appendix Table, available from www.cdc.gov/EID/content/15/1/107-appT. htm). One WUPyV-infected patient who exhibited acute respiratory failure had concomitant pneumonia caused by Pseudomonas aeruginosa infection. None of the 15 patients who were positive for KIPyV or WUPyV and tested for fungi had respiratory or blood samples positive for Aspergillus spp. (Table) . Lung or sinus imaging was assessed by computed tomography scan for 12 KIPyV-positive patients. Lung parenchyma abnormalities were noted in 9 patients, and sinusitis was diagnosed for 2 patients. Taking into consideration both the frequency of digestive symptoms in our patients and the former published detection of KIPyV in a stool sample, we looked for KIPyV infection in the available stool samples while the respiratory samples were being assessed for KIPyV (1). Strikingly This study shows the prevalence of KIPyV and WUPyV among immunocompromised patients with respiratory disorders. Previously, these 2 viruses had been observed mainly in young children (1) (2) (3) . Of the few adult patients with KIPyV or WUPyV infection mentioned in these studies, most were immunocompromised (3, 12) . Considering the seemingly higher prevalence of KIP-yV in our population (8%), immunocompromised patients may be more susceptible to this PyV, as they are to JC and BK viruses (13) (14) (15) . Results from previous reports suggest a similar frequency of both KIPyV and WUPyV infections being found in respiratory specimens, ranging from 1% to 3%. In contrast, in our series, we found a likely difference between the prevalence of KIPyV (8%) and WUPyV (1%), which suggests that the replication or reactivation of the 2 viruses in the respiratory tract may differ between immunocompromised and immunocompetent patients. However, this difference requires further investigation, in particular, by using similar real-time PCR assays. Notably, a significantly higher prevalence of KIPyV infection was found among HSCT patients, which suggests that a profound T-cell defi ciency may be a factor in facilitating KIPyV replication. As reported in other populations, our patients who yielded positive specimens for KIPyV or WUPyV had conditions ranging from a common cold to acute respiratory distress that required invasive ventilation. Respiratory coinfections, observed in other studies, had likely accounted for at least some clinical features. In the 7 of our patients in whom KIPyV was the sole pathogen detected in the respiratory tract, despite comprehensive screening for viruses, bacteria, parasites, and fungi, clinical and radiographic patterns were varied. Some of the patients had only upper respiratory tract infections, notably sinusitis, whereas others had lung parenchyma abnormalities as defi ned by computed tomographic scan imaging. However, due to the retrospective nature of the study, and therefore the lack of a control group of immunocompromised patients without respiratory symptoms, the association of KIPyV infection with the occurrence of respiratory disease cannot be stated defi nitively. In conclusion, the seemingly higher frequency of KIP-yV shedding in immunocompromised patients (as observed with other PyVs) and the detection of KIPyV as a single pathogen in respiratory disease (e.g., as cytomegalovirus recurrence can lead to pneumonia in immunocompromised patients) together support a reactivation hypothesis. Nevertheless, a reinfection hypothesis cannot be excluded due to immunocompromised patients' increased risk of acquiring viral infection from exogenous sources Controlled prospective studies of KIPyV shedding before and during immunosuppression will help determine the pathogenic role of this virus. The clinical implication of KIPyV detection in stools and the mechanisms underlying the concomitant presence in gastrointestinal and respiratory tracts also deserve further analysis. We thank F. Freymuth, P. Lebon, and A. Vabret for advice and technical assistance and for providing positive virus controls. Dr Mourez is a virologist in the Laboratory of Microbiology, Saint Louis University Hospital, Paris. His research interests include the development of tests for the diagnosis of emerging respiratory viruses and the study of the circulation and molecular analysis of human respiratory viruses in pediatric and immunocompromised patients.

On the significance of Surfactant Protein-A within the human lungs

Surfactant Protein-A (SP-A) is the most prominent among four proteins in the pulmonary surfactant-system. SP-A is expressed by alveolar epithelial cells type II as well as by a portion of non small cell lung carcinomas (NSCLC). The expression of SP-A is complexly regulated on the transcriptional and the chromosomal level. SP-A is a major player in the pulmonary cytokine-network and moreover has been described to act in the pulmonary host defense. By the use of cell culture or animal models the functional properties have been repeatedly shown in many aspects, often bearing surprising properties which strongly indicate the physiological importance of SP-A. To date SP-A is recognized as a molecule essential for pulmonary development, structure and function. An upcoming number of reports deals with the role of SP-A for pulmonary pathology. This article gives an overview about the state of knowledge on SP-A focused in applications for human pulmonary disorders and points out the importance for pathology-orientated research approaches using immunohistochemistry or in situ hybridization as promising methods to further elucidate the role of this molecule in adult lung diseases.

The role of the surfactant system for the development of the human lung is known to be essential. Since it is synthesized by humans starting in the 28 th week of pregnancy and reaching functional levels in the 34 th week, surfactantsubstitution-therapy is a fundamental part of the treatment of premature babies suffering from Infant Respiratory Distress Syndrome (IRDS) [1] . Pulmonary surfactant regulates dynamically the alveolar surface tension. The central role of the surfactant system for maintaining pulmonary function has been repeatedly shown by the use of cell culture or animal models [2] . Surfactant is a complex mixture of lipids, carbohydrates and four proteins (SP-A, SP-B, SP-C, SP-D). The initial descriptions of surfactant lead back to the 1950s, but little attention was given to the surfactant proteins until the 1980s [3] . The genes coding for these proteins are located on different chromosomes. SP-B and SP-C are similarly structured hydrophobic proteins participating in the adsorption of phospholipids at the alveolar border, which leads to rapid reduction of the surface tension. The hydrophilic proteins SP-A and SP-D are members of the collectins with C-type lectin domains. SP-D together with SP-A play a role in the pulmonary defense against Gramnegative bacteria [4] . SP-A is the major surfactant apoprotein exhibiting complex interactions and participation in processes fundamental for pulmonary structure and function with its expression restricted to alveolar epithelial cells type II. Moreover expression of SP-A was also described for a portion of NSCLC facilitating a diagnostic marker for these carcinomas [5, 6] . After characterization of the biochemical properties, a complex chromosomal organization of the genes coding for SP-A has been demonstrated [3] . The locus of the SP-A on the one hand consists of two functionally active genes and a pseudogene [7, 8] . The two active genes SP-A1 and SP-A2 on the other hand display several different alleles and splicing variants, moreover different oligomeric states have been described [3, 9] . During the development of the lung these two genes are regulated differentially, a process triggered by cAMP and glucocorticoids [10] . The SP-A1 and SP-A2 genes display a homology of 94% in their nucleotide sequences and even 96% homology in the amino acid sequences [11] . Fig. 1 , as one example, shows the transcriptional activity of the SP-A1 and SP-A2 genes determined by RT-PCR in homogenates of biopsies from NSCLC and corresponding tumor-free samples. The importance of the differential transcription of SP-A1 and SP-A2 for maintaining pulmonary function has repeatedly been demonstrated [12] . Phylogenetic analyses revealed that an ancestor proto-SP-A gene was diverged into SP-A1 and a second gene which subsequently emerged to SP-A2 and the SP-A pseudogene [3] . The high level of homology between SP-A1 and SP-A2 up to date prevents a differential analysis of the two gene products in situ. The expression of SP-A is also complexly regulated on the transcriptional level. Moreover the protein-turnover and the release of SP-A into the serum represents a further point of regulation [12] . This sophisticated regulation of the genetic activity is recognized as a further hint for the functional importance of SP-A. In recent years the role of defects in the expression of SP-A in context with different pulmonary diseases has become an issue of scientific investigations. Initially numerous studies have been performed to elucidate the role of surfactant substitution in pediatrics [2] . As one major function SP-A displays a protective role of the molecule in pulmonary host defense by interacting with various infectious agents such as bacteria, fungi and viruses. SP-A deficient knock-out mice -compared to wild type animals -are susceptible to infections with Pseudomonas aeruginosa [13] and the clearance of group B streptococcus is slower [14] . In accordance the defense of SP-A deficient mice against Respiratory Syncytial Virus (RSV) has been shown to be reduced and may be restored by exogenous SP-A administration [15] [16] [17] . By mediating the attachment of Mycobacterium tuberculosis to alveolar macrophages and promoting the phagocytosis of these bacteria, SP-A is important in the pathogenesis of tuberculosis [18] [19] [20] [21] . SP-A also functionally interacts with staphylococci [22, 23] Moreover, SP-A is involved in the complex pulmonary network of cytokines as a central player, for example interacting with TNF-alpha and several interleukins [31, 12, 14] . Therefore it is likely that defects in the expression of SP-A may be important in the course of non infectious pulmo-RT-PCR: Transcription of SP-A1 and SP-A2 in NSCLC tumors (T) and corresponding tumor-free tissues (TF) from the same cases in comparison to GAPDH nary diseases of adult patients. In the case of idiopathic pulmonary fibrosis, for example, low levels of SP-A (measured by ELISA) have been reported in broncho alveolar lavages (BAL), but elevated levels were found in the sera [32] [33] [34] . Immunohistochemical examinations of the expression of SP-A in pulmonary fibrosis demonstrated evident defects by using specimens from different diseases displaying fibrotic changes in the lungs. In good agreement with the results in BAL reduced levels of SP-A have been observed in fibrotic lungs. This reduced SP-A-expression in fibrotic lungs may be caused by two reasons: a limited number of the SP-A producing type II pneumocytes and by a clearly reduced SP-A expression of the remaining cells [35] . Keeping in mind that surfactant substitutes are widely available due to their application in pediatrics, a growing number of therapeutic possibilities may result from these findings. In sarcoidosis elevated levels of SP-A have been described [37] using BAL; the same applies for the sera from patients with alveolar proteinosis [38]. Since SP-A represents a central molecule in pulmonary immunoregulation as well as in host-defense it is obvious that defects in the surfactant system may have functional influence in the course of these pulmonary disorders. Another point of research concerning SP-A is the diagnostic value of this molecule, the expression of which is restricted to the lungs. It has been reported that SP-A levels in BAL or serum from patients suffering from idiopathic pulmonary fibrosis correlate with the progression of the disease and can be used to predict survival [34, 38] . In samples from airway secretions SP-A measurements are described to be useful also for the diagnostics of pulmonary edema where elevated levels have been found compared to healthy volunteers and ARDS patients [39]. By utilizing highly sensitive RT-PCR techniques the amplification of SP-A transcripts can be used for the detection of occult metastases in non small cell lung cancer patients [5, 40] . Comparative studies of different malignomas with pulmonary localization have shown the diagnostic properties of immunohistochemically determined SP-A [6, 41, 42] . In carcinomas of occult origin localized in the lungs the diagnosis has a crucial influence on the therapy. A positive detection of SP-A represents a clear hint for a primary location in the lung [43]. Fig. 2 as an example shows the immunohistochemical detection of SP-A in a moderately differentiated adenocarcinoma of the lung using the primary monoclonal antibody PE-10 and LSAB (AEC-substrate, × 100). The positive staining in the tumor cells (reddish) in this certain case helped to manifest the diagnosis as a primary carcinoma of the lung. However, the choice of a suitable SP-A antibody is highly important since approaches using polyclonals display cross reactions with other tumors [44] . This procedure has already become a part of pathological routine diagnosis, and -along with other markers such as the Thyroid-transcription-factor-1 (TTF-1) -the detection of SP-A (by PE-10) is a useful part of the immunohistochemical panel in pulmonary pathology. Immunohistochemical detection of SP-A even might be utilized for forensic purposes helping to distinguish between fatal drowning and postmortem immersion [45] . It is evident that SP-A is a molecule which already proves to be an interesting subject for medical research. However, the studies concerning the possible role of surfactantdefects in pulmonary diseases of adults have been performed mainly in different cell culture or animal models hardly analyzing adult human lung tissue. For these reasons SP-A is a promising target for histomorphological approaches using pathological specimens which exactly represent the scenarios of various diseases with all the different cell types involved which are difficult to simulate in models. With the modern tools of molecular pathology, the genetic activities of genes can be analyzed in situ, which provides evidence of the cellular activities in the context of a human native tissue. One example is shown Immunohistochemical detection of SP-A using the mono-clonal antibody PE-10 (LSAB, amonoethylcarbazole, 400×) Figure 2 Immunohistochemical detection of SP-A using the monoclonal antibody PE-10 (LSAB, amonoethylcarbazole, 400×). in Fig. 3 : a lung section hybridized with a digoxigeninlabeled SP-A probe to analyze the transcriptional activity in situ; the reddish signals of the transcripts are visible in the cytoplasm of type II pneumocytes. When analyzing the expression of SP-A in histological sections in context with other molecules of the pulmonary cytokine network one can expect further clues for the scenarios taking place in the course of interstitial lung diseases. Taken together, SP-A is a complexly regulated molecule with surprising properties and essential importance for pulmonary development, structure and function which is getting more and more into focus concerning various diseases of the adult lung. Thus, as an outlook, it will become an issue of pulmonary pathology which might provide promising perspectives for applications in research, diagnosis and therapy. In situ hybridization targeting using a 663 bp digoxigenated DNA-probe complementary to SP-A mRNA Figure 3 In situ hybridization targeting using a 663 bp digoxigenated DNA-probe complementary to SP-A mRNA. Detection was achieved by Anti-digoxigenin antibody conjugated to alkaline phosphatase with NBT/BCIP as a chromogen (400×). Publish with Bio Med Central and every scientist can read your work free of charge

Proteolytic processing of a precursor protein for a growth-promoting peptide by a subtilisin serine protease in Arabidopsis

Phytosulfokines (PSKs) are secreted, sulfated peptide hormones derived from larger prepropeptide precursors. Proteolytic processing of one of the precursors, AtPSK4, was demonstrated by cleavage of a preproAtPSK4–myc transgene product to AtPSK4–myc. Cleavage of proAtPSK4 was induced by placing root explants in tissue culture. The processing of proAtPSK4 was dependent on AtSBT1.1, a subtilisin-like serine protease, encoded by one of 56 subtilase genes in Arabidopsis. The gene encoding AtSBT1.1 was up-regulated following the transfer of root explants to tissue culture, suggesting that activation of the proteolytic machinery that cleaves proAtPSK4 is dependent on AtSBT1.1 expression. We also demonstrated that a fluorogenic peptide representing the putative subtilase recognition site in proAtPSK4 is cleaved in vitro by affinity-purified AtSBT1.1. An alanine scan through the recognition site peptide indicated that AtSBT1.1 is fairly specific for the AtPSK4 precursor. Thus, this peptide growth factor, which promotes callus formation in culture, is proteolytically cleaved from its precursor by a specific plant subtilase encoded by a gene that is up-regulated during the process of transfering root explants to tissue culture.

Phytosulfokines (PSKs) are a class of plant peptides discovered through the study of growth factors that mediate density-dependent growth in cell culture . isolated and identified growth factors from conditioned medium that promoted the growth at low density of asparagus mesophyll cells in tissue culture. They identified a sulfated pentapeptide [H-Tyr(SO 3 H)-Ile-Tyr(SO 3 H)-Thr-Gln-OH, abbreviated sYIsYTQ], named PSK-a, and a sulfated tetrapeptide [H-Tyr(SO 3 H)-Ile-Tyr(SO 3 H)-Thr-OH], named PSK-b, that were active in the asparagus cell system. Six genes encoding PSKs (AtPSK1-6) have been identified in Arabidopsis. Each encodes a preproprotein precursor of approximately 80 residues, with the YIYTQ peptide near their C-termini . generated transgenic Arabidopsis plants (AtPSK4ox) over-expressing one of the AtPSKs, and found that root growth and callus formation were slightly enhanced in over-expression lines, but otherwise the seedlings were phenotypically indistinguishable from wild-type. A PSK receptor was first identified in carrot (Daucus carota) as a leucine-rich repeat receptor kinase (LRR-RK) (Matsubayashi et al., 2002) . Sequence information from the carrot protein (DcPSKR1) was used to identify an ortholog in Arabidopsis, AtPSKR1 (At2g02220). described a mutant with a Ds insertion in AtPSKR1 (pskr1-1), and found that callus derived from the mutant was less sensitive to the growth-promoting effects of PSK in culture. However, they observed little difference in overall plant growth between wild-type, pskr1-1 and AtPSKR1ox, an over-expression line. The most prominent characteristic of pskr1-1 was that vegetative tissues in mature plants lost their ability to form callus. Unlike wild-type plants, leaf discs from the fully expanded leaves of pskr1-1 plants were less capable of producing callus, while unexpanded leaves retained callus-forming capacity. AtPSKR1 over-expressing plants (AtPSKR1ox) showed delayed senescence, and, as a result, leaves continued to expand, resulting in larger leaves than the wild-type. Little is known about the proteolytic processing of the PSK propeptide precursors, other than the fact that the precursors have conserved di-basic residues 8-10 amino acids upstream from the mature peptide sequence . Di-basic residues are characteristic of substrate sites for subtilases, subtilisin-like serine proteases (Barr, 1991) ; therefore, we wished to determine whether any of the proteases encoded by the 56 subtilase genes in the Arabidopsis genome (Rautengarten et al., 2005) are responsible for cleavage of the AtPSK4 precursor. In doing so, we identified a subtilase, AtSBT1.1, that is required for cleavage of the fusion protein AtPSK4-myc. Cleavage of AtPSK4-myc is induced in root explants, suggesting that release of the peptide hormone from the precursor protein is controlled, in part, by the proteolytic processing machinery. In studies of shoot regeneration in Arabidopsis tissue culture, we found that expression of a subtilase gene AtSBT1.1 (At1g01900) correlated with conditions for efficient shoot regeneration (Lall et al., 2004) . It was not clear what the causal connection might be between efficient shoot regeneration and expression of a gene encoding a serine protease. Most subtilases are predicted to be secreted proteins (Rautengarten et al., 2005) , so we reasoned that AtSBT1.1 could be involved in processing of an extracellular growth factor or receptor related to shoot regeneration. Growth factors that might require the action of AtSBT1.1 include peptide hormones, such as AtPSKs. AtPSKs promote callus formation in tissue culture , and are synthesized as preproproteins with signal peptides that target them to the secretory pathway and with prosequences that are processed during maturation of the peptide hormone. To determine whether AtSBT1.1 is involved in the proteolytic processing of AtPSKs, we developed a constitutively expressed, C-terminal 4 x myc-tagged construct of the AtPSK4 precursor, 35S:ppAtPSK4-myc, and studied its processing in vivo. PreproAtPSK4 will be referred to as ppAtPSK4 and proAtPSK4 as pAtPSK4. AtPSK4 was chosen for study because it is the most abundantly expressed PSK precursor . The predicted size of pAtPSK4 with the myc tag is 12.8 kDa, but we observed a band at approximately 19 kDa on Western blots, larger than the predicted size ( Figure 1a , wt, 0 time). To demonstrate, nonetheless, that this band is pAtPSK4myc, we analyzed the partially purified myc-tagged protein and identified three peptides derived from the fusion protein by MS/MS analysis ( Figure S1 ). Thus, the larger apparent size of pAtPSK4-myc may be due to anomalous gel migration behavior or post-translational modification of part of the precursor protein. (a) Root segments from wild-type transgenic seedlings expressing the construct 35S:ppAtPSK4-myc were explanted and incubated on callus induction medium (CIM) for 1-4 days and then transferred onto shoot induction medium (SIM). Root segments from sbt1.1-1 and sbt1.1-2 mutants expressing 35S:ppAtPSK4-myc were also explanted and similarly incubated. Arrows indicate the predicted migration position for processed AtPSK4-myc. The lane marked NT is an extract from roots of non-transformed seedlings. (b) Root segments from seedlings bearing 35S:ppAtPSK4-myc in a wild-type background were explanted and incubated on normal CIM or on B5 basal medium, without cytokinin or auxin hormones. (c) Time course following explantation for acquisition of capacity to process pAtPSK4-myc on CIM medium. We were surprised to find that pAtPSK4-myc was not cleaved in roots of intact transgenic seedlings (Figure 1a , wt, 0 time). We suspected that failure to detect processing may have been due to seedling growth or culture conditions. We had chosen to study AtSBT1.1 in the first place due to observations that we had made about shoot regeneration from root explants (Lall et al., 2004) . Therefore, we subjected root explants from 35S:ppAtPSK4-myc seedlings to tissue culture conditions for regenerating shoots. This involves pre-incubating root segments on an auxin-rich callus induction medium (CIM), and then transferring them after 4 days to a cytokinin-rich shoot induction medium (SIM). Under these conditions, a band appeared at approximately 7 kDa in Western blots (Figure 1a , wt, lanes 1d and 4d CIM or SIM), representing the myc-tagged cleaved peptide, AtPSK4-myc. The size of the processed protein was consistent with a cut at or near the cleavage site as determined by the in vitro experiments described below. Thus, cleavage of pAtPSK4myc appears to be induced by some aspect of the culturing process. Having identified conditions for pAtPSK4-myc cleavage, we wished to determine whether AtSBT1.1 was responsible for the proteolysis. To do so, we examined cleavage of pAtPSK4-myc in T-DNA mutants with insertions in AtSBT1.1. The gene encoding AtSBT1.1 (At1g01900) has a single intron, and sbt1.1-1 (SALK_111561) and sbt1.1-2 (SALK_108704) have T-DNA insertions in the first exon. The T-DNA lines were judged to be null mutants because AtSBT1.1 transcripts were not found in seedlings from either line ( Figure S2 ). The 35S:ppAtPSK4-myc construct was introduced into the two T-DNA mutant lines, and root explants were subjected to regeneration conditions (4 days incubation on CIM). The pAtPSK4-myc precursor was produced in these lines, but the cleavage product AtPSK4-myc peptide was not detected (Figure 1a , lanes sbt1.1-1 and sbt1.1-2). To demonstrate that the transgenic product in these lines was still cleavable, we out-crossed sbt1.1-1 bearing the 35S:ppAtPSK4-myc construct to wildtype, and demonstrated that processing was restored in root explants of the F 1 seedlings ( Figure S3 ). We concluded from these results that AtSBT1.1 is required for pAtPSK4-myc cleavage under these conditions. This finding is quite significant given that there are 56 different subtilases encoded by the Arabidopsis genome (Rautengarten et al., 2005) . As shoot regeneration in culture is dependent on cytokinin and auxin hormones, we determined whether induction of pAtPSK4-myc cleavage required hormone treatment during CIM pre-incubation. To test this, root segments were explanted to hormone-free B5 medium as well as to CIM, and it was found that pAtPSK4-myc was cleaved on basal B5 medium at 1 and 4 days after explantation ( Figure 1b ). We concluded that the processing activity was probably induced by the wounding or handling of root tissue during explanting. We then determined how rapidly processing was induced. Cleavage of pAtPSK4-myc was first detected approximately 8 h after explanting ( Figure 1c ). Thus induction is fairly rapid, but not immediate as one might expect for a post-translational activation mechanism. Up-regulation of AtSBT1.1 gene expression As pAtPSK4 cleavage was induced approximately 8 h following the explanting of root segments, we wished to determine whether expression of the gene encoding AtSBT1.1 was similarly up-regulated. Real-time quantitative RT-PCR was performed on extracts from root segments at various times following explanting. It was found that AtSBT1.1 was up-regulated approximately 3.5-fold starting at approximately 8 h after explanting root segments ( Figure 2a) . Thus, the up-regulation of AtSBT1.1 expression was similar to the kinetics for the acquisition of cleavage activity. The endogenous gene encoding AtPSK4 was also up-regulated after explanting, but not as much as AtSBT1.1 (approximately twofold, Figure 2a ). We also developed promoter:GUS constructs for AtSBT1.1 and AtPSK4, and found that the gene was expressed at the cut ends of the root segments where callus formation first occurs during shoot regeneration (4 days CIM, Figure 2b ,d). Later, AtSBT1.1 and AtPSK4 continued to be expressed most intensely at sites of callus and regenerative tissue formation, both at the ends and at other wound sites along the length of the root segments (6 days SIM, Figure 2c ,e). If pAtPSK4 is indeed a substrate for AtSBT1.1, then these proteins should occupy the same subcellular compartment. Both AtSBT1.1 and ppAtPSK4 have signal peptides, and both are predicted to be secreted proteins. To determine whether that is so, we constructed C-terminal YFP fusions and determined their location in root cells from Arabidopsis. Both AtPSK4-YFP (Figure 3a-c) and AtSBT1.1-YFP (Figure 3d-f) are associated with the periphery of the cell, coinciding with propidium iodide staining. Following plasmolysis (Von Groll et al., 2002) , most fluorescence from both YFP fusions remained associated with the extracellular matrix rather than the plasma membrane ( Figure S4 ). To confirm that AtSBT1.1 is indeed involved in the processing of pAtPSK4-myc, we tested whether pAtPSK4 can be cleaved by AtSBT1.1 in vitro. To do so, we developed a pull-down assay using a C-terminal myc-tagged version of AtSBT1.1 (AtSBT1.1-myc) and a fluorogenic peptide substrate. We employed a similar assay in a previous study to test the activity of another Arabidopsis subtilase, AtS1P (Liu et al., 2007a) . In both cases, the tagged subtilases were synthesized in transgenic Arabidopsis because we were not able to produce active enzyme at high levels in heterologous systems. To test AtSBT1.1-myc for activity against pAtPSK4, we generated a fluorogenic peptide called fpAtPSK4 with the sequence Abz-RRSLVLHTDY(NO2)D-OH [where Abz is a 2-aminobenzoyl fluorescent group and Y(NO2)D-OH is a 3-nitrotyrosine quencher], representing a probable subtilase recognition site in pAtPSK4. The site was chosen because di-basic amino acid residues (in this case, RR) are characteristic signatures of subtilase recognition sites (Barr, 1991) . When bead-bound protein from transgenic plants expressing AtSBT1.1-myc was incubated with fluorescence-quenched fpAtPSK4, a time-dependent increase in fluorescence was observed (Figure 4a ). Because the enzyme was produced in a homologous system, we were mindful of potential contamination by endogenous Arabidopsis subtilases in our preparations. To preclude that possibility, we thoroughly washed the beads and routinely ran controls in which the same pull-down procedure was conducted with extracts from non-transgenic seedlings and from transgenic seedlings expressing a tagged subtilase (S552A) that was mutated at the active site ( Figure 4a ). The kinetics of the reaction with the fluorogenic peptide substrate followed Michaelis-Menten kinetics, and the K m for the reaction was determined to be approximately 18 lM (Figure 4b ). The K m value lies within the range of affinity constants for fluorogenic peptide substrates with comparable subtilases, such as mammalian prohormone convertases (Basak et al., 2004 (Basak et al., , 2007 . The activity was inhibited by PMSF (Figure 4a ), an inhibitor of serine proteases. The pH optimum for the reaction was in the acidic range, centered on pH 6 ( Figure 4c) . To determine the site at which the peptide was cleaved, the reaction products were analyzed by MALDI-TOF analysis. The first N-terminal cleavage product that appeared had a molecular mass of 862.25, indicating a preferred cutting site at Abz-RRSLVLflHTDY(NO2)D-OH (Figure 4d ). MS/MS analysis of the 862.25 peak showed a spectrum of ions consistent with Abz-RRSLVL as the initial cleavage product ( Figure S5 ). It should be noted that this cut is three amino acid residues upstream of the N-terminus of the mature peptide as has been described for PSKs in Asparagus officinalis . If the mature Arabidopsis peptide is similar, then further N-terminal proteolytic processing, as well as C-terminal processing, probably occurs to generate the active peptide. Substrate specificity of AtSBT1.1 To determine whether AtSBT1.1 has specificity for the AtPSK4 recognition site, we conducted an alanine scan through fpAtPSK4, substituting one residue at a time for alanine (Figure 5a ). The reactions were carried out in triplicate, and the reaction rates for each of the substitutions were compared to that for the wild-type substrate. The cleavage site sequence was very sensitive to alanine substitutions. Substituting the first arginine in the di-basic residues (P6 position) resulted in a reaction rate that was less than half that of wild-type. The most sensitive positions were P2-P4, which are modestly conserved among the AtPSK precursors. We compared the sequence of AtPSK4 to some of the most closely related AtPSKs and to PSK-related sequences in other plants. The five amino acids representing the presumed mature PSK peptide in the six Arabidopsis sequences are nearly identical (Table 1) other plant PSKs in the GenBank database, except for AtPSK6. The upstream di-basic amino acids and the leucine (P3 position) and histidine and aspartate (P1¢ and P3¢, respectively) are also conserved in Arabidopsis. To compare the activity of AtSBT1.1 with other AtPSK precursors, we developed fluorogenic peptides representing putative recognition sites for the family of AtPSKs (fpAtPSK1, 2, 3, 5 and 6, see bold residues in Table 1 ). The putative recognition sites for pAtPSK2 and 5 are the same and are represented in the fluorogenic peptides fpPSK2 and 5. The cleavage reaction was slower for fpAtPSK2 and 5 than for fpAtPSK4, and was barely detectable or not detectable at all with the other fpAtPSKs ( Figure 5b) . Thus, the activity of AtSBT1.1 is fairly specific, with most activity directed toward cleavage of pAtPSK4, followed by pAtPSK2 and 5. The proteolytic processing of PSK4 appears to be highly regulated in the tissue culture system -in part by up-regulation of AtSBT1.1 gene expression. Among the 56 subtilases encoded in the Arabidopsis genome, AtSBT1.1 appears to be solely responsible for the release of PSK4 from its precursor in roots, because pAtPSK4-myc is not cleaved in AtSBT1.1 knock-out lines. PSKs are growth factors that stimulate callus formation in culture Sakagami, 1996, 2006) . Thus, up-regulation of AtSBT1.1 and the release of AtPSK4 from its precursor protein may be key factors in promoting the growth of cells and callus formation in tissue culture. AtSBT1.1 functions similarly to prohormone convertases that release peptide hormones and neuropeptides from protein precursors in animal cells (Steiner, 1998) . Prohormone convertases are subtilases in the secretory pathway that cleave substrates with mono-or di-basic amino acids in the general recognition motif R/K-Xn-R/Kfl (Seidah, 1997) . In Arabidopsis, AtSBT1.1 cleaves substrates in a fairly site-specific manner, cleaving pAtPSK4 at RRSLVLflHTDY and showing preference for cleavage of proAtPSK4 over other PSK precursors. Cleavage by AtS-BT1.1 at the preferred cleavage site of the PSK4 precursor leaves three residues on the N-terminus of the five-residue peptide. This is of concern because demonstrated that even the presence of the tripeptide GGG severely reduces the activity of PSKs in an asparagus suspension cell system However, the three residues (HTD) remaining on the amino-end of AtSBT1.1processed AtPSK4 are highly conserved among the PSKs, and it is possible that the sequence is part of the active, mature peptide or serves as a recognition signal for further processing by enzymes such as tripeptyl peptidases (Book et al., 2005) . Complete proteolytic processing of PSK4 probably involves a number of other steps, including trimming of the peptide at its C-terminus . The preferred AtSBT1.1 cleavage site within the recognition motif for proAtPSK4 is somewhat unconventional, because the di-basic residues are usually in the P1 and P2 positions, immediately upstream (on the N-terminal side) of the cleavage site (Barr, 1991) , The only caveat we have about the site is that it was identified from the cleavage product of a short fluorogenic peptide representing the putative subtilase recognition site. Other structural features of the intact Table 1 . AtPSK precursor may be important in determining the site of cleavage, but are not found in the peptide substrates that we used to characterize the cleavage reaction. We have shown that YFP fusions of AtSBT1.1 and PSK4 accumulate in the extracellular matrix, and it is probable that cleavage occurs there simply because AtSBT1.1 has a slightly acid pH optimum and the apoplast is acidic (Bibikova et al., 1998) . PSK4 is tyrosine-sulfated , and the protein is likely to be sulfated in the trans-Golgi as are other sulfated proteins in animal cells (Baeuerle and Huttner, 1987) . Therefore, it seems reasonable that the precursor is sulfated before it is cleaved. However, the precursor does not have to be sulfated to be cleaved, because AtSBT1.1-myc can cleave the unsulfated peptide, fpAtPSK4. Sulfation, however, is important for the function of the peptide, because the sulfated peptide binds to the PSK receptor (Matsubayashi et al., 2002) . Although there are six genes encoding PSKs in Arabidopsis, the mature peptides (YIYTQ) encoded by each gene are identical (with the exception of AtPSK6, YIYTH). Presumably, the peptides encoded by each gene should be able to bind and activate the single known receptor, AtPSKR1 . Each of the AtPSK precursors has the typical subtilase recognition site signature (di-basic amino acids) 8-10 residues upstream from the mature peptide (Barr, 1991) . However, the residues that are critical for AtSBT1.1 recognition (the four or five residues just downstream from the di-basic site) differ somewhat between AtPSK genes, and therefore AtSBT1.1 appears to be most specific for cleavage of PSK4. That suggests that other subtilases might be involved in the processing of other PSKs. Our interest in AtSBT1.1 stems from our earlier studies of shoot regeneration in Arabidopsis (Lall et al., 2004) . We found that higher expression levels of the gene encoding AtSBT1.1 (At1g01900) correlated with the presence of the superior allele at the major QTL conditioning shoot regeneration in Arabidopsis tissue culture (Lall et al., 2004) . Higher levels of AtSBT1.1 expression may not have anything to do directly with shoot regeneration. However, higher levels of AtSBT1.1 expression might promote the proliferation of callus from which shoots are derived. Similar reasoning was used by Hanai et al. (2000) to explain the stimulatory effects of PSKs on somatic embryo formation in carrot. They concluded that PSKs might promote the proliferation of cells giving rise to somatic embryos, rather than influencing the formation of somatic embryos. Two T-DNA insertion mutant lines for AtSBT1.1 were obtained from the Arabidopsis Biological Resource Center (ABRC, Columbus, OH). Seeds were surface-sterilized, rinsed with sterile water, and stratified at 4°C for at least 2 days in 0.1% agar. Seeds were germinated and grown vertically on agar plates containing Gamborg's B5 medium (Gamborg et al., 1968) . Root segments (5 mm) were cut and transferred to callus induction medium (CIM), which consisted of B5 medium with 5 g l )1 MES, 2.2 lM 2,4-dichlorophenoxyacetic acid, 0.2 lM kinetin and 0.8% agarose. Explants were incubated on CIM for 4 days under constant light conditions, and then transferred to shoot induction medium (SIM). SIM is prepared similarly to CIM except that it contains the hormones isopentenyladenine (5.0 lM) and 3-indoleacetic acid (0.9 lM). For AtSBT1.1-YFP and AtPSK4-YFP localization experiments, transgenic seedlings were grown on B5 plates for 7 days, and roots were used for confocal microscope examination. Screening for homozygous plants was carried out by PCR using left border (LB) T-DNA primers and the gene-specific primer pair ScrSBT1.1 (Table S1 ). The transcript level of AtSBT1.1 was evaluated by RT-PCR using primer pair sqRT-SBT1.1 listed in Table S1 . Plasmid construction ppAtPSK4 and AtSBT1.1 were amplified from root RNA of 1-weekold Arabidopsis seedlings by RT-PCR (primer pairs pSKMPSK4 and pSKMSBT1.1, respectively, Table S1 ), and cloned into the AscI and SpeI sites of pSKM36 in-frame with a 4 x epitope myc tag (EQKLISEEDLRN). cDNA clones with error-free copies were named ppSKAtPSK4 and pSKAtSBT1.1, respectively. A mutated form of AtSBT1.1 (S552A) was generated using a QuickChange site-directed mutagenesis kit (Stratagene, http://www.stratagene.com/) with primer pairs SDM1.1 and pSKAtSBT1.1 as template. YFP C-terminal fusions were created by inserting cDNAs from above at the AscI and SpeI sites of pSKY36. The clones were named 35S:AtSBT1.1-YFP and 35S:ppAtPSK4-YFP. Promoter-GUS constructs for AtSBT1.1 and AtPSK4 were generated by amplifying 984 and 934 nucleotides using the primers pCAMSBT1.1 and pCAMPSK4, respectively. The promoters were ligated into the BamHI and PstI sites of pCAM-BIA3300. Total protein was extracted from transgenic and wild-type plants using extraction buffer [0.1 M HEPES/KOH pH 7.0, 20 mM 2-mercaptoethanol, 0.1 mg ml )1 PMSF, 0.1% w/v Triton X-100, 1 mM EDTA, 20% w/v glycerol and protease inhibitor cocktail (Sigma-Aldrich, http://www.sigmaaldrich.com/)]. An aliquot of total protein was precipitated using trichloroacetic acid and quantified by the Bradford method (Bradford, 1976) . Reaction products were resolved by 12% SDS-PAGE and visualized by Western blotting using c-myc antibody (9E10; Santa Cruz Biotechnology, http://www.scbt.com) and an ECL kit (GE Healthcare, http://www.gehealthcare.com). Assay for AtSBT1.1 activity in vitro AtSBT1.1-myc was affinity-purified from transgenic Arabidopsis plants as follows: 500 g of seedlings were ground in liquid nitrogen and suspended in 25 mM Tris/HCl pH 7.2, 150 mM NaCl, 0.1% Nonidet P-40 (Calbiochem, http://www.emdbiosciences.com) and 10% glycerol. Anti-c-myc agarose conjugate (200 ll; Sigma) was added to the filtered lysate and incubated for 2 h at 4°C with continuous rotation. The agarose beads with bound AtSBT1.1-myc were recovered by centrifugation at 1000 g for 3 min at 4°C. The beads were washed three times with washing buffer (25 mM Tris/HCl pH 7.2, 150 mM NaCl) and suspended in 25 mM MES/sodium acetate buffer pH 6.0. The reactions were carried out at 32°C in the same buffer supplemented with 2.5 mM CaCl 2 . Parallel purification was performed using transgenic plants transformed with a mutated form (S552A) of AtSBT1.1-myc and the empty vector to obtain material for control reactions. For fluorogenic peptide assays, 40 ll of bead-bound AtSBT1.1-myc were added to a solution containing a final concentration of 50 lM fluorogenic peptide in a buffer consisting of 25 mM MES/ sodium acetate pH 6.0 supplemented with 2.5 mM CaCl 2 . Kinetic assays were performed at 32°C and monitored as fluorescence emission at 420 nm (10 nm slit) following excitation by 320 nm (10 nm slit) in a BioTek spectrophotometer (http://www.biotek. com). The reaction was carried out in 96-well plates (Nunc, http:// www.nuncbrand.com). Control reactions were performed using the same fluorogenic peptide with bead-bound mutated AtSBT1.1 (S552A) and bead-bound myc vector only. AtSBT1.1 incubated with 1 mM PMSF, peptide and reaction buffer, and buffer alone incubated in separate wells served as additional controls. To determine the pH optimum of the reaction, a tri-component buffer system of constant ionic strength was used (Ellis and Morrison, 1982) . This buffer comprised 25 mM acetic acid, 25 mM MES, 50 mM Tris/HCl and 2.5 mM CaCl 2 . Total RNA was isolated from ground plant tissues using an RNeasy kit (Qiagen, http://www.qiagen.com/), treated with RNase-free DNase I according to manufacturer's instructions (Qiagen), and quantified by 260/280 nm UV light absorption. A 1 lg aliquot of total RNA was reverse-transcribed using the Superscript III reverse transcription kit (Invitrogen, http://www.invitrogen.com/). Aliquots (2 ll) of cDNA were used for RT-PCR, and tenfold diluted cDNA was used for real-time quantitative RT-PCR. All primers are listed in Table S1 . For real-time quantitative RT-PCR, the efficiency of amplification of various RNAs was assessed relative to amplification of transcripts for two actin genes [actin2 (At3g18780) and actin8 (At1g49240)]. RNA samples were assayed in triplicate. Expression levels were calculated relative to actin using a comparative threshold cycle method with DDCt = DCt reference ) DCt sample , where DCt sample is the Ct value for the assay sample normalized to actin and DCt reference is the Ct value for calibration, also normalized to actin (Liu et al., 2007b) . Root segments (5 mm) incubated on CIM were harvested and further incubated for 6 h in GUS staining solution [100 mM Tris/ NaCl buffer, pH 7, 2 mM ferricyanide, 2 mM X-gluc (5-bromo-4-chloro-3-indolyl-D-glucuronide), 2 mM ferrocyanide, 10 mM EDTA and 0.1% Triton X-100] at 37°C in the dark. The staining solution was removed, and the tissues were dehydrated in an ethanol series from 70% v/v to absolute ethanol. Samples were visualized under the light microscope. MALDI-TOF MS and MS/MS analyses were performed using a QSTAR XL quadrupole TOF mass spectrometer (AB/MDS Sciex, http://www.appliedbiosystems.com) equipped with an orthogonal MALDI ion source. The mass spectrometer was operated in the positive ion mode. Mass spectra for MS analysis were acquired over m/z 500-2500. After every regular MS acquisition, MS/MS acquisition was performed against the most intensive ions. The molecular ions were selected by information-dependent acquisition in the quadrupole analyzer and fragmented in the collision cell. Subcellular localization was carried out using 35S:pSKYAtSBT1.1-YFP-and 35S:pSKYAtPSK4-YFP-expressing roots. The roots were stained with propidium iodide and examined with a laser scanning confocal microscope. Confocal microscopy was performed using a Nikon C1si confocal scanning system attached to a 90i microscope (http://www.nikoninstruments.com). The emission signals for YFP and propidium iodide were acquired using sequential scanning mode to eliminate emission signal bleed-through. The 488 line of the argon laser and 515/30 emission filters were used for acquisition of YFP images. The propidium iodide images of root cells were acquired using the 561 argon laser and 590/50 emission filter. Cells in root segments were plasmolyzed using the conditions described by Von Groll et al. (2002) . Epifluorescent and differential interference contrast images of plasmolyzed cells were acquired using a 5 megapixel Nikon DS-Fi1 camera and Elements BR software. A 200 W mercury light source and FITC filter cube were used for image acquisition. Additional Supporting Information may be found in the online version of this article: Figure S1 . MS/MS analysis of pAtPSK4-myc. Figure S2 . RT-PCR analysis of AtSBT1.1 transcripts in wild-type and sbt1.1-1 and sbt1.1-2 mutants. Figure S3 . pAtPSK4-myc processing is restored in F 1 hybrids resulting from out-crossing of 35S:ppAtPSK4-myc sbt1.1-1 to wildtype. Figure S4 . AtSBT1.1-YFP and AtPSK4-YFP are located in the apoplast. Figure S5 . MS/MS spectrum for the cleaved N-terminal fluorogenic peptide. Table S1 . Primer sequences used. Please note: Wiley-Blackwell are not responsible for the content or functionality of any supporting materials supplied by the authors. Any queries (other than missing material) should be directed to the corresponding author for the article.

Protein Domain Boundary Predictions: A Structural Biology Perspective

One of the important fields to apply computational tools for domain boundaries prediction is structural biology. They can be used to design protein constructs that must be expressed in a stable and functional form and must produce diffraction-quality crystals. However, prediction of protein domain boundaries on the basis of amino acid sequences is still very problematical. In present study the performance of several computational approaches are compared. It is observed that the statistical significance of most of the predictions is rather poor. Nevertheless, when the right number of domains is correctly predicted, domain boundaries are predicted within very few residues from their real location. It can be concluded that prediction methods cannot be used yet as routine tools in structural biology, though some of them are rather promising.

Computational/mathematical approaches, such as structural bioinformatics [1] , structural class prediction [2, 3] , molecular docking [4] [5] [6] [7] [8] [9] , molecular packing [10, 11] , pharmacophore modelling [12] , Mote Carlo simulated annealing approach [13] , diffusion-controlled reaction simulation [14] , graph/diagram approach [15] [16] [17] [18] [19] [20] [21] , bio-macromolecular internal collective motion simulation [22] , QSAR [23] [24] [25] , protein subcellular location prediction [26] [27] [28] [29] [30] , protein structural class prediction [31, 32] , identification of membrane proteins and their types [33] , identification of enzymes and their functional classes [34] , identification of proteases and their types [35] , protein cleavage site prediction [36] [37] [38] , and signal peptide prediction [39, 40] can timely provide very useful information and insights for both basic research and drug design and hence are widely welcome by science community. Several computational approaches aimed to the prediction of protein domain boundaries have been published during the last few years [41, 42] . Besides their intrinsic interest in genome analysis and evolution studies, they are tools that structural biologists may use to optimize the design of the constructs of the proteins, the three-dimensional (3D) structure of which must be determined [43] . While this is particularly important in structural genomics (SG), where the targets have, in general, not been deeply characterized with appropriate biochemical and biophysical tools, this can be important also for traditional hypothesis-driven structural biology projects, where a fine tuning of the construct that is inserted into the experimental pipeline -cloning, expression, purification, etc. -is often necessary in order to get suitable samples [44] . Several information about structure prediction methods are periodically published in the framework of the CASP *Address correspondence to this author at the Department of General Chemistry, Pavia University, Viale Taramelli 12, I-27100 Pavia, Italy; Tel: +43 1 4277 52208; E-mail: oliviero.carugo@univie.ac.at initiative, the main goal of which is to promote an evaluation of computational prediction methods [45] . This is a periodical exercise, performed every two years since 1994. During CASP experiment a series of protein sequences, the 3D structure of which was determined experimentally though it was not yet published, are distributed to research groups that develop computational methods for predicting protein structural features. It is thus a blinded test, where several methods of "in silico" structural biology techniques can be compared to the reality and to each other. Nevertheless, in each CASP run, the number of targets is obviously quite limited and a prediction method that performs very well in CASP is not necessarily better than other techniques in the reality. It is necessary to make additional investigations focusing on the possibility to use these prediction methods for practical application in structural biology. Although it is impossible to consider it a rule, it is generally easier to work with single-domain proteins than with multi-domain proteins, since the latter ones tend to be conformationally more flexible [46] . For example, the reciprocal orientation of the domains can vary and depend on the presence of other molecules. Multi-domain proteins may also be little prone to refold if, by chance, they had been overexpressed in cells lacking proper chaperones. This does not mean that multi-domain proteins cannot be studied but it implies that some care must be paid in structural biology experiments and that longer time and larger funding can be expected to be necessary to solve multi-domain proteins. It is thus extremely important to be able to predict, on the basis of its amino acid sequence, if a protein contains one or more structural domains. CASP is divided into several sections, ranging from prediction of conformational disorder to tertiary structure prediction. Protein domain boundary predictions began to be included in the CASP initiative in 2004. The dissection of a protein into separate structural domains is in fact not trivial at all [46, 47] . It is related to the ill-definition of what a protein domain is. An amino acid segment can be in fact consid-ered to be a structural domain if i) it is a compact ensemble of atoms/residues; ii) it is an ensemble of atoms/residues that behaves as a rigid body, in the sense that it can move relative to other protein moieties without changing its shape; iii) it is a self-folding subunit; iv) it is a polypeptide segment well conserved during molecular evolution. Given the ambiguity in any quantitative definition, the real domain boundaries were defined according to the CASP7 organizers and assessors [47] . They found a reasonable consensus definition for each investigated protein, which seems to be well suitable for a structural biology analysis. The present study is attempted to compare modern approaches for predicting protein domain boundaries and to define new prediction strategies. Here, we refer to the exercise named CASP7, organized in 2006, for which both predictions and experimental data are available on-line (http://www.predictioncenter.org/casp7/Casp7.html). In this manuscript, several tools, designed for predicting domain boundaries on the basis of the amino acid sequence, will be compared to the real domain architecture. The analysis of these data allows one to answer the following basic questions: i) Is it possible to predict, with the presently available bioinformatics tools, if a protein is made by a single domain or if it contains more than one domain? ii) What is the statistical significance of the available predictions? iii) How accurately can the domain boundaries be predicted in the cases where the presently available bioinformatics predictions work well? Data were obtained from the CASP7 web page (http://predictioncenter.gc.ucdavis.edu/casp7/). Table 1 shows the bioinformatics tools that are freely available and that were used in CASP7. Protein domain prediction methods can be classified into three main categories [42] : i) homology prediction; ii) domain recognition; iii) new domain prediction methods. The 14 prediction methods regarded in present study include all types of approaches. The homology prediction is presented by the chop [48, 49] methods that assign the query sequence to known PDB chains. Dsp [42] uses in addition more general properties of sequence conservation throughout the protein and it can be considered as lying between domain homology and new domain predictions. Domssea [42] belongs to the domain recognition approaches. It is based on the assumption that secondary structure is a more conserved feature of proteins with similar folds than sequence. Domssea aligns the secondary structure predicted for a query protein against a database of 3D domain structures and derives the domain boundaries from the known domain with the most similar secondary structure. Robetta [50] applies BLAST/PSI-BLAST for domain homology prediction and it uses FFAS03 and 3D-Jury to find remote homologues of known domain structure. Hhpred [51] is a server for remote homology detection and for structure prediction using pairwise comparison of profile hidden Markov models (HMMs). In the foldpro [52] method the structural relevance of the query-template pairs is extracted from global profile-profile alignments in combination with predicted secondary structure, relative solvent accessibility, contact map and beta-strand pairing using support vector machines. Distill [53] provides prediction of Contact Density defined as the Principal Eigenvector (PE) of a residue contact map. This information is an important intermediate step towards ab initio prediction of protein structure and is used to identify domains. Baker generates 3D protein models using the de novo prediction algorithm Rosetta and then assigns domain boundaries using Taylor's structure-based do- maopus http://sigler.bioch.bcm.tmc.edu/CASP7-DOM/ * metadp http://meta-dp.cse.buffalo.edu [54] main identification technique. Maopus performs a template screening with PSI-BLAST and FFAS03. The SKELEFOLD approach implemented in Maopus is a de novo folding algorithm that uses vector representations of secondary structural elements; domain boundaries are defined with three sequence-based filters. In the domfold method, the output from DomSSEA, DISOPRED and HHsearch is parsed to form a consensus. Metadp [54] and NNput are meta servers that comprise a number of domain prediction methods. Some of the bioinformatics methods provide multiple predictions. In this case, only the first, which is considered to be the more reliable, was retained for further analysis. Predicted domain boundaries were obtained from the CASP7 web page. The experimental domain boundaries were also obtained from the CASP7 web page, where they were generated by a group of expert scientists. 95 proteins are considered. Given that predictions were not deposited for each protein and for each prediction method, this results in a set of 1210 predictions [47] . To predict, on the basis of the protein length, that a protein contains one domain or it is a multi-domain construct, a threshold value can be used. If the protein is longer than the threshold value it consists of more than one domain. On the contrary, a protein, smaller than this threshold value, would be predicted to contain only a single domain. Consequently, a true positive (tp) is defined as a multi-domain protein, which is correctly predicted to be a multi-domain protein; a multi-domain protein that is predicted to contain a single domain is defined a false negative (fn); a single-domain protein predicted to be a multi-domain protein is defined a false positive (fp); and a correctly predicted single-domain protein is defined a true negative (tn). These four types of predictions can be used to estimate the reliability of this prediction methodology. A number of figures of merit have been used for that, like, for example, the Matthews correlation coefficient (mcc) [55] the values of which range from -1 to +1 (larger values indicate better predictions) and is little affected by sample heterogeneity (the number of single-domain proteins can be different from the number of multi-domain proteins). The prediction accuracy was validated with a Jack-knife procedure. In statistical prediction, the following three crossvalidation methods are often used to examine a predictor for its effectiveness in practical applications: independent test dataset, sub-sampling test, and Jack-knife test [56] . However, as elucidated in references [26] and [27] , amongst the three cross-validation methods, the Jack-knife test is deemed the most objective that can always yield a unique result for a given benchmark dataset, and hence has been increasingly used and widely recognized by investigators to examine the accuracy of various predictors [57] [58] [59] [60] [61] [62] [63] [64] [65] [66] . To compare the accuracy of different methods with a random prediction we estimated numerically the probability density functions of the indices used to measure the classification validity. This approach is based on idea that the problem of domain boundary prediction using the amino acid sequence is a classification problem. Each residue is in fact predicted to belong to a certain class and it cannot belong to two different clusters at the same time. In other words, a residue can be predicted to belong to a certain domain, to another domain, or to a linker segment. The comparison between a prediction and the reality or between two predictions can thus be performed by using statistical tools that are routinely employed to compare alternative classifications [67] and that are briefly described below. Given for example two classifications (C and K) of n residues, it is possible to count the number of cases in which residues i and j were classified in the same group in C and K (n_ss), the number of cases in which i and j were classified in the same group in C and in different groups in K (n_sd), the number of cases in which i and j were classified into two different groups in C and in the same group in K (n_ds), and the number of cases in which i and j were classified into two different groups both in C and in K (n_dd). On the basis of this description, it is possible to compute the Jaccard index (J), the Rand coefficient (R), and the Fowlkes-Mallows index (FM), which are defined as: FM = n _ ss n _ ss + n _ sd n _ ss n _ ss + n _ ds (4) where M = n _ ss + n _ sd + n _ ds + n _ dd . By definition, if the two classifications C and K are identical, all the indices (J, R, and FM) are equal to one. It is also important to observe that these indices can be computed independently of the fact that the classifications C and K contain the same number of clusters. This means that the values of J, R, and FM can be computed also if in one case, for example the classification C, all the residues were predicted to be in a unique domain while in the other case, for example the classification K, some residues were assigned to different domains. The only constraint to the computation of J, R, and FM is that both classifications C and K must include the same number of residues, and in the present case this is obvious. The computation of the values of J, R, and FM is elementary. The estimation of their statistical significance is less obvious [67] . For example, it is difficult to estimate the probability that a certain value of the index J was obtained by chance. From another point of view, if J CK > J DL , where J CK monitors the similarity between the classifications C and K and J DL difference between the classifications D and L, it is clear that C and K are more similar to each other than D and L. However, it is more difficult to estimate the statistical significance of the inequality J CK > J DL . In other words, it is more difficult to estimate the probability that C and K are really more similar to each other than D and L. This depends on the fact that the probability density functions of the indices J, R, and FM are unknown and must therefore be estimated numerically on the basis of adequate simulations. Therefore, we generated a series of simulated partitions, using a Metropolis-Monte Carlo approach, by mean of the following procedure. Each partition is characterized by a series of boundaries that separate a domain and a loop and that can be located also at the N-or at the C-terminus. Given a protein containing N residues, a boundary can be any integer k with 1 k N. A series of boundaries were generated iteratively. The first (k 0 ) was randomly selected in the range (1, N); the second (k 1 ) was randomly selected in the range (1, m 0 ), where m 0 = N -k 0 ; the third (k 2 ) was randomly selected in the range (1, m 1 ) where m 1 = m 0 -k 1 ; and so on, the i th boundary (k i ) was randomly selected in the range (1,m i-1 ), where m i-1 = m i-2 -k i-1 . Two constrains were imposed during the generation of random domain boundaries within a protein. We considered that a domain must contain more than 30 residues and a loop size must be smaller than 30 residues. 10,000 random partitions into domains were generated for proteins containing 75, 100, 125, ..., 550, 575, 600 residues. It was then possible to make 49,995,000 pairwise comparisons between two partitions and the 49,995,000 values of the coefficients J, R, and FM were retained in order to determine their distributions. As an example, Fig. (1) shows the distributions of the index R for some N values. It appears that the distribution dispersion decreases if N increases and that the maximum moves to higher R values for larger proteins. With these data, it is possible to estimate the probability pR to have R values higher than a given value Rx, simply by integrating the probability density curve from Rx to 1, and, analogously, it is possible to get the statistical significance for the other indices. The definition of what is a well predicted domain is obviously arbitrary and here the following conditions were used in order to select the predictions that can be considered to be satisfactory. If the domain contains N residues and it is predicted to contain M residues, and if C is the number of residues that are found in both the real and the predicted domain, a good prediction was defined as a case in which N M < 20 (6) and C min(M , N ) > 0.95 (7) For well predicted domains, we then computed the differences between the sequence position in which the domain is predicted to begin and the sequence position in which it begins in the reality (Delta_b). Note that a negative value of Delta_b indicates that the domain is predicted to begin before the real beginning along the protein sequence. Analogously, we also computed the differences between the sequence position in which the domain is predicted to end and the sequence position in which it ends in the reality (Delta_e). A positive value of Delta_e indicates that the domain is predicted to be slightly longer, at its C-terminus, than the reality. Fig. (2) shows the distributions of the protein dimensions, measured by the number of amino acid residues, for the single-and multi-domain proteins examined in the CASP7 experiment. As expected, single-domain proteins tend to be smaller than multi-domain proteins, though some overlap between the two distributions exists. Fig. (2) . Distribution of the number of residues (nres) in the singleand muti-domain proteins examined in the CASP7 experiment. It is thus easy to select a threshold value t and to predict that a protein contains only one domain if smaller than t and that it is multi-domain protein if larger than t. Table 2 shows the mcc values [see equation (1) ] observed at various threshold values and validated with a Jack-knife procedure for the proteins examined in the CASP7 experiment. It can be observed that the mcc values are obviously smaller for very small or large values of the threshold. On the contrary they are rather large (>0.6) for intermediate threshold values and the highest mcc (0.628) is observed with a threshold of 200 residues. This prediction approach is clearly very naive. It simply assumes that a protein domain has a little probability to be very large and, as a consequence, that larger proteins have a higher probability to contain two or more domains. A protein is predicted to contain a single domain if it contains less residues that t and it is predicted to contain more than one domain if it has a number of residues larger than t. Data are taken from the proteins examined in the CASP7 experiment. It is interesting to compare the results of this extremely simple prediction strategy with the results obtained within the CASP7 experiment, where several prediction methods were applied to about 100 proteins. Table 3 shows the mcc values computed on the basis of the predictions deposited by the participants to the CASP7 experiment. The same classification in tp, fp, fn, and tn, which is described in the Methods section, was used. This means that if protein P contains more than a single domain and it was predicted to contain more than a single domain by using the prediction method M, this was considered a true positive (tp). On the contrary, if it was predicted to contain only one domain by the method M, the prediction was considered a false negative (fn), etc. The data of Table 3 clearly show that most of the prediction methods are less reliable than the predictions based on the very simple assumption that a small protein has a high probability to contain a single domain and that a large protein is likely to contain two or more domains. Actually, only four methods (baker, foldpro, maopus and robetta) can predict a multidomain protein better than the simple predictor (Matthews correlation coefficient larger than 0.628). What does this mean? Are these bioinformatics tools useless in structural biology? The answer is no. First, some of them seem to be rather accurate. Second, these computational techniques were not specifically trained to identify multi-domain proteins and it is thus not surprising that some of them are not suitable to discriminate mono-and multidomain proteins. However, it is reasonable to suppose that these bioinformatics tools are still immature and progress should be expected in the future. Table 4 shows the average values of the J, R, and FM indices computed by comparing predicted and real partitions [see equations (2)-(4)]. All the values tend to be large, quite close to their maximal value of 1. However, the probabilities (pJ, pR, and pFM) to observe by chance values higher than these are quite large, ranging from about 30% to about 70%. Baker, foldpro, maopus and robetta are better in predicting a partition that is closer to the real one, with J, R, and FM values that are larger and have a minor probability to be observed by chance. Not surprisingly, they are the same methods that work better to identify multi-domain proteins (see the mcc values of Table 3 ). It must also be observed that matching between prediction and reality is slightly better for small proteins than for large proteins. For example, the probability pJ to find J values larger than those observed by comparing the reality and the predictions of the method "baker" is on average equal to 39%, it decreases to 33% for proteins shorter than 150 residues, and it increases to 43% for proteins containing more then 150 amino acids. This is actually not surprising, since it is easier to predict that a small protein contains a single domain, with, perhaps, two small N-and C-terminal segments protruding from the domain. However, it must be noted that, despite the fact that the pJ, pR, and pFM values can be used only as semi-quantitative indicators -since they are obtained from empirical statistical distributions -it is quite clear that the domain boundary predictions are still quite far from matching the reality. We have seen in the previous chapters that the bioinformatics tools are not yet mature enough to be used as routine instruments to design structural biology experiments. However, a very positive feature of these computational methods is that when they work [see equations (6) and (7)] they work very well. The following data are shown: the percentage of domains that are correctly predicted (see text for details) PC_C, the average deviation between the real and the predicted beginning of the domain Delta_b, and the average difference between the real and the predicted end of the domain Delta_e (standard deviations of the mean in parentheses). Table 5 shows the percentage of domains that are correctly predicted [according to equations (6) and (7) ] and the discrepancy between the real and the predicted boundary in the subset of domains that are correctly predicted. It appears that only a relatively modest fraction of the domains can be considered to be well predicted, according to the criteria defined by equations (6) and (7) . The percentage of good predictions is about 30-40%, with some prediction methods behaving considerably better than the others and able to well predict about 60% of the domains. The average values of Delta_b (see Methods) are close to and lower than 0 for all the prediction methods. Also the values of Delta_e are very small, though their absolute value tends to be slightly larger than that of Delta_b. Interestingly, the Delta_e values are positive, on average, for each prediction method. This clearly indicates that in the subset of good predictions the domain boundaries are located with very high accuracy. Actually, a deviation of 1-3 residues is probably a very minor mistake in the process of design a protein construct that has, on average, a high probability to be well folded and conformationally homogeneous. It is also interesting to observe that while the Delta_b mean values are negative, the mean Delta_e values are larger than 0, indicating that predicted domains tend to be slightly longer than real domains. In the present manuscript we have analyzed the reliability of the predictions that were made in the CASP7 experiment and that are publicly available. It was found that most of the bioinformatics tools are able to determine if a protein is made by a single domain or if it contains more than one domain, despite a similar reliability is reached by considering only the sequence length, a much simpler strategy. Using a standard and well known statistical test, we showed that most of the predictions that can be done are not impressively better than pseudo-random predictions. It was also observed that although the reliability of the prediction methods seems to be insufficient to make them routine tools in experimental structural biology, their performance can be extremely good. When the domain is correctly identified, its boundaries are very close, within one or two residues, to the experimental ones. In conclusion, these bioinformatics applications are not yet sufficiently accurate to be used as routine tools in experimental structural biology. It is rather probable that the use of more than a single prediction method by a sort of consensus approach might improve the reliability of the predictions. Although these bioinformatics tools are still immature, progress can be expected in the future. This work was supported by the Austrian GEN-AU project BIN-II. Björn Sjöblom and Kristina Djinovic are acknowledged for helpful discussions. Financial support by Putta None is also acknowledged. One reviewer is acknowledged for a series of references that deserved citation.

Enhancing Time-Series Detection Algorithms for Automated Biosurveillance

BioSense is a US national system that uses data from health information systems for automated disease surveillance. We studied 4 time-series algorithm modifications designed to improve sensitivity for detecting artificially added data. To test these modified algorithms, we used reports of daily syndrome visits from 308 Department of Defense (DoD) facilities and 340 hospital emergency departments (EDs). At a constant alert rate of 1%, sensitivity was improved for both datasets by using a minimum standard deviation (SD) of 1.0, a 14–28 day baseline duration for calculating mean and SD, and an adjustment for total clinic visits as a surrogate denominator. Stratifying baseline days into weekdays versus weekends to account for day-of-week effects increased sensitivity for the DoD data but not for the ED data. These enhanced methods may increase sensitivity without increasing the alert rate and may improve the ability to detect outbreaks by using automated surveillance system data.

S ince the late 1990s, the threats of bioterrorist attacks, the potential for outbreaks of natural disease such as severe acute respiratory syndrome and pandemic infl uenza, and the availability of computerized data have prompted the use of automated disease surveillance systems (1) . Sources of information include clinical data, such as records of hospital emergency department visits, and nonclinical information, such as sales of over-the-counter remedies (2) . However, human resources are limited for interpreting the large volume of available information. Thus, statistical algorithms are needed to fi lter large volumes of data, focus attention on potential public health problems, and provide an objective measure of increases in disease activity. BioSense is a US national automated surveillance system that receives data from various sources and makes them available for public health use. The data may be viewed simultaneously by local, state, and federal public health offi cials through the Internet-based BioSense Application, which may be accessed on a jurisdiction-specifi c basis through the Centers for Disease Control and Prevention (CDC) Secure Data Network (3) . Data received include coded fi nal diagnoses and free-text chief complaints, which are assigned as appropriate to >1 of 11 syndrome groupings representing general illness categories such as respiratory and gastrointestinal illnesses (4) and to >1 of 78 subsyndromes representing more specifi c categories such as asthma or cough (5) . To identify days when disease indicator activity is higher than expected, BioSense uses a modifi ed version of the C2 algorithm, 1 of 3 algorithms (C1, C2, and C3) developed for the Early Aberration Reporting System (EARS) (6, 7) . The C2 algorithm uses a sliding baseline of 7 consecutive recent days' counts to calculate a mean (μ) and SD (s t ). The test statistic is (x t -μ)/s t , the number of SDs by which the current value x t exceeds μ, or 0 if x t does not exceed μ. EARS uses a test statistic >3 to signal an alert (6, 7) . Owing to their simplicity, ease of implementation, and implicit correction for seasonal trends (only data from the prior 9 days are used), the EARS algorithms are widely used (8) (9) (10) . However, the algorithms do not perform optimally under all circumstances. First, because daily counts often vary by day of week, many alerts may be produced on high-count days such as Mondays and Tuesdays, and few may be produced on low-count days such as weekend days. Second, the short (7-day) baseline period may produce unstable values for the mean and SD; thus, the minimum daily count that triggers an alert may vary widely over a short period. Third, using simple count data does not account for the population at risk, which is generally unknown in these systems and which may vary, especially during crisis situations. Although C2 can be used on rates rather than counts, prior evaluations have not shown that using rates improves performance (L. Hutwagner, pers. comm.). Finally, occurrences of many disease indicators are rare, resulting in calculations for both expected values and SDs of 0; the EARS methods are not recommended in such instances. A minimum SD may be used to avoid division by zero, but if this minimum value is set to 0.2, a count of 1 will be 5 SDs above the mean and trigger a high-level alert. This article describes and evaluates modifi cations of C2 that retain its inherent advantages, address its potential limitations, and improve its performance. We used real daily syndrome counts from 2 sources as baseline data and assessed the ability of various algorithms to detect additional counts artifi cially added to the data. Because all analyses were conducted at a constant alert rate of 1%, improvements in sensitivity were not accompanied by an increase in alerts. Four algorithm modifi cations, designed to address shortcomings in the C2 algorithm, were tested. The fi rst modifi cation tested was stratifi cation by weekdays versus weekend days. Although many methods have been used to adjust for differing counts by day of week (11) , these methods may require customization to specifi c datasets and a long data history (up to several years). Our simple method is to stratify the baseline days used to calculate μ and s t into weekdays versus weekend days. This stratifi cation is denoted the W2 algorithm. For example, a 7-day W2 baseline for weekdays contains the most recent 7 weekdays. For unstratifi ed and stratifi ed analyses, the 2 days immediately before the index day were excluded from the baseline, a standard practice for C2, to avoid contamination with the upswing of an outbreak. The second modifi cation tested was lengthening the baseline period. Because a 7-day period may provide insuffi cient data for an accurate and stable calculation of μ and s t , we tested baseline periods of 7, 14, and 28 days. However, because we used data from <56 days before the index day, the stratifi ed 28-day baseline will include only ≈16 days for weekend days. The third modifi cation tested was adjustment for total daily visits. For the adjustment procedure, we used a formula in which n 0 = count of visits on the index day for the chosen syndrome (e.g., visits for the respiratory syndrome), and d 0 = the total number of facility visits on the index day, including visits that were both assigned and unassigned to any of the 11 syndromes. Σn i = total syndrome visits summed for all i baseline days. Σd i = total facility visits summed for all i baseline days. The formula for the adjusted expected value was e 0 = d 0 × Σn i /Σd i , which differed considerably from the mean of the n i if d 0 was high or low. Fewer visits for a given syndrome were thus expected on a day when the facility had fewer total visits. The estimated adjusted SD, s 0 , was taken as the mean absolute value of (n i -d i × Σn i /Σd i ) over i baseline days; that is, s 0 = Σ (abs (n i -d i × Σn i /Σd i ))/i. The test statistic adjusted for total visits was (n 0 -e 0 )/s 0 , analogous to the C2 statistic (n 0 -μ)/ s t , where μ and s t are the mean and SD of n i , the counts on baseline days. In the discussion below, we refer to this adjustment as the rate algorithm. The fourth modifi cation tested was increased minimum value for SD. We studied minimum values of 0.2 and 1.0. To test these modifi cations, 2 datasets were used: records of Department of Defense (DoD) facility fi nal diagnoses for September 2004-November 2007 and records of hospital emergency department (ED) chief complaints for March 2006-November 2007. The DoD data consisted primarily of data from outpatient clinics; however, ≈15% of the visits in this evaluation were from patients seen in emergency facilities and cannot currently be differentiated in the BioSense System. We studied the 11 syndrome groups designed to be indicative of infections resulting from exposure to pathogens plausibly used in a bioterrorist attack (4). The DoD data consisted of daily counts of patient visits with International Classifi cation of Diseases, 9th Revision (ICD-9)-coded diagnoses categorized into the 11 syndrome groups. The hospital ED data consisted of freetext chief complaints, which were fi rst parsed for a specifi ed set of keywords, abbreviations, and misspellings and then categorized into 10 of the syndrome groups (1 syndrome, specifi c infection, was used for diagnosis but not for chief complaint data). Some ICD-9 codes and chief complaints may be included in >2 syndromes. However, counts of different syndromes were analyzed separately, not added together, and therefore are not double-counted in the analyses. For both datasets, we analyzed counts aggregated by facility. We included facility-syndrome combinations that had mean counts >0.5 over all facility-syndrome days in the study period. Many DoD clinics are closed on holidays. Therefore, for the DoD data, 11 days (days on which federal holidays are observed and the day after Thanksgiving) were recoded as weekend days for purposes of stratifi ed algorithm calculations (5) . Because hospital EDs typically are open on these holidays, no recoding for holidays was performed for this dataset. The mean count for each facility syndrome was calculated and categorized as follows: 0.5 to <2, 2 to <4, 4 to <6, 6 to <8, 8 to <10, 10 to <20, 20 to <40, and >40. Empirical distributions of the test statistic (e.g., number of SDs by which the observed count exceeds the expected value) were conducted separately for each dataset, algorithm, and mean count category; the 99th percentile value for each of these distributions was used as the cutoff value to defi ne an alert rate of 1%. For example, for the standard C2 algorithm in DoD data with mean count 4 to <6, a cutoff value of 3.9 was used because 1% of the facility-syndrome days had a test statistic >3.9. Because no attempt was made to fi nd and exclude real outbreaks from the data, these cutoff values defi ne an alert rate rather than a false alert rate, the latter being equivalent to 1-specifi city (12) . At a constant alert rate of 1% for all methods, the sensitivity for detecting additional counts was calculated by performing the following steps: 1) running the algorithm to determine expected values and SDs for each facilitysyndrome-day; 2) fi nding the 99th percentile cutoff value for the test statistic for each dataset-algorithm-mean count category as explained above; 3) for each facility-syndrome day, determining whether the observed count plus additional counts is greater than or equal to the threshold value (threshold value = expected value + SD × 99th percentile cutoff value); and 4) calculating sensitivity as the percentage of days on which the additional counts would exceed the threshold value and therefore be detected. Using this method, a single computer run can calculate sensitivity for detecting single-day additional counts on all days in the dataset; if the additional counts are spread over multiple days, separate computer runs would be needed (7). The DoD diagnosis data contained 1,939,993 facilitysyndrome days from 308 facilities in 48 states with an overall mean of 7.7 counts per facility per day; of the 11 syndromes, respiratory visits comprised the highest percentage (16% of total facility-syndrome days) and had the highest mean count (26.0 visits per facility per day) ( Table 1 ). The hospital ED data contained 768,195 facility-syndrome days from 340 facilities in 21 states and had an overall mean of 7.8 counts per facility per day; no visits for lymphadenitis and severe injury and death were included because no facilities had a mean count >0.5 per day for these syndromes. The DoD data had a strong day-of-week effect; 16%-21% of total weekly visits occurred per day on weekdays, and only 3%-4% of visits occurred per day on weekend days and holidays ( Figure 1 ). The hospital ED data had a minimal day-of-week effect: 14%-16% of visits occurred per day on weekdays, and 14%-15% of visits occurred per day on weekend days. The accuracy of expected value calculation was evaluated by using mean absolute residuals. For lower residuals, expected values are closer to observed values than they are for higher residuals. Similarly, the expected value calculation is more accurate for lower residuals than for higher residuals. For the DoD data, lower residuals were seen with stratifi cation (W2) and the rate algorithm: mean residual 4.2 for unstratifi ed count algorithm versus 2.2 for stratifi ed rate algorithm (Table 2) . For the hospital ED data, residuals were lower for the rate algorithm, and stratifi cation had a minimal effect. Varying the baseline duration and minimum SD had no effect on the accuracy of expected value calculation (data not shown). The effect of modifi cations of the initial algorithm on the sensitivity for detecting additional counts was examined; each modifi cation was added consecutively (Table 3) . For the DoD data, sensitivity was 40.6% for the initial algorithm and increased to 43.9% when the rate method was used; 70.8% when the minimum SD was increased to 1.0; 79.4% when the baseline duration was increased to 28 days; and 82.0% when a stratifi ed baseline was used. Comparing the initial algoithm to the best algorithm showed a 41.4% increase in sensitivity. For the hospital ED data, sensitivity was 40.2% for the initial algorithm and increased to 64.8% for the best method (minimum SD = 1, 28-day baseline, rate method, unstratifi ed baseline); however, when the stratifi ed baseline was used, sensitivity decreased to 62.1%; the initial algorithm compared with the best algorithm showed a 24.6% increase in sensitivity. When these sensitivity calculations were stratifi ed by mean count for each facility-syndrome (data not shown), we found that the modifi cations increased sensitivity in all strata of the DoD data; for the hospital ED data, the rate method reduced sensitivity by 1.0% in the 8 to <10 count category and by 0.5% in the 10 to <20 count category, but increased sensitivity in other categories and overall. When we limited analysis to ED data with a mean count of 4 to <6 per day and explored sensitivity for detecting varying numbers of additional counts (Figure 2 ), we found, as expected, that as the number of additional counts increased, sensitivity increased. The difference between the initial and best algorithms was highest when sensitivity was ≈50% for the initial algorithm. That is, for 10 additional counts, sensitivity was 49.8% for the initial algorithm and 85.3% for the best algorithm, an improvement of 35.5%. However, if the initial C2 algorithm had either low or high sensitivity, the modifi cations had little effect. As an example, we analyzed fever syndrome data from 1 ED. The mean count was 4.9 per day, and the 99th percentile threshold values were 3.86 SDs for the initial and 3.55 for the best algorithm. Over 632 days, the sensitivity for detecting 8 additional counts was 47.2% for the initial and 70.9% for the best algorithm (23.7% difference). Data for a 2-month period showed that the calculated SD (Figure 3 , panel A) and the threshold value (i.e., count needed to trigger an alert; Figure 3 , panel B) varied substantially for the initial algorithm but were comparatively stable for the best algorithm. During the 2-month period, 8 additional counts would be detected by initial and best algorithms on 30 days, by only the initial algorithm on 2 days, and by only the best algorithm on 19 days; neither algorithm detected the additional counts on 10 days (Figure 3 , panel C). Our results demonstrate that simple modifi cations of the widely used C2 algorithm can substantially improve the ability to accurately recognize 1-day increases in disease syndrome activity. Depending on the dataset, mean count in the data, and the number of additional counts added, the enhanced methods may increase sensitivity by 20%-40%. These improvements were achieved without an increase in the alert rate, which was held constant at 1% for all methods. Although we chose a 1% alert rate for testing purposes, in practice, it is useful to vary the alert rate to fi t the circumstances, and the BioSense application enables the alert rate to be varied between 0.1% and 2%. Regardless of the alert rate used, the modifi ed methods have higher sensitivity. For the DoD and hospital ED datasets, sensitivity was improved by using a higher minimum SD of 1.0, a longer baseline duration of 28 days, and adjusting for total visits. Stratifying baseline days into weekdays versus weekends/ holidays increased sensitivity in the DoD data, which has a strong day-of-week effect, but modestly decreased sensitivity in the hospital ED data, which does not have such an effect. Thus, the best analytic methods depend on dataset characteristics, especially the day-of-week effect, and could be varied by manual or automated selection. These fi ndings can be used to improve both early event detection and situation awareness because accurate recognition of unusually high counts is needed for both uses. These modifi cations were apparently effective for the following reasons. Accounting for total visits to the facility (i.e., rate method) produces a more accurate expected value and lower residuals ( Table 2 ). Although number of total visits is not the ideal denominator, in general it is better than no denominator at all. An advantage of the rate method is that calculations may be made when only partial data for a given day are available. However, adjusting for total visits may reduce sensitivity slightly in some subgroups, as we found for the hospital ED data when the mean count was 8 to <20. Stratifi cation by weekday versus weekend day improves expected value calculations when a substantial day-of-week effect exists, such as in the DoD data. When such an effect is not present, stratifi cation causes days further from the index day to be used in the baseline period, therefore producing slightly less accurate expected values. Longer baseline durations have no effect on the accuracy of expected value calculation and improve sensitivity by producing more accurate and stable SD values. Using a higher minimum SD avoids nuisance alerts that may be prompted by small fl uctuations in the daily visit count. This method also changes the distribution of test statistic values, which results in a lower 99th percentile cutoff value, which increases sensitivity for detecting moderate-to-high numbers of added counts. Using a higher minimum SD is benefi cial if disease indicators with low and high counts are analyzed; an alternate approach is to use different methods for lowversus high-count data. The issues focused on by our suggested modifi cations may alternately be addressed by various sophisticated mathematical modeling approaches. However, health departments, which are generally limited in resources and in analysis expertise, may resist use of decision-support methods that are expensive, diffi cult to implement, or not transparent to human data monitors. For example, sophisticated Serfl ing-type regression models have long been used by CDC for tracking the progress of infl uenza season (13, 14) and have been used to analyze selected data in the Bio-Sense system. However, these models have both strengths and weaknesses and have not been widely embraced for daily disease surveillance. Even if the expertise and hardware capability for applying them were made available to local health departments, many time series are unsuitable for this approach. We present simple and easily understood and implemented enhancements to C2 to extend its applicability and improve its performance. These enhancements may be applicable to other control chart-based algorithms as well. Automated surveillance systems based on chief complaints and diagnoses have a number of uses: providing assistance in data collection; monitoring seasonal infl uenza (15) ; monitoring total ED visits during a crisis; and monitoring simple surrogates of infectious diseases, injuries, and chronic diseases during large outbreaks or disasters (16) . The utility of these systems has not been demonstrated for Additional counts Sensitivity Figure 2 . Sensitivity of detecting various numbers of additional counts, by using initial versus best algorithms for hospital emergency department chief complaint data, for selected BioSense data. Red line shows the initial algorithm (minimum SD = 0.2, 7-day baseline, count method, unstratifi ed baseline), and black line shows the best algorithm (minimum SD = 1.0, 28-day baseline, rate method, unstratifi ed baseline). monitoring small-or intermediate-sized outbreaks or illnesses defi ned primarily by laboratory testing. Even when using these suggested modifi cations, sensitivity for detecting additional counts at the facility level remains modest. However, the utility of automated biosurveillance will be expanded with the availability of better population coverage and more specifi c data, the use of multiple data types in combination, and improved detection algorithms, such as those proposed here. The limitations of this study include using only data with a mean count >0.5 per day; analyses of sparser data might show different results. We studied only facility-level aggregation of data, selected patient types (e.g., hospital inpatients were not studied), selected data types (e.g., ED diagnoses were not studied), and broadly defi ned syndromes (the more granular subsyndromes, which are likely to yield lower counts, were not studied). Although we evaluated only a simple time-series detection method, optimizing performance of simple methods is useful before they can be meaningfully compared with more sophisticated methods, such as regression. Also, we studied effects of additional counts on single days rather than multiday outbreak effects; however, because the C2 algorithm considers data from only 1 day at a time, this is a reasonable initial approach. These results must be confi rmed by trials of multiday signal injection and performance evaluated for multiple subgroups (e.g., syndrome, day of week, season). We adopted the approach of evaluating sensitivity at a fi xed 1% alert rate defi ned empirically for each algorithm and dataset, as used by Jackson et al. (12) . Our approach is in accord with a recent review that recommended basing alert thresholds on empirical data rather than on classical statistical theory (17) . A major strength of the study is that BioSense is a national system that provided access to 2 major datasets with differing characteristics and to data from hundreds of facilities in many states. The length, geographic spread, and syndrome variation of the study datasets lend weight to the results. The fi eld of electronic biosurveillance is in its infancy and is rapidly changing. Early work focused on attempts to detect outbreaks (early event detection) by using broadly defi ned syndromes (e.g., respiratory syndrome) based on chief complaints and diagnoses. Emphasis has recently shifted to monitoring for ongoing outbreaks (situational awareness) and for specifi c disease indicators (e.g., cough, dyspnea) called subsyndromes. The fi eld is now beginning to develop methods for case-based surveillance (i.e., automated application of a formal case defi nition using computerized data) (18) . Each data type and disease indicator may have unique characteristics that require modifi cations of standard data analysis methods. However, because the adaptation of time-series methods to recognize outbreaks will be an ongoing need, the enhanced methods identifi ed by this study are likely to have lasting usefulness. Figure 3 . Comparison of initial versus best algorithms for analysis of fever syndrome data at an example emergency department, October-November 2006. A) SD comparison. Count, fever syndrome counts; SD (initial), SD by using initial algorithm (minimum SD = 0.2, 7-day baseline, count method, unstratifi ed baseline); SD (best), SD by using best algorithm (minimum SD = 1.0, 28-day baseline, rate method, unstratifi ed baseline). B) Count threshold comparison. Count, fever syndrome counts; threshold 1, minimum count needed to trigger an alert by using initial method; threshold 2, minimum count needed to trigger an alert by using best method (for the best algorithm, which accounts for rate, 8 counts were added to total visits for calculating the threshold). C) Detection of 8 additional counts. Count, daily fever syndrome counts; count + 8, daily count plus 8 counts; both methods, 30 days with the additional counts detected by both the initial and best methods; initial only, 2 days with the additional counts detected by using initial method only; and best only, 19 days with additional counts detected by using best method only.

Molecular Mechanisms of Recombination Restriction in the Envelope Gene of the Human Immunodeficiency Virus

The ability of pathogens to escape the host's immune response is crucial for the establishment of persistent infections and can influence virulence. Recombination has been observed to contribute to this process by generating novel genetic variants. Although distinctive recombination patterns have been described in many viral pathogens, little is known about the influence of biases in the recombination process itself relative to selective forces acting on newly formed recombinants. Understanding these influences is important for determining how recombination contributes to pathogen genome and proteome evolution. Most previous research on recombination-driven protein evolution has focused on relatively simple proteins, usually in the context of directed evolution experiments. Here, we study recombination in the envelope gene of HIV-1 between primary isolates belonging to subtypes that recombine naturally in the HIV/AIDS pandemic. By characterizing the early steps in the generation of recombinants, we provide novel insights into the evolutionary forces that shape recombination patterns within viral populations. Specifically, we show that the combined effects of mechanistic processes that determine the locations of recombination breakpoints across the HIV-1 envelope gene, and purifying selection acting against dysfunctional recombinants, can explain almost the entire distribution of breakpoints found within this gene in nature. These constraints account for the surprising paucity of recombination breakpoints found in infected individuals within this highly variable gene. Thus, the apparent randomness of HIV evolution via recombination may in fact be relatively more predictable than anticipated. In addition, the dominance of purifying selection in localized areas of the HIV genome defines regions where functional constraints on recombinants appear particularly strong, pointing to vulnerable aspects of HIV biology.

Pathogens, and viruses in particular, are subject to strong selective pressures during infection and often have characteristically high degrees of genetic variation [1] . Recombination is an important evolutionary mechanism that contributes to this genetic diversification. By creating novel combinations of pre-existing genetic polymorphisms in a single replication cycle, recombination enables greater movements through sequence space than can be achieved by individual point mutations. As a consequence, recombination provides access to evolutionary ''shortcuts''. In addition, since recombination generally involves genes that already encode functional products, the probability of producing viable progeny is higher compared to the insertion of an equivalent number of random point mutations [2] . However, the generation of recombinant forms is not an unconstrained process. Genes and genomes generally evolve through the slow accumulation of point mutations, which often requires the progressive insertion of compensatory mutations at ''linked'' sites. This coevolution permits the preservation of epistatic interactions. By simultaneously introducing several substitutions, recombination has the potential to substantially perturb such coevolved intra-genome interaction networks [2, 3] , impairing the functionality of the genes involved. Thus, the balance between the advantages of taking evolutionary shortcuts and the risk of chimeras being dysfunctional [2] determines the role played by recombination in the evolution of a given gene or organism. Several studies have focused on the impact of recombination on the evolution of proteins, particularly in relation to directed evolution experiments [4, 5] . Two major factors have a large influence on the functionality of recombinants proteins. The first is the position of recombination breakpoint (the region where the sequence shifts from that of one parental sequence to the other) relative to the location of genetic polymorphisms within the gene. Recombinants involving a large number of non-synonymous substitutions will in fact have a low probability of being functional [2] . The second factor is the position of the breakpoints in relation to the boundaries of discrete protein folds. Breakpoints near the boundaries of these domains will in general have a smaller impact on protein folding, and hence protein function, than breakpoints occurring within them [3, 6, 7] . Recent work on Begomoviruses corroborated these findings by demonstrating that recombination events found in natural viral populations are significantly less disruptive of protein folding than randomly generated recombinants [8] . Adaptation of pathogens, either to on-going immune pressures within individual hosts or following transmission to new hosts of the same or different species, can result in infectious outbreaks that constitute major threats for public health [9] [10] [11] [12] . The human immunodeficiency virus (HIV) is an extremely recombinogenic pathogen in which recombination has been implicated in key aspects of viral pathogenesis such as immune evasion [13] , transmissibility [14] , the evolution of antiretroviral resistance [15, 16] and cross-species transmission [9, 12] . Indeed, the remarkable genetic flexibility of HIV is underlined by its large genetic diversity. The HIV-1 population is subdivided into three groups, named M, N and O, with group M (which is responsible for the vast majority of the infections worldwide) being further subdivided into nine subtypes (named A, B, C, D, F, G, H, J and K) [17] . Although recombination in HIV has been shown to occur at all phylogenetic levels (intra-and inter-subtype, as well as inter-group, reviewed in reference [18] ), the most widely noted impact of recombination on the genetic diversification of this virus is the frequent natural occurrence of inter-subtype recombinants in parts of the world where multiple subtypes co-circulate [19] [20] [21] [22] . When the same inter-subtype recombinant is transmitted between multiple individuals, and has therefore the potential to be of epidemiological significance, it is termed a Circulating Recombinant Form (CRF) [17] . As with the HIV-1 subtypes, CRFs form distinct clusters in phylogenetic trees and some of them contribute substantially to the pandemic. Sufficient inter-subtype recombinant sequences have been sampled to permit the detailed characterisation of variation in the locations of breakpoints both within individual genes [22, 23] , and entire genomes [24, 25, 32] . This makes HIV a particularly useful model for studying the forces that shape pathogen populations within the context of global epidemics. Here we focus on recombination within the envelope gene (env). This gene encodes two polypeptides (gp120 and gp41) that form a heterodimer at the surface of the viral particle. Trimers of these heterodimers are the functional units that are responsible for binding to the cellular receptors and co-receptors and ultimately lead to viral entry into target cells [26] . The two protein products of env are also the targets of all the neutralising antibodies identified to date [27] . By using a tissue culture system to characterise inter-subtype recombinants generated within env in the absence of selection, and assaying the functionality of recombinant genes, we produce an empirical model of HIV recombination that accurately describes recombination patterns found in viruses sampled throughout the HIV pandemic. We used different combinations of env sequences from primary HIV-1 isolates belonging to either different group M subtypes or group O (see Materials and Methods for the list of parental isolates used) to determine the distribution of breakpoints occurring within the HIV env gene in the absence of selection. We chose combinations of isolates belonging to subtypes that are cocirculating in regions of the world from which natural intersubtype recombinant forms have emerged [28] . In order to quantify variations in recombination rates across env we used a previously described experimental system where human T cells are transduced with HIV-1 replication-defective vectors pseudotyped with the Vesicular Stomatitis Virus (VSV) envelope [29] . As this system mimics a single cycle of viral infection in which reverse transcription products neither influence cellular survival, nor confer a specific phenotype to the transduced cells, recombinants that were produced during reverse transcription were not subjected to any selection. After cloning of the reverse transcription products in E. coli, the system enabled identification of the recombinants based on the presence of a lacZ reporter gene ( Figure 1 ). Given that known input sequences were used, such an approach enables the accurate and unambiguous localization of the breakpoint position to precise regions bounded by nucleotides that differ between the two parental sequences. The regions of the envelope gene that were studied were chosen so as to obtain 700 to 1,500 nucleotides overlapping windows, spanning the whole of env. For each of seventeen different combinations of parental sequence pairs (Figure 2A) , a recombination rate per nucleotide and per reverse transcription run was calculated within a 50 nucleotides sliding window (with 10 nucleotides step size). These were plotted as a function of the location of the window along the gene. To evaluate whether recombination-prone regions exist within the population, data from the 17 different pairs of parental sequences were pooled and an average recombination rate was computed for the different regions, and plotted as function of the position along the env gene ( Figure 2B , top panel). Peaks and troughs were apparent all along the gene, with regions refractory to recombination being more Recombination allows mixing portions of genomes of different origins, generating chimeric genes and genomes. With respect to the random generation of new mutations, it can lead to the simultaneous insertion of several substitutions, introducing more drastic changes in the genome. Furthermore, recombination is expected to yield a higher proportion of functional products since it combines variants that already exist in the population and that are therefore compatible with the survival of the organism. However, when recombination involves genetically distant strains, it can be constrained by the necessity to retain the functionality of the resulting products. In pathogens, which are subjected to strong selective pressures, recombination is particularly important, and several viruses, such as the human immunodeficiency virus (HIV), readily recombine. Here, we demonstrate the existence of preferential regions for recombination in the HIV-1 envelope gene when crossing sequences representative of strains observed to recombine in vivo. Furthermore, some recombinants give a decreased proportion of functional products. When considering these factors, one can retrace the history of most natural HIV recombinants. Recombination in HIV appears not so unpredictable, therefore, and the existence of recombinants that frequently generate nonfunctional products highlights previously unappreciated limits of the genetic flexibility of HIV. common in the gp120 coding portion than in the gp41 region. The probability that breakpoints were more or less clustered across env than could be accounted for by chance (given the null hypothesis that breakpoint positions occur randomly) was determined by a permutation test ( Figure 2B , bottom panel). Six major recombination-prone or ''hot'' regions (shaded light blue areas in Figure 2B ) could be defined as env regions where breakpoint clusters were bounded by statistically significant breakpoint ''cold spot'' (p,0.05). Each of the six identified breakpoint clusters contained at least one breakpoint cluster that constituted a statistically significant recombination ''hot-spot'' (p,0.01). While these recombination-prone regions covered only slightly more than half of the whole gene (55.3%), they included 81.6% of all the breakpoints (337/413) mapped. These six hot regions are areas where recombination occurs preferentially during HIV replication, irrespective of the parental strains involved. We next investigated the fate of these recombinants with respect to their establishment in the natural HIV-1 population. The fixation of a recombinant gene within a population is dependent on the interplay of multiple factors. Nevertheless, an obligatory component of evolution is undoubtedly the elimination by purifying selection of viruses that express dysfunctional proteins. To evaluate how profoundly this aspect of natural selection might influence the pattern of breakpoints generated by the mechanism of recombination, we determined the relative functionality of a subset of recombinant env genes. In addition to encoding the proteins that coat the viral membrane, env also encodes a well-known functional RNA structure, the Rev responsive element (RRE). For the recombinants containing breakpoints in the RRE region the functionality of this RNA module was therefore also tested. Being involved in the regulation of the timing and the balance among the various forms of unspliced and partially or completely spliced RNAs, RRE is essential for viral replication [30] . Failure to properly regulate this process results in either a decrease or complete halt in viral production [30] . The functionality of chimaeric RREs was tested by measuring viral titres obtained upon transfection of cells with a plasmid containing the proviral sequence of the molecular clone NL4.3 of HIV-1, in which we had replaced the native NL4.3 RRE with that of the various chimaeric RREs. To uncouple the effects on RNA-folding caused by the introduced RRE sequences from those altering the amino acid sequence of expressed proteins, we used a variant of NL4.3 that does not express Env (NL4.3-Env 2 ) [31] , and a plasmid encoding the wild-type Env was co-transfected to complement the production of gp120 and gp41 proteins. In order to increase the statistical power of the analysis, additional chimaeric RREs were constructed using parental sequences other than those employed in our cell culture recombinant generation experiments (following a PCR procedure described in Materials and Methods) and tested for their functionality. As can be seen in Table 1 , the viral titres obtained with every chimaeric RRE sequence we tested were both similar to those obtained with nonrecombinant parental RRE sequences and markedly higher than that observed when the RRE was replaced with a non-viral sequence (see Materials and Methods). This result therefore clearly indicated that recombinants generated by breakpoints within the RRE generally retain the functionality of this element. To determine the functionality of individual recombinant envelopes at the protein level, full-length recombinant envelope genes containing breakpoints of interest were constructed by successive PCR, as described in Material and Methods. Each fulllength recombinant gene was then cloned in the pcDNA3.1 expression vector, and used to transfect HEK 293T cells together with the pNL4.3-Env 2 -Luc plasmid, to generate viral particles pseudotyped with the recombinant envelope of interest. The functionality of the recombinant envelopes was then tested after transduction of HEK 293T-CD4 + -CCR5 + cells at a multiplicity of infection of 0.1, by measuring luciferase expression in these cells 48 hours after transduction. Since target cells cannot synthesize new viral envelope proteins, infection was limited to reverse transcription and, potentially, integration. The luciferase values observed therefore reflected the relative success of viral entry into the target cells. For this analysis recombinants derived from parental env sequences that yielded the strongest positive signals in this single cycle test were chosen (parental sequences A-Q461, C-CAP210, G-1033 and O-32, see Table 2 for their relative genetic distance) due to the higher reliability of the luciferase signal. The parental env sequences were used as controls. As for the functional analysis of the RRE, additional recombinants involving combinations of parental sequences -other than those involved in the experiments of recombination in cell culture, but carrying breakpoints in the same regions -were also tested. These additional recombinant env sequences were generated by PCR, as described for the reconstitution of the full-length env gene. Luciferase values determined for each recombinant were plotted as a function of the corresponding breakpoint position ( Figure 3 ). Recombinants with breakpoints falling within the six hot regions indicated in Figure 2B were preferentially characterized. It was apparent that most of the severely defective recombinants contained breakpoints in hot regions 2 and 3 of the recombination rate distribution (Figure 3 ). Given this data, we approximated a probability of Env functionality being disrupted by breakpoints falling within each of the six high recombination-rate regions. Since the parental sequences themselves were not uniformly functional ( Figure 3 ), a situation that is probably common in nature, for each recombinant an estimate of loss of functionality was calculated by dividing the luciferase value obtained with that recombinant by the one of the least functional parental sequence involved in its generation. Recombinants displaying values between those of the two parental sequences were considered to white-grey hatched box: partially deleted and therefore non-functional U3 sequence; white boxes: HIV envelope sequences studied, env X and env Y stand for sequences of two different isolates (isolate X on the donor RNA and isolate Y on the acceptor, respectively); black box: marker lacZ' bacterial gene; grey box: partial sequence of the bacterial gene malT, inserted in the reverse orientation. The approximate location of the BamHI site (present only on the donor RNA) and of the PstI site present on both RNAs is also indicated. The path followed by reverse transcription for generating the BamHI + /LacZ' + recombinants studied in the present work is indicated schematically. doi:10.1371/journal.ppat.1000418.g001 retain functionality (and assigned a functionality value of 1). Of note, none of the recombinants yielded functionality values higher than that of the most functional parent from which it was generated. Values from recombinants containing breakpoints within the same region of the six hot regions were pooled, and a functionality loss value for each region was averaged ( Figure 3 ). The most significant loss of functionality was observed in regions 2, 3, and 6. Natural recombination breakpoint distributions essentially mirror those of the functional recombinants generated in tissue culture Having defined a pattern of recombination in the absence of selection and the approximate probabilities of recombination events in various parts of env yielding fully functional products, we were interested in determining whether our experimental data could explain breakpoint patterns observed in circulating recombinants. The distribution along the whole HIV genome of 691 recombination breakpoints within HIV-1 group M full genome sequences from the LANL HIV Sequence Databases (http://hiv-web.lanl. gov/) was inferred. The same approach used in Figure 2B to define the probability that at any region of the genome the breakpoints were more clustered than would be expected by chance was used, with a 200 nucleotides window. A previous analysis of HIV recombinants modelled the distribution of breakpoints and indicated a significant clustering of breakpoints in the 59 and 39 ends of the envelope gene and a lack of breakpoints between these regions [32] . Our new analysis (Figure 4 ) confirmed the propensity for breakpoints to be located at the 59 and 39 ends of the env gene and the lack of breakpoints in the majority of its internal regions in recombinants from the database. In order to compare our experimentally determined breakpoint distribution to that found in recombinants from the HIV Sequence database, a higher-resolution view of the breakpoint distribution within the env gene was determined using the positions of 133 unambiguously unique recombination breakpoints detectable within 230 env sequences. Following the same procedure described above, but using a 50 nucleotides window to enable detection of breakpoint clusters at the same resolution as in our experimental system, we identified a series of recombination hot-and coldregions within the gene ( Figure 5A , purple graph). In a similar way to the breakpoint distribution detected in cell culture, various hot regions could be defined (light-purple boxes at the bottom of Figure 5A ), which corresponded remarkably well to recombination hot regions 1, 5 and 6 seen in cell culture (light-blue boxes). Whereas the other hot regions identified in cell culture had no corresponding counterparts in the natural breakpoint distribution, there was close correspondence between the cold-spots detected in both distributions. Next we used the SCHEMA-based method [8] to investigate whether or not this breakpoint distribution exhibits evidence of purifying selection acting on recombinants with disrupted protein folding. This analysis indicated that breakpoints observable in natural viruses tend to occur in regions within env that were predicted to have a significantly lower impact on protein folding than randomly placed breakpoints (p,1.0610 24 for gp120 and p = 8.9610 23 for gp41, see Protocol S1). To investigate whether accounting for variations in the functionality of recombinants might reconcile the natural and experimental breakpoint distributions, we first approximated the combined effects of mechanistic recombination rate variation ( Figure 2B ) and selection for fully functional recombinants ( Figure 3 ) on the distribution of breakpoints in cell culture. Selection ''corrected'' recombination rate estimates were then used to determine the distribution of 133 expected breakpoints. The resulting distribution was used to evaluate the probability of clustering of breakpoints (green graph in Figure 5A ). Only regions 1, 4, 5 and 6 remained areas of significant clustering (light-green boxes at the bottom of Figure 5A ), a pattern very close to that found in HIV recombinants sampled from nature, with the exception of region 4 for which there was substantially less evidence of recombination within natural recombinants than was expected based on our empirical model. Indeed, when compared to the Recombinants 115A/126D, 115A/89D, 120A/89D, and 126D/120A were described previously [33] . The number (n) of individual recombinants for which the position of the breakpoint has been mapped is given on the right of each graph. A map of gp120 and gp41 domains is given as a frame of reference at the top of the figure. (B) top graph: pooled distribution of recombination breakpoints across env, obtained as described in Materials and Methods. Bottom graph: the height of the black plot at any particular position represents the probability (determined by a permutation test with 10,000 iterations) that recombination breakpoint distributions are not more clustered than would be expected by chance within a 50 nucleotides window centred on that position. Assuming that breakpoints are randomly distributed, the dark and light grey regions represent degrees of breakpoint clustering expected due to chance in 95% and 99% of the examined windows, respectively. Whereas peaks emerging above the grey regions represent possible recombination hot-spots, troughs dipping below the grey regions represent possible recombination cold-spots. The paleblue shaded areas numbered from 1 to 6 correspond to breakpoint clusters, or hot regions, as defined in the text. doi:10.1371/journal.ppat.1000418.g002 distribution found for the 133 breakpoints encountered in the natural HIV recombinants ( Figure 5B ), a remarkable overlap was observed, with the discrete statistically significant breakpoint clusters being consistently recaptured by our empirical model of env recombination. The substantial difference of recombination rates in region 4 was also clear. Through the functional characterization of HIV envelope genes generated by recombination in the absence of selection, we retrace the early steps shaping patterns of inter-subtype env recombination found in the HIV-1 pandemic. We observe that the mechanism of recombination alone defines regions where recombination occurs at significantly higher rates than elsewhere along the gene. The existence of such regions is strongly suggestive of spatially conserved features in HIV genomes that either promote or restrict recombination between different isolates. The distribution of breakpoints within the gp120 encoding region ( Figure 2B ) is likely due to the distribution of conserved and variable regions, the latter restricting recombination because of the low degree of local sequence identity between the parental sequences [32, 33] . Within genomic regions where sequence identity is high, a trigger for recombination could be the presence of secondary The luciferase values that were determined for the four parental strains used to generate the recombinants are represented, as a frame of reference, by lilac bars on the right. The six recombination-prone regions defined in Figure 2B are shaded in pale blue and annotated accordingly above the graph. The value of loss of functionality approximated for each region is given in bold in the top part of the graph. Three M/O inter-type recombinants were also tested (AO456, AO7810, and AO8090, breakpoint positions 6508, 7810, and 8090, respectively, of the HXB2 reference strain), resulting in a complete loss of functionality, probably due to the higher sequence divergence between the parental isolates. In order to preserve the homogeneity of the dataset to intersubtype recombinants, these recombinants are neither presented in the figure, nor were considered for the calculation of the average functionality of the recombinants, described in the main text. doi:10.1371/journal.ppat.1000418.g003 structures [34] . The highest recombination peak within the second region in Figure 2B (corresponding with the C2 portion of gp120) coincides with a recombination hot-spot that is determined by the presence of a stable RNA hairpin structure [29, 35, 36] , while the fourth hot region ( Figure 2B ) corresponds to the RRE RNA structure that is highly conserved amongst all HIV isolates [37] . It is therefore possible that RNA secondary structures also contribute to the high rates of recombination observed at some of the other recombination hot regions. Noteworthy, the functionality of the RRE was retained even when crossing genetically distant isolates as for inter-group M/O recombinants (Figure 2A) , supporting the possibility that regions of the genome harbouring functional RNA structures, which are generally more conserved within the population, provide a mechanism for crossing distantly related retroviruses and are possibly important for recombination of RNA viruses in general. With respect to selection of recombinant genes at the protein level, experiments involving lattice proteins have shown that genes encoding proteins that tolerate mutations also tend to be recombination tolerant [2] . Since the env gene displays a degree of diversity between isolates from different HIV-1 group M subtypes ( [38] and references herein) that is two to three times higher than the genome average, we anticipated that the manifest mutation tolerance of env might predispose it to high recombination tolerance. However, we show that this is not the case with certain regions within the gp120 encoding portions of env (particularly region 2 described in the present work in Figure 3 ) tending not to tolerate recombination well. Viruses with small genomes (including all RNA viruses) tend to use overlapping genes expressed in different reading frames and to encode proteins that have multiple functions. The HIV envelope encodes for such proteins [26] , and the subtle biochemical equilibrium that regulates their functionality is very possibly limiting tolerance to recombination. The low recombination tolerance of the gp120-encoding region could only be imprecisely predicted based on computational estimates of recombination induced protein fold disruption using the SCHEMA algorithm [3] . This may have been due to either our SCHEMA analyses being based on incomplete gp41 and gp120 structures or the fact that the structures used only reflected a single conformation of these two proteins. Therefore this analysis neither takes into account the conformational changes required for Env functionality, nor the quaternary arrangement of the proteins within Env trimers. Despite these issues, the SCHEMA analysis indicated that, amongst the HIV env sequences sampled from nature, selection has been acting against recombinants with disrupted protein folding (Table S1 ). Unravelling the molecular reasons for the reduced functionality of certain recombinants could provide valuable insights into the nature of the molecular interaction networks required for proper Env function. The specific determinants of viral fitness (or in vivo replicative capacity) are complex and poorly understood at present. The fixation of a recombinant gene within a population is likely to depend on the interplay of multiple factors. Although combining cell culture functionality data with recombination rate heterogeneity is an oversimplified view of this process, the pattern of recombination predicted by our empirical model matches remarkably well the breakpoint distributions observed in nature ( Figure 5B ). The only major deviation from this was constituted by the fourth recombination hot region observed in cell culture, which was absent from the natural breakpoint distribution ( Figures 2B and 5B ). Determining the reasons for this discrepancy will improve our understanding of the mechanisms governing the success of recombinants in nature. Although the host immune response certainly plays a significant role in the selection of recombinant variants in vivo [13] , the similarities between the natural and experimental breakpoint distributions suggest that the forces responsible for the selection of recombinants in vivo only have limited impact on inter-subtype breakpoint patterns in env. This is most likely due to a combination of factors including mainly the complex epistatic interactions within env, the high density of fitness-determining loci within this gene, and the biochemical mechanism of recombination, which collectively constrain the fixation of genetic variability introduced by recombination. Negative fitness effects associated with recombination in env, however, should decrease with decreasing parental genetic distances [3, 6, 39] and therefore, in the context of intra-subtype recombination, the selective constraints on recombinants should be more relaxed than we have found them to be here. Considering recombination in env in the context of the rest of the HIV genome, it is apparent that env displays the most dramatically variable natural breakpoint distribution of all HIV genes [24, 32] , and it constitutes the only gene within which there is an extended region with limited recombination (Figure 4) . Nevertheless, although less marked, breakpoint distribution patterns reminiscent of those found in env, with alternate clusters and troughs are also identifiable in several other regions of the genome such as gag and pol [32] (Figure 4) . Although little information is presently available either on differential mechanistic predispositions to recombination across these regions, or on the functionality of the resulting products, it is tempting to speculate that underlying rules such as we have defined here for env may also be operational in these other cases. In conclusion, by experimentally reproducing the generation of HIV-1 recombinants, we demonstrate that the distinctive distribution of breakpoints found in natural viruses is strongly shaped by both the mechanism of recombination, and the relative functionality of the recombinant genes. Thus, HIV evolution might not be the relentlessly unpredictable process it sometimes seems, and exploiting this evidence to pre-empt and counter the most favoured evolutionary tactics of this virus may ultimately be an efficient means by which we can devise effective vaccines and improve drugs against the virus. Cell culture HEK 293T, and CD4 + CCR5 + 293T cells were grown in Dulbecco's modified Eagle's medium supplemented with 10% foetal calf serum, penicillin, and streptomycin (from Invitrogen, CA, USA), and maintained at 37uC with 10% CO 2 . MT4 cells were maintained in RPMI 1640 medium supplemented with 10% foetal calf serum and antibiotics at 37uC with 5% CO 2 . The parental isolates used in this study were A-115, A-120, A-899 [33] Single cycle recombination assays were performed using a system previously developed by our laboratory [29] . HIV-1 env fragments from group M subtypes A, C, D and G, and from group O viral DNA were amplified by PCR from infected PBMCs obtained from patients and cloned in plasmids (called genomic plasmids), which differ for the genetic marker present downstream (in the sense of reverse transcription) of the sequence in which recombination is studied (Figure 1 ). All constructs were verified by sequencing. The trans-complementation plasmids, pCMV R8.2 [40] encoding HIV-1 Gag, Pol, and accessory proteins, and pHCMV-G [41] encoding the Vesicular Stomatitis Virus envelope protein were co-transfected into 293T cells with the two genomic plasmids to produce defective retrovirus particles which were then used to transduce MT4 cells as previously described [29] . The reverse transcription products were purified from the cytoplasmic fraction of transduced cells using the method described by Hirt [42] . The purified double stranded DNA was digested with DpnI for 2 h at 37uC (in order to eliminate possible contaminating DNA of bacterial origin) prior to PCR amplification as previously described [29] . The amplified product was purified after electrophoresis on agarose gel, digested with PstI and BamHI, ligated to an appropriate plasmid vector and used to transform E. coli. Plating on IPTG/X-Gal containing agar plates allowed blue/ white screening of recombinant and parental colonies, respectively [29] . The frequency of recombination was determined by computing the number of blue colonies over the total number of colonies as described in reference [29] . Recombination breakpoints were identified by full-length sequencing of the env portion of the recombinant clones. The recombinant and parental sequences of each pair of isolates tested were aligned using CLUSTAL X [43] . The breakpoint location of each recombinant was determined as being the central position of the interval bounded by the two closest nucleotide sites that were characteristic of each of the parental sequences). Recombination rates were calculated as follows. We define each recombination window studied with each pair of parental sequences as RwXY a-b , for a recombination window involving isolates X and Y, spanning position a to position b of env (reference sequence HXB2); a 50 nucleotides window was then considered (XY a-b Sw i , for a sliding window starting at position i of env), beginning from the 59 border of the sequence studied and the number of breakpoints (indicated as XY a-b n i ) falling within the window was counted. The resulting recombination rate per nucleotide in the sliding window XY a-b Sw i is where XY a-b N is the total number of breakpoints characterized for the RWXY a-b pair, and 50 is the size in nucleotides of the sliding window, and F the frequency of recombination observed in the whole region studied, as defined in the previous chapter. The sliding window was then displaced with a 10 nucleotides increment (resulting in XY a-b Sw i+10 , XY a-b Sw i+20 , … ) across the recombination window, and XY a-b R i+10 , XY a-b R i+20 , … were computed. The various R values were reported in the graph as a function of the position of the midpoint of the window along the gene (i.e. the position of the 25 th nucleotide of each sliding window). For the pooled dataset reported in Figure 2B , the analysis based on the sliding window was repeated. If Swp i stands for the sliding window at position i for the pooled dataset, Rp i for the corresponding recombination rate, and q is the number of recombination window including position i, recombination rate at position i is calculated as To statistically test for the presence of recombination hot and coldspots in the experimentally determined recombination breakpoint distributions we used a modification of a permutation test described previously [44] . Unlike in analyses of natural recombinants, the breakpoint positions approximated in our experimental procedure were not subject to biases introduced by underlying degrees of parental sequence nucleotide variability and patchiness of parental sequence sampling. Rather than explicitly accounting for these biases when placing randomised recombination breakpoints as in the permutation test described by Heath et al. [44] , our modification of the test involved the completely randomised placement of recombination breakpoints. The test essentially involved the randomised recreation of 10,000 versions of our real dataset with each version containing exactly the same number of breakpoints between the same 17 parental sequence pairs observed in the real dataset. From breakpoint distributions determined for each of these 10,000 randomised datasets we were able to work out confidence intervals for expected breakpoint density variation given the completely random occurrence of recombination. For simulating the distribution of 133 breakpoints based on the combined effects of (i) the mechanistic recombination rate and (ii) selection for functional recombinants, local recombination rate data used to generate the graph in Figure 2B were first multiplied by the respective functionality scores given in Figure 3 for each corresponding region, yielding ''functionality corrected'' rates for each region. Once the expected breakpoint distribution of 133 unique recombinants determined by this method, the number of breakpoints present in a 50 nucleotides rolling window, sliding with a 10 nucleotides increment was calculated and plotted (in Figure 5B ) as function of the position along the gene. Deviations from expected degrees of breakpoint clustering given the null hypothesis of random breakpoint locations, was tested using the same modification of the Heath et al., [44] permutation test detailed above. Full-length sequences of recombinant env genes were reconstituted, using an overlapping PCR procedure. We separately amplified the region from the 59 end of the acceptor gene (using primer Topo59 annealing to positions 5966-5990 of the reference strain HXB2) to the breakpoint position (using a specific primer encompassing the region of the breakpoint) and from the 39 end of the donor gene (primer Donor39, HXB2 positions 8785-8819) to the breakpoint position (also in this case with a specific primer). These PCR products, overlapping by approximately 30 nucleotides around the breakpoint site, were mixed at equal ratios and used as templates to generate the full-length recombinant env gene using primer Topo59 and Donor39. All PCR reactions were run with Phusion DNA polymerase (Finnzymes, Finland) for 30 cycles. PCR products were gel purified and ligated to pCDNA3.1 Topo (Invitrogen, CA, USA). For RRE functionality assays, a portion of the envelope gene containing the RRE of pNL4.3-Env 2 -Luc (nucleotides 7646 to 8046) was replaced with the corresponding sequence of parental or recombinant envelope genes or, as a negative control, a 400 nt sequence from the Drosophila melanogaster desoxynucleoside kinase gene (DdNK). All constructs were verified by sequencing. HIV particles were produced by co-transfection of HEK 293T cells with an expression vector for a CCR5-tropic (ADA) HIV-1 envelope [45] kind gift of Dr. M. Alizon, together with a pNL4.3-Env 2 -Luc containing either a parental or recombinant RRE sequence or DdNK. Forty-eight hours post transfection, supernatants were filtered trough a 0.45 mM filter and p24 levels were determined using the HIV-1 p24 enzyme-linked immunoabsorbent assay kit (PerkinElmer Life Sciences, MA, USA). Reporter HIV-1 particles were produced by co-transfection of HEK 293T cells with pNL4.3-Env 2 -Luc and either an empty expression vector or an expression vector encoding either a parental or a recombinant env. For each individual recombinant variant, prior to their use for transfection, clones were verified by sequencing of the region encoding the recombinant gene as well as the vector-encoded promoter for its expression. Supernatants, containing virus stock, were harvested 48 h post transfection, and filtered trough a 0.45 mM filter. Production of viral particles was tested using an enzyme linked immunoassay for HIV-p24 antigen detection (Perkin Elmer, MA, USA) and 20 ng of p24 were used to infect 10 5 293T CD4 + -CCR5 + cells in 24 wells plates. Forty-eight hours later, cells were washed twice in PBS and lysed in 25 mM Tris phosphate, pH 7.8, 8 mM MgCl2, 1 mM dithiothreitol, 1% triton X-100, and 7% glycerol for 10 min in a shaker at room temperature. The lysates were centrifuged and the supernatant was used to measure luciferase activity using a GloMax 96 Microplate Luminometer (Promega, WI, USA) following the instruction of the luciferase assay kit (Promega, WI, USA). For samples that yielded negative results in the luciferase assay, plasmids from at least three independent bacterial clones were tested. The HIV-1 group M envelope sequence alignment was retrieved from the Los Alamos National Laboratory (LANL) HIV Sequence Database (http://hiv-web.lanl.gov/). The alignment was reduced to subtype reference sequences (3 strains for each where available), 53 CRF strains (2 strains for each where available) and finally 197 apparently unique recombinants. Recombination was analyzed using the RDP [46] , GENECONV [47] , BOOTSCAN [48] , MAXCHI [49] , CHIMAERA [50] , SISCAN [51] , and 3SEQ [52] methods implemented in the program RDP3BETA30 [53] . Default settings were used throughout except that: (1) only potential recombination events detected by four or more of the above methods, coupled with phylogenetic evidence of recombination were considered significant; (2) sequences were treated as linear; and (3) a window size of 30 variable nucleotide positions was used for the RDP method. Using the approach outlined in the RDP3 program manual (http:// darwin.uvigo.es/rdp/rdp.html), the approximate breakpoint positions and recombinant sequence(s) inferred for every potential recombination event, were manually checked and adjusted where necessary using the phylogenetic and recombination signal analysis features available in RDP3. Breakpoint positions were classified as unknown if they were (1) detected at the 59 and 39 ends of the alignment but could have actually fallen outside the analysed region; or (2) within 20 variable nucleotide positions or 100 total nucleotides of another detected breakpoint within the same sequence (in such cases it could not be discounted that the actual breakpoint might not have simply been lost due to a more recent recombination event). All of the remaining breakpoint positions were manually checked and adjusted when necessary using mainly the MAXCHI and 3SEQ methods (using three sequence scans and the MAXCHI matrix method) but also the LARD matrix method (generated by the LARD two breakpoint scan; [54] ), and the CHIMAERA method as tie breakers. The distribution of unambiguously detected breakpoint positions of all unique recombination events were analysed for evidence of recombination hot-and cold-spots with RDP3 as described by Heath et al. ([44] ; a window size of either 50 or 200 nucleotides and 10 000 permutations). A normalised version of the breakpoint distribution plot described in that study was used in which the local probability values of breakpoint numbers (determined by a permutation test that takes into account that local degrees of sequence diversity influence the delectability of recombination events) were plotted instead of absolute breakpoint numbers. PDB files detailing the three dimensional structures of both gp120 (PDB ID: 2B4C, determined by X-ray diffraction, resolution of 3.3 Å , 338 amino acids, [55] ), and gp41 (PDB ID 1AIK, determined by X-ray diffraction, resolution of 2 Å , 70 amino acids, [56] ) were obtained from http://www.rcsb.org. It is important to point out that these structures are partial and that we therefore only analysed a fraction of the structural interactions involved in Env folding. We performed SCHEMA predictions of recombination induced fold disruptions using the set of natural HIV env recombinants (described above) essentially as described in Lefeuvre et al. ([8] ; See Protocol S1, Supplementary Analyses, for a description of the SCHEMA method). This involved: (1) computing protein fold disruption, or E, scores for each natural recombinant with identifiable parents; (2) based on every pair of parental sequences identified for the observed set of recombinants, simulating every possible recombinant that could have been produced with these parental sequence pairs that involved the exchange of the same number of non-synonymous polymorphisms as were exchanged during the actual recombination events; (3) calculating E scores for each of these simulated recombinants; and (4) using a permutation test to determine whether mean E scores calculated for the natural recombinants were significantly lower than mean E-scores for the same set of recombinants randomly drawn from the simulated recombinant datasets (Table S1 ). If fewer than 5% of simulated datasets had an average E score lower than that of the actual dataset (p,0.05) then this was taken to indicate that predicted fold disruptions incurred by real events were significantly less severe than if the observed distribution of breakpoints was uninfluenced by negative selection acting against recombinants with disrupted protein folding. Protocol S1 Supplementary analyses. Schema analysis on the HIV envelope gene. Found at: doi:10.1371/journal.ppat.1000418.s001 (0.02 MB DOC)

Avian Influenza Virus Glycoproteins Restrict Virus Replication and Spread through Human Airway Epithelium at Temperatures of the Proximal Airways

Transmission of avian influenza viruses from bird to human is a rare event even though avian influenza viruses infect the ciliated epithelium of human airways in vitro and ex vivo. Using an in vitro model of human ciliated airway epithelium (HAE), we demonstrate that while human and avian influenza viruses efficiently infect at temperatures of the human distal airways (37°C), avian, but not human, influenza viruses are restricted for infection at the cooler temperatures of the human proximal airways (32°C). These data support the hypothesis that avian influenza viruses, ordinarily adapted to the temperature of the avian enteric tract (40°C), rarely infect humans, in part due to differences in host airway regional temperatures. Previously, a critical residue at position 627 in the avian influenza virus polymerase subunit, PB2, was identified as conferring temperature-dependency in mammalian cells. Here, we use reverse genetics to show that avianization of residue 627 attenuates a human virus, but does not account for the different infection between 32°C and 37°C. To determine the mechanism of temperature restriction of avian influenza viruses in HAE at 32°C, we generated recombinant human influenza viruses in either the A/Victoria/3/75 (H3N2) or A/PR/8/34 (H1N1) genetic background that contained avian or avian-like glycoproteins. Two of these viruses, A/Victoria/3/75 with L226Q and S228G mutations in hemagglutinin (HA) and neuraminidase (NA) from A/Chick/Italy/1347/99 and A/PR/8/34 containing the H7 and N1 from A/Chick/Italy/1347/99, exhibited temperature restriction approaching that of wholly avian influenza viruses. These data suggest that influenza viruses bearing avian or avian-like surface glycoproteins have a reduced capacity to establish productive infection at the temperature of the human proximal airways. This temperature restriction may limit zoonotic transmission of avian influenza viruses and suggests that adaptation of avian influenza viruses to efficient infection at 32°C may represent a critical evolutionary step enabling human-to-human transmission.

Influenza viruses circulating in the human population are predominately type A and B, with type A being more common [1] . All influenza type A viruses originate from aquatic birds and successful introduction of these avian viruses into the human population, by either direct adaptation or reassortment with already circulating human viruses, has led to influenza pandemics of historical significance (reviewed in [2] [3] [4] 5] ). Still, documented evidence of transmission of avian influenza viruses directly from birds to humans is rare, partly because species barriers restrict avian influenza virus infection of the epithelial cells of the human respiratory tract, the primary site of influenza virus infection and spread. Influenza A viruses possess a hemagglutinin (HA) attachment protein that binds sialic acid residues to facilitate infection of target epithelial cells. The HA of human influenza viruses preferentially binds to terminal sialic acid (SA) residues with a2,6 linkages, whereas avian influenza viruses preferentially bind to SA with a2,3 linkages [6] [7] [8] [9] . The prevalence of a2,6 SA but paucity of a2,3 SA in the human respiratory tract has been considered to restrict infection by avian influenza viruses [10] . Recent reports, however, have detected significant levels of a2,3 SA on human airway epithelium both in vitro and ex vivo, including in nasopharyngeal and tracheobronchial tissue [11] [12] [13] [14] . This SA distribution also correlated with avian influenza virus infection in vitro and ex vivo and raised the possibility that avian viruses could infect the upper airways in vivo. Therefore, although it is universally accepted that human-to-human transmission of avian influenza viruses requires adaptation of HA to switch from a2,3 to a2,6 SA usage, the cumulative data published to date indicate that SA linkages and their respective distribution in the human airways are not the sole barrier to avian influenza virus infection [15] [16] [17] . Other host factors and viral genes are likely also important determinants of infectivity. One such host factor that may limit zoonotic transmission is the difference in host temperatures between avian and human tissues that are susceptible to influenza virus infection. Avian influenza viruses are adapted for replication in the avian enteric tract at 40-41uC . While the surface temperatures of the human respiratory tract are variable, a temperature gradient clearly exists in which the surface temperature of the proximal large airways (i.e., nasal and tracheal) average 32+/20.05uC while temperatures of the smaller, distal airways (i.e., bronchioles) are closer to that of the core body temperature, 37uC [18, 19] . While multiple transmission routes have been described for influenza viruses, the proximal airways likely represent a predominant site for human influenza virus inoculation as they provide a large exposed surface area of virus-susceptible epithelial cells [20] . These cells are directly accessible by large droplet aerosols and by way of digital inoculation of the nasopharynx and conjunctival mucosa [12, 21] . Inefficient infection by avian influenza viruses, even in the presence of a2,3-linked SA, may be due to the cooler temperature of the proximal airways compared to that of the distal airways/lung regions where H5N1 avian influenza viruses appear to replicate efficiently [22] . Avian influenza viruses are attenuated at temperatures below 37uC and cold sensitivity of avian viral RNA replication in cell lines was linked to the presence of a glutamic acid at amino acid 627 in the avian virus polymerase subunit, PB2, instead of a lysine in the human virus PB2 [23] . Lysine substitution at residue 627 of H5N1 viruses improved virus replication in mice [24] . In addition to PB2, work utilizing human-avian reassortant viruses in MDCK cells provided initial evidence that avian glycoproteins, HA and neuraminidase (NA), may mediate temperature-dependent effects on viral growth [25] . To our knowledge, other viral genes have not been well characterized, nor the HA and NA further evaluated, in their contribution to temperature sensitivity of avian influenza viruses. To characterize the temperature dependency of avian vs. human influenza viruses in a relevant model of the target cell types of the human airways, we utilized an in vitro model of human ciliated airway epithelium (HAE). This model closely mimics the morphological and physiological features of the human airway epithelium in vivo and has been previously used to investigate infection by diverse respiratory viruses [26] [27] [28] [29] [30] . In humans, ciliated airway epithelium is present throughout the airways, extending from the nasal cavity and large proximal airways into the distal bronchiolar airway regions. Previously, we have shown that both human and avian influenza viruses replicate well in HAE and that human and avian influenza virus cell tropism correlates with the respective distribution of the specific sialic acid linkages [13] . However, these previous studies were conducted at 37uC, reflecting conditions encountered in the distal airways [13] . Others have also utilized these airway cell systems to characterize influenza virus replication of wild-type and recombinant viruses at 35uC [14, 31, 32] . In the present study, we utilize the HAE model, in combination with influenza virus reverse genetics, to investigate the influence of temperature on human and avian influenza virus infection, replication and spread. We demonstrate that, compared to human influenza viruses, avian influenza viruses are severely restricted for infection of human airway epithelium at the temperature of the human proximal airways. Then, using different strategies to 'avianize' human influenza viruses, we show that the temperature restriction of avian viruses is closely associated with the avian HA and NA glycoproteins. We and others have previously shown that human and avian influenza viruses infect and replicate in HAE [13, 14, 31] . Since our previous experiments were performed at 37uC, a temperature reflective of human distal airways, we have now compared human and avian influenza virus infection and growth in HAE at temperatures reflective of the proximal airways (32-33uC) and distal airways (37uC). HAE were inoculated at either 32uC or 37uC with a low multiplicity of infection (MOI; 0.01) of a representative human virus, A/Victoria/3/75 (H3N2), or an avian influenza isolate, A/Dk/Eng/62 (H4N6). Virus growth and spread throughout the epithelium at the two temperatures was measured and compared over time and infection further characterized with respect to virus-induced cytopathic effects (CPE). At the temperature of the distal airways (37uC), the growth kinetics and mean peak titers of A/Victoria/3/75 and A/Dk/ Eng/62 reached 2.3610 8 pfu/ml and 4.7610 7 pfu/ml, respectively, by 48 hours post-inoculation (hrs pi) ( Figure 1A ). At the temperature of the proximal airways (32uC), A/Victoria/3/75 showed a modest delay in replication but still reached maximal titer of 7.8610 7 pfu/ml by 48 hrs pi. In contrast, A/Dk/Eng/62 grew very slowly, with yields at time points up to 48 hrs pi reduced by 3 to 5 logs compared to growth for this virus at 37uC or A/ Victoria/3/75 at either temperature. In comparison to 48 hr titers, A/Victoria/3/75 titers at both temperatures and A/Dk/Eng/62 titers at 37uC were reduced at 72 hr pi and every time point thereafter, indicating reduced progeny virus production. A loss of titer was also observed for A/ Dk/Eng/62 at 32uC, but not before 120 hrs pi. To determine if loss of titer after reaching maximum levels correlated with increased CPE, we quantified adenylate kinase (AK) release by dead/dying cells into the apical compartment as a sensitive and global measure of cytotoxicity across the entire epithelial cell Influenza type A viruses are endemic in aquatic birds but can cross the species barrier to infect the human respiratory tract. While transmission from birds to humans is rare, the introduction of novel avian influenza viruses into immunologically naïve human populations has significant pandemic potential. Avian influenza viruses are adapted for growth at 40uC, the temperature of the avian enteric tract. However, the human proximal airways, the likely site of initial inoculation by influenza viruses, are maintained at a cooler temperature (32uC), suggesting that zoonotic transmission may be limited by temperature differences between the two hosts. Using an in vitro model of human ciliated airway epithelium, we show that avian influenza viruses grow well at 37uC, a temperature reflective of distal airways, but are restricted for infection at 32uC. A panel of genetically manipulated human influenza viruses possessing avian or avian-like surface glycoproteins were also restricted at 32uC, but not 37uC, suggesting that avian virus glycoproteins are not adapted for efficient infection at the temperature of the proximal airways. Thus, avian influenza virus infection is restricted in the human proximal airways due to the cooler temperature of this region, thus limiting the likelihood of zoonotic and subsequent human-to-human transmission of these viruses. culture surface. Figure 1B indicates that substantial increases in AK levels, indicative of the onset of CPE, are first detected at 48 hrs pi for A/Victoria/3/75 at 32uC and 37uC and A/Dk/ Eng/62 at 37uC. This induction of AK coincided with peak viral titer for these viruses under these conditions (compare Figure 1A and 1B) and suggested that the loss of titer correlated with the onset of CPE. Increasing levels of AK between 48 and 96 hrs pi were directly associated with continually decreasing viral titers, further supporting this claim. A relationship between the kinetics of virus growth in HAE and the level of CPE also suggested that CPE was a consequence of viral replication. This assertion is supported by the fact that trends in viral titers at a given time point are mirrored in AK levels detected 48 hrs later (e.g., compare viral titers at 48 hr pi ( Figure 1A ) to AK measurements taken at 96 hr pi ( Figure 1B) ). Since viral titer and AK levels could be related to the numbers of cells infected and/or the degree of virus replication within individual cells we compared titers of human and avian influenza viruses ( Figure 1A ) to the numbers of cells infected by each virus at the two temperatures over time. Immunodetection of viral antigen in inoculated HAE showed that human and avian influenza virus antigen was not detected 3 hrs pi, indicating that levels of antigen in residual viral inocula were below the limit of antibody detection (data not shown). For A/Victoria/3/75, a few isolated cells were positive for viral antigen by 6 hrs pi at 37uC, but by 24 hrs pi considerable numbers of antigen-positive cells were detected ( Figure 2A ). In agreement with our growth curves in Figure 1A , A/Victoria/3/75 infected slightly fewer cells at 32uC compared to 37uC at 24 hrs pi, but importantly, A/Victoria/3/75 spread efficiently within the epithelium at both temperatures and differences in infection at early time points became less significant over time ( Figure 2A) . In contrast to A/Victoria/3/75, A/Dk/Eng/62 antigen was detected in only a few cells 24 hrs pi at either temperature. However, it should be noted that antigen-positive cells in en face images are viewed linearly (Figure 2A ) whereas viral titers are shown on a logarithmic scale ( Figure 1A) . Thus, an apparently small difference in titer as is seen at 24 hrs pi between A/Victoria/ 3/75 and A/Dk/Eng/62 at 37uC may correspond to a larger difference in the number of cells positive for viral antigen. While our staining also confirmed previous data that avian influenza viruses infect fewer human airway epithelial cells in comparison to human influenza virus at 37uC (Figure 2A ; [13] ), the limited extent of A/Dk/Eng/62 antigen positive cells at 37uC by 24 hr pi was still unexpected given that titers at this time were slightly greater than those for A/Victoria/3/75 at 32uC. Whether this represents a difference in yield of infectious virus per infected cell between human and avian viruses is presently not clear. Overall, A/Dk/ Eng/62 grew and spread well at 37uC, but was severely restricted for growth at 32uC and antigen positive cells were barely detectable before 48 hr pi for this virus at lower temperature. HAE cultures infected with A/Victoria/3/75 at either 32uC or 37uC and A/Dk/Eng/62 at 37uC viewed en face exhibited loss of integrity of the epithelium although the extent of injury and time of onset varied ( Figure 2A ). Further evaluation of histological cross-sections indicated that A/Victoria/3/75 infection at 37uC, which had the highest and earliest induction of AK, resulted in the earliest evidence of morphological injury at 72 hrs pi. HAE infected with A/Victoria/3/75 at 32uC or 37uC or A/Dk/Eng/62 at 37uC all showed desquamation of the superficial layer of columnar epithelial cells with basal epithelial cells remaining attached to the matrix support by 120 hrs pi ( Figure 2B ). Similar cytopathology has been reported for A/Udorn/307/72 influenza virus infection of HAE in vitro and for clinical human influenza virus infection in vivo [29, 33] . The detection of AK in apical washes of A/Dk/Eng/62-infected HAE at 32uC suggested that this virus did eventually compromise cellular integrity at the lower temperature, but dramatic morphological effects were not seen at least for up to 120 hrs ( Figure 1B and 2B). It should be noted, however, that at 120 hrs pi, A/Dk/Eng/62-infected HAE at 32uC did display some morphological characteristics different from uninfected and infected HAE at earlier time-points. Preliminary assessment indicates that expansion of lateral spaces between the columnar epithelial cells had occurred. Although we do not know the significance of these morphological changes, we speculate these observations are the initiation of CPE that will ultimately result in similar cellular injury as seen for this virus at 37uC and human viruses at both temperatures. In sum, for both viruses at both temperatures, detection of maximal numbers of antigen-positive cells correlated with high titers (compare Figure 1A and 2A) and increasing CPE ( Figure 1B ). By 72 and 120 hrs pi considerable loss of cells from the culture was evident and this correlated with the drop off in viral titers at these time points ( Figure 1A ). Thus, we conclude that in the context of maximal infection in which there were no additional target cells available for infection within the finite surface area of the HAE culture, ongoing replication in antigen-positive cells shown at 48 and 72 hrs pi resulted in increased cell death. This CPE led to a reduction in the number of viable, virus-producing cells and in turn, to a reduction in progeny virus. Although A/Dk/Eng/62 induced CPE when sufficient titers were generated at 37uC, one consequence of restricted replication of this avian influenza virus at 32uC was a reduction in overt CPE in HAE, even at later time points associated with considerable viral titers. To determine whether other avian, but not human, influenza viruses display temperature dependent phenotypes, we performed multi-step growth curves with more human H3N2 isolates (A/ Eng/26/99 and A/Udorn/307/72) and A/Dk/Sing/97, an avian isolate of different subtype (H5N3). Growth of both humanderived influenza viruses tested, A/Eng/26/99 (H3N2) and A/ Udorn/307/72 (H3N2), was not significantly different between 32/33uC and 37uC ( Figure 3A and 3B). Indeed, these two additional human influenza virus strains showed even less difference in titer between temperatures than was determined for A/Victoria/3/75. Assessment of growth of avian influenza virus, A/Dk/Sing/97 (H5N3), over a 48 hr time course at 37uC showed similar growth kinetics to that of A/Eng/26/99 (H3N2), reaching titers of 7610 5 pfu/ml and 1.6610 6 pfu/ml, respectively ( Figure 3A and 3C). In contrast, at 32uC, A/Dk/Sing/97 (H5N3) failed to grow at all ( Figure 3C ). Clearly, the restriction of A/Dk/Sing/97 at 32uC compared to 37uC was an even more striking phenotype than A/ Duck/Eng/62. As the avian influenza virus strains used in this study were selected at random, with no selection for a temperature-dependent phenotype, we propose that low temperature restriction of avian influenza viruses, but not human influenza viruses, may be broadly characteristic of avian influenza viruses. The extent of restriction, however, may be variable between different virus strains. Since the avian virus isolates used in these experiments are neither derived from samples obtained from humans nor passaged in human cells in vitro, we next investigated whether growth attenuation at low temperatures would be retained in a highly pathogenic H5N1 (A/VN/1203/04) influenza virus isolated from a fatal human case [34] . We compared infection kinetics of H5N1 (A/VN/1203/04) at 33uC and 37uC on HAE using A/Udorn/ 307/72 in parallel cultures as a human influenza virus control. As described above, A/Udorn/307/72 grew with similar kinetics at 33uC and 37uC ( Figure 3B ). A/VN/1203/04, however, exhibited slower replication kinetics at 33uC when compared to that for 37uC ( Figure 3D ). Indeed, titers were significantly decreased at 33uC vs. 37uC at 24, 48 and 72 hrs pi. In addition, only at 37uC did A/VN/1203/04 approach similar peak titers as the human A/ Udorn/307/72 virus by the end of the 72 hr time course ( Figure 3D ). Histological analyses of A/VN/1203/04-infected HAE at either temperature showed absence of obvious CPE in sharp contrast to A/Udorn/307/72 that obliterated the epithelium by 72 hrs pi ( Figure 3E ). The lack of obvious CPE after H5N1 infection contrasts reports that H5N1 induced extensive apoptosis in mammalian airway cells [35, 36] . The fact that we did not observe obvious CPE with this highly pathogenic virus warrants further investigation but is in line with the limited cell damage shown following infection with A/Dk/Eng/62 for 72 hrs ( Figure 2B ). In sum, using diverse examples of human and avian influenza viruses we have shown that avian influenza viruses, but not human influenza viruses, are restricted for infection and growth in HAE at the lower temperature of 32uC. 'Avianization' of human virus polymerase restricts growth in HAE at both 32uC and 37uC Previously, the polymerase subunit PB2 has been shown to play an important role in host range restriction of avian influenza viruses in mammalian cells [37] [38] [39] . In influenza virus strains that circulate in humans, amino acid residue 627 in PB2 is a lysine, whereas in the majority of avian strains it is a conserved glutamic acid residue. The presence of glutamic acid at PB2 627 (avian-like) has been reported to account for the lower replication of avian influenza strains in mammalian cells and has been linked with reduced polymerase activity at lower temperature (33uC) in some cell systems [23, 24] . To assess the potential impact of this PB2 amino acid residue in restriction of avian influenza viruses at 32uC, we generated a recombinant A/Victoria/3/75 virus containing the PB2 K627E mutation and compared its growth with that of the isogenic wild-type virus in HAE at 32uC and 37uC. The K627E mutation resulted in restriction of the virus at both temperatures (Figure 4Ai ), and although titer at 32uC was 1.3 logs lower than at 37uC at 24 hrs pi, this difference was no greater than the differences in growth for wild-type virus at these temperatures (1.5 logs; Figure 4Ai ). Moreover, at the later time points analyzed, 48 and 72 hrs pi, the PB2 mutant did not show a significant difference in titer between the two temperatures. These data indicate that the K627E mutant virus was restricted for growth in HAE but that restriction was not temperature-dependent. Indeed, quantification of the numbers of infected cells identified by en face staining revealed that the K627E mutant virus infected a similar percentage of cells compared to wild-type virus at 24 hrs pi ( Figure 4Aii ) and that the mutant was capable of spread since new cells were infected by 48 hrs with similar kinetics to that of wildtype A/Victoria/3/75 at both 32uC and 37uC (Figure 4Aii ). Statistically, there was no difference between the wild-type and PB2 mutant viruses at either 32uC or 37uC at 48 hrs pi with respect to percent influenza virus-antigen positive epithelium. Together, these data suggest that the amino acid residue at PB2 627 influences viral fitness in HAE, but does not confer to a human virus the temperature-dependent phenotype of avian influenza virus infection in human ciliated airway epithelium. Human influenza viruses with avian-like glycoproteins display restricted replication and spread at 32uC in HAE Our initial phenotype indicated that A/Dk/Eng/62 was restricted in its ability to spread from cell to cell within the epithelium at 32uC (Figure 2A ). Several events in the viral life cycle that are critical for spread, including release of progeny virions from previously infected cells and attachment and entry into new target cells, are mediated by influenza virus glycoproteins. Thus, we hypothesized that glycoprotein function could be responsible for the restricted infection of HAE by avian influenza viruses at the lower temperature of 32uC. To test whether HA and/or NA contributed to the restricted phenotype of avian influenza viruses at 32uC, we used reverse genetics to generate mutant viruses genetically altered to confer avian virus-like glycoprotein specificities on the A/Victoria/3/75 background. First, mutations in HA previously shown to switch sialic acid usage from a2,6 to a2,3 linkages (L226Q, S228G) [40] were introduced to generate the Vic-226-228HA virus. Second, we generated a reassortant virus in which the Victoria NA was replaced by that of the avian virus A/ Chick/Italy/1347/99 to generate Vic+Chick N1. We again compared virus replication and spread of the recombinant viruses to that of wild-type A/Victoria/3/75 at the two temperatures. As stated above, replication measured for the wild-type virus was slightly compromised at lower temperature, noticeable at 24 hrs pi. Restriction at this time point was also observed during infection of HAE with Vic-226-228HA, as it had been for the PB2 mutant virus. Specifically, a 2.5 log decrease in virus growth was determined for Vic-226-228HA at 32uC compared to 37uC at the 24 hr time point (Figure 4Bi ). However, unlike the PB2 mutant virus, the difference between replication at 32uC and 37uC for Vic 226-228HA was also significant at the 48 hour time point. Moreover, this mutant virus with avian viruslike sialic acid usage spread less efficiently than wild-type at 32uC so that by 48 hrs pi the number of virus antigen-positive cells was significantly different (Figure 4Bii ). In contrast, at 37uC, Vic-226-228HA infected similar numbers of cells as the wild-type virus by 48 hrs; indeed, the mutant virus was able to spread significantly more efficiently at the higher temperature (Figure 4Bii) . Similarly, the reassorted virus Vic+Chick N1 displayed a 2 log decrease in viral titer in HAE at 32uC compared to 37uC at 24 hrs pi. Although this difference was not appreciably greater than the difference in titer between temperatures for either wild-type virus or the PB2 mutant, Vic+Chick N1, unlike wild-type A/Victoria/3/75 and Vic 627PB2, maintained the ,2-log difference in growth at 48 hrs pi (Figure 4Ci ), suggesting this virus was more restricted at the cooler temperature. Quantification of numbers of infected cells illustrated that, like Vic-226-228HA, Vic+Chick N1 was restricted for spread at 32uC which was significant at 48 hrs, but was capable of spread similar to wild-type A/Victoria/3/75 at 37uC (Figure 4Cii ). Together these data suggest that avianizing either the HA or NA glycoprotein of an otherwise human influenza virus limits spread and subsequent infection at 32uC compared to 37uC. We next generated a recombinant influenza virus containing both the 226-228HA and Chick N1 and tested infection and growth in HAE at 32uC and 37uC in comparison to wild-type A/Victoria/3/ 75. At 24 hrs pi, the double glycoprotein-altered virus exhibited similar restriction as observed for the other viruses. Nonetheless, an overall evaluation of the double glycoprotein-altered virus suggested that as infection proceeded, this virus was profoundly restricted at 32uC compared to 37uC (Figure 4Di ), exhibiting .2 log reduction in titer at 48 hrs. Notably, titers for the wild-type virus differed by less than 0.5 logs between temperatures at this time point. Furthermore, the double glycoprotein-altered virus was still significantly restricted at 72 hrs pi when titers at 32uC were compared to those at 37uC. The level of restriction observed for the double mutant was greater than that observed for either virus containing each of these mutations/ substitutions individually. Moreover, analysis of viral antigen positive cells at 72 hrs by en face staining of infected HAE indicated compromised spread of Victoria (226-228HA)+Chick N1 which was more severe at 32uC than 37uC (Figure 4Dii) . Determination of CPE during these experiments revealed that the double glycoprotein-avianized virus only produced CPE at 72 hrs pi when experiments were performed at 37uC, whereas wild-type human virus produced CPE earlier and at both temperatures (data not shown). These data are consistent with the levels of CPE observed for A/Dk/Eng/62 (H4N6) and A/ Victoria/3/75 (H3N2) in our initial studies ( Figure 1B) and suggest that altering the human virus glycoproteins to avian virus-like characteristics has profound effects on infection, spread and CPE in the environment of the human ciliated airway epithelium. One potential caveat of the recombinant viruses with avianized HA and/or NA utilized in our previous analysis was that they contained HA and NA pairs that had not co-evolved. To eliminate the possibility that the restriction we observed with these recombinant viruses was due to an imbalance between the activities of the surface glycoproteins that were not evolutionarily optimized, we next generated reassorted influenza viruses on a common genetic background, possessing human or avian glycoproteins with co-evolved pairings. This was achieved using human recombinant A/PR/8/34 (H1N1) in which the wild-type H1 and N1 glycoproteins were replaced by the H3 and N2 glycoprotein pair from A/Victoria/3/75 (generating PR8+Vic HA/NA) or the H7 and N1 glycoprotein pair from A/Chick/ Italy/1347/99 (generating PR8+Chick HA/NA, previously termed RD3) [41] . Since we and others have shown differential cell-type tropism between human and avian influenza virus in HAE [13, 14] , we next determined if avianizing the human virus HA by mutation or substitution (in the presence or absence of an avian NA) recapitulated the cell-type tropism exhibited by wholly avian influenza viruses in HAE. As shown by immunofluorescent detection in histological sections of infected HAE, PR8 containing A/Victoria/3/75 glycoproteins infected both ciliated and nonciliated cells in HAE with a tropism similar to wild-type A/ Victoria/3/75 ( Figure 5 ). In contrast, A/Victoria/3/75 with two avian-like amino acid substitutions in HA and PR8+Chick HA/ NA only infected ciliated cells, a tropism that was mirrored by wholly avian virus [13, 14] . These data clearly show that the ciliated cell tropism of avian influenza viruses is dictated by properties of the viral glycoproteins. These results correlate with the known increased sialic acid binding preference of avian HA for a2,3-linked SA, and to the presence of a2,3-linked SA on ciliated cells in HAE [8, 13, 14] . Growth kinetics in HAE of PR8+Vic HA/NA and PR8+Chick HA/NA inoculated at equal MOI (0.01) revealed that PR8+Vic HA/NA infection and growth was efficient at both 32uC and 37uC ( Figure 6A ). PR8+Chick HA/NA grew at 37uC to identical titers as PR8+Vic HA/NA at 32uC recapitulating our data obtained for wholly human (A/Victoria/3/75) and wholly avian (A/Dk/Eng/ 62) viruses. In contrast, PR8+Chick HA/NA was severely delayed in growth at 32uC and generated titers that were .2 logs less than titers obtained for this virus at 37uC at both 24 and 48 hrs pi. Indeed, PR8+Chick HA/NA, like A/Dk/Eng/62 avian influenza virus ( Figure 1A) , was significantly restricted for growth at 32uC at 12, 24 and 48 hrs pi compared to growth at 37uC and growth of PR8+Vic HA/NA at either temperature. As observed for wholly human and avian influenza viruses, peak titers were reached for PR8+Vic HA/NA at both temperatures and PR8+Chick HA/NA at 37uC by 48 hrs pi after which a decline in viral titer was apparent. Again, as noted in our observations with human and avian influenza viruses, the loss of viral titers with time correlated with the onset of CPE. While PR8+Chick HA/NA infection at 32uC did not result in substantial AK release until 96 hr pi, increased AK activity was detected in cultures inoculated with this virus at 37uC. AK activity measured in cultures at this temperature increased with similar kinetics and reached similar levels as AK measured in cultures inoculated with PR8+Vic HA/NA at either temperature. Furthermore, the kinetics of AK induction demonstrated that again, AK was virus and (Cii) N1 reassortant virus. Data obtained in parallel for wild-type A/Victoria/3/75 is repeated in each graph (striped bars) for comparison to the mutant (solid bars). Data shown represents the mean of the percentage of influenza virus antigen-positive epithelium across 10 different fields +/ 2SE. Differences in viral antigen positive epithelium between temperatures for each virus at 48 hrs pi is noted as significant (*p,0.05) or insignificant (NS). A one-way ANOVA model showed no significant differences between the wild-type virus and PB2 mutant at 32uC and 37uC at 48 consequential to viral replication and that, overall, CPE induced by reassortant viruses was reflective of CPE measured for human and avian influenza viruses. En face staining of HAE at 24 hr intervals after inoculation showed PR8+Chick HA/NA spread to additional target cells at 37uC at a rate similar to that of PR8+Vic HA/NA at 32uC and correlated with the titers measured for these two viruses under those conditions ( Figure 6C and 6D ). At 32uC, however, PR8+Chick HA/NA spread was severely compromised and resembled the infection characteristics shown for A/Dk/Eng/62 (H4N6) in Figure 2A . Thus, by replacing human glycoproteins with those from an avian virus isolate, we have recapitulated the effect of temperature on infection and growth kinetics as well as the degree of cytotoxicity produced by wholly avian influenza virus interactions in human ciliated airway epithelium. The relative contributions of reduced cell-cell spread and reduced CPE by avian-like influenza viruses at temperatures of the proximal airways to in vivo infection and pathology will, however, require further investigation. We have performed comparative studies of the infection kinetics of human and avian influenza viruses in a model of human ciliated airway epithelium at temperatures reflective of the human proximal and distal airways. Our data show that avian and avianized influenza viruses are restricted for infection and growth in HAE at 32uC but not 37uC, while human viruses infect and grow efficiently at both temperatures. Based on these data, we suggest that while the warmer temperatures of the distal airways enable comparable infection by both human and avian influenza viruses, the cooler temperatures of the human proximal airways only support efficient and robust infection of the ciliated airway epithelium by human influenza viruses. We speculate that the observed restriction for avian and 'avianized' viruses in HAE would render avian influenza viruses more susceptible to innate and adaptive immune responses that limit pathogenicity in vivo. These results have significant impact on our understanding of why avian influenza viruses rarely undergo zoonotic transmission and why, when the rare human case does occur, that avian influenza virus infection and pathology manifest predominately in the warmer distal airways and lungs. The inability of avian influenza viruses to replicate efficiently at cooler temperatures has been linked to the viral polymerase subunit, PB2 [23, 24] . In the present study, mutating position 627 in a human virus PB2 to an avian virus conserved residue resulted in growth restriction at both 32uC and 37uC, suggesting that this residue is important for general viral fitness in HAE, but is not responsible for the differences in infection seen at 32uC vs. 37uC. Two recent reports also found that viruses with 627E in PB2 were attenuated regardless of temperature in human bronchial epithelial cells and MDCK cells, respectively, although in other cell systems including human small airway epithelial cells, a temperature specific effect was found [24, 42] . It should be emphasized that those studies were performed in non-differentiated epithelial cells unlike our studies that use human differentiated airway epithelial cells. We and others have previously shown that differentiated airway epithelial cell models enable discrimination of attenuated phenotypes of respiratory virus infection whereas non-differentiated cells do not [26, 27, 43] . In addition, we also show using HAE, that the H5N1 strain A/VN/1203/04, which possesses a lysine at position 627 (human adaptation), is still restricted for growth at 32uC, albeit less so than avian influenza viruses that have never infected humans. The attenuation in HAE of this H5N1 isolate which possesses a ''human'' amino acid at residue 627 in PB2 suggests other residues in the polymerase subunit or other viral proteins altogether are involved in temperature sensitivity of avian influenza viruses. In our initial experiments, spread of avian influenza viruses from cell to cell at 32uC was compromised in cultures inoculated at low MOI, suggesting a potential role for the envelope glycoproteins, HA and NA, in mediating temperature restriction. Previous work by Kaverin and colleagues also demonstrated temperature Figure 5 . Cell tropism of human, avian and avianized viruses in HAE. Representative cross-sections of inoculated HAE, fixed 24 hrs pi, were probed for viral antigen (NP; green) and a2acetylated tubulin, a marker for ciliated cells (red). Notably, the staining pattern for wild-type A/Victoria/3/ 75 was identical to that of PR8+Vic HA/NA. Arrows mark ciliated cells infected with either wild-type A/Victoria/3/75 or PR8+Vic HA/NA; arrow-head denotes non-ciliated cells infected by these viruses. These data indicate that viruses with Victoria glycoproteins were able to infect both cell types previously shown to express a2,6 SA [13] . Viral antigen was detected only in ciliated cells in cultures inoculated with Vic-226-228HA (in the Victoria background with either endogenous N2 or avian N1 or PR8+Chick HA/NA). Scale bar equals 20 mm. doi:10.1371/journal.ppat.1000424.g005 effects on growth of human-avian reassortant viruses containing avian glycoproteins [25] , although this work was performed in non-polarized MDCK cells and did not investigate additional correlates of infection such as spread and CPE. In our study, we generated recombinant influenza viruses based on the A/Victoria/ 3/75 or A/PR/8/34 genetic backbone that were engineered to contain avian-like and/or avian glycoproteins and characterized infection in HAE. Kinetic studies showed that although human influenza viruses that possessed avian or avian-like surface glycoproteins were modestly restricted compared to wild-type viruses at 37uC, these mutant viruses were able spread like wildtype viruses throughout HAE at this temperature. Wide-spread infection throughout HAE was even observed for viruses in which their endogenous HA was replaced or mutated to preferentially bind a2,3 SA, restricting tropism to ciliated cells. Efficient replication of Vic-226-228HA at 37uC in our studies corroborates previous work by Matrosovich and colleagues in which little effect of HA-specificity 'switching' on replication was noted unless a very low MOI (0.00004) was used for inoculation [44] . In contrast, Wan and Perez described more profound differences in replication in HAE at 37uC with recombinant viruses that differed only in their receptor specificity [31] . However, it should be noted that their recombinant viruses were based on an H9N2 avian strain that yielded relatively low titers, and their initial infections were performed at 35uC before incubating at 37uC [31] . Compared to 37uC, viruses with a preference for binding to a2,3 SA, including Vic-226-228HA, were restricted for growth and spread in HAE at 32uC. Notably, the H5N1 strain examined in this study also maintains preference for a2,3 SA binding [45] ; thus, we may surmise that this characteristic of A/VN/1203/04 contributes to its attenuation observed in HAE. The contribution of a2,3 SA usage to replication of influenza viruses investigated by Hatta et al. in the upper respiratory tract of mice may have been masked in the mouse model (the 627 mutation in PB2 being more apparent) as mice express solely avian virus-like receptors (a2,3 SA) in their airways [46] . Restriction of a2,3 SA-binding viruses in HAE at 32uC was not due to a discrepancy in SA expression since HAE maintained at either 32uC or 37uC expressed similar levels of a2,6 and a2,3 SA (as detected by Sambucus nigra (SNA) and Maackia amurensis (MAA) lectin staining, respectively; data not shown). In conjunction with the HA, the sialidase activity of NA is crucial for successful virus penetration of mucus layers for initial infection and subsequent release of progeny virions from infected cells [47, 48] . This is especially critical both in vivo and in HAE models in which the luminal epithelial cell surface is robust with glycoconjugates displaying abundant terminal sialic acid moieties that may act as false receptors for influenza viruses [49] . Using standard laboratory assays that employ small monovalent soluble substrates for cleavage by NA (MUNANA), we were not able to demonstrate any temperature-dependent loss of NA activity associated with either human or avian virus (data not shown). However, the ability of the avian virus NA to cleave biologically relevant substrates present in HAE may be compromised at 32uC vs. 37uC restricting both initial infection and subsequent spread of the virus throughout the epithelium. This is supported by our data which demonstrate restricted growth and spread of reassortant viruses containing avian virus NA, including Vic+Chick N1 and PR8+Chick HA/NA in HAE at 32uC. In addition to their independent functions, the balance between the binding affinity of the viral HA and the sialidase activity of the NA is also critical for efficient infection. The ability of A/Victoria/ 3/75 viruses with mutations or substitutions in either the HA or NA alone to infect similar numbers of cells and replicate to comparable peak titers as for wild-type virus at 37uC implies that these viruses were not crippled by the mismatch between the specificities of their HA and NA. Replication and spread of influenza viruses that possess an avian HA paired with its ''matched'' NA was even more compromised than that of recombinant viruses with individual changes to levels seen with wholly avian viruses. Thus, viruses with co-evolved glycoprotein pairs exhibit restricted replication at low temperatures and both HA and NA genes contribute to the phenotype. Together, these data imply that in the complex environment of the luminal surface of the human ciliated airway epithelium, the viral surface antigens have a marked effect on the extent of virus infection and that temperature plays an important role in limiting avian, but not human, influenza virus infection and spread in the cooler proximal airway regions. Given these results, we draw attention to other recently published data using the HAE model in which mutations in viruses that are growth attenuated in vivo display similar growth attenuation in HAE but not in nondifferentiated cell lines, suggesting that HAE possess discriminating properties of attenuating phenotypes of mutants of respiratory viruses [26, 27] . Admittedly, in the present study, despite restriction in both growth and spread, wild-type avian viruses and human viruses with avian or avian-like glycoproteins did eventually reach high titer at 32uC at later time points. The efficiency of infection and replication of a virus that inoculates the airway epithelium, however, is likely a critical factor in determining whether the virus is capable of establishing infection in a host that normally possesses innate and adaptive immune systems that attempt to limit virus infection and spread. At temperatures of the distal airways, avian influenza viruses displayed similar infection kinetics as human influenza viruses and would therefore, in the case of sufficient inoculum reaching these distal regions, be as likely to establish infection. Indeed, the clinical pathology findings for humans infected with H5N1 do report distal airway infection in ciliated bronchioles and lung regions [22] . Under these conditions of inoculation and infection, avian influenza viruses present in the distal airways may still be unable to spread to proximal airway regions without additional adaptation to cooler temperatures. One caveat of this prediction is that virus may be transported to proximal airway regions by innate mucus clearance mechanisms indicating that caution is required when attempting to identify proximal infection by viruses in airway secretions obtained from tracheal swabs. In conclusion, the present study substantiates differential host temperature as a critical barrier for infection by avian influenza viruses. Since the ciliated airway epithelium of the proximal airways is a major portal for influenza virus infection and spread, accessible by multiple inoculation routes (e.g., ocular, nasopharyngeal or aerosol), the inability of avian influenza viruses to establish infection and spread in these regions would be predicted to reduce the frequency of successful zoonotic transmission. Furthermore, the ability of human influenza viruses to generate high viral titers in the human proximal airways is likely a factor in effective human-to-human transmission and the induction of airway epithelial cell cytotoxicity as shown in this study may increase particulate matter perhaps associated with virus that facilitates inoculation of new hosts. Rapid induction of cytotopathic effects by human, but not avian, influenza virus infection at the temperature of the human proximal airways may also contribute to the onset of other host defenses such as sneezing and coughing that facilitate clearance of particulate matter/virus from the airways and potentially promote transmission between human hosts. Human airway tracheobronchial epithelial cells isolated from airway specimens from patients without underlying lung disease were provided by the National Disease Research Interchange (NDRI, Philadelphia, PA) or as excess tissue following lung transplantation under University of North Carolina at Chapel Hill (UNC) Institutional Review Board-approved protocols by the UNC Cystic Fibrosis Center Tissue Culture Core. Primary cells derived from single patient sources were expanded on plastic to generate passage 1 cells and plated at a density of 3610 5 cells per well on permeable Transwell-Col (12-mm diameter) supports (Corning, Inc.). HAE cultures were grown in custom media with provision of an air-liquid interface for 4 to 6 weeks to form differentiated, polarized cultures that resemble in vivo pseudostratified mucociliary epithelium, as previously described [50] . Madin-Darby Canine Kidney (MDCK) cells were maintained in DMEM (Gibco-Invitrogen, Inc.) supplemented with 10% fetal bovine serum and 1% penicillin / streptomycin (Sigma-Aldrich, Inc.). Influenza virus A/England/26/99 (H3N2) was isolated at the Health Protection Agency, Colindale, London, UK, during the routine surveillance program and has been minimally passaged in MDCK cells [51] . A/Dk/Singapore/97 (H5N3) and A/Dk/ England/62 (H4N6) are typical avian influenza strains that have been passaged in both embryonated chicken eggs and MDCK cells during laboratory handling. Highly pathogenic A/VN/1203/ 04 (H5N1) was biologically derived and minimally passaged in embryonated chicken eggs. A/Udorn/307/72 (H3N2) was passed in baby hamster kidney (BHK) cells and represents a clone expanded once in embryonated chicken eggs. Recombinant viruses, including wild-type A/Victoria/3/75 (H3N2) and mutants in either the A/Victoria/3/75 (H3N2) or A/PR/8/24 (H1N1) background, were generated from cloned cDNA in 293T and MDCK cell co-cultures as previously described [52, 53] . Mutant viruses were generated in either the A/Victoria/3/75 (H3N2) or A/PR/8/34 (H1N1) genetic background as follows: 1) Vic 627PB2; A/Victoria/3/75 containing a lysine to glutamic acid amino acid substitution at position 627; 2) Vic-226-228HA; A/ Victoria/3/75 containing two amino acid substitutions in the HA gene (L226Q, S228G) that confer an avian-like receptor binding preference [6, 40] ; 3) Vic+Chick N1; A/Victoria/3/75 in which segment 6 containing the endogenous N2 NA gene was exchanged for the N1 NA gene from avian isolate A/Chick/Italy/1347/99; 4) Vic-226-228HA+Chick N1; A/Victoria/3/75 containing both L226Q and S228G mutations and the avian N1; 5) PR8+Vic HA/ NA; A/PR/8/34 in which the endogenous H1 and N1 were replaced with the H3 and N2 from A/Victoria/3/75 and 6) PR8+Chick HA/NA (RD3); A/PR/8/34 in which the endogenous H1 and N1 were replaced with the H7 and N1 from A/ Chick/Italy/1347/99. (RD3 was previously described as a candidate vaccine strain [41] .) The last two reassortant viruses were generated by substituting segment 4 and segment 6 from PR8 with those from either A/Victoria/3/75 (H3N2) or A/Chick/ Italy/1347/99 (H7N1). The multi-basic cleavage site in the avian H7 HA gene used in these studies was removed prior to rescue of these recombinant viruses for safety. Available accession numbers (GenBank: http://www.ncbi.nlm.nih.gov.libproxy.lib.unc.edu) are V01086 for A/Victoria/3/75 HA and CAD37074 for A/Chick/ Italy/1347/99 HA. HAE were rinsed with PBS to transiently remove apical secretions and supplied with fresh basolateral medium prior to inoculation. Virus inoculum was diluted in PBS and applied to the apical surface of HAE for 2 hrs at either 32uC, 33uC, or 37uC, as indicated. Following incubation, viral inocula were removed and cultures incubated at 32uC, 33uC or 37uC for the duration of the experiment. Viral growth kinetics were determined by performing apical washes with 300 ml of serum-free DMEM for 30 min at either 32uC or 37uC. Washes were harvested and stored at 280uC prior to analysis. Viral titers in the apical washes were determined by standard plaque assay or tissue culture infectious dose (TCID) 50 assay on MDCK cell monolayers as previously described [13, 52, 54] . At various points post-inoculation (pi), HAE were fixed in cold methanol-acetone (50/50) and stored at 4uC. Cultures were then permeabilized with 2.5% triton-X 100/PBS++ (containing 1 mM CaCl 2 and 1 mM MgCl 2 ) and blocked with 3% bovine serum albumin (BSA) in PBS++ before being probed with mouse antiinfluenza virus nucleoprotein (NP; Chemicon, Inc.; 1:100) and immunoreactivity detected with fluorescein isothiocyanate (FITC)conjugated anti-mouse IgG secondary antibody (Jackson Immu-noResearch Laboratories, Inc., 1:500). Fluorescent images were obtained using a Leica DMIRB inverted fluorescence microscope equipped with cooled-color charge-coupled-device digital camera (MicroPublisher; Q-Imaging, Burnaby, BC, Canada). The percentage of the epithelium positive for viral antigen as an index of percentage of infected cells was quantified over 5 images per culture by black and white pixilation of each image and computer calculation of percent black pixels after inverting the image. This technique determines percentage of black pixels in a defined area and does not account for differences in fluorescent intensity. Viral-induced cytotoxicity was determined by measuring adenylate kinase activity in apical washes using a commercially available assay (Lonza, Inc.). Apical samples were centrifuged prior to freezing to remove any cellular contaminants present in the wash. Luminescence detected in samples from infected HAE were normalized to uninfected HAE and expressed as fold change over AK measured in uninfected (mock) HAE. Morphological assessment of cytotoxicity in HAE was performed with paraformaldehyde (PFA, 4%)-fixed histological sections (5 mm) stained with hematoxylin and eosin. Detection of a2,3 and a2,6 linked sialic acids HAE maintained at either 32uC or 37uC for 72 hrs prior to sialic acid detection were washed, blocked with 3% BSA/PBS++ and probed with biotinylated SNA or MAA lectins to detect a2,6 and a2,3 SA, respectively (Vector Laboratories, Inc.; EY-Laboratories, Inc.; 1:100). HAE were then fixed in 4% PFA and incubated with streptavidin-alexafluor 488 (Molecular Probes, Inc.; 1:500) applied to the apical surface to detect lectin binding. Immunohistochemistry HAE fixed in methanol:acetone, were probed en face with antibody against viral NP (Chemicon, Inc.; 1:100) and FITCconjugated goat anti-mouse IgG1 and IgG2a (Jackson ImmunoResearch Laboratories, Inc., West Grove, PA; 1:500), then embedded in paraffin. Histological sections (5 mm) were prepared and reprobed for viral antigen using standard immunofluorescence protocols. Briefly, sections were bathed in 2.5% triton-X 100/ PBS++ for 30 min, blocked in 3% BSA/PBS++ and incubated with antibodies in 1% BSA/PBS++. Primary antibodies were antiviral NP (Chemicon, Inc., as above) and anti-alpha acetylated tubulin (Zymed Laboratories, Inc.; 1:2000), a marker for ciliated cells. Secondary antibodies were FITC-goat anti-mouse IgG2a and Rhodamine red-conjugated goat-anti-mouse IgG2b (Jackson ImmunoResearch Laboratories, Inc.; 1:500). Sections were prepared with FluorSave mounting media (EMD Chemicals, Inc.) and images captured using a Leica DMIRB inverted fluorescence microscope equipped with a cooled color chargecoupled-device digital camera (MicroPublisher; Q-Imaging, Burnaby, British Columbia, Canada). Linear mixed models were fitted to the repeated measurements of log-transformed viral titer over time that included effects for the four treatment groups (defined by virus and temperature), eight time points, and the interaction between treatment and time. We note that in a small number of cases, there were only two treatment groups (defined by temperature) and fewer than eight time points. A heterogeneous autoregressive correlation structure of order one was assumed for the repeated measurements. A joint test of the interaction terms (21 degrees of freedom) provides an assessment of the hypothesis of no differences among the four treatment groups with respect to viral titer growth (log scale). Provided this test was significant, indicating some differences among the four growth curves, pair-wise differences between the three treatment groups versus the a priori specified reference group (generally the avian strain at the lowest temperature) were carried out for each time point, and significant differences at the 0.05 level were noted. No adjustments for inflated Type I error due to multiple comparisons were made. Missing observations were assumed to be missing completely at random, based on the fact that the investigators determined a priori to remove samples at specific time points during the experiment.

A multidimensional classification of public health activity in Australia

BACKGROUND: At present, we have very limited ability to compare public health activity across jurisdictions and countries, or even to ascertain differences in what is considered to be a public health activity. Existing standardised health classifications do not capture important dimensions of public health, which include its functions, the methods and interventions used to achieve these, the health issues and determinants of health that public health activities address, the resources and infrastructure they use, and the settings in which they occur. A classification that describes these dimensions will promote consistency in collecting and reporting information about public health programs, expenditure, workforce and performance. This paper describes the development of an initial version of such a classification. METHODS: We used open-source Protégé software and published procedures to construct an ontology of public health, which forms the basis of the classification. We reviewed existing definitions of public health, descriptions of public health functions and classifications to develop the scope, domain, and multidimensional class structure of the ontology. These were then refined through a series of consultations with public health experts from across Australia, culminating in an initial classification framework. RESULTS: The public health classification consists of six top-level classes: public health 'Functions'; 'Health Issues'; 'Determinants of Health'; 'Settings'; 'Methods' of intervention; and 'Resources and Infrastructure'. Existing classifications (such as the international classifications of diseases, disability and functioning and external causes of injuries) can be used to further classify large parts of the classes 'Health Issues', 'Settings' and 'Resources and Infrastructure', while new subclass structures are proposed for the classes of public health 'Functions', 'Determinants of Health' and 'Interventions'. CONCLUSION: The public health classification captures the important dimensions of public health activity. It will facilitate the organisation of information so that it can be used to address questions relating to any of these dimensions, either singly or in combination. The authors encourage readers to use the classification, and to suggest improvements.

Classification of public health: working definitions of functions subclasses 26 The objective of the Public Health Classifications Project is to 'develop and endorse a higher-level classification that captures the breadth and scope of public health activity and provides a unified framework for multiple uses'. Such a unified framework will assist in improving the quality and consistency of reported information on public health activity, performance, investment and expenditure. The National Public Health Partnership funded the project in response to recommendations from the 2002 Public Health Performance Project. 1 During the early scoping stages of the Public Health Classifications Project, it became apparent that a simple, one-dimensional classification system for public health could not satisfy the needs, or reflect the diverse world-views, of its disparate potential users. To provide a single 'unified framework' for multiple public health uses, a multi-dimensional classification was needed. In the domain of public health, a flexible and inclusive approach offers particular advantages, because there are divergent (and strongly held) views regarding what is and is not 'in scope'. By making such issues explicit, the process of developing a public health classification potentially offers a way to move towards a common language to describe public health activity in Australia, and to develop a practical tool that will improve data collection processes and the utility of public health information. This report is the output of phase one of the Public Health Classifications Project. It introduces the concept of a multi-dimensional public health classification and describes the challenges encountered in developing it. The report presents version one of a classification of public health, outlines some potential practical applications, and proposes the next steps for phase two of the project. A Reference Group (see acknowledgements in Appendix B) oversaw phase one of the Public Health Classifications Project and provided ongoing expert advice and comment. The project used a formal methodology and supporting software. 2 A review of current public health definitions, concepts and relevant classifications was used to develop the scope, domain, and initial multi-dimensional structure. These were considered in a series of consultations with public health experts across Australia (see list of those consulted in Appendix B). Consultations involved both one-on-one meetings and larger group sessions. Experts identified important omissions, fine-tuned concepts, and nominated practical uses for the public health classification. The definition of public health adopted was as follows: Public health is the organised response by society to protect and promote health, and to prevent illness, injury and disability. The starting point for identifying public health issues, problems and priorities, and for designing and implementing interventions, is the population as a whole, or population subgroups. 3 The boundary between public health and clinical practice came up repeatedly as an issue in discussions about the scope of public health, with particular debate about whether preventive services delivered on a one-to-one basis to individuals should be considered in scope. Many agreed that immunisation was in scope because it is an activity that is 'organised' at a population level with benefits for both populations and individuals. More contentious, however, was the possible inclusion of interventions that are designed to prevent and manage chronic diseases, and that are delivered to individuals in primary care settings. Whether or not public health is a domain solely within health or whether it includes activities in other sectors (e.g. education, transport, local government) was also debated, particularly where the public health impact of the activities in these other sectors is incidental, rather than the primary purpose of the activity. The general approach adopted in producing the classification was to be inclusive, and to allow decisions about specific exclusions to be made at the later stages when developing individual applications and uses of the classification. The broad structure of version one of a classification of public health consists of six top-level classes as shown in Figure 1 . There was consensus among the public health experts consulted, that a public health classification should be multi-dimensional, and there was broad agreement on the top-level classes that should be included. Public health functions are defined as the purpose of public health interventions, actions, activities and programs. The 'functions' class was developed from the National Public Health Partnership public health core functions 4 and includes both primary and instrumental functions (shown in Table 1 There was reasonable agreement regarding the top levels of the 'health issues' class (although its name was the subject of some debate), and the 'determinants of health' and 'settings' classes. The remaining classes are less well developed and have had limited testing through consultations. As shown in Figure 2 , existing classifications (such as the international classifications of diseases, functioning and disability, and external causes of injuries; and various Australian standards) are available to classify the classes of 'health issues', 'settings' and 'resources'. The National Public Health Information Working Group has determined that the further development of classifications for the top-level classes of 'functions', 'determinants of health' and 'methods' are required and a priority. Other to be classified: A public health classification should facilitate the organisation of information to answer key public health questions that cannot currently be answered, such as 'How much was spent last year on the prevention of obesity?' It should assist in describing what public health is, and what its characteristics are, through the development of classes that capture the functions of public health, issues of public health concern (including determinants of health), the settings in which public health operates, the population groups targeted, resources available and so on. Potential practical applications for a public health classification include: Explaining what public health is; Organising information to answer key public health questions; Promoting consistency in describing public health; Improving data capture processes and the quality of reporting; Contributing to higher-level classification and standards activities; Lending structure to the design of public health information and communication systems; Auditing the spread of activity across the public health business cycle; Building models of good public health practice; and Linking research, policy and practice. Potential users of a public health classification include the various levels of government and other sectors that have an investment in public health, academics and students, researchers, evaluators, those involved in policy formulation, and anyone with an interest in public health. The Australian Institute of Health and Welfare has indicated an interest in the longerterm development and maintenance of a public health classification. It is recommended that phase two of the Public Health Classifications Project should: Focus on further developing the classes of public health 'functions', 'determinants of health' and 'methods'; Develop and release a web-based version of the public health classification with facilities for eliciting structured feedback and managing contributions to the further development and refinement of the classification; Develop a plan for ongoing development, support and governance of the public health classification; Further specify links or relations between the public health classification and relevant existing classifications and standards (with due regard for intellectual property rights); and Investigate inclusion of the public health classification in the Australian Family of Classifications. A classification is an 'arrangement of concepts into classes and their subdivisions to express the semantic relations between them'. 5 The essential characteristic of a classification is aggregation according to logical rules. Standardised, shared classifications are needed if we want to compare information about entities and discern their similarities and differences. The objective of the Public Health Classifications Project is to 'develop and endorse a higher-level classification that captures the breadth and scope of public health activity and provides a unified framework for multiple uses'. Such a unified framework will assist in improving the quality and consistency of reported information on public health activity, performance, expenditure and investment. This report is the output of phase one of the Public Health Classifications Project. It introduces the concept of a multi-dimensional public health classification and describes the challenges encountered in its development. It presents version one of a public health classification, outlines some potential practical applications, and proposes the next steps for phase two of the project. The National Public Health Partnership defines public health as: the organised response by society to protect and promote health, and to prevent illness, injury and disability. The starting point for identifying public health issues, problems and priorities, and for designing and implementing interventions, is the population as a whole, or population sub-groups. 7 As a sector, public health is largely funded by government. 8 In Australia the Australian Government is the major source of public health funding, while state and territory governments mostly apply the funds. 9 The public health workforce is as diverse as are its employers: there is no single or all encompassing occupation or industry group. 'general health and associated workers' who carry out aspects of public health functions on either a regular or occasional basis. 10 Some public health activities are carried out in sectors outside of health (e.g. local government, non-government organisations (NGOs), other government departments and agencies, including planning and environmental protection agencies). Some 'classic' public health functions are outsourced and funded well away from health and human services portfolios (e.g. sewage disposal, provision of safe potable water). Public health activity is costed at the program level, 11 and effectiveness and other measures are estimated at the aggregate level as theoretical constructs (e.g. population health status, potentially avoidable mortality). It is difficult to tell when public health effort and investment is effective, even over long periods of time; the small amount of work to this end is bedevilled by the poor quality of available data, the complexity of costing public health activity, 12 and lack of agreement about what should be included. 13 Costs to society when public health fails (e.g. cryptosporidium outbreak response, effect of SARS panic) may be easier to estimate. Available expenditure estimates suggest that there are relatively high overheads or indirect costs for public health programs and activities (e.g. design and coordination costs, costs of administering and managing complex operations). 14 Public health tends to exhibit large economies of scale and to be relatively insensitive to population size; hence unit costs may be lower in states with larger populations to absorb the fixed costs of overheads. 15 The National Public Health Partnership funded the Public Health Classifications Project in response to the 2002 Public Health Performance Project, 16 which recommended that the National Public Health Information Working Group undertake the development of a classification system for public health that could be used to further develop the categories used by the National Public Health Expenditure Project and 10 Employers include Australian, state and territory, and local governments; NGOs, Aboriginal Community Controlled Health Organisations, community services, environmental protection services, health promotion foundations, private sector organisations (e.g. pharmaceutical companies, pathology laboratories) (Riddout et al. 2002: 8) . 11 Even when program categories are artificially created, for example, state reporting against 'activity categories' in public health expenditure reporting (see AIHW 2004b). 12 Bennett 2003. 13 Abelson analysed the epidemiological and economic effects of five public health programs over decades (including programs to reduce: tobacco consumption, coronary heart disease -which some would dispute as a public health program -and road trauma), estimating costs of investment in public health interventions and benefits in terms of total return to society, and, savings to government. The 'net present value' to government of road safety programs and programs to reduce coronary heart disease was estimated as negative (expenditure greater than savings); while the benefit of immunisation for Haemophilus influenzae B disease was estimated at a 'marginal $10 million' (Abelson et al. 2003: 4) . 14 AIHW 2004b, nine public health programs in all jurisdictions. 15 Riddout et al. 2001. 16 Owen & Jorm 2002. performance monitoring by the National Public Health Partnership, and to inform a future review of the core functions for public health. The Public Health Performance Project used the public health core functions that were endorsed by the National Public Health Partnership in 2000, to develop performance indicators for public health. 17 These core functions differ from the categories used for national public health expenditure reporting 18 , resulting in difficulties in aligning data on performance with that on expenditure. More recently, a national report of health expenditure by disease groupings excluded expenditure on public health because this was not available 'by disease' 19 -further highlighting the inadequacy of current systems for capturing information about public health activities. The objective adopted by the National Public Health Partnership for the Public Health Classifications Project was to 'develop and endorse a higher-level classification that captures the breadth and scope of public health activity and provides a unified framework for multiple uses'. The project objective is to develop and endorse a higher-level classification that captures the breadth and scope of public health activity and provides a unified framework for multiple uses. During the early scoping stages of the project, it became apparent that a one-dimensional classification of public health might look very different, depending on its intended use, and user group. A simple, one-dimensional classification of public health could not satisfy all the needs, or gel with the diverse world-views, of its disparate potential users. To provide a 'unified framework for multiple uses', a multidimensional public health classification with explicit modeling of the relationships among dimensions is needed, rather than a single, mutually exclusive, hierarchy of categories. 20 This project used an ontology-building process to develop the public health classification. An ontology is an explicit formal specification of the concepts in a domain (in this case, public health), their attributes and the relations among them, which allows people to share a common understanding of the structure of information. 21 A multi-dimensional public health classification allows structure to be imposed on diverse material along different-but equally meaningful-dimensions, based on the way that public health experts and practitioners think about public health, and the ways in which they describe or classify it, or aspects of it, depending on their purpose. In the domain of public health, this flexible and inclusive approach offers particular advantages, because there are divergent (and strongly held) views regarding what is 17 Owen & Jorm 2002 : 8. 18 NPHP 1998 , NHPC 2004 , AIHW 2004b . 19 AIHW 2004c In practice, most classifications of complex domains are multi-dimensional, either implicitly so, or explicitly constructed as such. An example in the health field is the International Classification of Diseases (WHO 1992-94) , although the relationships among the dimensions are not all set out overtly. and what is not 'in scope'. By making such issues explicit, the process of developing a classification offers a way to move towards a common language to describe public health activity in Australia, and to develop a tool to improve data collection processes and the consistency of information about public health activity, performance, expenditure, effectiveness and returns on investment. A Reference Group (see acknowledgements in Appendix B) oversaw phase one of the Public Health Classifications Project and provided ongoing expert advice and comment. The project used the Ontology development 101 methodology 22 and the open source Protégé ontology-building software (from Stanford University 23 ) as the development tools. Ontology development 101 and Protégé were selected after a scan of available methods and software, because they were considered to be the most useful tools for the work, are openly available (i.e. do not require a commercial license), provide support for emerging Semantic Web standards, and have active communities of interest with strong representation from researchers and knowledge workers in health, biomedical and other related fields. The steps followed in the public health classification building process were: Step 1 Determine the domain and scope of the classification: -What is the domain that the classification will cover? -For what are we going to use the classification? -For what types of questions should the information in the classification provide answers? -Who will use and maintain the classification? Step 2 Consider reusing existing classifications. Step 3 Enumerate important terms in the classification. Step 4 Define the classes and class hierarchy. 24 Public health definitions and relevant classification systems, especially functional classifications, were reviewed for Step 2 and are available from the project. The Project Reference Group workshopped the preliminary material and drafted initial responses to Steps 1 to 4. Consultations with public health content experts in a sample of jurisdictions considered the class hierarchy and its top levels, and identified important omissions. They also identified further practical uses of the classification. Initial consultations were held in NSW from October 2004. Formal consultations were held in Brisbane, Melbourne, Canberra and Perth. Early consultations were informal, designed to seek the views of content experts on particular components (e.g. environmental health, health promotion). Later consultations were organised through Reference Group members representing various jurisdictions. Prepared material introducing the project was sent out to participants prior to each consultation. All consultations were face to face. The number of participants varied from one or two, to larger groups of up to fifteen, and the duration varied from one to three hours. In each formal consultation, an introduction and background to the project were given with the aid of a slide presentation, then an early version of a public health classification, rendered through a Web browser, was demonstrated, concluding with 22 Noy & McGuinness 2001. 23 For more information see http://protege.stanford.edu. 24 Noy & McGuinness 2001: 5-8. the definition of public health. This was followed by a live collaborative session using the Protégé software, during which changes and additions to the class structure were made in real time. Lastly, participants were asked to identify further practical uses for a unified public health classification. An example of the agenda and other pre-consultation material that was sent to participants is in Appendix B. The views, suggestions, and additional information captured in consultations were discussed by the Project Reference Group over a series of meetings and have informed the broad structure of the public health classification that is reported in Section 3 Results. The public health content experts who contributed are acknowledged in Appendix B. An earlier version of this report was presented to, and discussed by, the National Public Health Information Working Group in March 2005, and this version reflects the feedback and directions given by that Group. This section presents the results of the process of scoping the domain to be covered by a public health classification. A number of boundary issues are discussed, areas of likely agreement identified, and potential practical applications for the classification are outlined. Version one of the public health classification is presented. Issues for further consideration are highlighted in boxes. The existing National Public Health Partnership definition of public health was adopted, as follows: Public health is the organised response by society to protect and promote health, and to prevent illness, injury and disability. The starting point for identifying public health issues, problems and priorities, and for designing and implementing interventions, is the population as a whole, or population sub-groups. 25 Suggestions made during consultations that the Partnership definition should include references to 'evaluating' and 'measuring or achieving outcomes' were not adopted, as these were considered to be implicitly present in the definition. 26 Significant boundary issues were encountered in scoping the domain of public health, with disagreement among public health experts regarding where the boundaries are, or should be. While most public health experts agreed, when pressed, that accounting for public health should include the activities of, and investments by, the non-health portfolios of governments (such as education and transport), local governments and nongovernment organisations (NGOs), current public health expenditure reporting is largely limited to that by State, Territory and Australian Government health portfolios. 27 One view was that the activities of other (non-health) sectors should only be counted when public health is their primary purpose (e.g. immunisation organised by local government). In practice there are major difficulties in capturing information on public health activities and expenditure by non-health sectors. 28 The boundary between public health and clinical practice came up repeatedly in discussions about the scope of public health. In many situations preventive activities in clinical practice complement broader population-based activities. At what point do 25 NPHP 1998. 26 See Appendix C for additional information on this aspect of the Project. 27 With the exception of SA which has in the past included non-health expenditures by local government, etc, in public health expenditure reporting by AIHW. 28 A more fundamental difficulty is the time and expense to collect comparable information across all sectors. they become part of 'the organised response by society to protect and promote health'? Organised interventions for promoting health and preventing illness, injury and disability include those aimed at whole populations that do not necessarily require any particular action on the part of individuals (health protection activities, e.g. the provision of clean water, clean air, sewage disposal), and those organised and delivered at the level of the population or sub-group, but requiring individuals to modify their behaviour (health promotion activities, e.g. the range of activities to reduce smoking in the community-regulations, media campaigns, organised quit lines etc). There was general agreement that health protection and health promotion are public health activities. There was more debate in relation to preventive services delivered on a one-to-one basis to individuals. Such preventive services include screening, immunisation, and counselling and lifestyle advice to support healthy behaviour, as well as detection and management (through lifestyle changes or pharmacological means) of biological risk factors such as high blood pressure and high cholesterol. The perceived boundary between public health and clinical medicine is likely to change as new screening technologies and preventive medications become available. Some public health practitioners argued that those individual preventive services related to communicable diseases (e.g. immunisation, contact tracing, treatment for STIs) form part of public health practice because they help to protect the health of the whole population, through herd immunity and reducing the spread of infection. A minority argued that immunisation is only a legitimate part of public health activity when it is delivered as part of a publicly organised program, such as through local government or school health services. The corollary of this point of view is that immunisations performed in general practice are not a public health activity. Alternatively, it was argued that childhood immunisation in general practice is simply the implementation strategy for an organised national approach to immunisation, one which is supported by special payments to GP's and the Australian Childhood Immunisation Register, with follow-up of those parents who do not comply. With respect to non-communicable diseases, early detection through screening is a preventive service delivered one-to-one to individuals. Some public health practitioners felt that this is only a public health activity when it is delivered through an organised program such as BreastScreen. Cervical screening is largely delivered through general practice, although, as with immunisation, the delivery of services in the private sector is underpinned by the National Cervical Screening Program (State and Territory recruitment programs, Pap smear registers, follow-up and reminder systems). The question is whether the taking of smears (in general practice) and the reading of smears (in laboratories), which are largely in the private sector, but essential to the implementation of the program, should be considered as public health activities. On the other hand, opportunistic screening that is not part of an organised program, such as bone density screening for osteoporosis, was not generally considered to be a public health activity. Even more contentious was the issue of the prevention and management of non-communicable diseases, through one-on-one counselling about lifestyle risk factors (e.g. smoking, poor nutrition, risky alcohol use and lack of physical activity), and the early detection and management of biological risk factors such as high blood pressure and high cholesterol in the prevention of heart disease and stroke. Many public health practitioners regarded these activities as clinical practice. Some suggested that a distinction can be made on the basis of whether people have symptoms or signs of disease. For example, helping people to quit smoking would be considered a public health activity when they are symptom free, but part of clinical medicine if they have any symptoms or signs of disease or a history of previously diagnosed disease. Apart from any conceptual objections to such a distinction, it would be difficult to operationalise in practice. Further along in the disease continuum, most public health practitioners classified the effective management of chronic disease, with the goal of minimising disability and reducing complications and hospitalisations, as belonging firmly in the zone of clinical medicine. For example, the prescribing of cholesterol-lowering medication by a general practitioner, even in an otherwise healthy person, would not be considered a public health activity (although a media campaign urging people above a certain age to have their cholesterol levels checked by their GP might be regarded as public health). These boundary issues are set out for further consideration in Box 2 in Section 3.1.6. The Public Health Performance Project 29 envisaged that a unified classification for public health would be used to progress national public health expenditure reporting, 30 public health performance indicators, 31 and to build on the public health core functions developed by the National Public Health Partnership. 32 Potential uses and practical applications for a public health classification, identified during phase one of the current project, are summarised in Box 1. Explain what public health is Organise information to answer key public health questions Promote consistency in describing public health Improve data capture processes and the quality of reporting A public health classification will help to explain what public health is in a way that is recognisable and understood by the average person. It will allow description of the functions of public health, issues of public health concern, the settings in which public health activities occur, the population groups targeted by public health interventions, the resources available to public health, and so on. The process of developing a classification has the potential to unite the sector and improve understanding of the breadth of the public health effort. A public health classification can be used to organise information to answer key questions for public health that cannot be answered currently. While agreement on the scope of public health proved contentious during consultations, formulating questions that a competent public health classification should help to answer was somewhat easier. Questions like those shown in Box 2 set a practical test for the classification. Box 2 A public health classification should help answer questions like... What is public health? What are the characteristics of public health? How is public health relevant to components of the human services delivery system? Why do public health unit costs differ across jurisdictions? Can we describe screening in clinical settings (e.g. Pap smears taken in GP surgeries)? What are the nature and cost of public health partnerships between health and other sectors? Can we replicate the output of other models (e.g. current public health expenditure reporting)? How much was spent on social marketing last year? 33 As well as organising and integrating public health information, the development of a common classification will promote consistency in describing public health, through the standardisation of definitions and terminology. This will improve data capture processes and the quality of reporting (e.g. in expenditure and performance reporting). Promoting consistency will increase the ability to compare public health information over time and across jurisdictions. 34 There is potential for a public health 33 Additional questions include other advocacy-type questions, such as: what is the relative expenditure on specific risk factors or diseases? What is the difference in expenditure on prevention of HIV/AIDS relative to other preventable diseases? Has health funding to preventive/promotive investments increased?. There are also boundary questions such as: can we describe the hospital interface with public health interventions (e.g. screening in hospitals)? Can we calculate expenditures in specific areas (e.g. product safety and protection, public health emergencies, education as a health promotive activity)? Competency questions can be used as a 'litmus test' to help determine whether the classification contains sufficient information to answer them, and whether the answers require a particular level of detail or representation of a particular area (Noy & McGuinness 2001: 5) . 34 Recent reporting of public health expenditures over several years has enabled such analyses for the first time (AIHW 2004b). classification to be used to improve jurisdictional public health financial processes (e.g. budgeting, resource allocation) and accounting systems (e.g. through developing systems that can apportion public health activities to cost centres or to aggregate Treasury outputs). 35 A public health classification will contribute to higher-level classification and standards activities through the potential membership of the Australian Family of Classifications. The development of the classification could 'fill out' the public health cells and embed public health more firmly into the 'health and related classifications matrix'. 36 A public health classification can be used to structure and design information and communications (e.g. in designing websites, structuring resources, and planning report chapters). It has practical applications in building information systems, such as a database of public health projects, using the classification to create explicit, structured information to make meaning (as well as documents) accessible and shareable. One test application proposed was for a public health equivalent of the Semantic Web Environmental Directory (SWED). 37 This could be created through web-based, universally available tools, that make it easy for public health people to describe what they do, using the classes and terms from the public health classification. Other uses are based on a broad vision of a public health classification as signposting or semantically indexing a wide range of resources (including but not limited to: thesauri, dictionaries, terminologies and definitions, scientific papers, reports and other documents, legislation, policies, information databases and indexes, case studies, stories and vignettes). A further use for a classification identified in consultations is to audit the spread of public health activity, expenditure or investment, across the business cycle -from health problem identification and assessment to program or intervention planning and design, through to implementation and evaluation of results. This suggestion arose out of concerns that public health activity is too heavily weighted towards implementation, and that there is insufficient evaluation of interventions, and learning from and progressing beyond pilots. A related use is to examine the spread of all public health investments, for example, by Australian, state, territory, and local governments, NGOs and other investors; the links to employment and education; and public health investment by, and outcomes in, other sectors such as transport and housing. A classification can potentially be used to help build models of good public health practice that describe the program logic for public health activities, including specification of the links between activities, expenditure and outcomes. Another suggested use for a classification is in developing a continuous improvement model to ensure that public health learns from what it does. 38 Lastly, the classification was considered to have the potential to link public health research, policy and practice, by facilitating use of a common language, and the linkage of information across these domains. Potential users of a public health classification, who were identified along with the practical applications discussed above, are the various levels of government and other sectors that have an investment in public health. Other users include academics and students, researchers, evaluators, those involved in policy formulation, and anyone with an interest in public health. The Australian Institute of Health and Welfare has indicated an interest in the longerterm development and maintenance of a public health classification. During the development of the public health classification, the following principles of development were determined and agreed: The classification system should be multi-dimensional to be able to represent the multi-dimensional nature of public health. Different dimensions are of equal importance to public health and a range of the most important need to be considered and developed concurrently. Existing classification systems of relevance (including Australian and international standards) should be used wherever possible in the multi-dimensional structure of the classification system. The system should be inclusive (rather than exclusive) and deliberately broad at the top levels. Boundaries can be set (or moved) as needed for particular practical applications; they should not be used to restrict or hinder the development of a broad and inclusively scoped classification. In addition to the definitional issues raised and discussed in Section 3.1.2 above, public health experts consulted in phase one of the project raised the important issue of whether the name of the project domain should be 'population health' or 'public health'. 39 How is the domain that public health currently works in, best described? Is 'public health' subsumed in 'population health'? Or is a 'population health approach' merely one aspect of public health practice today? The concept of population health has its origins in the Canadian Lalonde Report in 1974, which promoted the (then radical) idea that health and well-being involve more than the health care system, and that the adoption of healthier lifestyles, and improvements in people's social and physical environments, would be the principal means of improving the health of Canadians in the future. 40 Population health, as a way of acting on the social and economic forces that structure health, builds on a tradition of public health and health promotion that goes beyond a focus on the medical, biological or lifestyle problems of individuals. 41 A population health approach can be defined as a subset of public health with a whole-of-population focus, 42 or as containing both public health and other health services. 43 Population health is not the only term that is sometimes misleadingly 39 An alternative would be to include both terms in the domain name. 40 Lalonde 1974. 41 Hayes & Dunn 1998. The population health approach is not without its critics, some of whom argue that it has been captured by the focus on the problems of individuals (e.g. overweight persons), while losing sight of the larger issues (e.g. obesogenic environments) (Raphael & Bryant 2002) . 42 Bennett 2003: 12. 43 For instance, a 'population health approach describes a comprehensive health system which ranges from public health at one end to individual health care at the other' (Buckett & Hunter 2004) . Fraser (2005) conceptualises population health as 'the health of a defined population, or a field of study that links health outcomes, determinants of health, and interventions' but notes that it is an 'ill-defined term' in the literature. The term public health has competing definitions, but is considered by many health professionals to be 'broader and more encompassing than population health' (Fraser 2005: 177) . The decision on what to call public health is partly semantic, as the domain called 'public health' has changed over time. 'Classical' or 'traditional' public health had an external, environmental focus and produced major infrastructure projects such as sewage and safe drinking water systems, and other improvements to the human environment. Figure 3 shows changes in the conceptualisation of public health over time in two axes: populationindividual and proactive-responsive. In the figure, quadrant D describes the 'new' public health (and 'social' health, with a health equity focus) as a proactive population approach. contrasted with public health and adds to the confusion about what public health is. Figure 4 shows how such confusion can arise from the intersection of public health with other perspectives on health-such as, a population health approach, and definitions of preventive health, and primary health care. Does it matter what the domain is called? The term 'population health' was preferred over 'public health' in several consultations during phase one of the project. A sampling of jurisdictional health departments showed that population health has overtaken public health in popularity as the name for the relevant organisational units (see Appendix E) . At other consultations, the term public health was strongly preferred to population health as the name of the domain (although a 'population health approach' was allowed as a method used by public health). There is also a widespread view among public health experts that the general public commonly confuses or equates public health with public hospitals, or the health system funded from the public purse. Some practitioners saw the rise in the popularity of the term 'population health' as an opportunity to gain agreement on an allencompassing definition and to replace the often misunderstood term 'public health'. As was noted in the discussion in Section 3.1.2 above, the boundary between public health and clinical medicine is contentious, and both the boundary and the components included in each are likely to change over time. There may never be complete agreement by all experts, but the act of making components and boundaries explicit can at least facilitate discussion on these difficult issues that are summarised as discussion points for further consideration in Box 2. 'Public health is the organised response by society to protect and promote health, and to prevent illness, injury and disability. The starting point for identifying public health issues, problems and priorities, and for designing and implementing interventions, is the population as a whole, or population sub-groups.' (NPHP 1998) What is the preferred name for the domain of public health today (population health, public health, public and population health)? How is 'organised response' defined? Is there agreement on the following examples of organised response? a. The breast cancer screening programme supervised by BreastScreen Australia; b. Screening for cervical cancer by GPs underpinned by registers, recall systems, and target population monitoring; c. GPs undertaking opportunistic screening for high cholesterol, in accordance with published National Heart Foundation guidelines, in patients consulting them for an unrelated matter? How is public health differentiated from clinical treatment services? When are treatment services -for example, treatment of sexually transmitted diseases or tuberculosis -part of public health? Does the place of delivery of services determine that a service is or is not a public health service? For example, is an immunisation delivered in a dedicated local government or school immunisation clinic different from an immunisation delivered in a hospital emergency department? Should the domain of public health be solely within health or should it include specific activities of other sectors (e.g. education, transport, local government) that have public health as a primary purpose? Or as a secondary purpose? One suggested response to these questions is for a checklist approach that operationalises the agreements realised in scoping the public health domain. This could be used to determine whether an activity is public health or clinical care, for instance. The checklist components could be weighted, so that an activity that meets one 'must-have' and two out of three other criteria is defined as public health. The checklist could test whether the activity is preventive, (e.g. primary or secondary reason for service is to prevent the need for acute care; treatment for sexually transmitted disease is to prevent transmission of disease); whether it benefits a population (this does not preclude services to individuals -the benefit could be to an individual and a population, e.g. immunisation); whether a public health response is required in addition to (any) individual treatment response required (e.g. assess area for contaminant after individual exposure, check cooling towers in response to case of Legionnaires disease, trace contacts of person diagnosed with infectious disease); whether it is an organised response, for instance, in response to a disaster, over time (e.g. immunisation register), or in scale (e.g. screening across the nation, quality assurance through pathology reference laboratories). The most important dimensions (or top-level classes) revealed in an analysis of the National Public Health Partnership public health core functions 44 were the functions of public health, 45 and the methods that public health uses to achieve those functions. A selection of other candidate top-level classes was made in order to focus the project. Those initially chosen for detailed examination were: public health functions and activities or programs that funds buy (e.g. public health expenditure activities); determinants of health, health risk and protective factors (e.g. socio-economic determinants, behavioural factors); disease, disability, and injury areas (e.g. vaccine preventable diseases) that determine intervention targets; and the public health 'toolkit'-methods, tools, and bodies of knowledge, both those specific to public health (e.g. epidemiology, health promotion techniques) and those used by but not specific to public health (e.g. management methods, policy development frameworks). These potential classes underwent extensive development and revision and are shown in Figure 5 as they stand at the conclusion of phase one of the project (working definitions are in Table 2 ). Potential classes that were identified but not selected for detailed examination are discussed in Section 3.2.5. Public health practitioners expressed both broad and narrow views of what a classification system for public health should include. These views reflect the range of practical applications they identified (detailed in Section 3.1.3), and their underlying requirements. For instance, for health expenditure reporting, mutually exclusive activity categories at meaningful expenditure levels are required. From a health promotion viewpoint, the ability to model the public health business cycle, and to identify gross expenditure proportions for different elements (e.g. design, implementation, evaluation) are equally important. There was however, consensus among the public health experts consulted, that a public health classification should be multi-dimensional, and there was broad agreement on the top-level classes that should be included. There was agreement that public health 'functions' form an important class, although there was some confusion regarding whether functions refer to the purposes of public health activities or the methods of intervention by which public health achieves its aims (see working definitions in Table 2) . 44 NPHP 1998. 45 The word 'function' is used here in the sense of 'the purpose, role or use of something'; thus, the function of public health is 'to protect and promote health, and to prevent illness, injury and disability' (NPHP 1998). There was also wide agreement that both 'health issues' and 'determinants of health' are central to public health, although there are differing views on the relative importance of individual determinants and how they should be structured at lower levels of the classification. The inclusion of a 'settings' class was also generally agreed. The project involved extensive discussion and work regarding how to define the practice of public health, the methods and strategies used in public health interventions, and the bodies of knowledge that these draw on. There were two strong perspectives on what should be included in a classification. One perspective was restrictive and would narrow the scope of a 'methods' class to those methods that are peculiar to -or only used by -public health (e.g. population-based epidemiology, health promotion, environmental risk assessment). The other focus was on capturing all methods used by public health, including those that, while not specific to it, are employed by public health workers in the normal course of their work (e.g. administration, management, policy development). A 'resources' class was elevated in importance when consultations reinforced the importance of the many types of infrastructure on which today's public health relies: physical infrastructure (e.g. sewers, public health laboratories), organisational infrastructure (e.g. partnerships, legislative and regulatory systems), logistical infrastructure (e.g. vaccine cold chains) -systems that are seen by some as 'joined up' resources. There were diametrically opposed views of whether infrastructure was a subclass of resources or vice versa. In the short term this has been dealt with by amalgamating the two into a 'resources and infrastructure' class. In consultations many public health experts wanted to add a 'policy' class. There are several elements to be described. One element is the public health work of developing healthy public policy. Whether or not policy is implemented, substantial work goes into its development, and its availability can provide a head start for action on a health issue that becomes of interest. Information on existing public policy that has an impact on public health is considered by some as important to collect and integrateespecially in the absence of a national public health policy. Some public health experts were comfortable with the concept of 'policy' as a resource or as part of the public health infrastructure, however others were strongly negative -they saw that as putting policy too low in the class hierarchy. This reveals the tendency to see the toplevel classes listed in Figure 5 as a hierarchy of the factors of most importance to public health, in which case, where is policy? Where are population groups? The discussion in Section 3.2.3, which is illustrated in Figure 7 , addresses these questions. Similarly, the addition of an 'outcomes' class was identified as important at almost every consultation, reflecting a view that outcomes (i.e. outcome indicators and their reporting 46 ) are necessary to 'close the loop' and complete the program logic for 46 For example, public health system performance measures and public health expenditure reporting. public health. This reflects a tendency to see the top-level classes listed in Figure 5 as a program logic or cycle (rather than a hierarchy of important factors) that requires information on outcomes to complete the cycle. An alternative view on the treatment of outcomes in a public health classification is that they are already captured in the classes of 'health issues' and 'determinants of health'. Section 3.2.3 also addresses these issues. Health issues Health, and well-being issues that affect health ('issues' includes: concerns, topics, problems). Health is defined (by the WHO) as 'a state of complete physical, mental and social well-being and not merely the absence of disease or infirmity'. Factors that influence health status and determine health differentials or health inequalities. They include, for example, natural, biological factors, such as age, sex and ethnicity; behaviour and lifestyles, such as smoking, alcohol consumption, diet and physical activity; the physical and social environment, including housing quality, the workplace and the wider urban and rural environment; and access to health care. 47 The methods used by organised public health interventions (actions, activities, programs, services) to protect and promote health and prevent illness, injury and disability, that are designed to change population exposure, behavioural or health status. Settings in which public health activities and interventions take place, institutional and social environments, partnerships, and locations (e.g. schools, local government, hospitals, workplaces). Resources and infrastructure, 'the means available for the operation of health systems, including human resources, facilities, equipment and supplies, financial funds and knowledge'. 48 It includes both person-time and calendar time. Considerable development of the functions 49 or purposes of public health took place during phase one of the project. As discussed above, the National Public Health Partnership (NPHP) public health core functions 50 were analysed and distilled into the individual functions of the 'functions' class of version one of a public health 47 Based on WHO 2005 , citing Lalonde 1974 Labonté 1993 . 48 WHO 1998a Function is defined as 'the purpose, role or use of something'; thus, the function of public health is 'to protect and promote health, and to prevent illness, injury and disability' (NPHP 1998) . 50 NPHP 1998. classification presented in this report (the proposed treatment of other components of the public health core functions is shown in Table 5 ). 51 The functions as scoped at the end of phase one of the project are shown in Table 3 and their working definitions are given in Table 4 . Both primary and instrumental functions are of importance in conceptualising public health. Primary functions are ends in themselves, while instrumental functions are means to those ends, as without primary functions there would be no need to 'ensure public health capability', for instance. Instrumental functions were also described in consultations as supporting, underpinning, or crosscutting functions, as all primary functions rely on them and they do not belong exclusively to any one of the primary functions. Although the instrumental 'build the evidence base...' function could be included in 'ensure public health capability', it is shown separately because building an evidence base and moving towards decisions informed by evidence are key features of the current context for public health. Other functional classifications of public health were explored during the course of the project, including that portion of the OECD System of Health Accounts that is relevant to public health. 52 The OECD classification has a similar mix of classes within functions as do the public health core functions, but excludes environmental 51 The public health core functions (NPHP 1998) are shown in Table 5 . 52 'Prevention and public health services' defined (in part) as services 'mainly of a preventive nature and ... publicly provided' which include 'special public health services such as blood-bank operation, public health service laboratories, and population planning services' (OECD 2000: 44) . health, and was structurally not helpful. 53 A functional division that followed the distinctions between primary, secondary and tertiary prevention was also explored but the classification was confusing and difficult to apply, and there are arguments that tertiary prevention in particular, has more relevance to clinical treatment services than to public health. A comparison of the public health functions of selected other nations (see Figure 6) shows that, in the UK for instance, both primary (e.g. health promotion and disease prevention programs) and instrumental (e.g. development and maintenance of a public health workforce) functions are prominent, while the public health functions of Canada and the Americas are limited to primary functions. 54 Both the UK core functions 55 and the USA essential public health services include a specific (instrumental) partnership function for public health. In Australia, the essential importance and defining nature of inter-departmental, inter-governmental, inter-sectoral and other partnerships, in the work of public health was made clear in the expert consultations. Accordingly, version one of a public health classification proposes 'build public health partnerships' as a subclass of the 'ensure public health capability' instrumental function (see Table 3 ). 53 Dimensions used by the OECD are: population groups, service types, disease types, and settings. 54 A recent review conducted by WHO (2003), describes comparable 'essential public health functions' as 'a set of fundamental activities that address the determinants of health, protect a population's health, and treat disease... public health functions represent public goods, and... governments would need to ensure the provision of these essential functions, but would not necessarily have to implement and finance them. They prevent and manage the major contributors to the burden of disease by using effective technical, legislative, administrative, and behaviour-modifying interventions or deterrents, and thereby provide an approach for intersectoral action for health [that] stresses the importance of numerous different public health partners. Moreover, the need for flexible, competent state institutions to oversee these cost-effective initiatives suggests that the institutional capacity of states must be reinforced' (Yach 1996 cited in WHO 2003 : 1, our italics). 55 The full description of this function is 'Creating and sustaining cross-Government and intersectoral partnerships to improve health and reduce inequalities' (CMO UK 2003; see chapter 3). The UK core functions also include a specific (instrumental) research function. In Australia, although 'public health research' is one of the nine core public health activities for which public health expenditure is reported, 56 there is no corresponding function in the NPHP public health core functions. Version one of a public health classification proposes 'conduct public health research' as a subclass of the 'build the evidence base for public health' instrumental function (see Table 3 ). A health surveillance function is common to both the UK and Canada. It is broadly specified in the UK as 'health surveillance, monitoring and analysis', while in Canada the function of 'population health assessment' is specified separately, in addition to 'health surveillance'. 57 In Australia, the first of the nine NPHP public health core functions is 'assess, analyse and communicate population health needs...' (and is proposed as 'assess health of populations' in version one of a public health classification -see Table 3 ), although expenditure on this public health activity is not currently reported in an identifiable manner. A quality assurance function is specific for public health in both the USA and the UK. Whether such a function is pertinent to public health in Australia is a matter for discussion and has not been canvassed in consultations. Working definitions of the public health functions proposed in Table 3 are given in Table 4 . The working definitions are based on NPHP public health core functions 58 and extensive discussion during the project. Some of the major strands that emerged in discussions, and their impact on the working definitions, are reported below. The 'promote health and prevent disease, disability, and injury' function was initially cast as two functions, with 'promote better health' separate. In examining the mission statements and goals of health promotion and prevention units across the jurisdictions it was clear that there was no hard boundary between the promotion of health and the prevention of disease, disability and injury. In consultations it was suggested that the two functions should be married together as the distinction is increasingly blurry in practice. They have thus been joined as one function at the top level. 'Develop healthy public policy' was initially classed as a subclass of a 'promote better health' function. In consultations it was pointed out that this function, method or strategy was cross-cutting, applying to all primary functions, and should not be singled out as belonging only to one function, or as separate to all other functions. 'Policy development' was thus classed as a 'method' of intervention so that it can be applied to any or all of the public health functions in an additional dimension. 56 AIHW 2004b. 57 The post-SARS Canadian view is that: 'Among the functions of public health are health protection (e.g. food and water safety, basic sanitation), disease and injury prevention (including vaccinations and outbreak management), population health assessment; disease and risk factor surveillance; and health promotion. The public health system tends to operate in the background unless there is an unexpected outbreak of disease such as SARS or failure of health protection as occurred with water contamination... Monitor health Monitor and analyse levels of health and its determinants in populations to identify and predict trends and emerging issues ('Assess health inequalities' would be a further subclass of this). Evaluate health risks and benefits Evaluate adverse and beneficial effects related to health and social policies and interventions, and environmental exposures. Assess health inequalities Assess inequalities in health (level and distribution) and health gain to target interventions to improve the health of the worst-off sub-populations. Protect from threats to health Protect from, and prevent, external threats to public health. Prepare for threats to health Identify and prepare for potential threats to health (including communicable diseases, environmental hazards, bio-terrorism and new patterns of exposures e.g. arising from ecological change). Respond to threats to health Respond to threats to health (including communicable diseases, environmental hazards, bio-terrorism and other disasters). Control and mitigate risks to health Minimise or reduce the severity of risks to health (includes setting and monitoring of standards for e.g. food, air and water quality and other potential hazards, also harm minimisation measures). Promote health and prevent disease, disability and injury Promote health and wellbeing, prevent the occurrence of disease, disability and injury; and detect disease in its early stages, through organised efforts that target populations. Promote health and wellbeing Promote better health and well-being as it affects health (e.g. community development and community empowerment initiatives clearly differentiated from 'Prevent the occurrence of...'). Prevent the initial occurrence of disease, disability and injury (e.g. population-level campaigns to promote physical activity, tobacco control, seat belt legislation). Detect disease, disability or injury early Detect disease, disability and risk of injury early and initiate prompt management or response (e.g. screening for cancers, newborn hearing screening). Ensure public health capability Ensure adequate public health capacity and responsiveness by maintaining and developing the public health workforce and infrastructure, and building partnerships with other sectors of society. Develop and maintain the public health workforce Train, maintain and develop the public health workforce. There was also a view that a 'promote better health' function should be expanded to 'enhance health and quality of life,' to incorporate the concepts of: (1) effort from non-health sectors that affects public health, and (2) quality of life and health maintenance (rather than improvement) where the presence of disease makes health improvement an inappropriate aim. Agreement on these definitional extensions was lacking in further consultations and they have not been adopted. The working definitions in Table 4 are shown as they stand at the conclusion of phase one of the project. They should be regarded as a work-in-progress and a point to move forward from, rather than the definitive last word on the public health functions. The relationship between the 'functions' class and other top-level classes in version one of a public health classification and the National Public Health Partnership (NPHP) public health core functions 59 is shown in Table 5 . The table illustrates how the public health classification can be used to achieve a functional equivalence to the several dimensions implicit in the NPHP public health core functions. The multi-dimensional core functions can be classified using different top-level classes of the classification (e.g. 'health issues', 'methods'), and instances (see Figure 7 ). For example, the function or purpose of core function two (shaded in Table 5 ) is to 'Prevent and control communicable and non-communicable diseases and injuries' using the public health intervention methods of 'risk factor reduction, education, screening, immunisation and other interventions'. Assess health of populations Protect from threats to health Promote health and prevent disease Risk factor reduction, education, etc classified as 'methods', and instances described as Interventions. Communicable and non-communicable diseases etc classified as 'health issues'. Promote health and prevent disease Ensure public health capability: Build partnerships Action with individuals, families, communities etc classified as 'methods', instances described as Interventions, and families, communities described as Population Group instances. All. Public policy measures classified as 'methods', and instances described as Interventions. Ensure public health capability Build the evidence base for public health Plan, fund, manage and evaluate classified as 'methods', and instances described as Interventions. Programmes described as instances of Public Health Activities. Promote health and prevent disease Ensure public health capability Consultation, participation and empowerment classified as 'methods', and instances described as Interventions. Protect from threats to health Promote health and prevent disease Actions described as instances of Public Health Activities. Promote health and prevent disease Healthy growth and development classified as a 'health issue' (e.g. 'health and well-being'). Life stages described in Population Groups. Protect from threats to health Promote health and prevent disease Individual vulnerable groups described as Population Groups classified by other classes (e.g. person-level demographic descriptors in 'determinants of health'). Actions described as instances of Public Health Activities. Core function nine (shaded in Table 5 ) 'Promote, develop and support actions to improve the health status of Aboriginal and Torres Strait Islander people and other vulnerable groups' identifies important target populations, rather than describing a separate function of public health. Functions, methods, and population groups thus form three distinct dimensions (among many) of interest in a multi-dimensional classification of public health. Few of the nine public health core functions have a one-to-one relationship with the functions of the public health classification, if the functional equivalence shown in Table 5 is accepted. Core function four 'Promote, develop and support healthy public policy, including legislation, regulation and fiscal measures' requires special mention, as it is shown as relevant to all the functions of the public health classification. It is proposed that public policy measures are methods to address all functions rather than a function in their own right. 'Public policy development' is thus shown as a separate method in Table 6 , as is 'legislation and regulation' (which some see as enacted policy). Public health activity using these methods can have a major impact on population health. Examples include the impact on population smoking rates of legislation, regulations, and fiscal measures implemented under the policy umbrella of the Tobacco Control Strategy. Although the multi-dimensional structure of the public health classification is quite different to the flat list structure of the public health core functions, its classes can be used in a functionally equivalent way to classify and describe the functions and other important dimensions of public health. The public health dimensions currently scoped, and their top-level classes are shown in Table 6 . The 'functions' class has been discussed in Section 3.2.2. While there was reasonable agreement among the public health experts consulted over the top levels of the classes of 'health issues' (although its name was debated), 'determinants of health', and 'settings', the remaining classes are in the early stages of development and have not yet been subject to detailed consideration. The 'methods' class, in particular, established to describe the methods of public health intervention, is at an early stage of development. While population groups are important, it was generally agreed that they are not a toplevel class in a public health classification. As the targets of public health interventions, instances of population groups can be described by other classes in the classification, such as the person-level demographic descriptors in the 'determinants of health' class (e.g. age, sex). There was also agreement that stakeholders and partners, although important in the work of public health, did not warrant their own top-level class. As with population groups, they may also be described by other classes in the classification. This distinction is illustrated in Figure 7 , which distinguishes between classes in the classification (circles) and items to be classified (heptagons). The latter include (but are not restricted to) public health activities and programs (centre), public health interventions, public policies, outcomes (indicators that are useful for public health purposes, and those that are nationally reported), population groups, partners and stakeholders in the public health effort. Figure 7 also shows whether suitable classifications exist for use by the top-level classes, or whether they need to be developed. Existing classifications (e.g. Australian standards, international classifications of diseases, functioning and disability, external causes of injury) are available to classify major parts of the 'health issues', 'settings' and 'resources' classes. The National Public Health Information Working Group has determined that further development of classifications for the 'functions', 'determinants of health' and 'methods' classes are a required priority for the second phase of the project. Not all public health experts will agree with the constituent parts of the classes as they stand, and some important parts are undoubtedly missing. The project anticipates feedback on these issues through making these results more widely available. During phase one of the project, some of the practical uses that had been identified were developed in a small way in order to test the usefulness of the classification. Two examples-of public health activity from national public health expenditure reporting, and details of public policies-are detailed below. Information on some recent developments of interest in public health classification in the UK can be found at the end of Appendix B. A selection of public health activities from public health expenditure reporting were classified using the top-level classes of the public health classification. The detail of an example public health activity is shown in Figure 8 . The symbol denotes classes and subclasses, while the symbol denotes 'instances' or individual cases, for example, an individual public health activity, partner, stakeholder, or population group. On the left of the figure is a list of public health activities extracted from the latest public health expenditure report 60 , and to the right are the details of a selected activity, characterised by a number of 'slots' or attributes of the activity. The selected activity is Queensland's 2000-01 health promotion initiatives, on which $18.7 million was expended. The example shows the variety of health issues and determinants addressed (sun protection, healthy diet, and so on) for population groups. Queensland's 2000-01 health promotion initiatives are classified by the public health expenditure core category of 'Selected health promotion', as used in national public health expenditure reporting, 61 while the (main) function or purpose is to 'Promote health and prevent disease, disability and injury' (using the public health functions developed in this project). Associated public health intervention methods used in the health promotion initiatives are also listed (e.g. intersectoral advocacy, community action). Partnerships and stakeholders are shown as test data. This classification of a public health activity is much better than a one-dimensional classification at answering the questions listed in Box 2 in Section 3.1.3 as a practical test for the classification. For instance, in response to the question 'How much was spent last year on the prevention of obesity?', Figure 8 shows that public health activities for which the function is 'prevention' and the health issue is 'obesity' can 60 AIHW 2004b . 61 AIHW 2004b easily be identified, and the values in the 'expenditure' slot (attribute) for these activities can then be summed. A selection of public health policies compiled from publicly available documents accessible on the internet were classified using the top-level classes of the public health classification. An example public health policy is detailed in Figure 9 . As previously, in Figure 9 , the symbol denotes classes and subclasses, while the symbol denotes 'instances' or individual cases, for example, a particular public health policy. Figure 9 shows detail on the Australian Government Draft National Injury Prevention Plan (NPHP 2004) and the health issues it addresses (external causes of injury, safe home environment, and so on). The plan is assigned to the function subclass 'Prevent occurrence of disease, disability, and injury'. Capture of the URL for the published policy allows rapid access to the policy through the internet. Details of the jurisdictions and/or portfolios that have endorsed the plan can be captured in additional slots. These examples do not completely illustrate the full power of a 'third generation' multi-dimensional classification for public health, developed using a formal ontologybuilding tool such as Protégé. While nothing can replace human knowledge and intelligence in the comprehensive collection, description (classification, indexing) and use of complex information, in the future it is envisaged that semantic tagging 62 of documentation and other written resources will allow much more meaningful information to be routinely made available to humans, through machine processing of this 'computable' information. More information on this aspect of the project is presented in Appendix D. In addition to those top-level classes discussed in detail above, other potential classes were identified in the first round of development. These included: Geography/access to health services (e.g. urban/rural/remote geographic classification). Intervention target or focus (e.g. target population defined by age, sex, ethnicity) and intervention type. Performance measures (e.g. the national health performance framework). Precepts, principles, philosophy (e.g. equity). Service production/provision (where service is produced/provided e.g. institutional health services, non-institutional health services) and service delivery/settings (where service is delivered e.g. school, workplace, community). Sources of funds (e.g. health/non-health; levels of government). Theories and models (e.g. 'harm minimisation', 'user pays'). Time (e.g. incubation periods, time-lags, investment periods, break-even points). Workforce (e.g. public health specialists, local government workers, school nurses). Contextual/macro-environmental/ecological factors that affect but are outside the influence of public health (e.g. factors that would be picked up in environmental scanning). Interventions as public health activities/strategies that are related but different to methods. Outcomes, including outcome indicators and reporting (e.g. national public health system performance measures), necessary to 'close the loop' and complete the program logic for public health. Policy including various views: policy development as an activity or 'method' or a cross-cutting component of all functions; policies as a class of things in existence (e.g. as in a policy register or library); policy as enhancing understanding of practice, cross-referenceable to other areas of interest. Population groups as defined in terms of attributes and characteristics from other classes (e.g. age, sex). Research/evidence allowing integration with the university sector, to link research and policy and practice, and to build the evidence base for public health. Risk factors (part of the 'determinants of health' class). Partners and stakeholders in the public health effort. Although the project focussed on only a few selected classes, many of the other areas listed above were considered in detail. In some cases the topic area suggested has been captured in the broad structure (e.g. 'settings' have been included among the toplevel classes). In other cases, the topic area has been built into the public health classification as attributes and characteristics of classes. Some are demonstrated in the examples of practical applications in Section 3.2.4. For instance, stakeholders and population groups are shown as attributes (slots) of 'public health activities' in Example 1. 'Policy' has been represented as a register or library of existing policies in Example 2. 'Research' should be identifiable through classification using the 'methods' class (which includes the subclass 'research and evaluation methods'). 'Workforce' and 'workforce development capacity' have been included as subclasses of the 'resources and infrastructure' class, as has 'time'. Definitional issues that were discussed during the project have been summarised as discussion points in Box 2. In Box 3 a range of other issues, raised throughout this report, are summarised for the further consideration of public health experts. Is there agreement with the principles of development: multi-dimensional, inclusive rather than exclusive, broad rather than narrow? Should public health classification be restricted to a domain solely within health or should it include specific activities of other sectors (e.g. education, transport, local government) that have public health as a primary or secondary purpose (e.g. immunisation organised by local government)? Is there agreement on the top-level classes? Are the public health functions appropriate? Are all important functions captured (e.g. is quality assurance a public health function in Australia)? Is the division between primary and instrumental functions clear and useful? Are there any important subclasses that are currently missing from the first two levels of the public health classification (see Table 7 )? What are the important characteristics of agreed top-level classes? Phase one of the project has produced version one of a public health classification, and achieved a degree of consensus among Australian public health experts regarding its major classes, and their structure at the top levels. The classes of public health 'functions', 'determinants of health' and 'methods' of intervention have been identified as priorities for further development. Many of the public health experts consulted during phase one of the project indicated that they were keen to continue their engagement. Most were positive about the project. They identified a range of practical applications for a public health classification that extended far beyond its uses for reporting public health activity and expenditure. The consultation process also brought to light a variety of issuesincluding areas of basic disagreement about the nature and boundaries of public health practice-that warrant more work. These are set out throughout this report in boxes. It is proposed that the second phase of the project will further extend the availability of, and seek feedback on, the public health classification through a web-based version, and develop a proposal for its future development and support. Because it attempts to capture the breadth of public health activity, and to serve multiple uses, the public health classification has a necessarily complex, multidimensional structure that is difficult to present adequately in paper-based forms. A web-based version, rendered in HTML, will allow interactive engagement and easier access to the structure, coverage and documentation (e.g. definitions). An early version of the classification was mounted on a test website and demonstrated in consultations with reasonable acceptance and understanding of its use as a navigation tool. A facility to collect structured feedback-rather than just adding large numbers of new classes and subclasses-and processes to compile and review this information will be needed to improve the utility of the classification for practical applications. Developing a plan for the ongoing development and support of the classification will involve consideration of governance and maintenance arrangements, as well as the issues of access, availability and intellectual property ownership and management. Maintenance of classification systems can be difficult, time-consuming and thankless work. International classifications, like that of diseases, rely on a lengthy consensual process of experts to identify and agree upon new entries. 63 However, new capabilities made possible by the Internet and the development of the Semantic Web present opportunities to distribute the maintenance burden across many contributors, and to dramatically speed up consensual agreement. 64 These will be explored as part of scoping the requirements for ongoing development and support of the classification. Further development of the classification will emphasise its relationships with classifications that are already in existence and widely used as standards. The public health classification, as it is currently structured, has subclasses that simply reference or point to relevant external classifications. These include (but are not limited to) Australian standards (e.g. geographical, industry, and occupational classifications, other standards promulgated by AIHW and ABS) and the international classifications of diseases, functioning and disability, and external causes of injury 65 (see Figure 7) . In a similar vein, it is proposed to investigate the possible inclusion of the public health classification in the set of standard classifications known as the Australian Family of Classifications. 66 It is recommended that phase two of the Public Health Classifications Project should: Focus on further developing the classes of public health 'functions', 'determinants of health' and 'methods' of intervention; Develop and release a web-based version of the public health classification with facilities for eliciting structured feedback and managing contributions to the further development and refinement of the classification; Develop a plan for ongoing development, support and governance of the public health classification; Further specify links or relations between the public health classification and relevant existing classifications and standards (with due regard for intellectual property rights); and A number of things regarded as forming one group through the possession of similar qualities; a kind; sort. (Delbridge & Bernard 1998) Classes are the focus of most ontologies. They describe concepts in the domain. For example, the class of public health 'functions' represents all public health functions. Specific functions, for example, 'protect from threats to health', are instances of this class. A class can have subclasses that represent concepts that are more specific than the superclass. For example, we can divide the class of all public health 'functions' into 'assess...', 'protect...' and 'promote...' functions. 67 Alternatively, we can divide the class of all public health functions into primary and secondary functions. (Noy & McGuinness 2001 : 3) An arrangement of classes in a taxonomic (subclass-superclass) hierarchy. A class hierarchy represents an 'is-a' relation, where a class X is a subclass of A if every instance of X is also an instance of A. A class hierarchy thus represents a set of classes related by inheritance. A class hierarchy is typically shown as a tree structure for single inheritance or as a lattice structure for multiple inheritance (where nodes represent classes and are connected by arcs to indicate inheritance relations). In an ontology there is no single correct class hierarchy for any given domain. The hierarchy depends on the possible uses of the ontology, the level of the detail that is necessary for the application, personal preferences, and sometimes requirements for compatibility with other models. (Noy & McGuinness 2001 : 6-8) A system for classifying things; in a library, a system of arranging items according to broad fields of knowledge and specific subjects within each field. To classify means to arrange or distribute in classes; to place according to class. Example: International Classification of Diseases (ICD) (WHO 1992-94). Computable information is information that can be readily manipulated and transformed by computers. Currently a great deal of information (on the Web and elsewhere) can be read by computers but not manipulated or understood by them. In the near future, the Semantic Web being developed by Sir Tim Berners-Lee, one of the founders of the World Wide Web, and others, will make information computable and connectable by adding semantic information, based on ontologies and classifications, to elements within text (Berners-Lee 2001). Determinants of health are factors that influence health status and determine health differentials or health inequalities. They are many and varied and include, for example, natural, biological factors, such as age, gender and ethnicity; behaviour and lifestyles, such as smoking, alcohol consumption, diet and physical exercise; the physical and social environment, including housing quality, the workplace and the wider urban and rural 67 See Section 3.2.2 for more information on the public health functions in a public health classification. environment; and access to health care (Lalonde 1974 , Labonté 1993 . All of these are closely interlinked and differentials in their distribution often lead to health inequalities (WHO 1998a). A part or aspect of something. For example, one dimension of public health is the settings in which public health work is carried out. A dimension is a property or construct whereby aspects of something can be distinguished (e.g. public health settings can be distinguished from public health functions and from public health methods). A dimension can also be described as a group of similar things that are from the same category of information (e.g. home and workplace settings are part of the settings dimension). Hence multi-dimensional, to have many aspects or dimensions (e.g. to provide a unified framework for multiple public health uses, a multi-dimensional classification is needed). Disease prevention refers to measures taken to prevent the occurrence of disease, to arrest or slow its progress and to reduce its consequences. Examples of disease prevention measures include risk factor reduction, screening and early intervention. Primary prevention of disease is directed towards preventing the initial occurrence of a disease. Secondary and tertiary prevention aim to arrest or slow the progression of existing disease and to reduce its effects through early detection of complications and appropriate treatment; or to reduce the occurrence of relapses and the establishment of chronic conditions through, for example, effective rehabilitation (WHO 1998a). The kind of action or activity proper to a person, thing, or institution (Delbridge & Bernard 1998: 452) . The function, purpose, role or use of something; for example, the function of public health is 'to protect and promote health, and to prevent illness, injury and disability' (NPHP 1998) . Machine readable-see computable information, Semantic Web Information about data. Metadata can describe the fields and formats of databases and data warehouses, documents and document elements such as Web pages or research papers. Metadata management is a functional component of an information management architecture. Example: the descriptive information provided in the 'META' tags in an HTML or XML document header that give information about the document. A model of a particular field of knowledge -the concepts relevant to that field (e.g. the field of public health), and their attributes, as well as the relationships between the concepts. In the Protégé ontology development software, 68 an ontology is represented as a set of classes that have associated slots (attributes). In philosophy, ontology describes a branch of metaphysics concerned with the nature and relations of being. The term has been redefined by the knowledge engineering and artificial intelligence communities to refer to a formalised description of the concepts and relationships that exist within a specific domain and all that can be represented about that domain. Ontologies can be mental models, computer models, or a combination of both. Ontologies provide a means by which characteristics of a specific representation can be assumed and behaviour predefined (Kemp & Vckovski 1998) . Multiple user views can be accommodated by providing translations between different ontologies. An ontology defines a common vocabulary for researchers who need to share information in a domain. It includes machine-interpretable definitions of basic concepts in the domain and relations among them (Noy & McGuinness 2001) . Ontologies are developed for the purposes of: Sharing common understanding of the structure of information among people or software agents, Enabling re-use of domain knowledge, Separating domain knowledge from operational knowledge, and Analysing domain knowledge. Example: The (US) National Library of Medicine's Unified Medical Language System (UMLS) 'knowledge sources' and associated lexical programs for system developers. The Meta-thesaurus is organised by concept or meaning. Its purpose is to link alternative names and views of the same concept together and to identify useful relationships between different concepts. Organised efforts focused on the health of defined populations in order to promote and maintain or restore health, to reduce the amount of disease, premature death and discomfort and disability due to disease. Programs, services and institutions here emphasize the prevention of disease and the health needs of the population as a whole. Among a broad scope of disciplines, various knowledge and skills are used, such as bio-statistics, epidemiology, planning, organisation, management, financing and evaluation of health programs, environmental health, application of social and behavioural factors in health and disease, health promotion, health education and nutrition. (IIME 2002) Preventable conditions include many chronic, non-communicable diseases such as cardiovascular disease, type 2 diabetes, obesity, chronic lung disease; conditions amenable to early detection and treatment such as breast and cervical cancer, high blood pressure; communicable diseases such as HIV/AIDS, food borne illness, vector borne diseases, vaccine preventable diseases; intentional and unintentional injuries; many mental health problems and related conditions such as substance abuse and family dysfunction. (Straton & Sindall 2001: 1) 68 Protégé is developed by Stanford University, see http://protege.stanford.edu. Prevention is characterised by activities that are taken to reduce the possibility that something will happen, or to minimise harm if it does occur. The prevention of illness or disability requires the identification of the factors that contribute to poor health and modifying, reducing or eliminating them, or, conversely, building and strengthening protective factors. Prevention is usually taken as a core responsibility of organised health systems-alongside the curative, restorative and palliative functions-and is a key element in achieving health improvement and the reduction of the burden of disease in society. Prevention is also an important component of many other branches of social policy (for example crime prevention, child abuse prevention), many of which also contribute, directly or indirectly, to health. It has been customary to categorise prevention at different levels, in terms of primary, secondary and tertiary prevention. Thus the goal of primary prevention is reducing the incidence of disease by preventing its occurrence, secondary prevention aims to prevent progression of disease though early detection, usually by screening at an asymptomatic stage and early intervention, 69 and the goal of tertiary prevention includes minimisation of the impact of established disease, and prevention of complications and further disability through effective treatment and rehabilitation. While the terminology used can vary in different fields (for example a slightly different set of categories is often used in relation to mental health 70 ), the basic concepts and objectives of prevention are essentially the same. It is often useful to think in terms of a hierarchy or spectrum of objectives for preventive activity, aimed at different points on the causal pathway, and for which there is often an important time dimension. For example, the short term aim of a preventive intervention at a certain point in time may be to change beliefs in the community about the risks of smoking; the intermediate objective may be to reduce uptake of smoking and smoking prevalence and the long term goal a reduction in rates of coronary heart disease and lung cancer. (Straton & Sindall 2001 : 1) Prevention and public health services comprise services designed to enhance the health status of the population as distinct from the curative services, which repair health dysfunction. Typical services are vaccination campaigns and programmes. (OECD 2000: 121) Primary prevention-see also prevention, disease prevention Primary prevention refers to the protection of health by personal and community wide effects, such as preserving good nutritional status, physical fitness, and emotional well-being, immunising against infectious diseases, and making the environment safe. There are no precise boundaries between the primary, secondary and tertiary levels of prevention. (IIME 2002) 69 A notable exception to this use of the term is found in the area of cardiovascular disease prevention and control where secondary prevention is commonly used to refer to prevention of a second heart attack. 70 In the mental health field primary prevention is further divided into approaches designated as universal, selective or indicated prevention, depending on whether they are applied to the whole population (universal) or sub-groups (selective) or those at an early stage of risk (indicated). A similar approach was used by the AIHW in development of the indicator framework for monitoring the National Health Priority Areas. Government-funded public health activity is described as an important part of the Australian health care system, with public health activities generally representing the organised response of society to protect and promote the current and future health of the whole population or of specific subgroups of the population, which can be viewed as a form of investment in the overall health status of the nation. (AIHW 2004b: 1) The nine public health core functions promulgated by the National Public Health Partnership (NPHP 1998) Public health has been defined by the World Health Organization as 'the art of applying science in the context of politics so as to reduce inequalities in health while ensuring the best health for the greatest number' (WHO 1998a cited in WHO 2003 . Public health expenditure reporting: core public health activities The core public health activities in public health expenditure reporting are defined as 'nine types of activities undertaken or funded by the key jurisdictional health departments that address issues related to populations, rather than individuals. Does not include treatment services.' (AIHW 2004b: 145) Government-funded public health activity is described as an important part of the Australian health care system, with public health activities generally representing the organised response of society to protect and promote the current and future health of the whole population or of specific subgroups of the population, which can be viewed as a form of investment in the overall health status of the nation. (AIHW 2004b: 1) Public health medicine is that branch of medical practice that is primarily concerned with the health and care of populations. It is concerned with the promotion of health and the prevention of disease and illness; the assessment of a community's health needs; and the provision of services to communities in general and to specific groups within them. (AFPHM 2002a) Research involving communities or populations, typically outside health care institutions, undertaken to identify the factors which contribute to ill-health in populations and ways of influencing these factors to prevent disease. It includes epidemiology, social and behavioural sciences, health services research on population-based health interventions, and evaluating the efficacy and effectiveness of preventive measures. (HMRSR 1998 : A6.4, Saracci 2004 Public health workforce The public health workforce is defined as those involved in protecting, promoting and/or restoring the collective health of whole or specific populations (as distinct from activities directed to the care of sick or frail individuals). (Rotem et al. 1995 cited in Riddout et al.2002 . Resource Description Framework (RDF) 'is a foundation for processing metadata; it provides interoperability between applications that exchange machine-interpretable information on the Web. RDF emphasizes facilities to enable automated processing of Web resources. RDF can be used in a variety of application areas; for example: in resource discovery to provide better search engine capabilities, in cataloguing for describing the content and content relationships available at a particular Web site, page, or digital library, by intelligent software agents to facilitate knowledge sharing and exchange, in content rating, in describing collections of pages that represent a single logical "document", for describing intellectual property rights of Web pages, and for expressing the privacy preferences of a user as well as the privacy policies of a Web site. RDF with digital signatures will be key to building the "Web of Trust" for electronic commerce, collaboration, and other applications' (W3C 1999). The Semantic Web Environmental Directory describes RDF as the 'equivalent of the language for writing Web pages, HTML (HyperText Markup Language), for the Semantic Web. The Semantic Web uses RDF as the basic language for representing metadata about any kind of resource on the Web' (SWED undated). Secondary prevention can be defined as the measures available to individuals and populations for the early detection and prompt and effective intervention to correct departures from good health. There are no precise boundaries between primary, secondary and tertiary levels of prevention. (IIME 2002) The Semantic Web provides a common framework that allows data to be shared and reused across application, enterprise, and community boundaries. It is a collaborative effort led by W3C with participation from a large number of researchers and industrial partners. It is based on the Resource Description Framework (RDF), which integrates a variety of applications using XML for syntax and URLs for naming. 'The Semantic Web is an extension of the current web in which information is given welldefined meaning, better enabling computers and people to work in cooperation' (Berners-Lee et al. 2001) . The Semantic Web and computable information are the visions of Tim Berners-Lee, the creator of the World Wide Web (familiar to us through Google 71 and other search engines), who views this future Web as a web of data, 'like a global database', where 'information is given well-defined meaning, better enabling computers and people to work in cooperation'. Making information on the Web 'semantic' (or meaningful) means much more efficient searching 'as though it were one giant database, rather than one giant book' (Berners-Lee 1998). The infrastructure of the Semantic Web will allow machines as well as humans to make deductions and organise information. The approach is to develop languages that express information in machine processable forms. The architectural components include semantics (meaning of elements), structure (organisation of elements), and syntax (communication). Abstract representation of data is being based on existing standards (eg RDF -Resource Description Framework) and standards yet to be defined, and is in development by the World Wide Web Consortium (W3C), in collaboration with researchers and industrial partners. A classification, especially in relation to its principles or laws; the department of science/s that deal with classification. A taxonomy is hierarchical, with the higher levels being larger, more inclusive and broadly defined, while the lower levels are more restrictive and specific. Example: the classification of plant and animal life into natural, related groups in descending order: phylum, class, order, family, genus, species. The system of terms belonging to a science, art, or subject; nomenclature. A controlled vocabulary contains metadata about terminology to make it easier to search and maintain knowledge management systems that integrate information from multiple sources and applications. Example: SNOMED CT ® -Systematized Nomenclature of Medicine-Clinical Terms (produced by the College of American Pathologists) is a comprehensive clinical terminology, and one of a suite of designated standards for use in US Federal Government systems for the electronic exchange of clinical health information, and is being implemented throughout the National Health Service in the UK. Tertiary prevention consists of the measures available to reduce or eliminate long-term impairments and disabilities, minimize suffering caused by existing departures from good health, and to promote the patient's adjustment to irremediable conditions. This extends the concept of prevention into the field of rehabilitation. There are no precise boundaries between primary, secondary and tertiary levels of prevention. (IIME 2002) A storehouse or repository, as of words or knowledge; a dictionary, encyclopedia or the like, especially a dictionary of synonyms and antonyms. Technical thesauri are used in search-language normalisation as they specify terms to be used (preferred terms), broader and narrower terms in the hierarchy, as well as related terms (nonhierarchically related, e.g. antonyms) and non-preferred terms (synonyms for the preferred term). Example: MeSH (Medical Subject Headings) -the (US) National Library of Medicine's controlled vocabulary, used to index articles for MEDLINE and PubMed. MeSH terminology provides a consistent way to retrieve information that uses different terminology for the same concepts. The wicked problem concept was originally proposed by Rittel and Webber (1984) in the context of social planning. They pointed out that in solving a wicked problem, the solution of one aspect may reveal another, more complex problem. Ten rules define the form of a wicked problem, including: 1. There is no definitive formulation of a wicked problem. 2. Wicked problems have no stopping rule. 3. Solutions to wicked problems are not true-or-false, but good-or-bad. Every wicked problem is essentially unique, and can be considered to be a symptom of another problem. (The last rule is that: The planner (designer) has no right to be wrong.) The continuing support of Richard Madden, and the additional assistance of John Goss, Tony Hynes, Daniel Aherne and Justine Boland of AIHW is gratefully acknowledged. On behalf of the project, a big thank you to all who made time to engage with the public health classification, and for your perspectives, reactions and suggestions for improvement. An example agenda and work-in-progress documentation used in consultations are shown in the following pages. Consultation on Thursday 10 th December, 10-12am Consultation Objective: to meet with content experts to model a unified public health classification that is useful and useable for multiple purposes. What is the domain that the classification will cover? Public health. Definition: Public health is the organised response by society to protect and promote health, and to prevent illness, injury and disability. The starting point for identifying public health issues, problems and priorities, and for designing and implementing interventions, is the population as a whole, or population sub-groups. (NPHP 1998) Principles: The classification should be inclusive, and deliberately broad at the top classes. For what are we going to use the classification? Generally, to develop a broad, generalisable public health classification that can be used to: organise information to facilitate answering key public health questions e.g. expenditure on prevention of obesity; reflect the full scope and breadth of public health activity, in expenditure and performance indicator reporting; articulate, describe and define public health, and promote consistency in describing public health (eg through standardised instructions); build in specific content expertise in different areas of public health; relate to other high level models of health (eg through interface and reference terms); structure and design information/communications e.g. in websites or report chapters. Specifically, a public health classification could be used to: promote standardised definitions, terminology and reporting of public health and public health functions to improve accountability across jurisdictions, eg through the development of a national Public Health Report describing public health in Australia; build systems such as a web-based database of public health projects that allows routine, bottom up, multi-dimensional reporting of public health projects; create semantic web documents that are 'marked up' for meaning (for the Semantic Web, the next generation of the world wide web) and which can be understood and manipulated by computers (e.g. computer agents can trawl semantic web documents for information to answer questions, eg what is the project expenditure, how many people work on the project, in what settings?). For what types of questions should the classification provide answers? Sample focus questions include: How much was spent on prevention of obesity? Other 'advocacy-type' questions, e.g. difference in expenditure on prevention of HIV/AIDS relative to other preventable diseases, relative expenditure on specific risk factors or diseases? Has health funding to preventive or promotive investments increased? What is public health? How is public health relevant to components of the human services delivery system? Why do public health unit costs differ across jurisdictions? Can we describe screening in clinical settings eg GP surgeries for pap testing? What did we invest in social marketing last year? Can we replicate the output of other models? (eg Public Health Expenditure Reporting, public health component of OECD Health Accounts) The top-level public health classes listed for examination, some of which have been examined in more detail to date, are: Selected views of the classification including a practical example from public health expenditure reporting, and another from the UK, follow. showing the main public health classes captured in the classification (public health activities derived from Public Health Expenditure Reporting and input from the Reference Group) Related recent developments in the UK include the development of a Public Health Information Tagging Standard -to provide website access to public health resources -and a National Public Health Language, incorporating other thesauri and vocabularies to improve web-based searching and retrieval for public health resources. A web-based system for the classification and retrieval of public health resources was conceived by Julian Flowers of the Eastern Region Public Health Observatory (ERPHO) in the UK, as there was no system specifically suitable for this purpose. The Public Health Information Tagging Standard (PHITS) was borrowed categories from a number of extant sources, 75 and took contributions from public health specialists nationwide(see Figure 10 ). Source: Eastern Region Public Health Observatory www.erpho.org.uk accessed November 2004. Figure 10 shows PHITS describing public health resources on the website of the ERPHO. Subjects or classes of interest can be selected from the 'Browse by subject' box on the left side of the underlying screen print. The overlying screen print shows the 'Services' class and its finer subclasses (e.g. 'Population based and preventive', 'Primary care'). The tabbed entries to the right show the types of resources available (e.g. all resources, data), and provides typical information on individual resources (e.g. 'A rapid mapping study of smoking 75 Sources included ICD10, MeSH, and SNOMED. projects', an 'ABC of smoking cessation'), including the URL of the resource for instant access. After its introduction on the ERPHO website, PHITS was adopted as a standard for use by all ten Public Health Observatories in England and Wales, as well as other public health organisations, such as Public Health Ireland. 76 Initially intended purely as a web site categorisation and retrieval system, PHITS has now become part of the development of a National Public Health Language for the UK. A National Public Health Language for the UK PHITS has been integrated with the UK Health Development Authority's Public Health Information Thesaurus 77 and two other controlled vocabularies, to create the National Public Health Language (NPHL) for the UK (Figure 11) . The development of a common public health language is intended to facilitate interoperability and improve the efficiency of searching for and retrieving, public health information and resources held on websites and in databases. All organisations that were already using PHITS have agreed to move to the NPHL when version one was available (December 2004 78 ). NPHL users will have both a public health biased classification system; and a powerful, thesaurus-driven, categorisation and searching mechanism for use on web sites. 79 Figure 11 shows the entry website for online access to the NPHL (left side) and top-level classes and their definitions (right side). UK National Public Health Language (NPHL) NPHL top-level classes initial ontology, in an iterative design process that continues through the whole of the ontology's lifecycle. 84 This iterative development style is a good fit for complex or wicked problems. Because public health is complex it is technically conceptualised as a 'wicked problem' 85 , meaning that there is no definitive formulation or solution, no 'right' or 'wrong', no absolute truth or perfect solution that holds for all cases-the best that can be achieved is a consensus of public health experts that it is good enough. In consultations it was clear that the conceptualisation of public health is time-specific (e.g. the 'old' and the 'new' public health), includes many contested definitions and terms, as well as fuzzy borders and boundaries. There is not even agreement on what it should be called, with the terms 'population health' and 'preventive health' currently challenging 'public health'. Two principles of development (see Section 3.1.5) address these difficulties: be inclusive; and, set rules and boundaries in applications, rather than in the development of the classification itself. Inclusiveness is a response to the divergence of views and definitions encountered in field consultations. The project took the position that a public health classification should not exclude elements that some (but not all) consider to be an important part of public health. It should actively seek to include divergent views since its usefulness as a unified classification depends on the best coverage of the breadth of public health. Rules and boundaries can and should be determined in practical applications rather than in the ontology. For instance, for the purposes of reporting health and public health expenditure, it may be determined that all one-to-one treatment services in clinical settings are not public health services. Another use might determine that some one-to-one clinical treatments, such as those for immunisations, sexually transmitted infections, or drug detoxification, are public health services. The decision to set a constraint or boundary for a particular application should not preclude the wider scope of a 'public health classification', which is developed as an ontology. A single ontology can be used to develop one or more classification systems, by developing specific rules and boundaries (developed as 'constraints' in the ontology) to organise classes into a hierarchy, and to assign elements to unique classes. Although defining and specifying classes (concepts within the domain of interest) is central to developing an ontology, the emphasis is on modeling the relationships among classes, rather than on hierarchy (broader classes contain the more specific) or mutual exclusion (an element cannot be in more than one class). An ontology allows elements to be assigned to more than one class. This is useful, for instance, for areas (of which there are many in public health) on which there is little agreement and competing views. In a classification system, with its emphasis on mutually discrete classes, it is not so easy to do this. 84 Noy & McGuinness 2001: 4. 85 The wicked problem concept in design was described by Rittel & Webber (1984) in the context of social planning. They pointed out that in solving a wicked problem, the solution of one aspect may reveal another, more complex problem. Ten rules define the form, including that there is no definitive formulation of a wicked problem (no stopping). Solutions to wicked problems are not therefore true-or-false, but good-or-bad. A concrete example is the categorisation of behavioural factors. Most public health experts would agree that as a determinant of health these contribute to health risk and/or protection; however some see behavioural factors as exclusively personal, while others see them as exclusively socio-economic, and some see them as both. Using an ontology, they can be classed under both categories, so that those who expect to find them under personal factors will do so, as will those who expect to find them under socioeconomic factors, as illustrated in Figure 12 . Thus all are satisfied (have found the category where they expected to), a practical result has been achieved, and an indecisive argument about where it is 'rightly' to be found has been avoided. Sophisticated software tools are available to assist in developing ontologies. These allow multiple inheritance (as described above), definition of relationships among classes, specifications of attributes of classes, and classification of elements (instances). Aspects of public health (characteristics, attributes, etc) can be described either textually as descriptions, mathematically as values, or in terms of other classes in the class hierarchy, and can be constrained by specific rules. Ontology development software is the backbone of the next generation of information tools. Increasingly, existing classification systems are being migrated to, or developed in, ontology building software such as Protégé 86 (used by this project). This software makes it easy to render form and content for the Web. As the Semantic Web develops and ontologies become more widely used in Web-based applications, the development of the public health classification in an ontology can be expected to produce major productivity gains in making existing information more available and better connected. 86 More information on Protégé, and the free, open source Protégé software, are available from Stanford University at http://protege.stanford.edu/index.html. The 'isa' relation arrows show the parent class or classes that each child class belongs to in the class hierarchy of the ontology.

Small Interfering RNA Targeting M2 Gene Induces Effective and Long Term Inhibition of Influenza A Virus Replication

RNA interference (RNAi) provides a powerful new means to inhibit viral infection specifically. However, the selection of siRNA-resistant viruses is a major concern in the use of RNAi as antiviral therapeutics. In this study, we conducted a lentiviral vector with a H1-short hairpin RNA (shRNA) expression cassette to deliver small interfering RNAs (siRNAs) into mammalian cells. Using this vector that also expresses enhanced green fluorescence protein (EGFP) as surrogate marker, stable shRNA-expressing cell lines were successfully established and the inhibition efficiencies of rationally designed siRNAs targeting to conserved regions of influenza A virus genome were assessed. The results showed that a siRNA targeting influenza M2 gene (siM2) potently inhibited viral replication. The siM2 was not only effective for H1N1 virus but also for highly pathogenic avian influenza virus H5N1. In addition to its M2 inhibition, the siM2 also inhibited NP mRNA accumulation and protein expression. A long term inhibition effect of the siM2 was demonstrated and the emergence of siRNA-resistant mutants in influenza quasispecies was not observed. Taken together, our study suggested that M2 gene might be an optimal RNAi target for antiviral therapy. These findings provide useful information for the development of RNAi-based prophylaxis and therapy for human influenza virus infection.

Influenza A virus (IAV) remains a scourge on human health [1, 2, 3] . Its antigen drifts and shifts are an ever-changing challenge for available vaccines [4, 5] . The appearance of drug resistance is the main hurdle for the development of antiviral drugs [6, 7, 8, 9] . Given the limitations of current anti-influenza A virus strategies, the need for novel strategies for prevention and treatment of IAV is evident [10] . In this regard, RNA interfering (RNAi) technology holds great promise to inhibit the replication of IAV, including H5N1 virus. RNAi is a form of posttranscriptional gene silencing mediated by short double-stranded RNA, known as small interfering RNA (siRNA) [11, 12] . In this process, the cellular complex Dicer cleaves a double-stranded RNA (dsRNA) molecule to yield doublestranded duplexes 21-25 nucleotides in length. These siRNAs then guide the RNAi induced silencing complex (RISC) to cleave target mRNAs that share sequence identity with the siRNA [13, 14, 15] . Since it was first demonstrated that adding exogenous, synthetic siRNA molecules to mammalian cells can induce RNAi, there have been rapidly expanding efforts to develop RNAi therapies that induce the degradation of target messenger RNA (mRNA) involved in genetically inherited diseases or acquired disorders [16, 17, 18, 19, 20, 21, 22] . IAV is an enveloped, negative-stranded RNA virus. The unique property of single-stranded RNA virus itself makes RNAi an attractive approach for development of anti-avian influenza therapeutics. The single-stranded viral genome, consisting of 8 segments contained at least 10 open reading frames (ORFs), serves as template for both viral genome replication and subgenomic mRNA synthesis. It has been reported that siRNAs respectively targeting to the viral genes of polymerase 1 (PB1), polymerase 2 (PB2), polymerase A (PA), nucleocapsid protein (NP), nonstructure proteins (NS1 and NS2), matrix proteins (M1 and M2), especially those specific for NP, PA and PB1, can potently inhibit replication of influenza A viruses [16, 23, 24, 25, 26] . However, it has been reported that HIV and HCV may develop siRNAresistant mutations quickly [17, 27, 28] , and therefore abrogated the further RNAi treatment. Thus, the evaluation of long term inhibition efficiency of designed siRNAs and screening of the emergence of siRNA resistance mutants are also an important research target. In the present study, we identified an effective siRNA targeting M2 gene (siM2), a highly conserved gene in IAV, as compared to a reported effective siRNA targeting NP gene (siNP). We further established cell lines which stably expressing the shRNAs by transducing lentiviral-shRNA vectors to Madin-Darby cannie kidney (MDCK) cells. Using these two cell lines, we evaluated long term antiviral effects of these siRNAs against IAV subtypes H1N1 and H5N1 and further screened the potential siRNA-resistant viral mutations. Our results showed that rationally designed siM2 conferred long term effective inhibition for IAV replication. It was further demonstrated that no siRNA-resistant viral mutation appeared in siM2 targeting sequence even after the virus was cultured in the shRNA expressing stable cell line for 40 passages. Two siRNAs targeting the M2 gene were rationally designed by siRNA target designer (the sequences of siRNAs are shown in the supporting information Table S1 ) and their effect in inhibiting the virus replication was assessed in MDCK cells. Two siRNAs targeting the NP gene were included in the experiments as controls. The results showed the siRNA M-950 exhibited a good inhibition effect with dose dependent manner, while another siRNA M-126 just slightly inhibited virus replication even at a concentration of 100 nM (Fig. 1A) . Fig. 1B showed that the siRNA NP-1496 could inhibit influenza virus replication, while siRNA NP-336 had no inhibition effect, which is consistent with the previous report [25] . The siM2 Exhibited Higher Inhibitory Effect of H1N1 Virus than siNP in Stable Cell Lines Based on the above results, the lentiviruses expressing the shRNAs M2-950 or NP-1496 were constructed and transduced into MDCK cells to establish two stable cell lines, shM2-MDCK and shNP-MDCK. MDCK cells and the MDCK cells transduced by blank lentivirus (Mock MDCK) were used as controls. The cell lines were infected with H1N1 virus at a moi of 0.005 and culture supernatants were harvested at indicated time-points to determine the virus titer by plaque assay. As shown in Fig. 2 , virus replication kinetics of Mock MDCK is similar with that of MDCK, indicating that lentivirus integration didn't influence virus replication. Virus titers in shNP-and shM2-MDCK cell cultures were 2 to 10 folds lower than the controls MDCK and Mock MDCK cultures, suggesting that virus replication had been suppressed by the expressed shRNAs in both shM2-MDCK and shNP-MDCK cells. Notably, siM2 exhibited a better inhibition effect, showing about 2-fold lower viral titer than siNP, although the expression levels of siM2 and siNP were similar (DCt siM2 = 6.68, siNP = 6.95). The siM2 Abolished not only M2 mRNA but also siNP mRNA Accumulation in the Stable Cell Lines We also measured the accumulation of mRNA for NP and M2 gene in infected MDCK, Mock MDCK, shM2-MDCK and shNP-MDCK cells. The mRNAs were extracted from the cells harvested at 1, 2, 4 and 24 hrs post-infection and tested by realtime RT-PCR. The mRNA expression level is normalized by copy To further confirm whether the suppression of NP mRNA in shM2-MDCK cells indeed affect NP protein expression, the NP protein level was tested by an indirect immunofluorescence assay. As shown in Fig. 4 , EGFP fluorescence, an indicator of shRNA expression, was detected in Mock MDCK, shNP-MDCK and shM2-MDCK but not in MDCK cells, while NP protein was detected in MDCK and Mock MDCK cells but not in shNP-MDCK and shM2-MDCK cells. The results were consistent with above viral mRNA results, indicating that siM2 indeed suppressed the NP protein expression. siM2 Provided More Potent anti-H5N1 Viral Effect than siNP in Stable Cell Lines We further tested whether siM2 could also inhibit the replication of a highly pathogenic H5N1 avian influenza virus. As shown in Fig. 5A , although numbers of plaques were similar in different MDCK cell lines, smaller size of plaques were only found in shM2-MDCK cells, suggesting that siM2 inhibited replication of H5N1 virus. The cell lines were also infected with different amounts of H5N1 virus and culture supernatants were collected at different time points to determine the virus titers by HA assay. The virus replication was significantly inhibited in shM2-MDCK cells at all time-points, but shNP-MDCK just offered a minor inhibition effect at early stage of the virus infection (Fig. 5B ). These results further confirmed that siM2 could provide a more potent protection than siNP against H5N1 infection. To test if siM2 siRNA-resistant virus mutant would quickly appeared when cultured in shM2-MDCK cells, H5N1 virus was continually cultured in shM2-MDCK cells for 40 passages. Every 10 passages, the culture supernatant was collected and tested by plaque reduction assay. No obvious larger size of plaque was found. Ten plaques with relative larger size were picked to further identify potential mutation in the siRNA targeting region by sequencing. The results showed that no mutation appeared in the siM2 targeting region even after 40 passages of the cultures (Fig. 6 ). The principal finding of this study is that rationally designed siRNA targeting influenza M2 gene (M-950) conferred effective long term inhibition against influenza A virus replication. Such high suppressive effect is not only against H1N1 influenza A virus but also against a highly pathogenic H5N1 subtype. In the previous related studies, Ge and his co-workers [25] screened siRNAs targeting to 6 conserved genes of influenza A virus and showed that NP-1496 was the best since it can confer a more than 200-folds inhibition of H1N1 virus. Li et al [29] and Tomkines et al [23] further confirmed that NP-1496 provided high anti-H5N1 effect. We therefore included NP-1496 as a positive control in this study. Our results showed that siRNA M-950 exhibited similar (Fig. 1 ) or even slight higher (Fig. 2) inhibitory effect against IAV replication as compared to that of NP-1496. A recent report by Zhou et al [30] also showed that several siRNAs targeting NP and M genes exhibited effective inhibition against influenza A virus replication in cultured MDCK cells and in animal models. However, sequences of their reported siRNAs targeting M2 gene are completely different from the siRNA M2-950. Furthermore, chemically synthesized siRNAs or plasmid based shRNAs were always delivered by transfection in previous related studies, whereas we used a lentivirus system to deliver selected shRNAs. Although the integration property of lentivirus has abrogated it to be used in human, it is helpful for our study purpose to successfully establish stable cell lines persistently expressing siRNAs. In this study we found that siM2 not only decreased the level of M2 mRNA but also the level of NP mRNA, suggesting that siM2 has a broad inhibition manner in the process of influenza virus replication. Ge et al have reported a similar broad inhibition of siRNAs [25] . In their study, NP-1496 and PA-2087 provided a broad inhibition to H1N1 influenza virus, which not only abolished the accumulations of specific NP or PA mRNAs but also inhibited the accumulations of mRNAs for M, NS1, PB1, PB2 and PA or NP genes. A possible explanation is that some double stranded siRNAs may result in IFN responses or activate a RNA degradation pathway, e.g. Phosphorylated protein Kinase R (PKR) [9, 31, 32] . However, the mechanisms of this broad inhibition of some siRNAs are still not very clear yet. From the standpoint of viral target choice in RNAi based antiviral therapy, NP protein is required for elongation and antitermination of nascent cRNA and vRNA transcripts [33, 34] . Without newly synthesized NP, further viral transcription and replication are blocked. While, M2 plays a critical role in the assembly of infectious virus particles. Thus, the potent antiviral effect of siM2 may be attributed to its broad inhibitory effect. Depending on the stringency of siRNA-target base pairing, siRNA treatment may cause selection of siRNA-resistant viruses, and this has been shown with HIV and HCV [17, 27, 28] , and therefore abrogated the further medication or treatments. Using lentiviral delivery system, we established stable cell lines persistently expressing shRNA, which provided a more convenient experimental approach to study long term inhibition effect of siRNAs and screen for siRNA resistant virus mutants in quasispecies in vitro. Our results showed that H5N1 virus cultured in shM2-MDCK were equally susceptible to siM2 as the original virus even after 40 passages. Moreover, sequencing of siM2 targeted region in 10 such independent plaque purified virus isolates revealed sequence identical to the parental one. The current data have shown no insertion, deletion and nucleotide substitution in the siRNA target sequence, therefore demonstrated siM2 possessed good long term inhibition effect for influenza virus replication without the problem of siRNA resistant mutants. Taken together, all the findings about effective RNAi target, lentiviral vector delivery and the establishment of stable shRNA expressing cell lines in our study provide rational information for the development of siRNAs as prophylaxis and therapy for influenza virus infection in humans. MDCK and Human embryonic kidney 293T cells were respectively maintained in MEM and DMEM (Invitrogene, USA) supplemented with 10% heat-inactivated fetal bovine serum (FBS) and antibiotics (100 U penicillin G/mL and 100 ug streptomycin/mL). Influenza virus strains A/New Caledonia/ 20/1999 (H1N1) and A/Hong Kong/486/97 (H5N1) used in these experiments were prepared in MDCK cells and virus titers were determined by TCID 50 . All experiments with H5N1 virus were performed in BSL-3 laboratory. The siRNAs targeting M or NP gene of influenza A virus were designed by siRNA target designer version 1.51 from Promega (http://www.promega.com/siRNADesigner/program/). The duplexes of designed and previously reported siRNAs were synthesized by Invitrogene (USA) (the sequences were shown in the supporting information Table S1 ). The siRNAs were reverse transfected to MDCK cells using Lipofectamine TM RNAiMAX (Invitrogene, USA) as described in company's instruction. After incubated the cells for 16,18 hrs, the cells were infected with the viruses and followed by detection of viral replication. 24 hours after infection, RNA were extracted from the cells and followed by real time RT-PCR to detect the relative quantities of replicated viral RNA. The H1-promoter-driven shRNA cassettes were constructed by annealing two primers containing the 19-nt sense and reverse complementary targeting sequences with a 9-nucleotide loop -TTCAAGAGA-and flanking Mlu1 and Cla1 cloning sites (the sequences of shRNA were shown in the supporting information Table S1 ), and then cloned into the 39-end of the H1 promoter in the LVTHM plasmid [35, 36] . The sequences of the insertions were confirmed by DNA sequencing. Lentiviral vectors with shRNA expression cassette were produced by calcium phosphate-mediated, three-plasmid transfection of 293T cells [37] . Briefly, 293T cells (2.5610 6 MDCK and the stable shRNA expressing cell lines in 24-well plates were infected with viruses at moi of 0.005,0.5 (2 mg/mL trypsin was used in the infection process of H1N1). After incubation for 1 hr, the infected medium was removed and MEM without FBS was added. Cell supernatants were collected at different time points. The viral load was detected by hemagglutination (HA) and/or plaque assays as described previously [39] . Briefly, the HA assay was carried out in U-bottom 96 well plates. Serial 2-fold dilutions of virus samples were mixed with an equal volume of a 0.5% suspension of turkey erythrocytes (Lampire Biologic Laboratories, Pipersville, USA) and incubated at room temperature (RT) for 45 mins. Wells containing an adherent, homogeneous layer of erythrocytes were scored as positive. For plaque assay, serial 10-fold dilutions of virus sample were added into a monolayer of MDCK cells. After 1 hr incubation, the virus was removed and the cultures were overlaid with 1% semi solid agar-MEM. Three days after infection, plaques were visualized by staining of crystal violent. Real-time RT-PCR was carried out as described previously [39] . Briefly, H1N1 or H5N1 virus infected MDCK, Mock MDCK, shNP-MDCK and shM2-MDCK were harvest at 1, 2, 4 and 24 hr after infection. Total RNA was extracted from the infected cell samples using RNeasy RNA isolation Kit (Qiagen, Germany) and reverse transcribed using Superscript II Reverse Transcriptase and Oligo dT primer (Invitrogene, USA), according to the manufacturer's protocol. Viral mRNA copies were measured by SYBR green M63000 Real-Time PCR System Indirect immunofluorescence assay was performed as described previously [40, 41] with some modification. MDCK, Mock MDCK, shNP-MDCK and shM2-MDCK cells grew on micro cover glasses (Thomas, USA) were infected with 1 moi of H1N1 virus for 6 hrs, After washed with PBS, the cells were fixed in 4% paraformaldehyde for 15 mins at RT and then permeabilized in 0.1% Triton X-100 for 3 mins at RT. After washed with PBS again, the cells were incubated with 1:50 diluted mouse anti-NP antibody (Abcam, UK) for 30 mins in dark at RT. The cells were washed three times in PBS with 1% FCS and incubated with 1:500 diluted Texas red-conjugated anti-mouse lgG (Abcam, UK) for 30 mins in the dark at RT. The cells were washed and mounted. Slides were viewed under an Olympus fluorescence microscope (Olympus, Germany). The screening of potential siRNA resistant mutants were performed in our established stable shRNA-expressing cell lines according to previously described protocols [42] with some modification. Briefly, the shM2-MDCK cells in a T25 cm 2 flask were infected with H5N1 virus. After cultured for 2 days, the supernatants were harvested. Part of the supernatants was inoculated to shM2-MDCK for next passage of the virus culture, another part was subjected for plaque assay to determine if potential siRNA-resistant virus appeared. Every 10 passages, ten bigger size of plaques in the plaque assay were picked for sequencing to detect any mutation in the siRNA targeting region using a pair of primers: forward, 59-AAG GCA GAT GGT GCA GGC AAT-39 and reverse, 59-TAC TCC AGC TCT ATG CTG ACA-39. Table S1 Found at: doi:10.1371/journal.pone.0005671.s001 (0.03 MB DOC)

Increased expression levels of the pvcrt-o and pvmdr1 genes in a patient with severe Plasmodium vivax malaria

BACKGROUND: There are increasing reports of severe clinical cases exclusively associated with Plasmodium vivax infections. Notably, this severity has been recently suggested to be associated with chloroquine resistance. PATIENTS: Two different patients presented at the Hospital Clinic in Barcelona with P. vivax malaria episodes. One patient had severe symptoms and the other mild symptoms. Both patients traveled through the Brazilian Amazon (Manaus) in 2007. For both patients the current diagnosis of malaria was the first. Two other patients with mild symptoms presented to the "Centro de Pesquisa em Medicina Tropical", also in the Brazilian Amazon (Rondônia) in 2000. METHODS: To exclude the possibility that the patient's severe symptoms were due to Plasmodium falciparum, a nested PCR was performed. A magnetic method was used to purify P. vivax free of human leukocytes. Quantitative real-time PCR was performed to compare the transcript levels of two main transporters likely to be involved in chloroquine resistance in P. vivax, namely the P. vivax chloroquine resistance transporter, pvcrt-o, and the P. vivax multidrug resistance transporter, pvmdr 1. RESULTS: Results demonstrated that the severe clinical symptoms were exclusively due to P. vivax. The patient presented acute respiratory conditions requiring admission to the intensive care unit. The magnetic method showed highly purified infected-reticulocytes with mature stages. In addition, it was found that parasites obtained from the severe patient had up to 2.9-fold increase in pvmdr1 levels and up to 21.9-fold increase in pvcrt-o levels compared to expression levels of parasites from the other patients with mild symptoms. CONCLUSION: This is the first clinical case of severe disease exclusively associated with vivax malaria in Spain. Moreover, these findings suggest that clinical severity could be associated with increased expression levels of parasite genes likely involved in chloroquine resistance. It is necessary to further explore the potential of pvmdr1 and particularly pvcrt-o expression levels as molecular markers of severe disease in P. vivax.

The renewed momentum for global malaria eradication has highlighted the need to further studies on Plasmodium vivax if eradication is to be achieved. Although the exact burden of disease is still a matter of debate, it is likely that it has been underestimated and that P. vivax is responsible for between 100 and 300 million clinical infections each year [1] . The emergence of worsening clinical severity and chloroquine resistance are two major factors responsible for this increasing burden. Plasmodium vivax infections have been associated with mild symptoms such as fever, headache, fatigue, chills, and musculoskeletal pain, in particular paroxysms. Recently, however, severe complications, including renal failure, jaundice, acute respiratory distress syndrome, cerebral malaria, seizures, anaemia, hyperparasitaemia, thrombocytopenia, pulmonary edema, splenic rupture and death, have been reported in exclusive association with P. vivax [2, 3] . Chloroquine resistance (CQR) is a major determinant of the present resurgence of malaria worldwide, including that of P. vivax [4] . Two main transporters have been studied in regard to CQR in P. vivax: the P. vivax chloroquine resistance transporter, Pvcrt-o, and the P. vivax multidrug resistance transporter, Pvmdr1 [5] [6] [7] . Interestingly, amino acid polymorphisms have not been associated with chloroquine resistance in pvcrt-o whereas pvmdr1 polymorphisms have been recently suggested to be associated with CQR in Southeast Asia [8] . This data indicates the involvement of other mechanisms in CQR in P. vivax. Likely candidates are gene amplifications and differential expression levels [9, 10] . The European Network on Imported Infectious Disease Surveillance, TropNetEurop, is an electronic network of clinical sites that monitors imported infectious diseases in Europe http://www.tropnet.net. Since its foundation in 1999, the network has recorded 8,374 cases of malaria, of which close to 11% (930) were due to P. vivax. Worth mentioning, TropNetEurop covers approximately 10% of all malaria cases reported in Europe. Moreover, according to data from the Spanish National Center of Epidemiology and the Autonomous Government of Catalonia there have been 266 cases of P. vivax in Spain in the last six years. In addition, three cases of severe symptoms due to P. vivax in Europe have been reported elsewhere in the literature [11] [12] [13] . These figures illustrate the clinical-epidemiological importance of this parasitic disease in a supposedly malaria-free region. Recently, CQR in P. vivax has been suggested to be associated, albeit not directly linked, with severe vivax malaria [14] . Here, the first clinical case of severe vivax malaria in Spain is presented. The data also indicates that clinical severity could be associated with increased expression levels of two parasite genes likely involved in chloroquine resistance, pvmdr1 and pvcrt-o. Two patients presented at Hospital Clinic in Barcelona with P. vivax malaria episodes. One had severe symptoms and the other mild symptoms. The patients had travelled through the Brazilian Amazon (Manaus) for 31 and 19 days, respectively, in 2007. The patient with severe symptoms was a 30-year-old Spaniard man who had previously travelled to Kenya, in 2006. Upon his return from Brazil, he presented to Hospital Clinic in Barcelona with high fever (39°C) and a Giemsa-stained thin blood film confirmed the presence of P. vivax at a parasitaemia of 1.8%. The patient presented acute respiratory conditions, anaemia and hyperbilirubinaemia, requiring admission to the intensive care unit. The patient with mild symptoms was a 31-year-old Spaniard man who had travelled without chemoprophylaxis and who had previously visited Mexico (2006), Vietnam (2005), and India (2004). There were no records of fever episodes from these previous trips and neither of the patients had ever been diagnosed with malaria. Two other Brazilian patients with mild symptoms presented at the "Centro de Pesquisa em Medicina Tropical, Rondônia, Brazil in 2000. Total RNA from parasites of these patients was extracted, pooled and stored in liquid nitrogen. Five mL of infected red blood cells were obtained from each patient. One mL was used to purify genomic DNA following standard methodologies. The remaining blood was processed to isolate total RNA using the Trizol reagent (Invitrogen) according to the manufacturers' instructions. A recently described magnetic method for the isolation of matures stages of malaria parasites was used to concentrate and purify P. vivax-infected reticulocytes [15] . Giemsa-stained smears showed an absence of human leukocytes, and all the reticulocytes were infected with mature stages of the parasite (Additional file 1). Eluents were centrifuged at 800 × g for 10 minutes, supernatants discarded, and pellets used to purify total RNA. The protocol for this study was approved by the Ethical Committee of Hospital Clinic and informed consent obtained from the patients. Nested polymerase chain reaction (PCR) was performed as previously described [16] to exclude P. falciparum infec-tions. Fragments were resolved and visualized on 2% agarose gels stained with sybr green. Amplification reactions were performed using Power SYBR Green PCR Master Mix (Applied Biosystems) and 45 ng of template cDNA prepared from each sample. Samples were set up in duplicate and experiments were repeated independently twice. PCR products were amplified and detected on an ABI Prism 7700 (Applied Biosystems). Cycling parameters for PCR were an initial denaturation step at 95°C for 10 minutes, followed by 40 cycles of 95°C for 15 seconds, and 60°C for 1 minute. To analyse the relative transcript levels, the threshold cycle value (Ct) of each sample was used to calculate and compare the ΔCt of each sample to that of the P. vivax housekeeping gene Sal I β-tubulin; the ΔΔCt was also calculated as in [17] to compare the transcript levels of pvcrt-o and pvmdr 1 in the patient with severe symptoms and in the patient with mild symptoms. During the month of August, 2007, a 30-year-old Spanish tourist traveled through the Brazilian Amazon region of Manaus, where, due to gastric disturbances, he took an incomplete chemoprophylaxis consisting of proguanil and chloroquine. Upon his return to Spain 30 days later, he presented to a health center with an eight-day history of fever, chills and dry cough. He was diagnosed with lower respiratory tract infection and treated with amoxicillin-clavulanic acid 875/125 mg every 8 hours for 24 hours (3 doses) without resolution. On presenting a week later to Hospital Clinic in Barcelona, he had high fever (39°C) and jaundice; the tip of the spleen was palpable and chest auscultation was unremarkable. A thin blood smear of peripheral blood stained with Giemsa revealed P. vivax infection with a parasitaemia of 1.8%. The patient presented acute respiratory conditions requiring admission to the intensive care unit (ICU). In the ICU, he was haemodynamically stable and blood tests revealed pancytopaenia (haemoglobin, 10 g/L; haematocrit, 29%; platelets 25 × 10 9 /L, leukocytes, 6.7 × 10 9 / L); hyperbilirubinaemia (8.3 g/dL); γ-glutamyltransferase, 146 U/L; alkaline phosphatase, 241 U/L; and prothrombin time, 76 seconds. Renal function tests were within normal limits. The C-reactive protein level was 16 mg/dL. Arterial blood gas measurement while breathing air revealed marked hypoxia (PaO2, 63 mm Hg), normocapnia (PaCO2, 32 mm Hg), low oxygen saturation (93.7%), and a blood pH of 7.49. A chest radiograph showed bilateral interstitial infiltrates and a computed tomography (CT) scan showed right midzone alveolar shadowing without parenchymal infiltrations ( Figures 1A and 1B) . To exclude the possibility that the patient's severe symptoms were due to P. falciparum, which is sympatric with P. vivax in Brazil, a nested PCR using P. falciparum-and P. vivax-specific primers, was performed [16] . The results demonstrated that coinfection with P. falciparum could be excluded (Figure 2A) . Moreover, sputum, blood, and nasopharyngeal swab samples obtained for culture were negative, as were serological tests for atypical respiratory pathogens, human immunodeficiency virus, histoplasma, coccidians, and paracoccidians. Together, these results demonstrated that the clinical symptoms were exclusively due to P. vivax infection. X-ray and computed tomography (CT) scans of patient with severe Plasmodium vivax malaria Due to the severe condition of the patient, a five-day treatment was initiated, starting with an intravenous loading dose of quinine (1,200 mg) on the first day, followed by oral quinine (600 mg) every 8 hours plus oral doxycycline (100 mg) every 12 hours. The treatment was well-tolerated and parasitaemia became negative within three days. Platelet, erythrocyte and leukocyte levels were all within normal ranges in the following controls. Normal blood levels of glucose-6-phosphate dehydrogenase activity were detected and oral primaquine (30 mg/day) was commenced for two weeks to prevent a relapse. The patient recovered well and the chest CT findings were normal at two months ( Figure 1C ). There have been no recurrences during the follow-up period. In view of the recently suggested association of severe disease in P. vivax with multidrug resistance [14] , expression levels of two genes suspected in having a pivotal role in mediating clinical resistance to chloroquine, pvcrt-o and pvmdr1, were determined. Significantly, the patient with severe symptoms had a 1.6-fold increase in pvmdr1 levels and a 21.9-fold increase in pvcrt-o levels compared to the other patient who presented to the same hospital with P. vivax infection and mild symptoms (Figure 2b ). This finding was validated by analysing parasite material obtained previously from two other patients from the Brazilian Amazon who also had mild symptoms (Additional file 2). There are increasing reports of severe clinical cases exclusively associated with P. vivax infections, involving severe anaemia, renal failure, jaundice, cerebral malaria, seizures, respiratory failure, multi-organ failure, and death [2, 3] ; all these complications are generally believed to be exclusively associated with severe forms of falciparum malaria. Although these same clinical severity criteria are being used for P. vivax, it would be desirable to conduct prospective studies to establish a precise definition of clinical severity in P. vivax. Expression levels of chloroquine resistance genes in severe and mild Plasmodium vivax malaria Figure 2 Expression levels of chloroquine resistance genes in severe and mild Plasmodium vivax malaria. In panel A, nested PCR to identify Plasmodium species. Amplification of gDNA from P. falciparum 3D7 strain and P. vivax from severe patient, respectively, using specific P. vivax primers (lanes 1 and 2) . Amplification of gDNA from P. falciparum 3D7 strain and P. vivax from severe patient, respectively, using specific P. falciparum primers (lanes 3 and 4) . Molecular weight ladder (lane L). A positive reaction is noted when primers for P. falciparum and P. vivax produce amplification products of 205-bp and 120-bp, respectively [16] . Molecular weight in base-pairs (bp). In panel B, relative quantification of pvcrt-o and pvmdr1 transcripts in total RNA obtained from parasites from the severe patient vs total RNA obtained from parasites from a patient from Brazil with P. vivax and non-severe symptoms also presenting to our hospital. The following oligonucleotide primers were designed for the realtime experiments using the Primer Express program (Applied Biosystems). Primers: F-pvcrt-oRT 5'-ATGTCCAAGATGT-GCGACGAT-3';R-pvcrt-oRT 5'-CTGGTCCCTGTATGCAACTGAC-3'; F-pvmdr1RT 5'-AAGGATCAAAGGCAACCCA-3'; R-pvmdr1RT5'-TCAGGTTGTTACTGCTGTTGCTATT-3'; F-pvtubulinRT 5' CCAAGAATATGATGTGTGCAAGTG 3'; R-pvtu-bulinRT 5' GGCGCAGGCGGTTAGG 3'. The patient with severe symptoms had acute lung injury according to the definition of the American-European Consensus Conference on ARDS [18] , with acute onset, bilateral changes in chest radiography, a PaO2/FiO2 ratio of ≤ 300, and absence of clinical left ventricular failure. Importantly, these respiratory complications appeared before initiation of anti-malarial drug treatment, an observation also reported in two other severe clinical cases of lung injury due to P. vivax [19, 20] . Two main hypotheses have been proposed to explain lung damage in P. vivax malaria infections. The first suggests that there is no sequestration of P. vivax-infected reticulocytes in the deep capillaries of internal organs, postulating instead an inflammatory process due to an increase in capillary permeability associated with cytokine-induced damage in the pulmonary epithelium [21] . The second suggests cytoadherence of P. vivax-infected reticulocytes in lung capillaries, causing obstruction of blood flow and reduction of respiratory function before treatment and alveolar capillary damage and inflammation 24 to 48 hours after initiation of anti-malarial drug treatment [22] . The possibility that antibiotics given 24 hours before respiratory failure might have destroyed P. vivax-infected reticulocytes, inducing lung damage and favouring the second hypothesis, cannot formally rule out. Yet, it is clear that the lack of sequestration in P. vivax needs to be re-evaluated, as cytoadherence has been hypothesized to occur in both the spleen [23, 24] and the lungs [22] . Chloroquine is currently the first-line treatment for P. vivax, but resistance has been rapidly increasing since it was first described in two cases of treatment failures in Papua New Guinea [4] . The clinical case reported here originated in Manaus, Brazil, where P. vivax chloroquine resistance and severe disease are now being reported [25, 26] . Interestingly, after the appearance of chloroquine resistance in P. vivax, reports of clinical severity exclusively associated with this human malaria parasite started to appear [2] . Moreover, multidrug resistance has been recently suggested to be associated, albeit not directly linked, with severe disease in P. vivax [14] . Remarkably, higher expression levels of pvmdr1 and pvcrt-o, in particular, were found in parasites from the patient with severe clinical symptoms compared to three patients with mild symptoms. The levels were increased by up to 2.9 fold in the case of pvmdr1 (when including all samples) and up to 21.9 fold in the case of pvcrt-o. Confounding effects due to sample concentrations or contamination with P. falciparum or human material were excluded by using P. vivaxspecific primers and β-tubulin as an internal control and calibrator. This data thus indicates that severe vivax disease could be associated to molecular markers. There are increasing reports of severe disease exclusively associated with vivax malaria and this study presents the first one in Spain. The use of PCR excluded incompetent microscopy, cryptic mixed infections or sequestered P. falciparum reinforcing the need of using PCR as a new diagnostic tool to avoid default diagnosis of severe malaria as due to P. falciparum. The finding on increased expression levels of parasite genes likely involved in chloroquine resistance supports to furthering explore the potential of pvmdr1 and particularly pvcrt-o as molecular markers of severe disease in P. vivax. Publish with Bio Med Central and every scientist can read your work free of charge

Health workers' views on quality of prevention of mother-to-child transmission and postnatal care for HIV-infected women and their children

BACKGROUND: Prevention of mother-to-child transmission has been considered as not a simple intervention but a comprehensive set of interventions requiring capable health workers. Viet Nam's extensive health care system reaches the village level, but still HIV-infected mothers and children have received inadequate health care services for prevention of mother-to-child transmission. We report here the health workers' perceptions on factors that lead to their failure to give good quality prevention of mother-to-child transmission services. METHODS: Semistructured interviews with 53 health workers and unstructured observations in nine health facilities in Hanoi were conducted. Selection of respondents was based on their function, position and experience in the development or implementation of prevention of mother-to-child transmission policies/programmes. RESULTS: Factors that lead to health workers' failure to give good quality services for prevention of mother-to-child transmission include their own fear of HIV infection; lack of knowledge on HIV and counselling skills; or high workloads and lack of staff; unavailability of HIV testing at commune level; shortage of antiretroviral drugs; and lack of operational guidelines. A negative attitude during counselling and provision of care, treating in a separate area and avoidance of providing service at all were seen by health workers as the result of fear of being infected, as well as distrust towards almost all HIV-infected patients because of the prevailing association with antisocial behaviours. Additionally, the fragmentation of the health care system into specialized vertical pillars, including a vertical programme for HIV/AIDS, is a major obstacle to providing a continuum of care. CONCLUSION: Many hospital staff were not being able to provide good care or were even unwilling to provide appropriate care for HIV-positive pregnant women The study suggests that the quality of prevention of mother-to-child transmission service could be enhanced by improving communication and other skills of health workers, providing them with greater support and enhancing their motivation. Reduction of workload would also be important. Development of a practical strategy is needed to strengthen and adapt the referral system to meet the needs of patients.

Prevention of mother-to-child transmission (PMTCT) has been considered a simple intervention: just giving medication to prevent viral transmission from mother to child. Now, though, PMTCT is recognized as a comprehensive set of interventions requiring capable health workers. It starts with testing pregnant women for HIV, preferably during their first antenatal visit. When giving the test result, health care workers should provide good counselling, including information about PMTCT options. The health system should ensure that HIV-positive women receive the PMTCT services that they choose and should provide postnatal care. All along the timeline from finding out their serostatus to getting treatment for HIVrelated problems, women and their children should be followed closely. The need for comprehensive and longterm care for HIV-infected women has become a challenge for health systems, particularly where lack of coordination among different facilities is common [1, 2] . Viet Nam's HIV epidemic is still in a concentrated phase, with the highest seroprevalence among high-risk key populations, including injecting drug users (IDU), female sex workers (FSW) and men who have sex with men (MSM). Hanoi is one of 10 provinces/cities reported to have the highest number of HIV infections per 100 000 inhabitants. After the first case of AIDS was identified, in 1992, the HIV epidemic in Hanoi increased sharply by 1994. The capital has 12 628 persons living with HIV/AIDS (PLHIV), mostly IDU from poor families. Currently, there are 3623 AIDS patients and 2081 who had died of AIDS in the city. Although HIV is predominantly concentrated among IDU and FSW, it is gradually spreading among the general population. In 2007, HIV prevalence among pregnant women attending antenatal care (ANC) clinics in the Hanoi was 0.34% [3] . Hanoi was selected as the study site because comprehensive PMTCT care is theoretically available there. Hanoi had 45 hospitals, 290 surgeries and five international hospitals. The national hospitals in Hanoi serve as referral centres for the northern half of Viet Nam. HIV testing and counselling for pregnant women are offered at health facilities at district or higher level, but often only after the 28 th week of gestation. HIV-positive women are referred to provincial or national hospitals for ARV prophylaxis, delivery and postnatal care. Hanoi health services have sufficient supplies of prophylactic ARV to meet the demand. Antiretroviral therapy for adults is available at district level or higher. HIV-exposed infants are offered polymerase chain reaction (PCR) testing and free infant feeding formula for at least six months, while free paediatric ARV is available for children three years of age or older. The extensive health care system in Hanoi reaches the commune level, but multisectoral and cross-programme collaboration to link the pillars of the World Health Organization's (WHO) comprehensive approach to PMTCT are weak [4] . For example, there is little collaboration between the programmes for HIV/AIDS and family planning. Our previous work suggested that a large number of HIV-infected pregnant women remain undetected by the health system [5] . In addition, a number of barriers result in failure to access PMTCT during pregnancy and delivery [6] [7] [8] [9] [10] [11] . Among the weak points identified were that HIV-infected women received inadequate information about postnatal care, but even when they had knowledge, many expressed fear of stigma and discrimination that reduced their access to care; HIV testing is not available via health services at commune level, where many pregnant women go for care and delivery; and women feared lack of confidentiality of HIV test results [4, 12] . Our previous studies on the experiences and views of women about the provision of PMTCT in Hanoi included criticisms about the quality of services provided by health workers [4] . Other studies in Asia found that health workers were unwilling to provide appropriate care for HIVpositive pregnant women, often because of their own fear or lack of knowledge, or because of high workloads and lack of staff [13, 14] . Inadequate health care delivery may be caused by a variety of factors, but we need to identify the main issues before planning interventions to strengthen it. We report here the health workers' perceptions on the factors that lead to their failure to give good quality PMTCT. The findings should inform the development of a more effective programme for the fourth prong of the WHO-recommended comprehensive PMTCT programme. Sampling of study sites was based on availability of services in Hanoi and their level and function in the health care system. All hospitals in Hanoi that provide ARV, PMTCT and opportunistic infections (OI) treatment were selected, including all health facilities providing HIV testing. Additionally, all health facilities in one high-prevalence and well-resourced district were selected. In that district, interviews were carried out at all levels of the health system involved in PMTCT: district health centre, district maternity ward, district committee for family planning and mother and child health, and preventive medicine centres, including HIV testing sites. The research team also visited commune health stations in the district. Details are presented in Table 1 . Selection of respondents was based on their function, position and experience in the development or implementation of PMTCT policies/programmes. Interviewees were first screened to check that they had appropriate positions and at least one year of experience with PMTCT, so that they could provide insightful information. They included doctors, nurses, midwives, counsellors, laboratory technicians and programme managers. Detailed information on interviewees is presented in Table 2 . We conducted semistructured interviews with these 53 health workers about their experience in implementation of services for PMTCT, their point of view about users of their services and their perception about the challenges they faced in providing good PMTCT services in their health facility. In addition, unstructured observations were made in nine health facilities, in waiting rooms, counselling rooms, ANC examination wards, delivery wards, postnatal wards, outpatient and inpatient clinics for ARV and OI facilities, and laboratories. The interviewers were four trained public health and social science researchers. Institutional ethical approval was obtained from the Scientific Committee of Hanoi Medical University and written informed consent was obtained from all interviewees, who were invited to participate voluntarily. The interviews were conducted privately and anonymously. A code book was developed focusing on key findings and terminologies. The transcripts of the semistructured interviews were coded, entered and analysed by means of N-VIVO software. Many hospital staff explained their reasons for not being able to provide good care; the most frequently heard was high workload. Observation confirmed that the national and provincial hospitals' ANC caseload was high: the wards were always crowded. One of the main obstetrics hospitals provides ANC to between 200 and 400 pregnant women daily. None of the health facilities offering PMTCT services had recruited additional staff to provide counselling. However, the workload in ANC facilities at district and commune level was not so high, according to the respondents. "There are only very few health staff with only basic information on counselling in this hospital. Poor knowledge and skill is common problem here." Counsellor, ANC provincial hospital Especially at district or lower level, knowledge is limited regarding ARV prophylaxis and on follow-up care, such as continuing replacement feeding (RF) supplies, infant testing and services for HIV-infected mothers and exposed infants. Consequently, health workers cannot provide adequate counselling on these issues before women are discharged. "What I can do is to provide information about her HIV test result. I know that there is a medication to prevent transmission of HIV from mother to child, but I don't know exactly. Our common practice is to refer her to provincial hospital." Midwife, ANC commune level The staff in most health facilities reported having had limited training on PMTCT in general and on counselling in particular; that they apply knowledge and skills gained from observing colleagues conducting counselling sessions; and that refresher courses are rare. The duration of the training varied from two days to two weeks. After the training, counselling is added to their regular ANC or maternity work. "What we do now is only to inform. [Counsellors] lack knowledge. If they will be trained, who will train them?" Manager, ANC provincial hospital "We counsel from our experience. To me, our counselling may be not complete. We don't even know what might be our shortcomings". Nurse, ANC commune level An important point of the comprehensive PMTCT approach is that HIV-infected women should be provided with several different services -such as ARV prophylaxis, formula, counselling and HIV testing for exposed children -provided by different facilities. However, there are no inter-or intra-hospital linkages to make the PMTCT comprehensive. For example, family planning services at a national obstetric hospital are not linked to other departments, including the infectious disease department that provides ANC and delivery services for HIV-infected women. Women are seldom referred to ARV sites for clinical staging or immunological assessment. Referral to postnatal care and social support for both mothers and children is not available at the hospital exit point. "There is no linkage with obstetric hospitals; they never directly inform us. It is very difficult to know which HIV+ patient or child has been referred to this paediatric hospital for follow-up by what hospital. We don't treat the mothers here. We only provide counselling to them. Support groups? There are some but we don't know where they are." Doctor, paediatric ARV national hospital Many health workers stated that their task is to provide services available in their own facilities but not services provided by other departments or facilities. Lack of medications was another reason given for not being able to provide services for HIV-infected women. In many health care facilities, the ARV was not consistently available. Often even single-dose Nevirapine (NVP) was lacking, for women who were tested only at the time of labour. Even in the two PMTCT sites in the city, shortage of ARV for adults was observed to happen every few weeks. One problem with the NVP for children was that it is provided as a large bottle (200 ml) of syrup. Once it has been opened, it cannot be kept for long, but very few HIVexposed children were identified each day. That means that each bottle was not fully used, and that later, drugs were lacking when supplies ran out. "In some periods, there is a shortage of SD-NVP for PMTCT. We could not do anything in that situation. In practice, the NVP syrup for children is very inconvenient to use. On one day, we have no more than two children to treat; we have to open one bottle for them and the rest of the syrup is unused. The syrup quickly runs out, and then we don't have medication for another child." Manager, ANC provincial hospital "We also counsel them to use condoms. If someone asks me what they should do to avoid unwanted pregnancy then I tell them. But I do not have condoms to give them for free for family planning." Counsellor, ANC district level Another issue is that there are no national guidelines on counselling and testing. Observation showed that facilities at provincial and national levels had counselling and PMTCT guidelines and protocols developed by the projects that support those facilities, but most facilities at district or lower level do not have guidelines or even access to them. Moreover, health workers at all levels often complained about the lack of attention to the needs of health workers when they have to work in a high-risk environment. "Among 1,000 health workers, how many want to provide care and treatment for HIV-infected patients? There is no good compensation regimen to support staff working with HIV-infected patients. There is no benefit to save the life of patients in the late period, so how could we be enthusiastic?" Doctor, paediatric ARV national hospital "We receive extra pay for providing treatment for HIVinfected people. But it is just for one health staff while all [12] staff in my department provide service. We have to share among us." Doctor, adult ARV district hospital Fear of infection Many respondents admitted that they were afraid of HIV transmission from patients, either because they feared being injured by the patients or through an occupational accident, because they lack protective equipment. Observation at adult and pediatric ARV sites supported this finding. Health workers confront their fear of infectionTo reduce both fear and risk of infection, health workers often find ways to protect themselves, either by trying to identify which patients might be infected with HIV, by using protective equipment, or by avoid exposure to patients as much as possible. "It is easier for us to prevent transmission if we know who among the patients is infected with HIV." Midwife, ANC provincial hospital Observation revealed good practice of precaution when health workers assisted deliveries for HIV-positive women, but not always for HIV-negative women. All health workers said they knew how to protect themselves against occupational exposure to HIV and did so very carefully if they knew who was infected with HIV. "Staff wear protective uniforms, maintain all hygiene practices and disinfection procedure on all equipment used." Nurse, adult ARV district level However, even if health workers want to protect themselves by using protective equipment, not all health facilities can provide these means for them. "Health workers do not have enough protective clothes. We have to use cloth coats and short gloves when assisting deliveries for HIV-infected pregnant women. So we often wear a raincoat on top of the cloth coat." Nurse, ANC provincial hospital Although hospital managers reported that occupational exposure is rare, among the study population we found five health workers who claimed to have had an exposure to blood that they thought might have put them at risk for HIV infection, either because of needle-sticks or blood that went to their eyes. They all informed us that ARV prophylaxis for prevention of occupational exposure is free of charge at an ARV site at district hospital. But only two of them took medication because the others turned out not to need prophylaxis after closer assessment of their exposure. Prevention of HIV transmission became a good excuse for health workers to avoid taking care of HIV-infected women, or if they had to provide care, to isolate the HIVinfected women for easier control and management. "We offer counselling, family planning, nutrition, delivery and care after delivery at home, vaccinations for tetanus. We offer this for normal pregnancies including those with hepatitis B. Women who have high risk with HIV are referred. It is not our business." Doctor, ANC site "Not everyone understands thoroughly about stigma and dread. The more they know, the more they fear and they try to push responsibility to others." Manager, PMTCT site Many hospital managers emphasized that although the number of HIV-infected women attending their hospitals has been increasing, the number may still account for quite a small number of women in community. Observation at all the health facilities revealed that HIVinfected women were often placed in a separate room or area. Even at a high-level hospital, where two or three patients often had to share one bed, there was still an empty bed in the room for HIV-infected patients. The manager in an ANC facility explained: "HIV is an infectious disease, more severe than hepatitis B. Therefore patients should be controlled carefully to avoid transmission to other patients and staff." Not only fear of HIV infection influenced the quality of care provided to HIV-positive women. Some of the weaknesses in providing service were related to the views of health workers regarding HIV-infected people, who are often seen as drug users or sex workers, or as having a "strange appearance". The real or expected behaviour of such people also induces fear and other negative emotional responses in the health workers. "They often have tattoos and never dress well. There are spots on their arms. Easy to recognize their bluish black lips." Doctor, ANC national hospital "They inject in our department. How can we have a good attitude toward them?" Nurse, ANC national hospital Health workers did admit that not all HIV-infected people behaved badly towards them. However, a few bad experiences could give all staff a negative attitude about HIVinfected people in general. "Almost all HIV-infected patients cannot be trusted. For instance, when they know they have opportunity to have care and support, they often find ways to get as much as possible. Or if they want to leave the hospital, they often lie to the nurse that the doctor already agreed. Or when they have to pay the hospital fee, they often tell the lie that they will pay tomorrow but after that they disappear." Nurse, paediatric ARV national hospital "When you have to work with them [HIV-infected patients], you will see the difficulties. It's already hard to gain trust from normal patients. Now we have to serve the very scoundrel social class and at the same time, we receive very low salary. We have to provide service because it's our responsibility but we are not happy because they [HIV-infected patients] are drug users, they are very rude. My experience shows that health workers should not be too good to HIV patients." Doctor, ARV district level Some of these attitudes are based on real experiences, but many are also based on prejudicial expectations, and women wishing to access PMTCT services will be victims of that stigma, too. Most of the health workers agreed that the quality of care could be improved by several interventions addressing both individual and structural issues. Reducing workload and providing better compensation for working with risks were mentioned by almost all health workers at provincial and national level as important solutions. In the interviews, hospital managers suggested that it is very difficult to hire new staff because of the limited budget allocated from the government. A better solution would be to rearrange the services in a more logical structure, for example, to replace individual pretest counselling (which is often not offered anyway) by group counselling, to use peer counsellors to provide counselling and follow-up of care (reducing the burden for health workers and improving the quality of counselling) and to improve the quality of services at a lower level to reduce the burden on the higher levels. "If we have more equipment, we can deal with our workload with even few staff. For instance, if we are provided a video, leaflets on HIV and PMTCT, and a room with table and chairs, we can do group counselling for pregnant women." Manager, ANC national hospital "I have seen in a hospital in Thailand that peer counsellors can work in hospital to help health workers doing some administration work, providing counselling, making appointments. We may need to think about how to apply their experience." Manager, ARV district level However, many doctors and nurses are still unconvinced about the involvement of peer counsellors in the health service, because they feel that peer counsellors do not have enough capacity and the appropriate attitude to do this work, or even bring the potential for crime and corruption into the hospital. "You can find good peer counsellors in other countries but not in Viet Nam. They have low education. They may become drug dealers or whatever, we don't know." Doctor, paediatric ARV national hospital Another possible intervention is training and updating information for health workers. Midwives and nurses said they needed to improve their basic knowledge on HIV/ AIDS and PMTCT because they received less training than doctors, while doctors preferred to have advanced PMTCT training. All of them asked to be updated routinely on PMTCT. "Although training has often been provided to doctors but not nurses, in fact training should be conducted more often for both of them." Doctor, paediatric ARV national hospital Regarding the system, hospital managers admitted that much needs to be done to improve the quality of service -for example, improving ARV procurement, developing detailed PMTCT guidelines taking into consideration the staff's high workload, increasing availability of high technology equipment and providing sufficient protective equipment for health workers. "Although ARV is sufficient for all HIV-infected women in Hanoi, in the provincial hospital, there are some periods when they lack medicines. The provincial hospital should make a plan on how much medicine they need. The procurement facility should make administrative procedure short and easy for hospitals to get the medicines." Manager, ANC national hospital "We have implemented PMTCT programme for more than five years without a PMTCT guideline to instruct us. I think a PMTCT guideline should be developed with involvement of all planners and implementers." Manager, ANC provincial hospital Improvement of the referral system is seen as an important task that needs to be addressed in the near future. Hospital managers propose to have regular meetings of the health network that provides different services for HIV-infected people. "When there are not a lot of patients, I think it's appropriate to divide tasks among different hospitals. But we also need to link all hospitals. In practice, the linkage is loose: that makes difficulties for patient access to services. It is due to lack of information about services provided by other places, lack of active coordination to link between different services, people are too busy to think about other services besides the treatment that we can provide. Some health workers are not aware of the need to refer patients, or if they do, patients don't want to go and disclose their positive status in other hospitals, or may not go even because of lack of referral forms." Manager, ARV district hospital Some health workers also proposed that all services should be offered at one point for better coordination. "Now we have Global Fund project providing ARV at district health centre. Why don't we provide PMTCT and paediatric treatment at the same facility? I have also heard that there are some self-help groups working at district level that can provide further support for HIV-infected people. The Ministry of Health should think about how to make one facility able to provide all services. That could help to avoid loss to follow-up, which is common among HIV-infected people." Manager, ANC national hospital The health workers did perceive not only problems but also solutions and seemed to have some motivation to solve the problems and to provide better services for HIVpositive women seeking PMTCT services. Studies in Viet Nam have demonstrated that both HIVinfected and non-infected women had many criticisms of ANC and delivery services, about provision of information about PMTCT and counselling, and about stigma displayed by health care workers [3, 15] . At present, the health service has not yet addressed this gap. Access to comprehensive PMTCT is still very poor, even in such a well-resourced setting as Hanoi [16] . Because the health care workers are subjected to many accusations about their performance in this context, this study was undertaken to find out their opinion regarding these gaps and weaknesses. Health care workers usually want to do a good job and provide good care for patients. However, they are often unable to provide as good care as they would like, particularly in facilities with an overload of clients [17, 18] . Results of a survey among women who had been pregnant in Hanoi revealed that they attended and paid for ANC services at higher-level health facilities (provincial and district hospitals) rather than go to commune health stations where ANC services are free of charge [4] . High workloads were observed at provincial hospitals, while district hospitals and commune health stations appeared to have more time and space for pregnant women. Studies of ANC services in Vietnam have identified a number of weaknesses. Staff shortages and staff motivation can significantly affect the quality of service, especially for counselling, which takes a long time to do well [12, [19] [20] [21] [22] . In the case of HIV and PMTCT, additional reasons for the unsatisfying performance included inadequate knowledge and skills due to lack of training, medicines, protective equipment and practice guidelines. Health care workers had poor knowledge about HIV and about prevention of occupational exposure to HIV, especially at district or lower levels. Even in the expected sources of expertise, the medical schools around Vietnam, 70% of medical students and 48% of lecturers recapped used needles by hand, while two thirds always cleaned their hands with antiseptic after contact with blood. Sixty percent of medical students and 37% of lecturers had been exposed directly to blood or body fluids and were worried about HIV transmission. However, 15% of the respondents recommended antibiotics for post-exposure prophylaxis, while one third proposed ARV prophylaxis [23] . These results reveal a disturbing lack of knowledge and awareness about HIV, even among the medical profession. Lack of practical needs can become an excuse for health care workers to justify their fear of HIV infection and their reluctance to provide good services for HIV-infected people [7, 24] . As the HIV epidemic has evolved in Viet Nam, both governments and international donors have given priority to prevention and surveillance activities. The main reason is that Viet Nam has had success up to now using surveillance and containment to control infectious diseases such as polio, SARS and, more recently, avian flu. HIV/AIDS policy and practice also aims foremost at controlling the spread of the virus and has paid less attention to providing care and treatment to individuals already affected. In keeping with past experience in other epidemics, health staff perceived HIV-infected persons as sources of contamination, who should be isolated. Health care workers are the key providers of medical care. Stigma from health care workers can reduce patients' ability to manage their infection and gain access to health care [7, 25] . Persons infected with HIV are often grouped with drug users and sex workers as marginalized, discriminated-against and criminalized elements in society, also by health workers. Stigma towards HIV-infected persons has been documented in health care settings all over the world [26, 27] . Showing a negative attitude during counselling and provision of care, treating in a separate area and avoidance of providing service at all are perceived as enacted stigma by HIV-infected patients. On the other hand, from the health workers' point of view, these actions result from a combination of factors: high workload and personal priority-setting influenced by fear of being infected as well as distrust towards almost all HIV-infected patients because of the association with antisocial behaviors [28, 29] . When health care workers have fear and lack knowledge, they can find reasons not to focus their attention on the HIV-positive patients and give those reasons for not providing service as they think/ know they should. Moreover, health care workers are not only a source of stigma from the perspective of HIVinfected people, but can also become recipients of stigma from colleagues and family because of their exposure to HIV-infected patients. The best and most feasible solution is to provide training and reference materials for health workers, to inform them about HIV transmission routes, universal precautions and post-exposure prophylaxis. Reduction of workload would also be important [24] . The results of this study also suggest that the quality of PMTCT service could be enhanced by improving communication and other skills of health workers and providing them with greater support and motivation. A positive atmosphere in the health facilities should be promoted by normalizing HIV-related services, and undertaking behaviour-change communication campaigns aimed at staff of the health facilities. Feedback from service users could be used as one way to evaluate the quality of service. In addition, health facilities should make ARV continuously available. The health workers' fear could also be reduced by ensuring that they have and use the protective clothing they need to maintain good hospital hygiene. It will be more difficult to address the issues of fear and stigma towards drug users and sex workers. Self-help groups of both drug-using and non-drug-using women in Hanoi and other countries were able to improve the relationship and communication between health care staff and patients/clients; peer counsellors and a buddy system led to improvements in the health care provided to and accessed by the women [3, 30, 31] . Successful examples of this intervention have been documented among clubs for tuberculosis patients [32] , alcoholics, cancer patients and patients with chronic illnesses and mental problems [30] . Continuous care and support for HIV-infected mothers after delivery was often not seen as a need to be addressed [12, 33] . In practice, the fragmentation of the health care system into specialized vertical pillars including a vertical programme for HIV/AIDS is a major obstacle to providing a continuum of care. Medical treatment, including ARV provision and medications for OI, is increasingly available but is often not accessible to PLHIV because of a weak referral system and social stigma [3, 7] . The vertical organization of the health care system and the contradictory mandates between sectors obstruct the effective collaboration and referral between different services that the women and their families need. A lack of multisectoral collaboration is a barrier to effective information exchange for patients between national staff in different facilities [34] . Providing information about topics such as abortion, clean needle exchange programmes and condoms is also politically sensitive in voluntary counselling and testing sites. The study suggests that information on available services should be made known to health workers. Frequent meetings between different service sites should be organized, with the involvement of high-level health staff that can make decisions, to update information on services available and provide feedback on the quality of the referral system. Development of a practical strategy is needed to strengthen and adapt the referral system to meet the needs of patients. For example, comprehensive services for HIVinfected people should be provided at one service site at district level [2]. As information was collected by means of qualitative methods, the identified factors that lead to their failure to give good quality PMTCT could not be quantified and be representative for the health care worker population in Hanoi. In conclusion, the results of this study show that health care workers also face a number of barriers in their efforts to provide good PMTCT services at different levels of the health services in Hanoi. These include a high workload, a lack of equipment and materials, lack of training and skills updating, the common fear of the type of patients who may present with HIV, and little support to improve their performance. These weak points can be addressed by a number of feasible interventions. Results of the study contribute to the picture of the PMTCT programme not only in low-prevalence settings, as in Asian countries, but also in high-prevalence settings with weak health care systems, such as African countries, and may require different interventions to improve the quality of the service.

Inhibition of Lassa Virus Glycoprotein Cleavage and Multicycle Replication by Site 1 Protease-Adapted α(1)-Antitrypsin Variants

BACKGROUND: Proteolytic processing of the Lassa virus envelope glycoprotein precursor GP-C by the host proprotein convertase site 1 protease (S1P) is a prerequisite for the incorporation of the subunits GP-1 and GP-2 into viral particles and, hence, essential for infectivity and virus spread. Therefore, we tested in this study the concept of using S1P as a target to block efficient virus replication. METHODOLOGY/PRINCIPAL FINDING: We demonstrate that stable cell lines inducibly expressing S1P-adapted α(1)-antitrypsin variants inhibit the proteolytic maturation of GP-C. Introduction of the S1P recognition motifs RRIL and RRLL into the reactive center loop of α(1)-antitrypsin resulted in abrogation of GP-C processing by endogenous S1P to a similar level observed in S1P-deficient cells. Moreover, S1P-specific α(1)-antitrypsins significantly inhibited replication and spread of a replication-competent recombinant vesicular stomatitis virus expressing the Lassa virus glycoprotein GP as well as authentic Lassa virus. Inhibition of viral replication correlated with the ability of the different α(1)-antitrypsin variants to inhibit the processing of the Lassa virus glycoprotein precursor. CONCLUSIONS/SIGNIFICANCE: Our data suggest that glycoprotein cleavage by S1P is a promising target for the development of novel anti-arenaviral strategies.

Lassa virus (LASV) belongs to the family Arenaviridae, which are enveloped, single-stranded RNA viruses distributed worldwide. Based on their antigenic relationships and geographic distribution, arenaviruses are divided into two major groups. The Old World group includes the prototype of this family, lymphocytic choriomeningitis virus (LCMV), and LASV, which is endemic in West African countries and causes every year thousands of human infections with hemorrhagic fever as a severe clinical manifestation [1] . The New World group includes among others Machupo, Junin, Guanarito and Sabia viruses which can cause viral hemorrhagic fever (VHF). With the exception of the New World virus Tacaribe, which was isolated from Artibeus bats, arenaviruses are rodent-borne viruses [2] . Over the past few years great efforts have been made to find potential therapeutic and vaccination approaches in the arenavirus field (reviewed in [3, 4, 5] ). Until now there is no specific and effective treatment available to combat hemorrhagic fevers caused by arenaviruses. Administration of convalescent plasma has been reported to reduce the mortality rates of patients with Argentine hemorrhagic fever, however, 10% of immune-plasma recipients developed a late neurological syndrome of unknown origin [6] . The only existing drug used to treat Lassa fever and certain South American hemorrhagic fevers is the broad-spectrum antiviral agent ribavirin, a ribonucleoside analogue, which has shown to be partially effective if given early in the course of illness [7, 8, 9, 10] . Even though the drug is relatively inexpensive for patients in highdeveloped countries, it is still unaffordable for many of those living in West Africa and South America. Moreover, several adverse effects have been associated with ribavirin therapy in patient studies and animal models [11, 12, 13, 14, 15] . The lack of effective disease control measures as well as the discovery of new fatal arenavirus species that pose a risk of epidemic potential [16, 17] , emphasize the need for novel therapeutic interventions. Lassa virions are pleomorphic lipid-enveloped particles that contain two single-stranded RNA segments, designated L (large) and S (small), encoding four viral proteins in a unique ambisense coding strategy. The L segment encodes the viral RNA-dependent RNA polymerase (L) and the small zinc finger matrix protein (Z) [18] ; the S segment encodes the virus nucleoprotein (NP) and the virus surface glycoprotein precursor (preGP-C) [19] . preGP-C is cleaved co-translationally into a stable signal peptide and GP-C [20] . Post-translational maturation cleavage of GP-C by the proprotein convertase site 1 protease (S1P, [21] ), also known as subtilisin kexin isozyme-1 (SKI-1, [22] ), leads then to the generation of the distal receptor-binding subunit GP-1 and the transmembrane-spanning fusion competent subunit GP-2 [23] . Together with the signal peptide these subunits form the tripartite glycoprotein spike complex on the viral surface [24, 25] . The glycoproteins of the Old World arenaviruses LASV and LCMV were the first viral glycoproteins that were shown to be proteolytically processed by S1P [23, 26] , which normally plays important physiological regulatory roles in cholesterol metabolism, ER stress response, cartilage development and other cellular processes [21, 27, 28, 29, 30, 31] . Using systematic mutational analysis of the LCMV GP cleavage site, the consensus motif R-(R/K/ H)-L-(A/L/S/T/F) was determined, which is conserved in the glycoprotein sequences of the Old World viruses LASV, Mopeia and Mobala, as well as the New World virus Pichinde, suggesting that all arenavirus glycoproteins are cleaved by S1P [26, 32] . Indeed, more recently Rojek et al. reported that glycoproteins from the New World hemorrhagic fever viruses Junin, Machupo and Guanarito are also processed by S1P, although Guanarito possesses a protease recognition motif that differs from known arenavirus GP consensus cleavage sequences, indicating a broader substrate specificity of S1P than previously anticipated [33] . Proteolytic activation of LASV GP-C by S1P is not necessary for transport of GP-C to the cell surface, where budding of arenaviruses occurs, but is essential for incorporation of the cleaved subunits into virions, and thus, for the formation of infectious viral particles. In the absence of GP-C cleavage, enveloped non-infectious LASV-like particles are released containing L, NP, Z protein and viral RNA but are devoid of viral glycoproteins [23] . Similar results were described for LCMV and New World hemorrhagic fever viruses [33, 34] . In addition to its important role in the arenaviral life cycle, S1P is critical for the infectivity of Crimean-Congo hemorrhagic fever virus (CCHFV), a member of the Bunyaviridae family, through processing of the glycoprotein Gn [35, 36] . These findings make the inhibition of S1P particularly interesting for the development of a novel antiviral therapeutic that will target pathogenic viruses known to be processed by S1P. A successful approach to inhibit proprotein convertases involves genetically engineered antitrypsins, which are derived from a 1antitrypsin (a 1 -AT). a 1 -AT is a serine protease inhibitor (serpin) with a characteristic exposed reactive center loop (RCL), which mediates binding to the active site of its target protease. The exploration for the potential use of modified antitrypsins with an altered inhibitory spectrum has been guided by the discovery of a natural variant of a 1 -AT, known as Pittsburgh (a 1 -AT-PIT), found in a patient who had a severe bleeding disorder caused by mutation of the P1 reactive center residue of antitrypsin from methionine to arginine [37] . This substitution changed its specificity from elastase to thrombin and other coagulation proteases. Due to the introduction of a second mutation from alanine to arginine at P4 of the RCL, the engineered a 1antitrypsin variant Portland (a 1 -AT-PDX) showed high affinity for furin [38] . a 1 -AT-PDX efficiently inhibited the formation of infectious HIV, measles virus, and human cytomegalovirus progeny by blocking furin-dependent processing of glycoproteins gp160, F0 and gB, respectively [38, 39, 40, 41] . Pullikotil and coworkers used this approach for the generation of highly selective a 1 -antitrypsin variants specific for S1P by introducing various S1P recognition motifs into the RCL of a 1 -antitrypsin [42] . The adaptation of a 1 -antitrypsin towards S1P efficiently inhibited the processing of the S1P substrates SREBP-2 (sterol regulatory element binding protein), ATF6 (activating transcription factor 6) as well as CCHFV glycoprotein [42] . However, the effect of these inhibitors on CCHFV infection was not analyzed in that study. To block cleavage of the LASV glycoprotein, we generated here recombinant a 1 -antitrypsin variants mimicking the S1P recognition motifs RRIL, RRVL and RRYL that exhibited the greatest inhibitory potential based on immunoblot quantification. In addition, we used an a 1 -AT construct that contains the LASV GP cleavage motif RRLL in its RCL. Using a doxycycline regulated expression system we demonstrate that S1P-adapted a 1antitrypsin variants efficiently block proteolytic maturation of the glycoprotein precursor GP-C, whereas a furin-specific a 1 -AT had no effect on GP-C processing. Virus replication of both a replication-competent recombinant vesicular stomatitis virus expressing the LASV glycoprotein GP-C (VSVDG/LASVGP) and authentic LASV was significantly inhibited in the presence of S1P-specific a 1 -antitrypsins. The degree of inhibition of viral replication correlated with the ability of the different a 1 -antitrypsin variants to inhibit the processing of LASV GP-C. Since glycoprotein processing by the endoprotease S1P is not only critical for virus infectivity of LASV [23] , and other arenaviruses causing hemorrhagic fever [33] , but also for members of the Bunyaviridae family [36] , further optimization based on our findings could lead to a potent and specific S1P inhibitor with the potential treatment of certain VHFs. cDNA of the open reading frame of rat a 1 -antitrypsin (Gene Bank Accession Number NM_022519) (a kind gift from Dr. G. Thomas, Vollum Institute, Oregon Health & Science University, Portland, USA) was inserted into pSG5 and used as a template to generate S1P-specific a 1 -antitrypsin variants by recombinant polymerase chain reaction (PCR) using overlapping oligonucleotides [43] . The sequences of the oligonucleotides used are listed in Table S1 . The resulting full-length PCR products were digested The virus family Arenaviridae includes several hemorrhagic fever causing agents such as Lassa, Guanarito, Junin, Machupo, and Sabia virus that pose a major public health concern to the human population in West African and South American countries. Current treatment options to control fatal outcome of disease are limited to the ribonucleoside analogue ribavirin, although its use has some significant limitations. The lack of effective treatment alternatives emphasizes the need for novel antiviral therapeutics to counteract these life-threatening infections. Maturation cleavage of the viral envelope glycoprotein by the host cell proprotein convertase site 1 protease (S1P) is critical for infectious virion production of several pathogenic arenaviruses. This finding makes this protease an attractive target for the development of novel antiarenaviral therapeutics. We demonstrate here that highly selective S1P-adapted a 1 -antitrypsins have the potential to efficiently inhibit glycoprotein processing, which resulted in reduced Lassa virus replication. Our findings suggest that S1P should be considered as an antiviral target and that further optimization of modified a 1 -antitrypsins could lead to potent and specific S1P inhibitors with the potential for treatment of certain viral hemorrhagic fevers. Effect of S1P Inhibition on LASV Replication www.plosntds.org with BamHI and NheI and cloned into the tetracycline (Tet)controlled inducible mammalian expression vector pTRE2hyg (Clontech). The accuracy of all constructs was confirmed by DNA sequencing. To generate stably expressing cell lines, Chinese hamster ovary (CHO)-K1 Tet-On cells (Clontech) were transfected with pTRE2hyg containing the a 1 -antitrypsin constructs using Lipofectamine 2000 (Invitrogen) according to manufacturer's instructions. Cells were then cultured for 2 weeks under selective pressure in the presence of 500 mg/ml Hygromycin B, the selection agent for the a 1 -antitrypsin expressing plasmid, and 500 mg/ml G418, the selection agent for the rtTA (reverse Tet-controlled transactivator) cassette. The selective media were replaced every 3 days. Well-separated antibiotic-resistant cell clones were individually isolated with cloning cylinders (Sigma). Therefore, a small volume of Trypsin-EDTA (Sigma) was added and the culture dish was incubated briefly at 37uC until cells detach. Cells were then collected from inside the cylinder and transferred to individual wells of a 24-well plate for further growth in selective medium. When grown to confluence, cells were transferred to larger flasks. Protein expression was induced with 1 mg/ml doxycycline (Clontech) and analyzed by Western Blot and immunofluorescence. Stable cell lines showing similar expression levels of the various a 1 -antitrypsins were chosen for further experiments. Vero E6 cells (green monkey kidney) were cultured in Dulbecco's modified Eagle medium (DMEM, Gibco) and CHO-K1 Tet-On cells in DMEM/F12 (Gibco), both media containing penicillin (100 U/ml), streptomycin (100 mg/ml), and L-glutamine (2 mmol/l) (all from Invitrogen) as well as 10% fetal bovine serum (PAN Biotech). S1P-deficient SRD-12B cells (a generous gift from Dr. J. L. Goldstein, Department of Molecular Genetics, University of Texas Southwestern Medical Center, Dallas, USA) were maintained as CHO cells but supplemented with 5 mg/ml of cholesterol (Sigma), 1 mM sodium mevalonate (Sigma), and 20 mM sodium oleate (Sigma) [44] . The vesicular stomatitis virus reverse genetics system (VSV, Indiana serotype) was kindly provided by Dr. J.K. Rose (Department of Pathology, Yale University School of Medicine, New Haven, USA) and was described in detail earlier [45, 46, 47] . Recombinant VSV expressing the glycoprotein GP-C of Lassa virus (LASV, strain Josiah) designated as VSVDG/LASVGP and wild-type VSV (VSVwt) were propagated in Vero E6 cells as described previously [48] . Influenza virus A/FPV/Rostock/34 (H7N1), designated as fowl plague virus (FPV), was propagated in embryonated hen eggs and stored at 280uC until further use. Virus titration of FPV was described previously [49] . All experiments with infectious FPV were done under biological safety level 3 conditions. VSVDG/LASVGP titration was performed using a microplate format plaque assay with subsequent immunostaining as described before [50] . In brief, virus dilutions were incubated on Vero E6 cells with an overlay of 3% carboxymethylcellulose (CMC) during plaque formation. Infected cells were visualized after cell fixation with paraformaldehyde (PFA, 4%) and permeabilization with 0.3% Triton-X 100 using a specific LASV GP-C/GP-2 antibody followed by incubation with horseradish peroxidase-labeled secondary anti-rabbit antibody (DAKO). Finally, cells were stained with True Blue Peroxidase substrate (KPL). For virus spread experiments, CHO cell lines were seeded into 96-well plates in the presence or absence of doxycycline. 24 h after induction, cells were infected with VSVDG/LASVGP or FPV and were grown without solid overlay. Cells were fixed at different time points post-infection and immunostaining was performed as described above using rabbit sera against VSV (kindly provided by Dr. G. Herrler, Institut für Virologie, Zentrum für Infektionsmedizin, Stiftung Tierä rztliche Hochschule Hannover, Germany), for the detection of VSVDG/LASVGP infected cells, and against FPV, for cells infected with FPV, respectively. Virus titration of LASV (strain Josiah, Gene Bank Accession Number NC_004297 and NC_004296) was performed by defining the 50% tissue culture infectious dose (TCID 50 ). For this, Vero cells were grown in 96-well plates to 30 to 40% confluence. Cells were inoculated with 10-fold serial dilutions of supernatants from LASV-infected CHO cell lines grown in the presence or absence of doxycycline. The assays were evaluated at 7 to 9 days postinfection. TCID 50 values were calculated using the Spearman-Karber method [51] . All experiments involving LASV-infected samples were performed under biological safety level 4 conditions at the Philipps-University Marburg. At 24 h post-infection, cell culture supernatants from infected cells were cleared from cell debris and pelleted in an SW-60 rotor through a 20% sucrose cushion at 52000 rpm at 4uC for 2 h. The pellet was then resuspended in PBS buffer and mixed with SDS-PAGE sample buffer. To control the intracellular expression level, cell lysates were collected simultaneously. Samples were analyzed by SDS-PAGE and Western blotting using protein-specific antibodies as indicated. Proteins were separated by SDS-PAGE using 10% polyacrylamide gels. Immunoblotting was performed as described previously [52] . Antiserum against Lassa virus GP-C/GP-2 was also described previously [32] . Polyclonal rabbit anti-ß-tubulin antibody was purchased from Abcam (UK), and monoclonal mouse anti-Flag antibody from Sigma-Aldrich. Secondary antibodies labeled with Alexa680 or IRDye800 were from Molecular Probes Invitrogen and Biomol, respectively, and were used for visualization and quantification of detected proteins using the Odyssey Infrared Imaging System (LI-COR Biosciences). CHO cell lines were grown on coverslips and 24 h after doxycycline-induction, cells were washed with PBS and fixed with 4% PFA in DMEM for 30 min. The fixative was removed, and free aldehydes were quenched with 100 mM glycine in PBS. Then, samples were washed with PBS and permeabilized for 10 min with PBS containing 0.1% Triton X-100. Cells were incubated in blocking solution (2% bovine serum albumin, 0.2% Tween 20, 5% glycerol, and 0.05% sodium azide in PBS) and subsequently stained with a primary mouse-anti-flag antibody (1:400) and a secondary anti-mouse antibody coupled to rhodamine (1:200, Jackson Immunoresearch). Cell nuclei were stained with DAPI (49,69-diamidino-2-phenylindole, Sigma). Microscopic analysis was performed with a Zeiss ApoTome/ Axiovert 200 M microscope using a magnification of 1:40. Replication-competent recombinant vesicular stomatitis virus (rVSV) expressing foreign envelope glycoproteins has been demonstrated to be a suitable model system to study the role of Effect of S1P Inhibition on LASV Replication www.plosntds.org viral glycoproteins in the context of virus replication [47, 53, 54] . In the present study, we took advantage of a rVSV expressing the LASV glycoprotein GP (designated VSVDG/LASVGP) [48] . In this system biosynthesis and processing of GP was shown to be authentic compared to LASV [48] . In an initial experiment we wanted to determine whether CHO-K1 cells are susceptible to VSVDG/LASVGP infection. The reason we chose CHO-K1 cells for our studies is the availability of a site 1 protease-deficient CHO cell line (designated SRD-12B cells), in which GP maturation is abolished and only GP-deficient non-infectious LASV particles are released [23] . Thus, this cell clone provides an ideal control for inhibition studies. Vero E6, CHO-K1, and SRD-12B cells were infected with either VSVDG/ LASVGP or wild-type VSV (VSVwt) as a control. Aliquots of cell culture supernatants were collected at different times after infection and were analyzed by plaque assay. Growth kinetics revealed that VSVDG/LASVGP grows to similar titers in CHO-K1 cells compared to Vero E6 cells which have been used in earlier studies (Fig. 1A) [48] . These data demonstrated that CHO-K1 cells support efficient VSVDG/LASVGP replication, and thus are useful tools for further investigations. As expected, VSVDG/ LASVGP lacks efficient replication in SRD-12B cells, whereas virus growth of VSVwt remained unaffected in these cells (Fig. 1A) . The reason for the low but detectable virus titers in the supernatant of VSVDG/LASVGP-infected SRD-12B cells is currently not known but has been also observed for LASV ( [23] and present study), LCMV [34] and New Word arenaviruses [33] . Glycoprotein activation by a yet unknown protease though with only very low efficiency might explain this phenomenon. To mimic the conditions of short-term treatment, we decided to use the inducible doxycycline-dependent Tet-On expression system, which allows regulated expression of the protein of interest [55] . To determine whether treatment of cells with doxycycline interferes with viral replication, we cultivated VSVDG/LASVGPinfected CHO-K1 Tet-On cells in the presence or absence of doxycycline (1 mg/ml) for 24 h and 48 h, respectively. As shown in Fig. 1B , CHO-K1 Tet-On cells treated with doxycycline produced a virus titer comparable to cells that were cultivated in the absence of doxycycline, indicating that these conditions used in our experiments have no influence on efficient virus replication. Generation of S1P-adapted a 1 -antitrypsin expressing cell lines Pullikotil and colleagues recently reported that various antitrypsins mimicking S1P recognition motifs are able to block processing of the S1P substrates SREBP and ATF6, although to different degrees [42] . In addition to the a 1 -AT variants shown to be most effective in that study we have chosen the LASV GP-C cleavage motif RRLL to investigate whether they also inhibit LASV GP-C cleavage. Therefore, we generated various S1Pspecific a 1 -ATs, and as a specificity control, a furin-adapted a 1 -AT, by recombinant PCR technology using the rat a 1 -AT-PIT as a template (Fig. 2A) . To facilitate their detection, we introduced a flag epitope at the C-termini of the constructs. Stable cell lines were generated and individual clones were isolated and screened for a 1 -antitrypsin expression after doxycycline induction by immunoblotting and immunofluorescence analysis. Cell lines that showed similar expression levels of a 1 -antitrypsins were chosen for further experiments (Fig. 2B and 2C ). To test the inhibitory potential of S1P-specific a 1 -antitrypsins on proteolytic processing of LASV GP, stably transfected CHO-K1 Tet-On cells, and non-transfected wild-type CHO-K1 Tet-On cells as well as SRD-12B cells were infected with VSVDG/ LASVGP at an MOI of 0.2 in the presence or absence of doxycycline. To allow only one replication cycle, cell lysates were analyzed 10 h post-infection for detection of LASV GP cleavage by Western blot analysis using a GP-specific antiserum that recognizes both the precursor GP-C and the cleaved subunit GP-2. In CHO-K1 Tet-On cells LASV GP was efficiently cleaved, regardless of whether doxycycline was present or not. In contrast, virtually no detectable cleavage of GP was observed in SRD-12B cells that are deficient in S1P (Fig. 3A, lanes 1-4) . Without expression of the various antitrypsins efficient cleavage was detected in these stably transfected cell lines, similar to the processing of GP in wild-type CHO-K1 Tet-On cells (Fig. 3A , lanes 1, 5, 7, 9, 11, and 13) . In contrast, cells expressing the S1Padapted a 1 -antitrypsins inhibited proteolytic maturation of LASV GP (Fig. 3A, lanes 6, 8, 10, and 12) . Furthermore, our results show Effect of S1P Inhibition on LASV Replication www.plosntds.org that the presence of a furin-specific a 1 -AT did not influence LASV GP-C processing, demonstrating the specificity of the generated S1P-adapted a 1 -antitrypsins (Fig. 3A, lanes 13 and 14) . Quantification of GP-2 cleavage revealed that the a 1 -AT variant RRIL exhibited the greatest inhibitory effect on GP processing (.80% inhibition) followed by a 1 -AT RRLL (.60% inhibition), which possesses the amino acid cleavage motif of the LASV GP-C. Also a 1 -AT variants RRVL and RRYL were found to be inhibitory, although to a lesser extent (inhibition less than 50%) than the variants RRIL and RRLL (Fig. 3B) . Taken together, these data clearly demonstrate that S1P-specific a 1 -antitrypsins efficiently block the maturation cleavage of LASV GP, however, they differ in regard to their inhibitory potential. We have shown earlier that S1P-mediated cleavage of GP-C is absolutely required for incorporation of the glycoprotein subunits into the virion envelope and thus for production of infectious LASV [23] . Therefore, we addressed the question of whether a S1P-specific a 1 -AT has the potential to prevent GP incorporation by blocking glycoprotein processing. To this end, a 1 -AT RRIL cells were infected in the presence or absence of doxycycline with either VSVDG/LASVGP or VSVwt as a control. At 24 h postinfection, viral particles released into the cell culture supernatant were purified over a 20% sucrose cushion and analyzed by means of immunoblotting. In viral particles from supernatants of noninduced a 1 -AT RRIL cells and CHO-K1 Tet-On control cells cleaved GP-2 was readily observed, whereas in the particulate material isolated from the supernatant of a 1 -AT RRIL expressing cells no glycoprotein was detected (Fig. 4A) . However, Western Blot analysis for VSV proteins revealed the release of these viral proteins into the supernatant of a 1 -AT RRIL expressing cells, which is consistent with our earlier findings that, in the absence of GP-C cleavage, enveloped non-infectious LASV-like particles containing the matrix protein Z and the ribonucleoprotein (RNP) complex, but devoid of viral glycoproteins, are still released [23] . The lower amount of VSV proteins observed in the cell lysate and supernatant of a 1 -AT RRIL expressing cells reflect lower levels of viral replication, which is due to less efficient virus spread (Fig. 4A) . In contrast to its ability to efficiently block incorporation of LASV GP into virions, the presence of a 1 -AT RRIL had no effect on the release of glycoprotein G containing wild-type VSV particles. The amount of VSV proteins detected in the supernatant from a 1 -AT RRIL expressing cells was similar to the amount of viral proteins observed in supernatants of non-induced cells and CHO-K1 cells, indicating efficient viral replication and cell-to-cell spread of VSVwt despite the presence of a 1 -AT RRIL (Fig. 4B) . Taken together, these data demonstrate that S1P-specific a 1 -antitrypsins have the potential to prevent LASV GP incorporation by inhibiting glycoprotein cleavage, which is an essential prerequisite for infectious progeny. Virus spread is reduced in the presence of specific a 1antitrypsins Next, we wanted to know whether the observed inhibition of LASV GP processing correlates with the ability of the different a 1antitrypsin variants to inhibit virus spread. To investigate this, we established a 96-well plate assay in which infected cells are immunostained with True Blue substrate as described in Materials and Methods. Virus spread can be monitored by the appearance of characteristic comet-shaped foci, showing that the virus progeny is carried over the cell monolayer, while prevention of virus spread results in limited radial growth, due to infection of only neighbouring cells. This assay allows rapid screening of potential inhibitors [50] . Effect of S1P Inhibition on LASV Replication www.plosntds.org type cells (Fig. 5A, upper panel) . In contrast, virus spread was significantly diminished in cells expressing a 1 -AT specific for S1P (Fig. 5A, lower panel) . These data indicate that S1P-adapted a 1antitrypsins have the potential to specifically inhibit the processing of LASV GP, which in turn is required for efficient virus spread. It should be noted that the infectious foci observed in a 1 -AT RRIL expressing cells were larger compared to SRD-12B cells in which virtually no virus spread of VSVDG/LASVGP was observed, resulting in only single infected cells (Fig. 5A) . Although similar inhibition values were observed by means of immunoblot quantification (Fig. 3) , a few remaining non-detectable cleavage events may count for this limited cell-to-cell spread in a 1 -AT RRIL expressing cells. Cells expressing the furin-adapted a 1 -AT variant RVKR did not prevent virus spread. At first glance, we rather observed an enhancement of infectivity compared to noninduced cells, which might be due to an increase in the LASV cellular receptor a-dystroglycan on the cell surface [56] . To further confirm the specificity of the a 1 -AT variants, we used fowl plague virus (FPV), which contains a hemagglutinin with a multibasic cleavage motif recognized by furin [57] . Thus, the furin-adapted a 1 -AT should prevent virus spread of FPV, while virus spread in the presence of S1P-specific a 1 -antitrypsins should remain unaffected. Fig. 5B clearly demonstrates that the most potent S1P-specific a 1 -AT variant RRIL had no effect on FPV replication, and that virus spread was found to be similar to that observed in wild type CHO-K1 Tet-On cells. In contrast, in cells expressing the furin-adapted a 1 -AT variant RVKR virus spread of FPV was drastically reduced, whereas FPV replication occurred efficiently under doxycycline-free conditions in these cells. These results demonstrate that the generated a 1 -AT variants exhibit high specificity for the corresponding proteases, which are essential for virus spread in cell culture. To further elucidate the effect of the different a 1 -AT variants on multicycle replication, viral titers were determined. To this end, cells were infected with VSVDG/LASVGP at an MOI of 0.02 in the presence or absence of doxycycline. Cell culture supernatants were collected 24 h and 48 h post-infection and virus titers were determined by plaque assay. As shown in Table 1 , non-induced S1P-specific a 1 -AT cell lines permitted unaffected growth of Effect of S1P Inhibition on LASV Replication www.plosntds.org VSVDG/LASVGP to comparable titers, whereas virus titers were reduced in cells expressing the S1P-specific a 1 -AT variant. At 24 h post-infection virus production decreased about 100 fold in cells expressing the a 1 -AT variant RRIL compared to non-induced control cultures. The presence of a 1 -AT variant RRLL reduced the virus titer in the supernatant about 10 fold, followed by a 6.2 fold reduction of virus production in a 1 -AT variant RRVL expressing cells. The presence of the a 1 -AT variant RRYL only exhibited a very moderate inhibitory effect on viral replication (inhibition ,2 fold). Again, the presence of the furin-adapted a 1 -AT variant RVKR did not affect VSVDG/LASVGP replication compared to non-induced control cells. Our results indicate that the various S1P-adapted a 1 -antitrypsins exhibit different inhibitory potentials, due to their different recognition motifs. However, the degree of inhibition of virus replication correlated well with the inhibitory potential of the various S1P-adapted a 1 -antitrypsin variants to block LASV GP processing. Interestingly, following the inhibition of virus progeny over a time period of 48 h only the S1P-adapted a 1 -AT variants RRIL and RRLL sustained their inhibitory capacity, whereas in cells expressing a 1 -antitrypsin variants RRVL and RRYL virus production was found to recover although the initial expression levels of a 1 -antitrypsin variants were similar (Table 1 ). These data indicate that the inhibitory potential of the a 1 -AT variants RRVL and RRYL is not sufficient to efficiently suppress the formation of infectious particles by effectively blocking LASV GP-C cleavage, whereas the a 1 -AT variants RRIL and RRLL seem to be appropriate candidates for efficient inhibition of LASV propagation. Finally, we wanted to investigate the impact of blocking S1Pmediated GP processing on virus progeny of authentic LASV. Therefore, we assessed the inhibitory potential of the most potent variant, a 1 -AT RRIL, on the multiplication of LASV (strain Josiah). For this purpose a 1 -AT RRIL cells and, as controls, CHO-K1-Tet-On and SRD-12B cells were infected with LASV at an MOI of 0.1. To induce a 1 -AT expression, a 1 -AT RRIL cells and, as a control for off-target effects, CHO-K1 Tet-On cells were cultivated in the presence of doxycycline. To determine virus titers, infectious virions released into the cell culture supernatant were analyzed by defining the 50% tissue culture infectious dose (TCID 50 ) at various times post-infection, as indicated. In noninduced a 1 -AT RRIL cells, LASV revealed a growth kinetic similar to that observed in CHO-K1 Tet-On control cultures, while expression of a 1 -AT RRIL resulted in an average 2 log10 reduction in viral titer (Fig. 6) . The difference between infectious LASV titers in the supernatant of a 1 -AT RRIL expressing cells and SRD-12B cells correlated with the limited virus spread observed in a 1 -AT RRIL expressing cells compared to single cell infections in S1P null cells (Fig. 5) . Taken together, this result highlights the inhibitory activity of modified a 1 -antitrypsins against LASV and demonstrates that inhibition of endogenous S1P is a potent strategy to reduce the production of infectious LASV progeny. Current drug treatment of Lassa fever and certain New World hemorrhagic fevers is limited to the guanosine analogue ribavirin [7, 8, 9] . Although ribavirin therapy can reduce the mortality rates of severe clinical cases, its unavailability to most patients in West Africa and South America as well as its association with severe adverse effects including anaemia [11, 13] , teratogenicity and embryo lethality [12] , argues for the development of new alternative treatment options. In principle, every step in the viral life cycle is a potential target for antiviral inhibitors. While current antiviral strategies in the arenavirus field mainly target virus entry [58, 59, 60, 61] or replication and assembly [62, 63, 64, 65, 66, 67] , inhibition studies of the glycoprotein activating endoprotease and its impact on viral replication are largely unexploited. Due to its central role in the arenavirus life cycle [23, 26, 33, 34] , S1P should be considered as a cellular target for antiviral drug development. In the present study we analyzed the inhibitory effect of S1P-adapted a 1 -antitrypsins on proteolytic processing of LASV GP-C and its consequences for viral replication. To our knowledge, this is the first report that Effect of S1P Inhibition on LASV Replication www.plosntds.org addresses the impact of protein-based S1P inhibition on LASV GP-C cleavage and multicycle replication. Furin-adapted a 1 -ATs have been shown to efficiently inhibit the formation of infectious progeny of other viruses (e.g. HIV, measles virus and human cytomegalovirus) [38, 39, 40, 41, 68, 69] . Using a replication-competent recombinant VSV pseudotyped with the LASV glycoprotein GP [48] , we demonstrate that proteolytic maturation of the precursor GP-C is sensitive to S1Padapted a 1 -ATs. Mutagenesis of the reactive centre loop (RCL) into the S1P recognition motif RRIL resulted in an abrogation of GP-C processing similar to that observed in S1P-deficient SRD-12B cells. The inhibitory activity of the a 1 -AT variant RRIL on LASV GP cleavage described here is in agreement with a previous study showing its inhibitory potential on the processing of the natural S1P substrates SREBP-2 and ATF6 [42] . Also an a 1 -AT variant that contains the LASV GP-C cleavage motif RRLL exhibited a high S1P inhibitory potential and was found to drastically reduce GP processing. Interestingly, this variant exhibited a 100% inhibition activity on maturation cleavage of an artificial pro-PDGF (precursor of platelet-derived growth factor) mutant that is processed by S1P due to introduction of a RRLL cleavage site, but failed to inhibit cleavage of endogenously expressed SREBP-2 [42] . These data indicate that various substrates differ in their sensitivity towards S1P inhibition. The outcome of severe illness increased significantly with the level of viremia in Lassa fever patients [70] . Therefore, the extent of multicycle replication of LASV and thus, the load of infectious particles in its host organism have an important impact for the progress of disease. Our studies revealed that a 1 -AT variants RRIL and RRLL have a potency sufficient to sustain their inhibitory capacity during multicycle replication, which resulted in a significant reduction of virus infectivity. Inhibition of viral replication correlated with the ability of the a 1 -AT variants RRIL and RRLL to efficiently inhibit the processing of the LASV glycoprotein precursor. Although our data demonstrated that inhibition of glycoprotein cleavage by a 1 -AT RRIL reduced incorporation of the subunits GP-1 and GP-2 into virions to below detectable levels, the viral titer from a 1 -AT RRIL expressing cells was found to be greater than that obtained from S1P null cells. Based on this observation, we consider that even the most potent a 1 -AT variant RRIL failed to entirely inhibit S1P activity. However, given that S1P has important biological functions in the regulation of various cellular processes, a complete inhibition of the catalytic activity of S1P is not desirable. For a 1 -AT variants RRVL and RRYL, we observed similar inhibition values by immunoblot quantification analysis as described for CCHFV GP cleavage [42] . Though, their inhibitory activity on LASV GP-C cleavage was not sufficient to efficiently reduce virus replication of VSVDG/LASVGP. These results should be taken into consideration for experimental setups in future studies that address the impact of S1P inhibition in arenavirus replication. The most potent a 1 -AT variant RRIL revealed a similar inhibitory potential on virus release of authentic LASV to that observed with the corresponding VSVDG/LASVGP pseudotype. Therefore, this study also demonstrates that the replicationcompetent VSV expressing the LASV glycoprotein is an excellent surrogate model for analyzing potential antivirals that target the biological function of GP and its consequence for virus replication. These studies can be performed under biosafety level 2 laboratory conditions that would otherwise require biosafety level 4 laboratory conditions [71] . Taken together, our data indicate that S1P-adapted a 1 -antitrypsins may represent a promising lead compound for the development of a new class of anti-arenavirus inhibitors. In recent years improvements were made in the application of bioengineered serpins to combat bacterial and viral infections [39, 72] . For example, the addition of exogenous a 1 -AT-PDX, a potent and selective furin inhibitor, was found to efficiently block human cytomegalovirus infection [39] . However, in contrast to furin, which is known to recycle between the plasma membrane and the TGN via endosomal compartments, membrane-bound S1P is localized in the secretory pathway and can be sorted to endosomal compartments but not to the cell surface [73, 74, 75] . Follow-up studies with small synthetic peptides, which are derived from S1P-specific a 1 -antitrypsins described in the present work, are currently in progress and will address cellular delivery and organelle specific targeting, as well as their inhibitory potential on authentic LASV replication. In analogy to inhibition strategies of the eukaryotic subtilase furin, we previously designed and developed a cell-permeable peptidyl chloromethylketone S1P inhibitor, which contained the LASV GP-C cleavage site, designated dec-RRLL-cmk [76, 77, 78] . This irreversible inhibitor efficiently blocked the processing of LASV GP at nanomolar concentrations, however, because of cell type-dependent toxicity observed by us and others, its potential in vitro use requires further investigation [79, 80] . Due to the essential roles of S1P in cholesterol metabolism and fatty acid synthesis, this enzyme has attracted great attention by the pharmaceutical industry. Research efforts are currently directed towards the development of S1P inhibitors that may be used in the treatment of dyslipidemia and a variety of cardiometabolic risk factors associated with diabetes and obesity [81] . Identification of specific S1P inhibitors in this therapeutic area may also be beneficial in treatment of hemorrhagic fevers caused by viruses known to be processed by S1P. Future studies Effect of S1P Inhibition on LASV Replication www.plosntds.org will have to elucidate the anti-viral efficacy of these and other novel S1P inhibitors that have been developed [82, 83] . While most conventional antiviral drugs target proteins that are virus-encoded, cellular proteins essential for viral replication are currently considered to be alternative potential targets for antiviral therapy [84, 85, 86] . With the exception of Ebola virus, whose glycoprotein cleavage by the proprotein convertase furin is not essential for virus replication in cell culture and virulence in nonhuman primates [71, 87, 88, 89] , maturation cleavage of surface glycoproteins of several virus species by endoproteases is a key determinant for host cell tropism and pathogenicity [90] . Thus, the emergence of viral escape mutants that confer resistance due to targeted inhibition of an endogenous protease is rather unlikely. In S1P-deficient SRD-12B cells, which were persistently infected with Junin virus vaccine strain Candid 1, no virus escape variants possessing a cleavage motif other than a S1P recognition motif have evolved, indicating a low potential of arenaviruses to develop de novo a different glycoprotein maturation pathway [33] . This observation together with our findings that inhibition of S1P significantly affects LASV GP processing and virus infectivity should encourage the development of S1P inhibitors as a potential drug target to counteract infections caused by pathogenic arenaviruses. Alternative Language Abstract S1 Translation of the abstract and author summary into French by Stephane Daffis. Found at: doi:10.1371/journal.pntd.0000446.s001 (0.05 MB PDF)

VirHostNet: a knowledge base for the management and the analysis of proteome-wide virus–host interaction networks

Infectious diseases caused by viral agents kill millions of people every year. The improvement of prevention and treatment of viral infections and their associated diseases remains one of the main public health challenges. Towards this goal, deciphering virus–host molecular interactions opens new perspectives to understand the biology of infection and for the design of new antiviral strategies. Indeed, modelling of an infection network between viral and cellular proteins will provide a conceptual and analytic framework to efficiently formulate new biological hypothesis at the proteome scale and to rationalize drug discovery. Therefore, we present the first release of VirHostNet (Virus–Host Network), a public knowledge base specialized in the management and analysis of integrated virus–virus, virus–host and host–host interaction networks coupled to their functional annotations. VirHostNet integrates an extensive and original literature-curated dataset of virus–virus and virus–host interactions (2671 non-redundant interactions) representing more than 180 distinct viral species and one of the largest human interactome (10 672 proteins and 68 252 non-redundant interactions) reconstructed from publicly available data. The VirHostNet Web interface provides appropriate tools that allow efficient query and visualization of this infected cellular network. Public access to the VirHostNet knowledge-based system is available at http://pbildb1.univ-lyon1.fr/virhostnet.

Eukaryotic cells express a large panel of proteins that coordinately participate to the cellular machinery through a highly connected and regulated network of protein-protein interactions (1) . Physical architecture of model organisms and human cellular protein networks exhibits a strong robustness against random failures, and strikingly a high sensitivity to targeted attacks on highly connected and central proteins, also called 'hubs' (2, 3) . Cellular protein network is not static and its robustness may change dynamically according to various factors like tissue and cell-line origins, signals received by cellular environment or more specifically during viral infections (4) . Replication and pathogenesis of viruses depend on a complex interplay between viral and host cellular proteins both acting through a complex network of protein-protein interactions. In order to evade the cell innate immune response and/or to favour their own replication and transmission, viruses have developed strategies to hijack central functions of the cell (5) (6) (7) . Viruses also use intra-viral, i.e. virus-virus, protein-protein interactions for virion assembly or viral egress from the cell. Accumulation of functional perturbations associated with such virus-virus and virus-host protein-protein interactions may lead to severe and complex diseases, like the development of cancers (8, 9) . From a systems biology perspective, a deeper understanding of infectious diseases may rely on an exhaustive characterization of all potential interactions occurring between proteins encoded by viruses and those expressed in infected cells (10) . Thus, integration of all protein-protein interactions into an infected cellular network, or 'infectome', is a great challenge that may provide a powerful framework for virtual modelling and analysis of viral infection. The first draft of the human cellular network, also referred to the human interactome, has been explored at the proteome-wide level by the mean of high-throughput experiments such as yeast-two hybrid screens (11, 12) or tap-tag procedure (13) . The overall quality and completeness of this human cellular network has been significantly improved thanks to systematic approaches based on text-mining and literature-curated interactions extracted from low-throughput experiments. Many generalized and specialized databases are involved in the integration of these protein-protein interactions, such as BIND (14) , MINT (15) , INTACT (16) , HPRD (17) , DIP (18) , BIOGRID (19) , REACTOME (20) , GENERIF (21) and NETWORKIN (22) . However, the low redundancy of interactions found between these databases has raised the need to unify such data resources for human and model organisms (23) . Concerning virus-virus and virus-host protein-protein interactions, few high-throughput experiments have been achieved, except some yeast-two hybrid screens completed for Herpes viruses (EBV, KSH, VZV, HSV-1) (24) (25) (26) and SARS (27) . Although some generalist databases like BIND, MINT, INTACT and HIV-GENERIF provide access to virus-virus and virus-host protein-protein interactions, no systematic approach has been reported to exhaustively mine and curate all interactions that have accumulated in scientific publications. In this context, we have developed VirHostNet (Virus-Host Network), a public knowledge-based system specialized in the management, analysis and integration of virus-virus, virus-host and host-host interactions as well as their functional annotations in the cell. Based on an extensive scientific literature expertise, VirHostNet provides a high-confidence resource of manually curated interactions defined for a wide range of viral species. The content of this high-confidence dataset has been illustrated by the analysis of cellular functions and pathways enriched in proteins targeted by one or many viruses. An integrated cellular network has also been reconstructed from public data and combined with viral data to provide the first draft of the infected cellular network. In addition, an original Web interface has been developed, which provides multi-criteria query and visualization tools for infection network navigation. The utility of the visualisator has been exemplified by network representation of the mTOR pathway and its interplay with viruses. A bioinformatics pipeline was developed to fully integrate virus-virus, virus-host and host-host protein-protein interactions gathered from a wide range of public databases, with those mined from scientific literature and curated by VirHostNet experts (Figure 1 ). In addition to the management of this large protein interaction resource, VirHostNet integrates contextual information concerning interacting proteins, like structural and functional annotations of proteins: Gene Ontology term (28) , KEGG pathway (29) , INTERPRO domain (30) . All these data were integrated into a knowledge-based system implemented by using PostgreSQL DataBase Management System (release 8.2.6). The low level of redundancy observed among available databases involved in molecular interactions management has emphasized the need to integrate these heterogeneous data sources (23) . Virus-virus, virus-host and host-host protein-protein interactions and meta-data related to experimental procedures or publications were extracted from 10 databases (BIND, MINT, INTACT, HPRD, DIP, BIOGRID, REACTOME, GENERIF, HIV-GENERIF, NETWORKIN) ( Figure 1 ). Due to the heterogeneity of protein sequence identification found across these databases (i.e. gene identification number, gene name, protein accession number, protein name), NCBI and ENSEMBL protein sequence databases were chosen to unify virus and host proteins respectively (see Supplementary Table 1A) . Towards this end, the IPI database system (31) was chosen to cross-reference all the human protein sequences to ENSEMBL protein accession numbers. In addition, viral protein sequences defined at EMBL and UNIPROT were mapped on NCBI protein sequences by using BLAST Alignment software (32) . Protein cross-referencing led to the definition of nonredundant protein-protein interactions that were in many cases defined in different databases, publications or supported by distinct experimental procedures (see Supplementary Table 1B) . Thus, all information associated with non-redundant interactions, like database origin, experimental procedure description in PSI-MI 2.5 standard format (33) or PUBMED identification (PMID) number, were retrieved in VirHostNet to provide the most documented interactions. This compilation of interaction meta-data will facilitate data quality filtering based on the number of databases, methods or PMIDs used (34, 35) . An automatic text-mining pipeline was developed and plugged into the VirHostNet system in order to prioritize scientific papers for protein-protein interaction curation. As a first step, all abstracts containing keywords related to both viruses and experimental procedures used for interactions identification (mainly yeast-two hybrid, co-imunoprecipitation, pull-down and tandem affinity purification) were extracted for an in-depth expertise. During curation, protein-protein interactions were carefully annotated according to: (i) the protein accession numbers of each of the protein interactor, the human and/or viral proteins being respectively referenced to ENSEMBL and NCBI accession numbers; (ii) the molecular interaction methods based on the PSI-MI 2.5 ontology vocabulary; and (iii) the PMIDs. Based on 1174 selected PMIDs, literature curation led to the annotation of 2186 redundant interactions in 723 papers (Supplementary Table 2 ). This effort significantly complemented data from public databases with 1297 new non-redundant protein-protein interactions. In order to provide a higher level of data accuracy, virus-virus and virus-host protein-protein interactions from public databases were also carefully inspected. From 2294 PMIDs for which at least one protein-protein interaction was defined, database curation led to the validation of 2261 redundant interactions found in 789 papers, corresponding to 1374 confirmed non-redundant proteinprotein interactions (Supplementary Table 2 ). Strikingly, our experts confirmed 20% of BIND and GENERIF (HIV) against 90-95% for MINT and INTACT data. One reason is that all protein-protein interactions defined by functional associations and/or genetic interactions between proteins were discarded from BIND and HIV-GENERIF. To our knowledge, VirHostNet provides the largest and the most confident infected cellular network. This network is composed of 2671 virus-virus and virus-host non-redundant protein-protein interactions concerning 180 distinct viral species. The curated protein-protein interactions were mainly defined by low-throughput and high-throughput yeast two-hybrid screens (40%), co-immunoprecipitation (24%) and pull-down (21%) (Figure 2A ). Even if only 65% of interactions rely on a single experimental procedure, a total of 944 proteinprotein interactions (35%) were defined by at least two independent methods, in good agreement with other high-confidence databases (36) ( Figure 2B ). All these interactions were defined in 36 distinct viral families, underlying the broad taxonomical diversity provided by VirHostNet ( Figure 3A ). In addition, the distribution of interactions observed among viral Baltimore groups should allow large-scale comparative study of virus-virus ( Figure 3B ) and virus-host networks ( Figure 3C ). In the infection network, the virus-host interactions occurred between 407 viral proteins and 1012 human proteins, suggesting the strong tendency of viruses to interact with a large number of cellular proteins. In order to characterize cellular functions targeted by the viral machinery, we performed functional enrichment analysis of host proteins interacting with viruses, by using Gene Ontology and KEGG databases and the same methodology described by Zheng and Wang (37) (Supplementary Tables 3 and 4 , respectively). The results showed that viruses interact significantly with a large panel of cellular functions (e.g. cell cycle, apoptosis, cell communication, protein transport) and with canonical signalling pathways (e.g. Jak-Stat, Toll-like Receptor, MAPK, TGF-b, mTOR). The majority of these functions and pathways have already been described to participate in either viral infectious cycle, cellular anti-viral mechanisms or viral associated diseases (38) . Interestingly, analysis of KEGG pathways revealed cellular mechanisms poorly documented in the case of viral infections. One example is focal adhesion, a pathway involved in cell contact with the extracellular matrix and in many other cellular processes including invasion, motility, proliferation and apoptosis (39) . Indeed, on 202 protein members of the focal adhesion pathway, more than 25% (59) were found significantly targeted (exact Fisher test, Benjamini-Hochberg multiple correction test P-value < 0.05) by at least one viral protein in 36 distinct viral taxons. This may suggest the central role of focal adhesion during viral infections and its potential impact on viral induced cancer development that might be associated for instance to the loss of cellular adhesion. Although cellular functions of proteins are far from being completely known and/or annotated in public databases, based on the 'guilty by association' concept the human protein-protein network may serve as a template to complete our understanding on cellular functions perturbed during viral infection. In order to include virusvirus and virus-host interactions in their cellular context, a human-human protein interaction network containing roughly 70 000 non-redundant protein-protein interactions and 10 000 proteins was built from public databases (details on interaction methods distribution are given in Supplementary Figure 1) . Thus, based on roughly 40 000 unique proteins annotated in ENSEMBL, 25% (10 000/ 40 000) are connected within the human protein network. Analysis of the infection network revealed that surprisingly 88% (881/1012) of targeted human proteins interact with at least one cellular protein. Thus, targeted proteins tend to physically interact in the cell and may probably participate in cross-linked functions and pathways. Based on protein neighbourhood or sub-networks, the human protein-protein interaction network may help to elucidate new protein regulators or modular functions associated to viral or cellular anti-viral strategies. A user friendly and powerful Web interface based on PHP, JAVA and AJAX technologies was developed. This interface is intended to facilitate: (i) protein and contextual based queries (ii) protein-protein interaction quality filtering and display; (iii) protein-protein interaction network query (viral and host neighbours, virus-virus, virus-host, host-host sub-networks); and (iv) protein network graphical visualization. Description and examples of the database features are available in the Wiki page of the VirHostNet Web site (http://pbildb1.univ-lyon1.fr/virhost net/wiki). Once logged-in, VirHostNet users can directly query the knowledge base by using a wide range of information concerning viral (e.g. NCBI protein name or accession number) or human proteins (e.g. ENSEMBL gene or protein accession number, NCBI gene name, REFSEQ protein accession number and UNIPROT primary and secondary accession numbers) ( Figure 4A ). AJAX technology was incorporated to control protein name and accession number availability in VirHostNet. Another important feature of the interface is batch query. It allows in-depth analysis of interaction profiles with cellular and/or viral proteins from a list of proteins defined for instance in high-throughput studies (microarray, yeasttwo hybrid). A list of genes or proteins of interest can also be assessed by the mean of contextual information, such as taxonomical information, Gene Ontology terms, KEGG pathways and INTERPRO protein domains. These properties offer a unique access to protein-protein interaction networks: (i) associated to a specific virus taxon or (ii) underlying canonical sub-cellular localization, cellular functions and pathways. To access protein-protein interactions from a list of proteins ( Figure 4B ), users have: (i) to select all or a subset of proteins of interest; (ii) to define the kind of interactions to retrieve (virus-virus, virus-host and hosthost) and their database origin and (iii) to select the mode of navigation to perform, either protein neighbours or protein subgraph ( Figure 4C ). Neighbours are viral or cellular proteins interacting directly with a protein of interest. A subgraph (or subnetwork) is a graph made of all interactions between a set of proteins. The resulting host-host, virus-host and/or virus-virus protein-protein interactions are then given into a tabulated format in three independent tabbed panels ( Figure 4D ). For each protein-protein interaction, users have a privileged access to interaction meta-data ( Figure 4E ) and a colour code highlights interactions that have been checked by VirHostNet experts. Beside table representation of protein-protein interactions, a more dynamic and interactive network visualization tool was specifically developed for graph representation of infection networks. This new network visualisator was fully implemented in Jung 2.0 (http://jung.sourceforge.net) as a Java Web applet. It efficiently takes into account viruses and host nodes dichotomy for both graph rendering (colour of nodes) and navigation (host and viral neighbours). The visualisator provides also sliders to dynamically filter graphs based on the number of PMIDs and experimental procedures used to identify interactions. Additional features are also provided to draw protein node size according to the number of viral or host interacting partners (i.e. their degree into the virus-virus, virushost, host-host networks) or to highlight targeted proteins. As a case study, we built a protein sub-network view of the mTOR KEGG signalling pathway that has been found significatively enriched in targeted proteins (Supplementary Table 4 ). Indeed, the modulation of PI3K-Akt-mTOR signal transduction pathway by viruses has been shown to play a crucial role in inhibition of apoptosis, cell survival, cell transformation, viral replication and viral assembly (40, 41) . To identify and compare how viruses interplay with this network, virus-host protein interactions annotated by VirHostNet were added. This mTOR-infection network is composed of 42 cellular interacting proteins (blue nodes), 10 viral proteins (coloured nodes according to viral taxonomy), 84 host-host (blue edges) and 14 virus-host (red edges) physical proteinprotein interactions ( Figure 5 ). Protein network visualization showed that cellular proteins of this pathway are highly inter-connected in contrast to the classical representation given by the KEGG pathway, underlying the extreme complexity and regulation of this pathway. Moreover, graph visualization allows identifying viral proteins targeting multiple cellular proteins (e.g. NS5A protein interacting with AKT1, PDPK1 and PIK3CB) and reciprocally cellular proteins interacting with multiple viral proteins (e.g. HIF1A interacting with LANA of Human Herpes Virus 8 type P and X protein of Hepatitis B Virus). Hence, the VirHostNet interface allows users to visualize protein interaction networks associated to any kind of GO term, KEGG pathway, list of proteins or keywords and to analyse how they interplay with viruses. VirHostNet provides now a public access to the largest known resource of integrated virus-virus, virus-host and host-host protein interaction networks. Literature-and database-curated interactions have led to the definition of an original and high-confidence protein-protein interactions dataset. We have briefly illustrated the need of this high-confidence dataset for the characterization of cellular functions targeted by viruses. This resource may also be crucial for network-based analysis of molecular mechanisms involved during viral infections, such as cellular network properties disturbed after the connection of viral proteins. VirHostNet will also provide a backbone for automatic screening of specific protein domains or peptides motifs associated to virus-host interactions and hence may help to delineate at the proteome-wide scale footprints in both viral and host proteins sequences. VirHostNet will allow systematic prediction of virushost protein-protein interologs based on sequence homology criteria between closely related viral proteins. The knowledge-based system is also intended to integrate virus-host protein-protein interactions data derived by our team from high-throughput yeast-two hybrids experiments (Orthomyxovirus, Paramyxovirus, Flavivirus . . .). Thus, the availability of virus-virus and virus-host networks for a broad range of viruses will encourage comparative analysis and will be very helpful for the identification of molecular interactions associated to viral pathogenesis or virulence. As virus-host and virusvirus protein-protein interactions curation is one of the central features of the VirHostNet knowledge base, one of our missions is to keep these data up to date Users can select one or more proteins from the list of proteins (by default, all proteins are selected). Then, from the preferences panel, they have to choose the kind of interaction to retrieve (virus-virus, virus-host or host-host) or the database origin. Finally, they have to choose an operation ('neighbours' or 'sub-graph'). The 'neighbours' button allows users to retrieve all direct protein-protein interactions from a single protein or a list of selected proteins. The 'sub-graph' button allows users to retrieve only protein-protein interactions occurring between selected proteins (see result in Figure 4D ). (D) List of the 89 protein-protein interactions occurring between proteins of the mTOR pathway. For the interacting proteins, NCBI taxonomy name, ENSEMBL gene acc (host)-NCBI protein acc (viruses) and official gene name (host)-product name (viruses) are given. The '?' link at the end of each interaction line allows users to access interaction annotations ( Figure 4E ). In the protein-protein interactions research panel, in addition to 'neighbours' and 'sub-graph' buttons, the 'visualize' button allows from a list of interactions the graphical visualization of the resulting network (see result in Figure 5 ). (E) More detailed information concerning protein-protein interactions are available. PMID, PSI-MI method accession number and database origin of the selected interaction are given. continuously from data published in scientific journals. The update of public databases will occur at least once or twice a year in order to keep the data as current as possible. In the next future, integration of other host species, such as mammals or insects, is envisaged. This will facilitate comparison of interaction profiles among different hosts and thus may help to elucidate the molecular basis underlying the ability of some viruses to overcome the inter-species barrier. Efforts will be made to facilitate data exchange with other generalist databases (MINT, INTACT) and to add Web2.0 capabilities to the Web interface (save, comparison and analysis of user customized networks). Altogether, VirHostNet provides an entry gate for proteome wide analysis of the virus-host system and will greatly help scientists willing to take advantage of functional genomic and systems biology to decipher viral infection, evolution and pathogenesis mechanisms and/or to rationalize anti-viral drug design. Public access to the VirHostNet knowledge base is available at http://pbildb1.univ-lyon1.fr/virhostnet. Access can be made either anonymously (by default) or by creating a personal account (register in the account menu). On simple request, this personal account allows users to participate to the literature-curation effort. Literaturecurated and Database-curated protein-protein interactions flat files are available in a tabulated format on request. Contact V.N. (navratil@univ-lyon-1.fr) for more information. Figure 5 . VirHostNet Visualisator. Graph representation of the mTOR-infection network is drawn (right of the applet). Only interacting proteins of the mTOR pathway are represented (cellular proteins = blue nodes; host-host protein-protein interactions = blue edges). Viral interacting proteins are also added to define the infection network (viral proteins = coloured nodes according to viral taxonomy; virus-host protein-protein interactions = red edges). Cellular proteins of the network targeted by at least one viral protein (red stroke) are highlighted. The width of the edges is roughly proportional to the number of PMIDs used to describe the interaction. The width of the nodes is roughly proportional to the cellular degree, i.e. the number of cellular partners in the whole network. The taxonomical classification of viral proteins interacting with the mTOR network is given (left panel of the applet). The menu of the applet (top of the applet) allows users to define nodes and edges preferences (width, colour, labelling), interaction filtering (according to the number of methods, number of PMIDs or both), network navigation (expand viral or host neighbours) and graph layout. To obtain this figure, from the list of the 89 selected protein-protein interactions obtained in Figure 4E , users have to click on 'visualize' button. Then from the applet, in picking mode, they have to select all proteins and ask for viral proteins neighbours (in expand menu).

Rfam: updates to the RNA families database

Rfam is a collection of RNA sequence families, represented by multiple sequence alignments and covariance models (CMs). The primary aim of Rfam is to annotate new members of known RNA families on nucleotide sequences, particularly complete genomes, using sensitive BLAST filters in combination with CMs. A minority of families with a very broad taxonomic range (e.g. tRNA and rRNA) provide the majority of the sequence annotations, whilst the majority of Rfam families (e.g. snoRNAs and miRNAs) have a limited taxonomic range and provide a limited number of annotations. Recent improvements to the website, methodologies and data used by Rfam are discussed. Rfam is freely available on the Web at http://rfam.sanger.ac.uk/and http://rfam.janelia.org/.

Rfam is a database of sequence families of structural RNAs, including non-coding RNA genes as well as cisregulatory RNA elements. Rfam release 9.0 contains 603 families, each represented by a multiple sequence alignment of known and predicted representative members of the family, annotated with a consensus base-paired secondary structure. This so-called SEED alignment is used to build a covariance model (CM) with the Infernal software (1) . Each Rfam covariance model is searched against a nucleotide sequence database, producing a list of putative hits. Matches that score above a curated threshold are then aligned to the CM to produce a so-called FULL alignment. This process is outlined diagrammatically in Figure 1 . The Rfam database was developed as a generic approach to the annotation of structured RNA families on genomic sequences (2, 3) , but it has been widely used as a source of reliable alignments and structures for the purposes of training and benchmarking RNA sequence and secondary structure analysis software. All Rfam models are searched against an underlying nucleotide sequence database, known as RFAMSEQ, which is derived from the EMBL nucleotide sequence database (4) . Prior to release 9.0, RFAMSEQ represented only the various species sections of EMBL. These sections contained only sequences that were considered to be of finished quality and excluded sequences from many important genomes. With release 9.0, RFAMSEQ has been expanded to include the whole genome shotgun (WGS) and environmental sequence (ENV) divisions. These changes have increased the number of sequences in RFAMSEQ by more than an order of magnitude (2 225 030 sequences in Rfam 8.0 versus 29 574 458 sequences in Rfam 9.0). In order to make it feasible to search more than 120 gigabases of sequence with hundreds of covariance models in a reasonable time, we use sequence-based filters to prune the search space prior to applying the more accurate and more computationally expensive CMs. One of the primary limitations of the Rfam annotation pipe-line has been the use of BLAST-based sequence filters, which are likely to compromise search sensitivity. In order to address this issue at least partially, NCBI-BLAST has been replaced with a WU-BLAST search, which has been tuned for high sensitivity and low sequence similarity. A benchmark of several homology search tools has shown WU-BLAST to be the more accurate of the two methods on nucleotide data (5) . Additionally, in order to make the BLAST filters more *To whom correspondence should be addressed. Tel: +44 1223 494 983; Fax: +44 1223 494 919; Email: pg5@sanger.ac.uk ß 2008 The Author(s) This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/ by-nc/2.0/uk/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. similar to profile HMMs, a sequence mask has been applied to each sequence in the alignment. Any nucleotide in an alignment column that has either a low frequency or is an insert relative to the majority of the rest of the sequences is 'soft masked' and not used for the BLAST word matches. These masked nucleotides do, however, still contribute to alignments that were seeded in the flanking regions. This approach has resulted in many fewer spurious hits with no detectable cost to sensitivity (data not shown), thus allowing E-value thresholds to be further relaxed. These observations together mean that the BLAST filters have been improved in terms of specificity and sensitivity. In order to improve the species and sequence depth of individual Rfam families, more than 370 families have been expanded by an 'iteration' process, in which some sequences in the FULL alignment are chosen for promotion to the SEED alignment. The sequences selected from FULL alignments for inclusion in the SEED must pass a series of stringent quality control requirements and be manually approved by a curator. The quality control steps include: ensuring that the sequence and secondary structure are consistent with the existing SEED sequences; the sequence identity with existing SEED sequences falls within 60-95%; the sequence is not truncated with respect to the SEED alignment. The curator also ensures that the new sequences make phylogenetic sense before allowing them to be incorporated into an updated SEED. An example of the utility of iteration is the snoRNA U103 SEED from Rfam 8.1 (accession: RF00188), which contained just three sequences and spanned two eutherian mammals (human and mouse). The SEED in Rfam 9.0 after iteration contains 42 sequences and spans Eutheria, Teleost (ray-finned fishes), Iguanidae, Monotremes, Marsupials, Placentals, Aves and Chondrichthyes (cartilaginous fishes). Phylogenetic trees have been estimated for both the SEED and FULL alignments. For the majority of the alignments we produced the trees using an accurate maximum-likelihood approach, which included models of indels (6) . However, the computational complexity of tree estimation meant that maximum-likelihood was not always possible and hence, where the number of sequences in the alignment was greater than 64, a neighbour-joining method was used instead (7). Large alignments and trees are problematic, both in terms of the computational cost of generation and the challenges of displaying them. Therefore, where the number of sequences in the alignment was greater than 1024, the highly similar sequences were filtered by sequence similarity, resulting in relatively sparse and easily presented trees that required comparatively little computing power to generate. We are currently developing a new Rfam website, with the aim of improving the presentation of Rfam data and providing more and better tools for searching the data. The new site is now available from http://rfam.sanger.ac.uk/ and can be used to access Rfam 9.0 data. The new site lacks some features of the old site, but we aim to add all existing features and add many new ones over the coming months. Note that, at time of writing, the new website was available only at the Wellcome Trust Sanger Institute (http://rfam.sanger.ac.uk/). The two mirror sites will be updated to run the same website to coincide with the release of Rfam 9.1. The new site provides detailed overviews of Rfam families, including: a snapshot of the latest community-contributed annotation from Wikipedia (see below); tools for viewing and downloading the sequence alignments in various formats; representations of predicted secondary structure (see below); the taxonomic tree for the family; and phylogenetic trees for the SEED and FULL alignments. Additionally, we provide several search tools in the new site. We currently support interactive searches, allowing a single RNA sequence to be searched against the whole Rfam database, and a batch search tool for searching multiple sequences against Rfam, the results of which are returned to the user via email. A new taxonomy search tool allows the user to find Rfam families that are specific to a given taxonomic level, or those found in a set of taxonomic levels that are specified by a complex, boolean query. For example, the query 'Drosophila AND Caenorhabditis NOT Mammalia' returns the two Rfam families (RF00047 and RF00533) that contain sequences from both drosophila and caenorhabditis but no sequences from any mammalian species. New graphical representations of secondary structures have been added to the Rfam website, based on software from the Vienna RNA package (8) . We now annotate several statistics directly on secondary structure diagrams, including sequence conservation, covariation, base-pair conservation and the maximum CM scores (Figure 2 ). The sequence conservation metric uses a metric computed for each column in the alignment; this is the frequency of the most common nucleotide in each column (Figure 2A ). The covariation metric is based upon that used by RNAalifold (9) . For each base pair in the consensus structure and for each pair of sequences in the alignment, the difference in structurally consistent and inconsistent mutations is taken. Each mutation is weighted using a treeweighting scheme (10) and this value is then normalized by the number of possible mutations ( Figure 2B ). The base-pair conservation metric is the fraction of canonical base pairs (Watson-Crick and G:U) in any two columns that correspond to a base pair in the consensus structure ( Figure 2C) . The maximum covariance model score and corresponding nucleotide/base pair is computed for each node in the CM. The resulting sequence, structure and bitscores are used to produce a marked up secondary structure ( Figure 2D ). The Rfam website now draws textual annotation of RNA families directly from the scientific community, through the online encyclopedia Wikipedia. Any updates to relevant Wikipedia articles are downloaded on a nightly basis using the MediaWiki API, verified by members of the consortium and presented on the Rfam site (11) . We consider the resulting articles to be a great improvement on the original static text because they are frequently updated, provide cross links to related articles and are generally considerably more comprehensive and informative than the original Rfam annotations that they replace. The rate of discovery of new RNA families is accelerating rapidly, facilitated by advancements in new sequencing technologies (12, 13) and targeted computational screens (14) (15) (16) (17) . Keeping abreast of these updates whilst still ensuring the quality of alignments and secondary structures is an ongoing challenge for Rfam. We continue to evaluate new technologies and techniques as they emerge and will adopt new procedures for building and checking Rfam families as necessary. We have been actively updating Rfam families and database crosslinks using more specialized RNA databases such as miRBase (18) , IRESite (19) , Pseudobase (20), snoRNABase (21) , the plant snoRNA database (22) , TransTerm (23) and the Yeast snoRNA database (24) . As a result of these efforts, the next release of Rfam (version 9.1) will contain more than 700 entirely new families, bringing the total number of Rfam families to over 1300. A new version of Infernal (v1.0) is now available (http://infernal.janelia.org) and we plan to use this 0 Sequence conservation latest version to prepare the next major release of Rfam. Testing suggests that, compared with the version used for Rfam 9 (v0.72), v1.0 is faster and slightly more sensitive, whilst introducing for the first time E-values for hits returned from database searches. Although the speed increase will not be sufficient to obviate the need for BLAST filters in the Rfam production pipeline, this remains a major goal for Infernal development. Importantly, Infernal v1.0 is not compatible with the Rfam 9 CM files. Rfam/Infernal users may wish to generate new CMs from Rfam 9 SEED or FULL alignment files. We have mapped a subset of three-dimensional RNA structures found in the Protein DataBank (PDB) (25) (primarily SRP and ribosomal RNAs) to corresponding sequences in Rfam. In an initial feasibility study, we have demonstrated that RNA sequences can be retrieved from PDB files and mapped to Rfam sequences reliably. The mapping is currently performed using BLAT (26) to detect local regions of high similarity with high specificity. The positions of matches to Rfam entries are transferred to the PDB sequences, allowing us to colour threedimensional structures as in Figure 3 . We intend to rollout this mapping across all Rfam families and PDB entries using both local similarities and global matches to Rfam models. This sequence-to-structure mapping will allow us to use determined tertiary structures to calculate secondary structure as a quality control for existing families, and catalogue interactions between RNA-RNA and RNAprotein families. A further area of active research at Rfam is how best to distribute genome annotations. We plan to make annotations available in a variety of formats including the distributed annotation service (DAS) (27) , General Feature Format (GFF) (http://song.sourceforge.net/gff3.shtml) and EMBL format, together with links to relevant genome browsers, e.g. ENSEMBL, UCSC and Genome Reviews.

New Respiratory Enterovirus and Recombinant Rhinoviruses among Circulating Picornaviruses

Rhinoviruses and enteroviruses are leading causes of respiratory infections. To evaluate genotypic diversity and identify forces shaping picornavirus evolution, we screened persons with respiratory illnesses by using rhinovirus-specific or generic real-time PCR assays. We then sequenced the 5′ untranslated region, capsid protein VP1, and protease precursor 3CD regions of virus-positive samples. Subsequent phylogenetic analysis identified the large genotypic diversity of rhinoviruses circulating in humans. We identified and completed the genome sequence of a new enterovirus genotype associated with respiratory symptoms and acute otitis media, confirming the close relationship between rhinoviruses and enteroviruses and the need to detect both viruses in respiratory specimens. Finally, we identified recombinants among circulating rhinoviruses and mapped their recombination sites, thereby demonstrating that rhinoviruses can recombine in their natural host. This study clarifies the diversity and explains the reasons for evolution of these viruses.

H uman rhinoviruses (HRVs) and enteroviruses (HEVs) are leading causes of infection in humans. These 2 picornaviruses share an identical genomic organization, have similar functional RNA secondary structures, and are classifi ed within the same genus (www.ictvonline.org/virusTax-onomy.asp) because of their high sequence homology (1) . However, despite their common genomic features, these 2 groups of viruses have different phenotypic characteristics. In vivo, rhinoviruses are restricted to the respiratory tract, whereas enteroviruses infect primarily the gastrointestinal tract and can spread to other sites such as the central nervous system. However, some enteroviruses exhibit specifi c respiratory tropism and thus have properties similar to rhinoviruses (2) (3) (4) (5) . In vitro, most HRVs and HEVs differ by their optimal growth temperature, acid tolerance, receptor usage, and cell tropism. The genomic basis for these phenotypic differences between similar viruses is not yet fully understood. HRVs and HEVs are characterized by ≈100 serotypes. Recently, molecular diagnostic tools have shown that this diversity expands beyond those predefi ned serotypes and encompasses also previously unrecognized rhinovirus and enterovirus genotypes. As an example, a new HRV lineage named HRV-C was recently identifi ed and now complements the 2 previously known A and B lineages (6-8) (N.J. Knowles, pers. comm.). The C lineage has not only a distinct phylogeny (9) (10) (11) (12) (13) (14) (15) (16) but is also characterized by specifi c cis-acting RNA structures (17) . In this study, we screened a large number of persons with acute respiratory diseases by using assays designed to overcome the diversity of both rhinoviruses and enteroviruses circulating in humans. Whenever possible, we systematically sequenced 5′ untranslated region (UTR), capsid protein VP1, and protease precursor 3CD regions of strains. Our goals were 1) to characterize the diversity of circulating rhinoviruses and, to a lesser extent, enteroviruses, to identify putative new picornavirus variants, and 2) to assess whether recombination may drive HRV evolution, which has not been shown in natural human infections (18) . Reverse transcription-PCR (Superscript II; Invitrogen, Carlsbad, CA, USA) was performed on RNA extracted by using the HCV Amplicor Specimen Preparation kit (Roche, Indianapolis, IN, USA), TRIzol (Invitrogen), or the QIAamp Viral RNA Mini kit (QIAGEN, Valencia, CA, USA). Real-time PCR specifi c for HRV-A, HRV-B, and HEV (19) , and a generic panenterhino real-time PCR (forward primer 5′-AGCCTGCGTGGCKGCC-3′, reverse primer 5′-GAAACACGGACACCCAAAGTAGT-3′, and probe 5-FAM-CTCCGGCCCCTGAATGYGGCTAA-TAMRA-3′), were performed in several cohort studies (Table) . Picornavirus-positive samples were detected from patients enrolled in cohort studies in different regions of Switzerland during 1999-2008. The main characteristics of these populations, type of respiratory specimens, and screening methods are shown in the Table. The rhinovirus serotypes used for 3CD sequencing were obtained from the American Type Culture Collection (Manassas, VA, USA). Sequencing was performed directly from the clinical specimen except for samples selected by routine isolation methods on human embryonic (HE) primary fi broblast cell lines (Table) or for HRV reference serotypes. Primers used to amplify the 5′-UTR and the VP1 and 3CD regions are listed in online Technical Appendix 1 Table 1A (available from www.cdc.gov/EID/content/15/5/719-Techapp1.pdf). Full-length genome sequences of CL-1231094, a related clinical strain of enterovirus, and partial sequences of CL-Fnp5 and CL-QJ274218 were obtained as follows. RNA extracted by using the QIAamp Viral RNA Mini kit (QIAGEN) plus DNase treatment or with Trizol was reverse transcribed with random-tagged primer FR26RV-N and amplifi ed with the SMART RACE cDNA Amplification kit (Clontech, Mountain View, CA, USA) with a specifi c forward primer and FR20RV reverse primer (online Technical Appendix 1 Table 1B ) (23) . Amplifi cation products were separated by electrophoresis on agarose gels and fragments (0.6-2.5 kb) were extracted by using the QIAquick Gel Extraction kit (QIAGEN). Purifi ed products were cloned by using the TOPO TA cloning kit (Invitrogen). Minipreps were prepared from individual colonies and clones with the largest inserts were chosen for sequencing. Sequences obtained were used to design a new forward primer (online Technical Appendix 1 Table 1 ) to advance toward the 3′ end of the genome. PCR products of 3′ genomic ends were obtained by using the BD Smart Race cDNA amplifi cation kit (Becton Dickinson, Franklin Lakes, NJ, USA) according to manufacturer's instructions. All PCR products were purifi ed by using microcon columns (Millipore, Billerica, MA, USA) and sequenced by using the ABI Prism 3130XL DNA Sequencer (Applied Biosystems, Foster City, CA, USA). Chromatograms were imported for proofreading with the vector NTI Advance 10 program (Invitrogen). Overlapping fragments were assembled with the contigExpress module of the vector NTI Advance 10. Alignments were constructed by using MUSCLE (24) with a maximum of 64 iterations. (For detailed analyses, see http://cegg.unige.ch/picornavirus.) Multiple FastA was converted into PHYLIP format (for tree building) with the EMBOSS program Seqret (25) . Trees were built with PhyML (26) by using the general time reversible model, BIONJ for the initial tree, and optimized tree topology and branch lengths. Trees with <50 species and larger trees used 16 and 8 rate categories, respectively. Transition/transversion ratios, proportions of invariant sites, and shape parameters of the γ distribution were estimated. To investigate the hypothesis of recombination and map the breakpoints, we adapted the bootscanning method (27) as follows. The alignment was sliced into windows of constant size and fi xed overlap and a 100replicate maximum-likelihood (using HRV-93 as an outgroup) was computed for each window. From each tree, the distance between the candidate recombinant and all other sequences was extracted. This extraction yielded a matrix of distances for each window and for each alignment position. A threshold was defi ned as the lowest distance plus a fraction (15%) of the difference between the highest and lowest distances. The nearest neighbors of the candidate recombinant were defi ned as sequences at a distance smaller than this threshold. This distance ensured that the nearest neighbor, as well as any close relative, was always included. Possible recombination breakpoints thus corresponded to changes of nearest neighbors. Serotypes included in this analysis represented serotypes close to CL-013775 and CL-073908 on the basis of 5′-UTR and VP1 phlyogenetic trees (online Technical Appendix 2 Figure 1 , panels A, B, available from www.cdc.gov/EID/ content/15/5/719-Techapp2.pdf), as well as serotypes close to CL-135587 on the basis of VP1 and 3CD phlyogenetic trees (online Technical Appendix 2 Figure 1 , panels B, C) and whose full-length sequence was available. Distance matrices were computed from alignments with the distmat program in EMBOSS (http://bioweb2. pasteur.fr/docs/EMBOSS/embossdata.html) by using the Tamura distance correction. This method uses transition and transversion rates and takes into account the deviation of GC content from the expected value of 50%. Gap and ambiguous positions were ignored. Final values were then converted to similarity matrices by subtracting each value from 100. Persons enrolled in several cohorts of children and adults with respiratory infections (Table) were screened for picornavirus by culture isolation on HE cell lines, real-time PCR specifi c for HRV-A and HRV-B (19) , or by a panenterhino real-time PCR designed to theoretically detect all rhinoviruses and enteroviruses with publicly available sequences. Of 1,592 respiratory samples tested by real-time PCR, 248 were virus positive (Table) . The 5′-UTR sequences were obtained for 77 real-time PCR or culture-positive samples and VP1 and 3CD sequences for 48 of these (Table; online Technical Appendix 1 Table 2 ). In parallel, the 3CD sequences were identifi ed for all reference serotypes. The results of this screening are summarized in online Technical Appendix 1 Table 2 , and all sequences are available from GenBank (accession nos. EU840726-EU840988). On the basis of these results, respiratory infections caused by HRV-B might be less frequent than those caused by HRV-A, and HRV-A infections are distributed among the whole library of reference serotypes. A specifi c real-time PCR used to detect enteroviruses in respiratory specimens from some of the cohorts studied indicated that these viruses are rare in children (2.5% vs. 6.3% for HRV) and even rarer or absent in adults (0% vs. 24% for HRV) (28) . To include all 99 HRV reference strains and new di- Table 2 ). Characterization of HRVs newly identifi ed during 2006-2008 showed that they all belong to the same HRV-C species (9) (10) (11) (12) (13) (14) (15) (16) . Recently, Lee et al. (13) identifi ed another cluster of viruses (HRV-C′; Figure 1 , panel A) and suggested that this group was phylogenetically distinct from all other HRVs on the basis of analysis of their 5′-UTR sequences. To defi ne the phylogeny, we adapted a previously described method (23) to complete the genome sequence directly from our clinical strains (CL-Fnp5 and CL-QJ274218) that showed a similar divergent 5′-UTR (online Technical Appendix 2 Figure 1, panel A) . A condensed version (Figure 1, panel B) of the phylogenetic tree based on VP1 sequences (online Technical Appendix 2 Figure 1, panel B) indicated that CL-Fnp5 clustered with the new HRV-C clade, a fi nding further confi rmed by CL-QJ 274218 partial sequences. This fi nding supports the view that new HRVs variants described since 2006 (9-16) all belong to the same lineage. As shown in Figure 1 , panel A, the panenterhino realtime PCR enabled detection of a new HEV strain phylogenetically distinct from all previously known HEV species and associated with respiratory diseases. Enterovirus-specifi c real-time PCRs or reference VP1 primer sets routinely used to type enteroviruses (primers 222 and 224 and nested primers AN88 and 89) (29,30) did not amplify this new genotype. We could not grow this virus on HeLa and HE cell lines. Consequently, we applied the method described above to complete the genome sequence directly from the CL-1231094 (EU840733) clinical specimen. VP1 and fulllength genome sequences showed that, albeit divergent at the 5′-UTR level, this new variant belonged to the HEV-C species (Figure 1, panels B, C) . Full-length genome phylogenetic tree ( Figure 2 ) and VP1 protein identity plots (online Technical Appendix 2 Figure 2 ) with all members of the HEV-C species indicated that this virus represents a new HEV-C genotype that shares 68%, 66%, and 63% nucleotide and 77%, 75%, and 68% amino acid sequence identity, respectively, with coxsackieviruses A19 (CV-A19), A22, and A1, the closest serotypes. This new virus was named EV-104 (www.picornastudygroup.com/types/ enterovirus_genus.htm). Specifi c primers (Ent_P1.29/P2.13 and Ent_P3.30/ P3.32; online Technical Appendix 1 Table 1C) were then designed to amplify the VP1 and 3D regions of the 7 other samples of this cluster collected from children with acute respiratory tract infections and otitis media. VP1 nucleotide homology among these strains was 94%-98%, except for 1 distantly related sample (74%-76%), which may represent an additional genotype. Additional sequencing is ongoing to verify this assumption. At the 5′-UTR level, the strain described by Lee et al. (13) and EV-104 diverged from other members of HRV-C and HEV-C species, respectively. Thus, the 5′-UTR-based phylogeny was inconsistent with that based on VP1 sequences and suggested possible recombination events (Figure 1, panels A, B) . Because the 5′-UTR is the target of most molecular diagnostic assays, this sequence divergence needs to be taken into account in future studies. Other studies have provided sequences of clinical strains, but genetic characterization was often limited to 1 genomic region. Our goal was to sequence 3 genomic regions for each analyzed strain to determine defi nitively whether recombination events could represent a driving force for the evolution of rhinoviruses in their natural environment. Although recombination events have been suggested for reference serotypes, they have never been shown for circulating clinical strains (18, 31, 32) . In contrast, recombination is well established as a driving force of enterovirus evolution. Thus, we completed the 5′-UTR, VP1, and 3CD sequences of 43 clinical strains by using a pool of adapted and degenerated primers (online Technical Appendix 1 Table 1A) . Independent phylogenetic trees (online Technical Appendix 2) and similarity matrices were constructed for the 3 genomic regions. Since the last common ancestor and as depicted on the distance matrices and highlighted by boxplots of maximum-likelihood branch length distributions (online Technical Appendix 2 Figure 3 ), there are more mutations fi xed in the VP1 region than in the 3CD region, and more in the 3CD region than in 5′-UTR, which is indicative of a variable rate of evolution in these regions. Accordingly, VP1 sequences enabled genotyping of all but 3 clinical strains analyzed (online Technical Appendix 2, Figure 1 , panel B). These strains may represent rhinovirus genotypes only distantly related to predefi ned reference serotypes. In contrast, genotyping based on 3CD and 5′-UTR was less accurate, as expected. These results confi rmed that molecular typing of rhinoviruses, similarly to other picornaviruses, must use capsid sequences. Phylogeny of the 5′-UTR, VP1, and 3CD of reference serotypes showed many incongruities caused by insufficient tree resolution or recombinant viruses as previously proposed (18, 31) . As an example, 2 VP1 clusters including HRV-85/HRV-40 and HRV-18/HRV-50/HRV-34 (online Technical Appendix 2 Figure 1 , panel B) were reorganized as HRV-85/HRV-18/HRV-40 and HRV-50/HRV-34, respectively, on 3CD (online Technical Appendix 2 Figure 1 , panel C). The differential cosegregations between these virus strains suggested recombination events. When available, full-length genome sequence bootscanning applied to all serotypes will give an estimate of the number of reference strains with mosaic genomes. Similarly, the noncoding region, VP1, and 3CD trees showed major phylogenetic incongruities for 3 clinical isolates (online Technical Appendix 2 Figure 1 ). Two of these isolates (CL-013775 and CL-073908) were typed as HRV-67 on the basis of VP1 sequence and were closest to this serotype in 3CD, whereas the 5′-UTR cosegregated with HRV-36 (see 5′-UTR recombinant; online Technical Appendix 2 Figure 1 , panels A-C). These viruses were isolated by cell culture from 2 epidemiologically linked cases and thus represented transmission of the same virus. To confi rm the recombination, we completed the sequencing by obtaining the 5′-UTR, VP4, and VP2 sequences (EU840918 and EU840930) and compared them with HRV-36, HRV-67, and other closely related serotypes. Bootscanning analysis ( Figure 3 , panel A) enabled mapping of the recombination site within the 5′-UTR, just before the polyprotein start codon. Sequence alignment mapped recombination breakpoints more precisely between positions 524 and 553 with reference to HRV-2 (X02316). The other incongruent isolate (CL-135587) was typed as HRV-76 on the basis of VP1 sequence and was closest to this serotype in the 5′-UTR, but 3CD cosegregates with HRV-56 (3C recombinant; online Technical Appendix 2 Figure 1, panels B, C) . Similarly, we completed the fulllength sequence of this isolate (EU840726) and HRV-56 (EU840727). The same approach enabled mapping of the recombination site at the N terminus of protein 3C between positions 1511 and 1523 with reference to HRV-2 ( Figure 3, panel B) . These results demonstrate that recombination occurs among clinical rhinoviruses. In our analysis of 40 rhinovirus-positive samples collected over 9 years (3 additional samples were duplicates of 2 different viruses; online Technical Appendix 1 Table 2) for 3 genomic regions, 2 of the analyzed viruses appeared to be recombinants. The 2 documented recombinations occurred in members of the HRV-A species. The design of this study and technical issues (e.g., inability to sequence low viral loads) limited the ability to calculate a recombination rate, particularly for HRV-B and HRV-C. Our genomic analysis of picornaviruses associated with upper or lower respiratory diseases in adults and children indicates that rhinoviruses circulating in the community are widely diverse. The large number of circulating genotypes supports the view that rhinoviruses do not circulate by waves or outbreaks of a given dominant genotype, which might explain the high frequency of reinfection during short periods. As expected, the observed variability is higher for surface capsid proteins, the targets of most immune pressure, and this region remains the only accurate one for genotyping and defi ning phylogeny. Technical constraints such as the limited amount of clinical specimens, the use of different screening methods, and the need to sequence an unknown target of extreme variability might have limited the representativeness of our sequence collection. Therefore, our study should not be considered as an exhaustive epidemiologic analysis of rhinoviruses and enteroviruses associated with respiratory diseases. By using a systematic approach, we have identifi ed a new enterovirus genotype (EV-104) that has a divergent 5′-UTR. Undetectable by conventional methods, EV-104 could be detected by using a more generic real-time PCR assay designed to match all known available rhinovirus and enterovirus sequences. Such diagnostic tools have and will lead to constant discovery of new picornavirus genotypes (9) (10) (11) (12) (13) (14) 16, (33) (34) (35) (36) . These genotypes may represent viruses, in most instances, that have remained undetected because of insensitive cell cultures or overly restrictive molecular tools. In addition, enterovirus genotypes causing respiratory infections, such as EV-68 and CV-A21, might be underrepresented because enteroviruses are usually searched for in fecal specimens (37) . EV-104 belongs to the HEV-C species: CV-A19, CV-A22, and CV-A1 are its closest serotypes. These HEV-C subgroup viruses are genetically distinct from all other serotypes of the species. These viruses show no evidence of recombination with other HEV-C strains and, similar to EV-104, do not grow in cell culture (29) . On the basis of our epidemiologic data, we conclude that EV-104 was found in 8 children from different regions of Switzerland who had respiratory illnesses such as acute otitis media or pneumonia. Future studies using adapted detection tools will provide more information on the range of this virus. On the basis of its genomic features and similarities with coxsackieviruses and poliovirus, EV-104 could theoreti- cally infect the central nervous system (2, 38) . Detection of new subtypes of picornaviruses indicates that viruses with new phenotypic traits could emerge, and conclusions on tropism of new strains should be substantiated by extensive experimental or clinical investigations (39) . By completing the sequence of a seemingly divergent rhinovirus (13), we assigned this virus to the new HRV-C species, thus limiting currently to 3 the number of HRV species. For the sake of simplicity, we propose to consider this virus as a member of the HRV-C clade. Finally, we demonstrated that rhinovirus evolves by recombination in its natural host. Known to be a driving force of enterovirus evolution, rhinovirus recombination among clinical strains has never been observed. Two clinical isolates of 40 viruses analyzed resulted from recombination events and their breakpoints were identifi ed within the 5′-UTR sequence and the N terminus of protein 3C, respectively. These fi ndings are consistent with the fact that recombination breakpoints in picornaviruses are restricted to nonstructural regions of the genome or between the 5′-UTR and the capsid-encoding region (40) . Our observations provide new insight on the diversity and ability of rhinovirus to evolve in its natural host. The fact that only 2 of 40 analyzed viruses over a 9-year period were recombinants is suggestive of a lower recombination frequency in rhinoviruses than in other picornaviruses (32, 40) and might be related, but not exclusively, to the short duration of rhinovirus infection (18, 31, 32) . Recombination events occurred between HRV-A genotypes, but whether they can occur in species B and C remains unknown. Interspecies recombination is rare in picornaviruses and is mainly the result of in vitro experiments. For rhinoviruses, the different location of cre elements in each species might be an additional limiting constraint (17) . In summary, we have highlighted the large genomic diversity of the most frequent human respiratory viral infection. Our phylogenetic analysis has characterized circulating strains relative to reference strains and has identifi ed a previously unknown enterovirus genotype. We have shown that recombination also contributes to rhinovirus evolution in its natural environment.

Low-Cost HIV-1 Diagnosis and Quantification in Dried Blood Spots by Real Time PCR

BACKGROUND: Rapid and cost-effective methods for HIV-1 diagnosis and viral load monitoring would greatly enhance the clinical management of HIV-1 infected adults and children in limited-resource settings. Recent recommendations to treat perinatally infected infants within the first year of life are feasible only if early diagnosis is routinely available. Dried blood spots (DBS) on filter paper are an easy and convenient way to collect and transport blood samples. A rapid and cost effective method to diagnose and quantify HIV-1 from DBS is urgently needed to facilitate early diagnosis of HIV-1 infection and monitoring of antiretroviral therapy. METHODS AND FINDINGS: We have developed a real-time LightCycler (rtLC) PCR assay to detect and quantify HIV-1 from DBS. HIV-1 RNA extracted from DBS was amplified in a one-step, single-tube system using primers specific for long-terminal repeat sequences that are conserved across all HIV-1 clades. SYBR Green dye was used to quantify PCR amplicons and HIV-1 RNA copy numbers were determined from a standard curve generated using serially diluted known copies of HIV-1 RNA. This assay detected samples across clades, has a dynamic range of 5 log(10), and %CV <8% up to 4 log(10) dilution. Plasma HIV-1 RNA copy numbers obtained using this method correlated well with the Roche Ultrasensitive (r = 0.91) and branched DNA (r = 0.89) assays. The lower limit of detection (95%) was estimated to be 136 copies. The rtLC DBS assay was 2.5 fold rapid as well as 40-fold cheaper when compared to commercial assays. Adaptation of the assay into other real-time systems demonstrated similar performance. CONCLUSIONS: The accuracy, reliability, genotype inclusivity and affordability, along with the small volumes of blood required for the assay suggest that the rtLC DBS assay will be useful for early diagnosis and monitoring of pediatric HIV-1 infection in resource-limited settings.

It is estimated that 33.2 million people were infected with HIV-1 at the end of 2007; 2.5 million were children under 15 years of age, the majority of whom acquired infection through mother-tochild transmission (MTCT; [1] ). Antiretroviral therapy (ART) is effective at blocking MTCT and can markedly improve the outcome of pediatric HIV-1 infection. However, efforts to provide widespread access to ART have been hampered by the limited availability of infant diagnostic methods. [2] [3] [4] [5] [6] Methods to diagnose and monitor HIV-1 infection in resource-poor settings are usually limited to serologic assays and CD4/CD8 counts. [2, 7, 8] However, antibody based assays can reliably guide diagnosis and management only after 18 months of age following clearance of passively transferred maternal antibodies. [9] PCR based nucleic acid amplification and quantification of HIV-1 is the gold standard for early HIV-1 diagnosis in infants and for evaluating ART efficacy. [2, 5, [10] [11] [12] [13] [14] [15] [16] Several commercial nucleic acid based molecular tests are available for HIV-1 diagnosis and viral load quantification. [15] These commercial assays require relatively large volumes of blood that need to be processed for plasma, stored, and transported under special conditions. The cost, run-time, training, maintenance and infrastructure needed to perform these assays have also limited their use in low resource settings. [6, 17, 18] Practical and reliable methods to obtain, store, and transport blood samples are necessary to develop cost effective early diagnostic assays in limited-resource settings. Spotting of whole blood onto filter paper offers technical and economic advantages over conventional venipuncture methods since it simplifies sample collection and transport to reference laboratories for diagnostic testing and viral load quantification. [17, [19] [20] [21] [22] [23] [24] In the present study, we describe a LightCycler-based real time PCR assay (rtLC) to quantify viral loads using RNA extracted from filter paper containing dried blood spots. DBS prepared either under controlled conditions in a laboratory setting by research technicians or in a clinical setting by health-professionals were successfully evaluated. This assay has comparable reproducibility, diagnostic accuracy, specificity and broader linear range at a lower cost compared to probe-based commercial assays. The rtLC HIV-1 DBS assay is also capable of detecting and quantifying different clades of HIV-1 without sacrificing sensitivity. Study Participants. In this prospective study, blood samples were drawn by trained healthcare professionals from 33 HIV-1 positive children who presented to the UMass Mother-Child HIV Program clinic for routine care between May 2005 and September 2008. Study participants ranged in age from 5 to 21 years. Four HIV-1 positive adults who were enrolled in a study of viral kinetics between 1999 and 2002 were also included in the study. Thirty patients were of U.S. origin and were infected with clade B HIV-1, while 7 were of non-U.S. origin, infected with non-B. Nineteen (51%) of the 37 HIV-1 patients were on ART and eight (22%) had undetectable RNA by Versant HIV-1 RNA 3.0 branched DNA (bDNA) assay (Siemens Healthcare Diagnostics) or Ultrasensitive Amplicor HIV-1 Monitor v1.5 assay (Roche Diagnostics) ( Table 1) . These tests were performed by trained technicians. Blood samples were also drawn from 44 HIV-1 negative individuals (27 adults, 17 children). Individual patient consent was obtained from study participants and guardians and no adverse events were associated with drawing of blood for the study. The Human Subjects Committee of the University of Massachusetts Medical School approved all studies. In addition, DBS samples were collected between February 2008 and January 2009 by heel-stick from 19 infants born to HIV-1 positive mothers at the HEAL Africa Hospital (Goma, Democratic Republic of Congo). Informed consent for testing was obtained from the infants' guardians. These studies were exempt under the guidelines of the Human Subjects Committee of the University of Massachusetts Medical School since the samples were received in the lab as coded, de-identified DBS with no traceable patient information. Preparation of DBS. At the UMass clinic, whole blood was drawn by venipuncture and collected in tubes with EDTA. DBS were prepared by spotting 50 ml of whole blood onto filter paper (Whatman no. 903; formerly SS 903, Schleicher & Schuell, Keene, NH). For 6 of the 23 patients, cryopreserved plasma was spiked into donor HIV-1 negative blood to prepare DBS since whole blood was not available. The spotted filter papers were allowed to dry for at least 4 hours at room temperature in a hood and placed in individual ziplock bags containing a silica desiccant (Whatman, Schleicher & Schuell, Keene, NH). DBS were stored at 220uC until further processing and testing. The plasma was recovered from the remaining blood sample by centrifugation and stored at 280uC for quantification of viral load using commercial (Roche) RNA assays. At the HEAL Africa Hospital (Goma, Democratic Republic of Congo) site, DBS samples collected by heel stick from infants born to HIV-1 positive mothers were stored and shipped to the laboratory at ambient temperature; upon receipt in the laboratory, the DBS samples were stored at 220uC until ready to be analyzed. To better investigate whether the rtLC DBS assay could detect and quantify across clades, in addition to the Congo DBS specimens, we prepared dried spots of whole blood or plasma from 7 non-US origin patient samples (clades A, C, D, AE, AG) spiked into HIV-1 negative donor blood. Genotype Inclusivity Studies. Genotypic clade determination of patient viruses was done using RT-PCR followed by nested amplification of were performed to amplify env and/or gag genes from plasma HIV-1 RNA. Amplified products were cloned for bidirectional DNA cycle sequencing using an ABI model automated sequencer. Phylogenetic analyses based on env and gag nucleotide sequences were used to determine the clade specificity. Globally Prevalent Strains. Reference viruses (13 CCR5-tropic and 1 CXCR4-tropic) were obtained from the NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH (Catalog #11412). [25] This panel represented 6 major globally prevalent HIV-1 clades (A, B, C, D, and circulating recombinant forms CRF01_AE and CRF02_AG). RNA from these reference viral isolates was extracted using the QIAamp Viral RNA mini kit, according to manufacturer's instructions (Qiagen, Valencia, CA), eluted in 60 ml of elution buffer and then diluted 1:1000. A DBS panel was prepared by mixing 60 ml of each virus isolate with 240 ml of seronegative donor whole blood and then spotting 50 ml of this spiked donor blood onto 903 filter paper. Using the protocol described below, this dried blood spot panel was extracted and run in the rtLC DBS assay. Preparation of Standard Curve. To ensure uniformity and reproducibility in the DBS preparation and extraction process, we prepared customized DBS standards with known viral copies. An HIV-1 clade B infected patient isolate with known viral load was selected and spiked into HIV negative donor whole blood to prepare DBS in five fold serial dilutions. A series of 5 independent extractions were performed and quantified using the rtLC DBS assay in a total of 25 independent runs. RNA Extraction. Each DBS containing 50 ml of whole blood was cut into 4 equal pieces and incubated for 5 minutes in Tris-EDTA buffer (1.0 M Tris-HCl, 0.1 M EDTA) at room temperature. HIV-1 RNA was extracted from the filter paper using Trizol reagent as lysis solution according to the manufacturer's instructions (Invitrogen, Carlsbad, CA). Glycogen (200 mg) was added to the lysis reagent as a carrier to facilitate RNA precipitation for each DBS extraction. After removing the residual filter paper, 1-bromo-2-chloropropane (BCP) was used to separate the extracted RNA from the organic phase. RNA was ethanol precipitated, eluted and further reconstituted in 40 ml of Rnase-free water containing 40 units of Rnasin Plus, an Rnase inhibitor (Promega, Madison, WI). DNA Extraction. One spot each of DBS filter paper was processed to extract DNA using the previously described resinbased Chelex (Bio-Rad, Hercules, CA) method. [26] Briefly, the paper punched from one entire DBS was immersed in 1 mL of whole blood specimen wash solution (Amplicor HIV-1 Monitor test v1.5, Roche Molecular Systems) and rocked for 2 hours at room temperature. After a quick spin for 5 minutes at high speed, the red-tinged buffer was removed and discarded. The DBS paper was washed once, and then immersed in 250ul of Chelex resin resuspended in DNA dilution buffer (10% v/v in 10 mM Tris buffer, pH 8.3 containing 50 mM KCl). DNA kit Internal Control (3.3ul) was added to the DBS sample, mixed frequently to keep the resin in suspension, and was heated to 100uC for 1 hour, with a 10 second vortex after the first 30 minutes of heating. After a spin for 3 minutes to pellet the resin, the supernatant containing the extracted DNA was removed and stored at 280uC until HIV-reactivity of the extract was determined using Roche Amplicor HIV-1 DNA Amplicor 1.5 v kit according to a previously published protocol. [26] The cellular equivalents of DNA in each test sample were determined in a real-time Taqman PCR assay by probing for CCR5 copy number using forward primer 59-GCTGTGTT-TGCGTCTCTCCCAGGA-39 and reverse primer 59-CTC-ACAGCCCTGTGCCTCTTCTTC-39, and the corresponding fluorogenic probe 59FAM-AGCAGCGGCAGGACCAGCCC-CAAG-TAMRA 39. Known copies of plasmid carrying the CCR5 gene were used to generate the standard curve, from which the number of CCR5 copies and cellular equivalents were determined for each sample. Real-time PCR analyses on RNA extracted from infant DBS was performed as described below. Real-time PCR. Real-time PCR amplification of HIV-1 RNA was performed in a one-step, single-tube closed system of 32 sample format using the LC-32 Roche LightCycler (Roche, Indianapolis, IN). All the samples were tested in duplicate in a 20 ml total reaction volume containing 16 ml of PCR reaction mix (Quantitect SYBR Green RT-PCR kit [Qiagen, Valencia, CA]; 0.5 mM each of forward and reverse oligonucleotide primer pairs) and 4 ml of the template. The primers were specific to a conserved region of HIV-1 LTR: 59-GRAACCCACTGCTTAASSCTCAA-39 (LTR sense; position 506 of HxB2) and 59-TGTTCGGGCGCCACTGCTAGAGA-39 (LTR antisense; position 626 of HxB2). [27] The PCR reaction was performed according to the following cycling parameters: 1) Reverse-Transcription: 50uC 20 minutes, ramp 20uC/second; 2) Activation: 95uC 15 minutes, ramp 20uC/second; 3) Amplification: 50 cycles a) 94uC 10 seconds, ramp 20uC/second, b) 52uC 20 seconds, ramp 20uC/second, and c) 72uC 20 second, ramp 2uC/ second (single data collection); 4) Melting: a) 92uC 0 second, ramp 20uC/second, b) 57uC 15 seconds, ramp 20uC/second, and c) 92uC 0 second, ramp 0.1uC/second (continuous data collection); 5) Cooling: 40uC 30 seconds, ramp 20uC/second. HIV-1 specific amplicons were detected using SYBR Green technology according to manufacturer's instructions (Qiagen, Valencia, CA). The number of HIV-1 RNA copies in each test template was measured by its threshold cycle (C t ) and determined from the standard curve of serially diluted customized DBS standards using software for data analysis (Sequence Detector version 1.6, PE Applied Biosystems). For each experiment, a standard curve was generated from serial endpoint dilutions (586,000 to 37 copies) of the customized DBS standards. The threshold cycle values were plotted against copy numbers to construct the standard curve. Quantification of HIV-1 RNA in each test sample was back calculated and viral load was expressed as copies/ml. Using the protocol described above, DBS was extracted, run and analyzed in the rtLC DBS assay by personnel with molecular and virology expertise who were blind to the results. Rnase-free water (4 ml) was routinely added instead of test sample to 16 ml of the master mix and used as a no-template control for every run. At the end of the assay, the specificity of each amplified product was ascertained by means of melting curve analysis. This eliminated false positive detections due to primer dimers or nonspecific amplicons. To confirm that there was no DNA contamination in the input RNA and to assess the specificity of the reverse transcribed rtLC DBS products, initial assays included PCR reactions without the addition of reverse-transcriptase enzyme. Gel electrophoresis of the amplicons was initially performed to further confirm the specificity of the products. Each sample was tested in replicates and a second DBS was independently extracted and run to verify a positive result. The test was repeated if the sample had indeterminate results, and if necessary a new spot was extracted and tested if the results of the repeat test were also indeterminate. Quality control for each experiment was assessed by the performances of the standard curve and the negative control. All DBS prepared by trained personnel were tested for this study as long as multiple filter spots were available for each patient, and the blood spots were good quality with no hemolysis or clots. To determine whether viral load data using the Roche LightCycler system were comparable to viral load data obtained by using MyiQ TM Single-Color Real-Time I-Cycler PCR Detection System, (Bio-Rad, Hercules, CA), cDNA synthesis was carried out with the iScript cDNA Synthesis Kit (Bio-Rad, Hercules, CA). Select patient and standard DBS-RNA were used to synthesize cDNA, and then amplified and quantified using the iQ SYBR Green Supermix Kit (Bio-Rad, Hercules, CA) by using these parameters: 3 min @95C, 50 cycles of 10 sec @94C, 20 sec @52C, 20 sec @72C, and final extension for 20 sec @72C. Results were analyzed using MyiQ software, version 1.0.410 (Bio-Rad, Hercules, CA). Sensitivity and Linearity of the Assay. Linearity of the assay was evaluated using serial dilutions of the customized standard DBS described earlier (586,000 copies to 37 copies). Statistical Analyses. The customized DBS standards serial dilutions were tested in 25 separate runs to determine the threshold, inter-assay precision and linearity of the rtLC DBS assay. Trained statisticians performed data analyses. Probit analysis was used to determine 95% and 99% detection limits. The likelihood ratio test was used to evaluate the effect of extractions and runs; the fitness was measured by R-squared value. Spearman correlation coefficients were calculated to determine the relationship between the HIV-1 RNA levels quantified by the Roche Ultrasensitive and bDNA assays with the rtLC DBS technique. The sensitivity of the rtLC DBS assay was calculated as the number of positive results divided by the total number of samples from infected patients who had plasma viral load above the threshold level of detection of commercial assays (bDNA and Roche Ultrasensitive) expressed as a percentage. The specificity was calculated as the number of negative samples divided by the total number of known negative specimens obtained from normal healthy donors and uninfected infants expressed as a percentage. Paired Wilcoxon tests were used to determine the effect of temperature and time on RNA stability and assay performance. The linear dynamic range of the rtLC DBS assay was initially assessed using a 5 log 10 dilution series of the customized standard DBS. The assay was shown to be linear over the entire range of 586,000 to 37 copies. A linear regression of the rtLC DBS customized standards copies on true concentrations yielded a correlation coefficient of 0.984 (P,0.001) and the fitted model is shown in Figure 1 . The slope of 1.037 closely approximates the theoretical maximum amplification efficiency of 100%; the fitted slope is slightly greater than 1 due to the several undetected cases in low range. To estimate the detection limit of the assay, we performed 22 extra runs at low concentrations (7, 17, 37, 83, 187 copies), which spread evenly in log scale and contained two concentration 37 and 187 that were already in design. We used all available data at concentrations from 7 to 937 to fit the probit model to get estimations for 95% detection limit as 135, with 95% confidence interval [82, 223]; 99% detection limit was estimated as 292, with 95% confidence interval [147, 583] ( Table 2) . To evaluate the inter-assay precision of the assay, the data on the customized DBS standards of known concentrations (586,000 to 37 copies) were analyzed across 25 independent experiments. The results of the statistical analysis for the standard deviation (SD) and percent coefficient of variation (%CV) at each concentration level are shown in Table 3 . The coefficient of variation increased dramatically when the true concentration was approximately equal to or below the assay's detection limit. The clinical sensitivity of the assay was determined using DBS specimens of 32 patients with known infection and viral loads above the detection threshold of the bDNA and Roche Ultrasensitive assays (Table 1 ). All 32 samples were positive using the rtLC DBS assay, yielding a clinical sensitivity of 100% for this cohort. An additional 5 DBS samples drawn from infants of indeterminate HIV-1 infection status tested positive on the rtLC DBS assay; all 5 were also positive on the commercial DNA assay. The clinical specificity of the rtLC DBS assay was determined using DBS specimens from 27 healthy HIV-1 negative adult donors and 17 HIV-1 negative infants born to HIV-1 positive women. All of these samples were negative for HIV-1 RNA, resulting in a clinical specificity of 100%. The information on the clinical specificity of rtLC DBS assay in HIV negative donors is provided in Table S1 in the supplementary section. rtLC Assays Using Samples Collected in the Field DBS samples from infants born to HIV positive women were also prepared at a clinical site in Goma, Congo. The range of cell equivalents across these Congo DBS samples was 2,769 to 201,116 Table 2 . Calculation of the 95% and 99% detection limit of the rtLC DBS assay using probit analysis. The rtLC DBS assay successfully detected HIV-1 RNA in each of the DBS samples prepared from 14 viral isolates representing an international HIV-1 reference panel (clades A, B, C, D and CRF-AE and CRF-AG). The rtLC DBS assay also successfully detected and quantified HIV-1 RNA in DBS samples prepared in our lab using whole blood samples from 23 patients (US origin) infected with clade B virus (Table 1) , and 6 patients (non-US origin) infected with nonclade B virus (Table 1) . Finally, the rtLC DBS assay successfully detected HIV-1 nucleic acids in the blood of all 5 HIV-1 positive Congo infant DBS samples that were prepared in the field and shipped at ambient temperature ( Table 4 ). Plasma viral RNA copy numbers determined by the rtLC DBS assay were compared to results obtained using commercial assays (bDNA and Amplicor). The Spearman coefficients of rtLC DBS with the Roche and bDNA assays were 0.91 and 0.89 respectively. In a pilot study using the i-Cylcer real-time PCR system to quantify viral load from select patient and standard DBS RNA preparations, and using Sybr Green dye for detection, we demonstrated comparable performance to that of the LightCycler system ( Figure S1 ; Supplementary section). A significant correlation between i-Cycler and LightCycler based viral loads was observed in DBS specimens (Spearman Ranks correlation, p,0.0001) suggesting that our LightCycler-based DBS assay will have universal applicability. To evaluate the performance of this assay in limited resource settings where storage and shipment of DBS at ambient/roomtemperature (.25uC) is the norm, we investigated the effect of 37uC temperature on the stability of DBS/RNA, by comparing the detection of HIV-1 RNA in DBS samples stored at 220uC or 37uC. We also evaluated 7 days of storage at 37uC to emulate shipment of DBS from the point of preparation (clinic) to a tertiary referral center (reference-laboratory), and compared the assay results to those of identical DBS stored promptly at 220uC. No statistically significant difference in viral load was observed for 12 samples stored either at -20uC or 37uC (Wilcoxon signed rank test, p = 0.06) ( Figure 2 ). Further, to evaluate for potential loss of RNA over time in DBS samples stored at 220uC and 37uC, viral load was determined on RNA samples extracted day 1 and day 7 postpreparation of DBS. The data (available for 8 patients) for days 1 and 7 were comparable irrespective of the storage temperature (paired Wilcoxon test; data not shown). The fitness measured by R-squared value in the simplest model, which includes only the true concentration as predictor, is 0.955, and R-squared values from models that include extraction and run effects were 0.9959 (extraction and run) and 0.9948 (extraction), respectively. The improvement of fitness by including extra extraction specific parameters and run specific parameters are marginal, and we conclude that true concentration explains most (95.5% percent) of the variation in observed concentration, and the influence of extractions and runs are limited. Measurement of Plasma HIV-1 RNA over time in Patients on ART The rtLC DBS assay was used to monitor plasma HIV-1 RNA levels in two patients for up to one year on ART (Figures 3a and 3b) . The rtLC DBS assay showed good correlation with the Roche Ultrasensitive assay for the longitudinal follow-up of these two patients, suggesting that it may also be useful for monitoring viral load in patients on ART. In this study, we used a nucleic acid based real-time PCR assay to detect and quantify plasma HIV-1 copy numbers on samples from 56 HIV-1 infected patients utilizing the DBS format. The assay results were comparable and correlated well with commercially available viral load assays (Siemens bDNA, Spearman r = 0.89 and Roche Ultrasensitive Amplicor, Spearman r = 0.91). In the patient cohort analyzed, the assay successfully detected all positive samples. The calculated specificity using known negative samples was 100%. The estimated 95% detection threshold was 136 copies and the dynamic range of the assay was 5 log 10 . Finally, the assay successfully detected four major subtypes and 2 CRF of HIV-1. A little over a decade ago, we demonstrated the utility of early combination antiretroviral therapy (ARV) in infants. [28, 29] HIV infection is associated with particularly high morbidity and mortality in limited-resource settings and a recent randomized trial conducted in South Africa demonstrated reduced morbidity and mortality in infants treated with early combination ARV. [30] In a recent WHO consultants' meeting, revision of current guidelines was recommended to include routine early diagnosis and treatment of HIV positive infants under 1 year of age. [31] Nucleic acid based testing is the gold standard for early diagnosis. Commercially available assays (bDNA, Roche Amplicor Ultrasensitive, and Cobas) are relatively expensive and require significant infrastructure and technical expertise to allow transfer of technology to resource-limited settings. [2, 8] Hence access to nucleic acid testing in these settings is currently very limited. [6, 17, 18, 32, 33] Studies to detect and quantify HIV-1 have traditionally involved two-step, nested or probe based PCR [34] [35] [36] [37] ; more recently, real time PCR has been used for HIV detection and quantification. [6, 22, 24, 34, 38, 39] Multiple investigators have documented the utility of DBS sample collection for early HIV-1 detection in infants, viral load monitoring, and surveillance of seropositivity and drug resistance in laboratories and clinics that lack facilities for refrigeration or sample processing. [8, 17, 20, [22] [23] [24] 38, [40] [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] The DBS format greatly facilitates the logistics of sample collection, processing, and shipping for limited resource settings. Whole blood saved as DBS can be transported or mailed to reference laboratories without refrigeration and has low biohazard risk. Optimal storage conditions for DBS and long term stability of DNA and RNA from DBS under different storage conditions have been extensively documented. [23, 44, 47, 51, 53] Our pilot studies with a small cohort of patient samples to assess the effect of temperature and time, albeit one week, on the stability of RNA demonstrated no difference, and were in agreement with previous larger studies. [39, [54] [55] [56] Comparable efficacy using multiple extractions and runs of our customized standards support the reliability of the DBS format and the rtLC DBS assay. The rtLC DBS assay described herein is a one-step, walk away technique. Automation of this assay provides potential for high throughput with very small sample volumes, which makes the assay suitable for use in infants and children from whom one often has access to only small blood volumes. It is an assay system which will be cost-effective and easily adaptable to limited-resource settings, where the majority of new HIV-1 infections are seen today. The detection of PCR products by SYBR Green ensures good sensitivity. SYBR Green is relatively inexpensive compared to probe-based detection, and in general, SYBR Green detection is one cycle threshold or so more sensitive than probe-based assays. Several previous studies have utilized SYBR Green for HIV-1 diagnosis and are reviewed by Espy et al. [57] The efficiency of this dye-based assay is also comparable with the currently available diagnostic assays on HIV. Aside from use in HIV diagnosis and quantification, SYBR Green has been widely used to detect and quantify diverse human pathogens. [58] [59] [60] [61] [62] [63] [64] [65] The data in these reports strongly support the utility and reliability of SYBR Green for detection of specific PCR amplicons above the background. The use of degenerate LTR primers in the rtLC DBS assay allows for a wide range of genotype inclusivity. Compared to commercial assays, the rtLC DBS assay is rapid and cost-effective. The equipment used in the rtLC assay is self- contained, occupies minimal bench-space and doesn't require accessory equipments (such as a plate washer, optical density reader, and incubator; Table 5 ). Aside from the initial cost of obtaining the LightCycler instrument, a comparative analysis of assay costs and technician time reveals a 40-fold decrease in cost as well as 2.5 fold decrease in technician time (4.5 hours) associated with the rtLC DBS when compared to commercial kit based PCR assays ( Table 6 ). The reduced equipment requirements, personnel hours, and costs compared to the commercial 'gold' standard assays make the rtLC DBS assay attractive for transfer to and use in resourcelimited settings. The success of the pilot field study on DBS from Congo emphasizes the utility and applicability of our assay although further studies with larger sample sizes are definitely warranted. In summary, we have utilized a real-time LightCycler based PCR assay on small volumes of whole blood dried on filter paper to successfully detect and quantify viral loads across different HIV-1 clades. The use of dried blood spots provides a simple and inexpensive means for obtaining blood samples for analysis that minimizes the risk for contamination while maximizing the ability to obtain timely results. A major advantage of the rtLC DBS assay is that the amplification, real-time detection and quantification, and confirmation of amplicon-specificity by melting curve analysis are performed in a one-step, closed-tube format. In addition, viral load tests by rtLC DBS assay are substantially less expensive and logistically less intensive than commercial assays. Preliminary data suggest the adaptability of the assay into other real-time systems. Validation of these results with larger field studies would constitute a more robust evaluation and will have major implications for early diagnosis, disease management, and epidemiological-or resistance-surveillance studies in limited resource settings. Table S1 Found at: doi:10.1371/journal.pone.0005819.s001 (0.04 MB DOC)

A Statistical Framework for the Adaptive Management of Epidemiological Interventions

BACKGROUND: Epidemiological interventions aim to control the spread of infectious disease through various mechanisms, each carrying a different associated cost. METHODOLOGY: We describe a flexible statistical framework for generating optimal epidemiological interventions that are designed to minimize the total expected cost of an emerging epidemic while simultaneously propagating uncertainty regarding the underlying disease model parameters through to the decision process. The strategies produced through this framework are adaptive: vaccination schedules are iteratively adjusted to reflect the anticipated trajectory of the epidemic given the current population state and updated parameter estimates. CONCLUSIONS: Using simulation studies based on a classic influenza outbreak, we demonstrate the advantages of adaptive interventions over non-adaptive ones, in terms of cost and resource efficiency, and robustness to model misspecification.

Epidemiological interventions generally remove susceptible individuals or apply some form of treatment to infected individuals in order to prevent further spread of a disease. The susceptible population may be culled, as in the case of foot-and-mouth disease [1, 2] , in which case the total population size is permanently reduced. The infected population may be quarantined, as in the case of SARS [3] , in which case total population size is unchanged but the fraction of infecteds that may be in contact with susceptibles is reduced. Most commonly, susceptibles are vaccinated (cf influenza or smallpox [4, 5] ), in which case the total number of susceptibles, but not the total population size, is reduced. Each of these interventions incurs a quantifiable cost: culling results in additional deaths; medical treatments or quarantines result in monetary expenses; vaccination incurs both monetary expenses, and in some cases additional vaccine-induced infections. Additionally, in many situations the costs associated with each of these actions can depend upon the state of the disease within the population of interest. For example, per-dosage prices of vaccine can increase as resources become scarce as a result of an aggressive vaccination campaign. Similarly, vaccine efficacy can decrease as a result of selection for drug resistance. Such observations raise the question of how to find optimal interventions that adaptively depend on the state of the epidemic. A key challenge to calculating optimal intervention strategies involves devising ways to characterize and explore the space of intervention policies. Most existing work on optimal intervention has required various limiting assumptions about the forms of such strategies. Ball and Lyne [6] considered optimal vaccination in terms of the allocation of vaccine doses to households of various sizes in an explicitly structured population model. Patel et al [7] considered optimal vaccination in terms of the allocation of vaccine doses to different age classes in an explicitly age-and geographically-structured population model. Tildesley et al [1] describe optimal vaccination strategies for a foot-and-mouth epidemic in which the optimized parameter is the size of the radius surrounding a point of infection within which all livestock are to be vaccinated. These methods are primarily concerned with pre-emptive interventions that can be completed before the arrival of the pathogen. Under such scenarios, there is no need to consider adaptive or sequentially updated interventions because as soon as the intervention policy is triggered, the threat of epidemic is eradicated. In real scenarios, such widespread vaccination may not be achievable. Moreover, these methods traditionally involve calculations that assume no uncertainty in key model parameters such as transmission rate, recovery rate, and mortality rate. Recently Elderd et al [8] , using Bayesian methods, demonstrated the importance of explicitly quantifying such underlying uncertainty when forecasting the expected efficacy of trace versus mass vaccination policies. Their findings demonstrate that accurate propagation of parameter uncertainty can sometimes reveal deep and troubling consequences of a proposed vaccination strategy, and they suggest that incorporation of such uncertainty could impact policy decisions. Here we address the question of how to dynamically propagate uncertainty in order to respond to an emerging epidemic while simultaneously and continuously learning about its underlying transmission dynamics. Estimation of model parameters is facilitated by regarding the transmission dynamics as stochastic processes rather than deterministic solutions to a structural equation model. This allows us to explicitly account for uncertainty in both model parameters and disease transmission. We consider a very general class of vaccination strategies defined by a fraction of the current susceptible population to be targeted for vaccination, and a threshold number of susceptibles such that once the number of susceptibles falls below this threshold, the vaccination campaign is called off. We demonstrate the calculation and application of optimal strategies of this form when coupled with iteratively updated parameter estimates using simulations based on a well-studied influenza outbreak [9] . Our emphasis is not on the realism of the underlying SIR model (though it has been shown that even simplistic transmission models can provide good fit to actual data [10] ), but rather to describe an effective approach for combining estimation and policy calculation. Permitting greater flexibility in the form of the possible intervention renders calculation of optimal intervention strategies analytically intractable, thus requiring evaluation by Monte Carlobased methods. Once in a Monte Carlo-based framework, it becomes straightforward to couple the evaluation of intervention strategies with Bayesian procedures for performing on-line estimation of parameters of the underlying epidemic model, thereby propagating parameter uncertainty through to policy decisions. The policies produced by this framework are optimal in that they minimize the expected cost of the epidemic and adaptive in that the optimal policy changes as a function of the state of the epidemic and the degree of uncertainty in underlying model parameters. Using extensive simulation studies we compare the distribution of costs accrued under adaptive intervention to those arising from non-adaptive policies in a variety of scenarios. Our studies show that adaptive policies perform similarly to nonadaptive policies based on perfect parameter estimates, and significantly better than nonadaptive policies based on imperfect parameter estimates. Additionally, we show that adaptive online estimation affords the method some robustness to model misspecification. These results further demonstrate the importance of accounting for such underlying uncertainties in dynamic settings and indicate the utility of adaptive policies in settings where perfect estimates and a true model do not exist. All computational methods used herein have been made freely available through the amei (Adaptive Management of Epidemiological Interventions) R package [11] . A classic study of Murray's [9] describes the spread of influenza through the population of a British boarding school. During the course of the epidemic, which was traced to the arrival of a single infectious student, all 763 students were eventually infected. The epidemic conforms to many standard assumptions of SIR models: a population essentially closed to immigration and emigration, recovery with immunity, and nearly homogeneous mixing of susceptibles and infectives. Viewing the transmission dynamics as a discrete time stochastic process rather than a deterministic system of coupled differential equations implies a distribution of possible outcomes for the epidemic. By conditioning on parameter values and initial conditions (S 0~7 62, I 0~1 ), Monte Carlo simulation can be used to explore the distributions of numbers of susceptible, infected, and recovered individuals, as well as total accrued cost, as functions of time. Murray provides estimates of the transmission rate (b b~0:00218) and recovery rate (v v~0:4), which we regard as the ''true'' underlying parameter values in our simulations. Additional aspects of the transmission function are discussed in the Methods section. We assume that all costs can be expressed in a common monetary cost unit. Other choices of cost functions that address the issue of nonconformable costs (e.g. lives vs dollars) are mentioned in the Discussion. Setting the unit cost to be that of maintaining a single infected individual for one time step (cost per infected, c t~1 ), repeated forward simulation of the epidemic (Figure 1 ) indicates that the mean total cost over 40 time steps is approximately 2100 cost units ( Figure 2 ), attributable entirely to the cumulative cost of maintaining a large population of infected individuals until recovery. We consider a relatively simple but flexible class of intervention strategies that involve vaccinating a target fraction (a) of susceptible individuals at each time step. After a round of vaccination, if the number of remaining susceptibles is less than a designated threshold (c), the vaccination campaign is discontinued. Policies defined in this way provide effective target population sizes, to which post-hoc corrections can be applied in light of knowledge of the population structure. We assume that in a single time unit there is an upper bound on the maximum targetable fraction of susceptibles. In our simulations we set this bound to be 30%, so that several time units are required to vaccinate the majority of susceptibles. We also assume there is a period of time after the arrival of the initial infection before intervention can begin. In our examples, we assume this lag time to be 7 time units. These values are chosen purely for the purpose of demonstration, and can be assigned any value in the amei software. The optimal variable stop time vaccination strategy can be found by searching the policy space (i.e. pairs of fractions-tovaccinate a and stopping thresholds c) for the policy that most frequently produces the lowest expected cost. The calculation of the optimal policy therefore explicitly accounts for uncertainty associated with the disease transmission and recovery processes (see Methods) under a given valuation of the model parameters. Assuming a value of 2 cost units per dose of vaccine (c v~2 ), we use Monte Carlo simulation to estimate the expected cost surface associated with variable stop time policies based on the true parameter values (Figure 3 ). The minimum expected cost is achieved under a policy of maximum (30%) vaccination and a stopping threshold of 150 individuals. Repeated simulation of the epidemic under this policy shows that in the average case (dashed line), the policy amounts to 4 time units of maximum vaccination as soon as the initial lag is over ( Figure 4 ). In situations where the number of susceptibles remaining after the lag is already below 150 individuals, no policy is implemented. The 95% central interval for the final distribution of total vaccine units dispensed is (339,581), representing variation in the total size of the epidemic at the time of the vaccination sweep, and the numbers of new infections after vaccination begins. Figure 5 shows the distribution of total costs accrued under this policy. After the end of the vaccination campaign, the uncertainty bands widen, representing variations in the costs associated with maintaining the remaining population of infected individuals until their natural recoveries. The mean total cost at time 40 is 1652 cost units, approximately a 21% reduction in total cost compared to no-intervention. The intervention calculated in the previous section represents a gold-standard for this particular scenario because the vaccination strategy was calculated using the same parameter values and the same SIR model formulation as the simulated disease process. In most settings it will be natural to regard the transmission model parameters as unknowns to be estimated from incoming count data describing the sizes of the susceptible, infected, and recovered subpopulations. In this section we describe the procedure for performing adaptive management of an emerging epidemic, in which we account for parameter uncertainty and its impact on vaccination strategies. An epidemic can be effectively summarized by the disease state of the population (i.e. the current numbers of susceptible and infected individuals) and by the SIR model parameters that define the dynamics of transmission, death, and recovery. In adaptive management, the former is used to perform inference on the latter. Each time new data are collected, Markov chain Monte Carlo (MCMC) is used to sample from the current posterior distribution on model parameters. The optimal variable stop time strategy associated with each set of sampled parameter values is calculated, and the policy that most frequently minimizes the total expected cost (over all sampled parameter values) is enacted at the next time step. The fundamental difference between the adaptive policies calculated here and those calculated in the previous section is that here, the vaccination policy is a dynamic function of the current disease state and the current distribution of each parameter, whereas before, the policy was a static function of the initial disease state and the initial point estimate of each parameter. The effectiveness of this approach can be similarly explored by repeated simulation of epidemics under adaptive management. As before, we assume an initial lag time of 7 time units before vaccination begins. Here we also introduce a cost associated with deaths (c d~4 ). Even though the ''true'' model does not include mortality, the fitted model includes a mortality parameter (m). This allows examination of the degree to which adaptive management strategies are robust to model misspecification. Initial uncertainty regarding parameter values is expressed in the form of vague/noninformative prior distributions, as specified in the Methods. The choice of prior distributions in Bayesian models is of fundamental importance, and other possible choices are mentioned in the Discussion section. At each time step, the state of the epidemic is advanced one time step using the same ''true'' parameter values used in the previous section. Intervention strategies, however, are calculated based on the current parameter estimates. Figures 6 and 7 show the distributions of susceptible, infected, recovered, and vaccinated individuals, and total accumulated costs for repeated simulation of the epidemic under adaptive management. These dynamics can be compared to those in The tighter bound about a smaller mean is due to the ability of the adaptive strategies to methodically diminish the vaccination campaign as a function of the epidemic state. This can be seen in Figures 8 and 9 , which display the distributions of implemented vaccination strategies for each time step during the course of adaptive management. In the average case (dashed line), the maximum policy is enacted for 3 time steps, followed by a round of 20% vaccination. The uncertainty surrounding the implemented strategies indicates the degree to which the the adaptive policies are adjusted in light of data. In epidemics associated with the upper 97.5 percentile of vaccination strategies (top solid line in Figures 8 and 9 ), the adaptive policy calls for 4 rounds of maximum vaccination followed by a round of 20% vaccination, followed by a final round of 5% vaccination. In this way, the adaptive nature of the interventions enables more efficient use of vaccine resources than achieved under nonadaptive policies. The distribution of total cost associated with the adaptive intervention simulations (Figure 7) is essentially equivalent to the distribution of costs achieved under static intervention with perfect information (Figure 5 ), indicating that even the short period of data collection prior to action produces parameter estimates that are sufficient for accurate prediction of the disease dynamics. Figure 10 shows the final posterior distributions on the four model parameters estimated from the data during one simulation of the epidemic under adaptive management. True values are indicated with a circle, mean values are indicated with an 'x', and the central 95% region of each distribution is shaded. The prior densities of each parameter for the same interval are shown in red. As mentioned above, the inference model is misspecified relative to the model being used to simulate the epidemic, in that the inference model includes a mortality parameter (m, see Methods), even though no deaths were observed in the simulated outbreaks. By coupling the policy calculations with an inference framework, the effect of such model misspecification appears to be reduced. We can further demonstrate the utility of the adaptive approach in situations of more severe model misspecfication. To do so, we construct a simulation experiment in which the inference model upon which the adaptive management is based is as described here, but in which the underlying transmission model through which new infecteds are generated is an entirely different, non- nested transmission model with a latent infective resevoir (see the amei vignette on CRAN [11] for details). This situation more closely resembles one that may be encountered in practice, where new infections are arising from an actual disease transmission process whose dynamics are at best approximated by any mathematical characterization. Table 1 compares summaries of the posterior distribution of cumulative cost arising under adaptive management to those predicted under the optimal static policy using parameters estimated for the misspecified model based on a completely observed epidemic. It is important to recognize that the adaptive policy is at a severe disadvantage, basing its actions on parameter estimates produced simultaneously during the course of a single epidemic (and using vague prior distributions) while the static policy conditions on parameter estimates obtained from a completely observed epidemic. In spite of this, the adaptive policy achieves nearly identical costs. We have now shown the near equivalence of the adaptive and static policies in two different scenarios. These situations indicate that the proposed methodology is efficiently and with sufficient accuracy estimating the parameters of the transmission model, such that adaptive strategies based on these on-line estimates produce equivalent outcomes to those static strategies based on full retrospective analyses. Moreover, it is simple to demonstrate that static control measures based on reasonable but imperfect parameter estimates can lead to substantially worse outcomes/ higher costs than the adaptive policies ( Table 2 ). In real situations, where actions must be based on parameter estimates made from incomplete or limited information, the practice of iterative refinement of estimates and policies is likely to result in significantly improved outcome. We have demonstrated a novel adaptive management strategy based on a relatively simple characterization of the underlying SIR model and the epidemiological cost function. In principle, this methodological framework can readily accommodate more complicated disease dynamics such as immigration, latent infected states, missing data, and vector-communicated diseases, as well as more complicated intervention strategies that allow combined vaccination and quarantine. However, the incorporation of such features is likely to impose a heavy computational burden, and so model complexity should only be increased when additional parameters are supported (and identified) by the data and demanded by the biology. As in all Bayesian analyses, care must be taken when choosing prior distributions. In this study, our primary interests required the use of vague/noninformative prior distributions, in order to demonstrate the estimability of model parameters. In practice, informative, even pessimistic priors (i.e., overestimated infectiousness and mortality, underestimated recovery) may provide useful reference points for the adaptive policy calculations, especially in situations of acute infections for which the duration of the epidemic may be too short for incoming data to dominate the prior information. In such situations, the adaptive approach still provides the opportunity for data to inform parameter values if it becomes available, while basing interventions on current parameter estimates as determined by their prior distributions. There is an important choice to be made in assigning costs to the various actions that comprise an intervention strategy. A monetary valuation scheme is the most straightforward, but it may be difficult to construct such a scheme that adequately represents all aspects of the decision. One alternative would be a valuation in which each cost is chosen to represent a probability of mortality. In this way, the cost to be minimized would be the expected total loss of life for the epidemic under a given intervention strategy. By assuming that the removal rate can be expressed as r~mzv, where m is the rate of disease-induced mortality and v is the rate of natural recovery from the infected state, we can set c i~1 {e r ð Þ m r , so that the cost associated with maintaining a given number of infected individuals for a unit of time is the number of infected individuals that are expected to die in a unit of time. Similarly, situations exist where it is reasonable to assign a probability of mortality to the removal of susceptibles, as in the cases of smallpox vaccination or the culling of livestock. A related extension to this framework would involve applying a monetary constraint to a loss-of-life cost function. If we were to assume p i and p r to be, respectively, the probabilities of mortality associated with untreated infected individuals and the removal of susceptibles, and define d to be the monetary resources available for the intervention, then within this framework it is possible to find the intervention that minimizes the total loss-of-life subject to the total spending constraint d. Similarly, it would be possible to optimize with respect to some selective criterion in order to preserve vaccine efficacy rather than select unnecessarily for drug-resistant pathogens. Also note the possibility of calculating policies based on minimization of some quantile of the realized cost rather than the mean cost. This would lead to minimization of costs associated with worst case scenarios, rather than that associated with the average case scenario. These and other alternative formulations of the underlying optimization problem can be easily accommodated in the framework presented here. The utility of adaptive interventions is especially evident in situations of an emerging pathogen with which the host population has no previous experience. In such a situation, vaccines will not be immediately available at the onset of the epidemic, and so a methodology for combining currently available actions while anticipating the future availability of vaccines would be of great use. Effective epidemiological intervention requires swift decision in consideration of the various direct and indirect costs of intervention. The methodological framework described here provides a decision theoretic basis for automating this process. All statistical and computational methodology described here has been implemented in a freely available R package called amei (Adaptive Management of Epidemiological Interventions), which can be downloaded at http://cran.r-project.org/web/packages/ amei/index.html [11] . We consider a standard Susceptible-Infected-Removed (SIR) model [10, 12] with no loss of immunity but with mortality. In this model, the dynamic variables at time t are the number of susceptible individuals, S(t); the number of infected individuals, I(t); the number of recovered individuals, R(t); and the number of removed/dead individuals, D(t). We assume that the population is closed to immigration such that S(t)+I(t)+R(t)+D(t) = N is constant, and any three of the dynamic variables define the fourth. To characterize the transmission of the disease, we adopt the negative binomial form for the transmission function [13] , so that the model parameters are the transmission rate b, the overdispersion parameter k, the death rate m, and the rate of recovery to the immune class n. Under these assumptions, the SIR model is described by the following system of differential equations [12, 13] : The negative binomial transmission function implies that disease transmission occurs following a Poisson process in which encounters between infected and susceptible individuals are Poission distributed with the encounter rate varying according to a gamma distribution with coefficient of variation k {1=2 . Via the parameter k, the negative binomial transmission function can account for social interactions and/or network factors in disease transmission, without requiring explicit characterization of the population structure. The SIR model formulation also leads immediately to a natural discrete time approximation for the numbers of infections (Ĩ I), recoveries (R R) and deaths (D D) arising in the unit time interval from t to t+1. Holding the total number of infected individuals I constant and integrating Equation 1 over a unit time interval gives Here lower case denotes the realized value of the associated capital letter random variable. In this discrete time approximation we have assumed a particular ordering of events, namely that recoveries occur first, followed by deaths from among those infected individuals who did not recover, followed by new infections. Simulation studies indicated that these assumptions, as well as other possible orderings, resulted in system dynamics that were approximately equal in expectation to deterministic solutions of the continuous time SIR model. In all forward simulations of the disease dynamic (except where noted) we assume the ''true'' underlying parameter values to be those estimated by Murray [9] , with the exception of the negative binomial overdispersion parameter k. Thus, b = 0.00218, n = 0.4, and m = 0 (no disease-related mortality). We set the overdispersion parameter to be k = 0.1, in order to produce epidemics that, without intervention, have run their course by 40 time units but such that there is variation in the size of the outbreak. We formulate the total expected cost of the epidemic in terms of the underlying costs associated with maintaining infected individuals until recovery, suffering death, and administering vaccinations. Let c 1 a,c,s ð Þ denote the cost associated with interventions involving susceptibles when S(t) = s. Here a is the fraction of susceptibles that are moved directly into an immune/recovered class, as by vaccination, and c is the threshold below which the intervention is discontinued. Letting c v denote the cost per unit, then We let c 2 i ð Þ denote the cost associated with interventions involving infecteds when I(t) = i. This component includes the costs associated with maintaining the non-recovered infected individuals and costs associated deaths, as in where c t is the cost per treatment/maintenance of a non-removed infected individual, and c d is the cost per death. Assuming the initial epidemiological state is S 0 ð Þ~s 0 , I 0 ð Þ~i 0 , the expected total cost of the epidemic under intervention strategy (a,c) can be expressed recursively as where E C t f g denotes the expected cost accumulated from time t onwards. The optimal intervention strategy (a,c) is the one that minimizes the total accumulated cost over the course of the epidemic. Two methods for calculating such strategies are as follows. The total expected cost depends on the parameter values and the initial epidemiological state s 0 ,i 0 ð Þ. Thus, conditional on a set of parameter values, Monte Carlo simulation can be used to search over values of a and c in order to find the combination that minimizes E C 0 f g: For each combination of a and c, with a ranging from 0 to 0.7 and c from 0 to 750 in increments of 50, we conduct 100 simulations of the epidemic, using the true parameter values, in order to estimate the mean cost associated with the intervention (a,c). The strategy producing the lowest mean cost is defined to be the optimal intervention. As above, the expected cost surface associated with a given set of parameter values (as obtained by MCMC, described below), can be explored using standard Monte Carlo methods. At each time step, MCMC is used to produce 10,000 samples from the current posterior distribution on model parameters. These samples are thinned to 100 samples, and for each of these 100 samples the optimal variable stop time vaccination strategy is calculated as described above. The adaptive strategy to be implemented at that

Bench-to-bedside review: Angiopoietin signalling in critical illness – a future target?

Multiple organ dysfunction syndrome (MODS) occurs in response to major insults such as sepsis, severe haemorrhage, trauma, major surgery and pancreatitis. The mortality rate is high despite intensive supportive care. The pathophysiological mechanism underlying MODS are not entirely clear, although several have been proposed. Overwhelming inflammation, immunoparesis, occult oxygen debt and other mechanisms have been investigated, and – despite many unanswered questions – therapies targeting these mechanisms have been developed. Unfortunately, only a few interventions, usually those targeting multiple mechanisms at the same time, have appeared to be beneficial. We clearly need to understand better the mechanisms that underlie MODS. The endothelium certainly plays an active role in MODS. It functions at the intersection of several systems, including inflammation, coagulation, haemodynamics, fluid and electrolyte balance, and cell migration. An important regulator of these systems is the angiopoietin/Tie2 signalling system. In this review we describe this signalling system, giving special attention to what is known about it in critically ill patients and its potential as a target for therapy.

Critical illness is a life-threatening disease by definition. Patients treated for critical illness in the intensive care unit have underlying causes such as infection, trauma, major surgery, hemorrhagic shock, pancreatitis and other major insults. Despite maximal supportive care, severely ill patients treated in intensive care units are still likely to die, usually after an episode of increasing failure of multiple organs [1] . The mechanisms that underlie multiple organ dysfunction syndrome (MODS) are not known [2] , although several have been proposed, including overwhelming infection or immune response, immune paralysis, occult oxygen debt and mitochondrial dysfunction [3] [4] [5] . Although these potential mechanisms have features in common, it is not clear whether MODS is a final common pathway or when it is engaged. The innate and adaptive immune systems, coagulation, and hormonal and neuronal signalling are undoubtedly involved and are all connected. For example, the hypoxic response is linked to innate immunity and inflammation by the transcription factor nuclear factor-κB (NF-κB) [6] . It is no coincidence that the few interventions that appear to be of benefit, although this is still under debate, have pleiotropic mechanisms of action [7] [8] [9] . Thus, it seems reasonable to study the intersections between and within cellular and molecular systems to elucidate the interactions and to develop therapeutic options. One of the central cellular players in this system is the endothelial cell (EC). Once thought to serve as an inert vascular lining, ECs are highly heterogeneous and constitute an active disseminated organ throughout the circulatory system. ECs form the border between every organ and the bloodstream and thus with the rest of the body. The EC receives and gives signals, stores active substances of multiple systems, and regulates the passage of fluids, electrolytes, proteins and cells. The EC has a time and place dependent phenotype that is dynamically controlled, and its reactions to stimuli are specific to organ and vascular bed [10] [11] [12] [13] . The EC merits robust investigation in critical illness, as in vascular medicine [14] . genesis and organogenesis in normal physiology and in wound repair, but it is considered pathologic in tumour growth and diabetes [15] . Second, in the adult organism, ECs help to maintain homeostasis, including fluid, electrolyte and protein transport, and cell migration into and out of the vessel, and to regulate blood flow. Third, ECs react and respond to disturbances of homeostasis (for example, in inflammation, coagulation and hypoxia/reperfusion). All three functions are involved in MODS, in which ECs are shed, blood flow regulation is hampered, vessels become leaky, cells migrate out of the vessel and into the surrounding tissue, and coagulation and inflammation pathways are activated [16] . The machinery involved -receptors, signalling pathways and effectors -is largely the same in each function, but the net effect is determined by the balance between the parts of the machinery and the context [15] . The angiopoietin/Tie2 signalling system (Ang/Tie system) appears to be crucial in all three functions [17, 18] . The Ang/Tie system, which was discovered after vascular endothelial growth factor (VEGF) and its receptors, is mainly restricted to EC regulation and is the focus of this review. Accumulating evidence suggests that this system is nonredundant and is involved in multiple MODS-related pathways. All components of potential pathophysiological mechanisms in MODS should be viewed within their own context, because all systems are mutually dependent. Thus, examination of the Ang/Tie system might offer insight into the mechanisms underlying MODS and provide opportunities for therapeutic intervention. The notion that the Ang/Tie system contributes to disease pathogenesis is supported by clinical studies and studies in animal models, and by the relation between symptoms of critical illness and disturbances in this system. In mice, Ang-2 over-expression in glomeruli causes proteinuria and apoptosis of glomerular ECs [19] . In a rat model of glomerulonephritis, Tie2 is over-expressed by ECs, and Ang-1 and Ang-2 are over-expressed by podocytes in a time-dependent manner during the repair phase [20] . Therefore, Ang/Tie might be involved in renal failure and repair. Lung dysfunction is common in critical illness, and evidence of Ang/Tie involvement has been found in animal models. In a rat model of acute respiratory distress syndrome, Ang-1 reduces permeability and inflammation, whereas Tie2 deficiency increases damage [21] . In an experimental model of asthma, Ang-1 mRNA was decreased, and Ang-1 supplementation decreased alveolar leakage and NF-κB-dependent inflammation [22] . In hypoxia-induced pulmonary hypertension in rats, decreased activity of the Tie2 pathway contributed to right ventricular load, and this effect was antagonized by Ang-1 [23]. On the other hand, a causative role for Ang-1 in pulmonary hypertension has also been suggested [24] . In hyperoxic lung injury, Ang-2 is involved in lung permeability and inflammation [25] . Ang/Tie also may contribute to critical illness in patients with pulmonary conditions. Ang-1 and Ang-2 concentrations in sputum from asthma patients correlated with airway microvascular permeability [26] . In patients with exudative pleural effusion, the Ang-2 level was increased whereas Ang-1 was unchanged [27] . Ang-2 levels are associated with pulmonary vascular leakage and the severity of acute lung injury. Plasma from patients with acute lung injury and high Ang-2 concentrations disrupts junctional architecture in vitro in human microvascular ECs [28, 29] . Patients with cardiovascular disorders also exhibit changes in the Ang/Tie system. Circulating Ang-1 concentrations are stable in patients with atrial fibrillation, but Ang-2 concentrations are increased, along with markers of platelet activation, angiogenesis and inflammation [30] . Patients with hypertension resulting in end-organ damage have increased levels of circulating Ang-1, Ang-2, Tie2 and VEGF [31] . Congestive heart failure is associated with elevated plasma levels of Ang-2, Tie2 and VEGF, but normal levels of Ang-1 [32]. A similar pattern is seen in acute coronary syndrome [33] . Circulating levels of components of the Ang/Tie system have been measured in patients admitted to the critical care unit. In trauma patients plasma Ang-2, but not plasma Ang-1 or VEGF, was increased early after trauma, and the level correlated with disease severity and outcome [34] . In children with sepsis and septic shock, Ang-2 levels in plasma were increased and once again correlated with disease severity, whereas Ang-1 levels were decreased [35] . The same Ang-1/ Ang-2 pattern is seen in adults with sepsis [28, 29, [36] [37] [38] [39] . The results of studies of the Ang/Tie system in humans are summarized in Table 1 . In sepsis, VEGF and its soluble receptor sFLT-1 (soluble VEGFR-1) are also increased in a disease severity-dependent manner [40] [41] [42] .The picture that emerges from these studies is that the Ang/Tie signalling system appears to play a crucial role in the symptoms of MODS. Findings in animal models and in patients suggest that Ang-1 stabilizes ECs and Ang-2 prepares them for action. The close relation with VEGF is also apparent. four (Ang-1) and two (Ang-2) subunits [48, 49] . The dissimilarity between Ang-1 and Ang-2 signalling lies in subtle differences in the receptor binding domain that lead to distinct intracellular actions of the receptor; differential cellular handling of both receptor and ligands after binding and signalling initiation may also play a role [49, 50] . The receptors are Tie1 and Tie2 [51] . Tie2 is a 140-kDa tyrosine kinase receptor with homology to immunoglobulin and epidermal growth factor [47, 52] . Tie receptors have an amino-terminal ligand binding domain, a single transmembrane domain and an intracellular tyrosine kinase domain [51] . Ligand binding to the extracellular domain of Tie2 results in receptor dimerization, autophosphorylation and docking of adaptors, and coupling to intracellular signalling pathways [47, [53] [54] [55] . Tie2 is shed from the EC and can be detected in soluble form in normal human serum and plasma; soluble Tie2 may be involved in ligand scavenging without signalling [56] . Tie2 shedding is both constitutive and induced; the latter can be controlled by VEGF via a pathway that is dependent on phosphoinositide-3 kinase (PI3K) and Akt [57] . Shed soluble Tie2 can scavenge Ang-1 and Ang-2 [56] . Tie1 does not act as a transmembrane kinase; rather, it regulates the binding of ligands to Tie2 and modulates its signalling [58] [59] [60] . Ang-1 is produced by pericytes and smooth muscle cells ( Figure 1 ). In the glomerulus, which lacks pericytes, Ang-1 is produced by podocytes [61] . Ang-1 has a high affinity for the extracellular matrix, and so circulating levels do not reflect tissue levels, which in part is probably responsible for the constitutive phosphorylation of Tie2 in quiescent endothelium [62] [63] [64] [65] . Ang-2 is produced in ECs and stored in Weibel-Palade bodies (WPBs) [66, 67] . The release of Ang-2 from WPBs by exocytosis can be regulated independently of the release of other stored proteins [68] . Tie2 is expressed predominantly by ECs, although some subsets of macrophages and multiple other cell types express Tie2 at low levels [69, 70] . In ECs, Tie2 is most abundant in the endothelial caveolae [71] . The Ang-1 and Ang-2 genes are located on chromosome 8. Functional polymorphisms have not been identified in the Ang-1 gene, but three have been identified in the coding region of Ang-2 [72] . In ECs under stress, Ang-2 mRNA expression is induced by VEGF, fibroblast growth factor 2 and hypoxia [44, 73] . Upregulation of Ang-2 induced by VEGF and hypoxia can be abolished by inhibiting tyrosine kinase or mitogen-activated protein kinase [73] . Ang-2 mRNA expression can be downregulated by Ang-1, Ang-2, or transforming growth factor [74] . After inhibition of PI3K by wortmannin, Ang-2 mRNA production is induced by the transcription factor FOXO1 (forkhead box O1) [75] . EC-specific Ang-2 promoter activity is regulated by Ets-1 and the Ets family member Elf-1 [76, 77] . Because Tie2 signalling is required under circumstances that usually hamper cell metabolism, its promoter contains repeats that ensure transcription under difficult circumstances, including hypoxia [78] . Tie2 is present in phosphorylated form in quiescent and activated ECs throughout the body [62] . Signalling is initiated by autophosphorylation of Tie2 after Ang-1 binding and is conducted by several distinct pathways [54, 71, 79, 80] . Tie2 can also be activated at cell-cell contacts when Ang-1 induces Tie2/Tie2 homotypic intercellular bridges [65] . In human umbilical vein endothelial cells (HUVECs), Ang/Tie signalling resulted in 86 upregulated genes and 49 downregulated genes [81, 82] . Akt phosphorylation by PI3K with interaction of nitric oxide is the most important intracellular pathway [51, [83] [84] [85] [86] ; however, ERK1/2, p38MAPK, and SAPK/JNK can also participate in Ang/Tie downstream signalling [71, 81, 84, [87] [88] [89] [90] . Endothelial barrier control by Ang-1 requires p190RhoGAP, a GTPase regulator that can modify the cytoskeleton [80] . The transcription factors FOXO1, activator protein-1, and NF-κ B are involved in Ang/Tie-regulated gene transcription [75, [91] [92] [93] . Ang-1induced signalling is has also been implicated in cell migration induced by reactive oxygen species [94] . ABIN-2 (A20-binding inhibitor of NF-κB 2), an inhibitor of NF-κB, is involved in Ang-1-regulated inhibition of endothelial apoptosis and inflammation in HUVECs [93] . However, the downstream signalling of Tie2 varies depending on cell type and localization and whether a cell-cell or cell-matrix interaction in involved, which results in spatiotemporally different patterns of gene expression. For example, Ang-1/Tie2 signalling leads to Akt activation within the context of cell-cell interaction, but it leads to ERK activation in the context of cell-matrix interaction. The microenvironment of the receptor in the cell membrane plays a central role in this signal differentiation. Adaptor molecules such as DOK and SHP2 and the availability of substrate determine which protein is phosphorylated [95] . After binding of Ang-1, and to a lesser extent Ang-2, Tie2 is internalized and degraded, and Ang-1 is shed in a reusable form [50] . VEGF is an important co-factor that can exert different effects on Ang-1 and Ang-2 signalling [88] . Ang-2 is antiapoptotic in the presence of VEGF but induces EC apoptosis in its absence [96] . Autophosphorylation and subsequent signalling are inhibited by heteropolymerization of Tie1 and Tie2 [59] . Although the Ang/Tie system appears to play its role mainly in paracrine and autocrine processes, its circulating components have been found in plasma. The significance of this finding in health and disease has yet to be determined. The Ang/Tie system is an integrated, highly complex system of checks and balances ( Figure 1) [45,54]. The response of ECs to Ang-1 and Ang-2 depends on the location of the cells and the biological and biomechanical context [97, 98] . It is believed that PI3K/Akt is among the most important downstream signalling pathways and that VEGF is one of the most important modulators of effects. Below we describe in more detail how this system responds to changes in homeostatic balances under various conditions of damage and repair. Angiogenesis, inflammation and homeostasis are highly related, and the Ang/Tie system lies at the intersection of all three processes [99, 100] . The Ang/Tie system is critically important for angiogenesis during embryogenesis, but in healthy adults its function shifts toward maintenance of homeostasis and reaction to insults. Except for follicle formation, menstruation and pregnancy, angiogenesis in adults is disease related. Neoplasia-associated neoangiogenesis and neovascularization in diabetes and rheumatoid arthritis are unfavourable events, and improper angiogenesis is the subject of research in ischaemic disorders and atherosclerosis. Finally, failure to maintain homeostasis and an inappropriate reaction to injury are detrimental features in critical illness. A schematic model of the angiopoietin-Tie2 ligand-receptor system. Quiescent endothelial cells are attached to pericytes that constitutively produce Ang-1. As a vascular maintenance factor, Ang-1 reacts with the endothelial tyrosine kinase receptor Tie2. Ligand binding to the extracellular domain of Tie2 results in receptor dimerization, autophosphorylation, docking of adaptors and coupling to intracellular signalling pathways. Signal transduction by Tie2 activates the PI3K/Akt cell survival signalling pathway, thereby leading to vascular stabilization. Tie2 activation also inhibits the NF-κB-dependent expression of inflammatory genes, such as those encoding luminal adhesion molecules (for example, intercellular adhesion molecule-1, vascular cell adhesion molecule-1 and E-selectin). Ang-2 is stored and rapidly released from WPBs in an autocrine and paracrine fashion upon stimulation by various inflammatory agents. Ang-2 acts as an antagonist of Ang-1, stops Tie2 signalling, and sensitizes endothelium to inflammatory mediators (for example, tumour necrosis factor-α) or facilitates vascular endothelial growth factor-induced angiogenesis. Ang-2-mediated disruption of protective Ang-1/Tie2 signalling causes disassembly of cell-cell junctions via the Rho kinase pathway. In inflammation, this process causes capillary leakage and facilitates transmigration of leucocytes. In angiogenesis, loss of cell-cell contacts is a prerequisite for endothelial cell migration and new vessel formation. Ang, angiopoietin; NF-κB, nuclear factor-κB; PI3K, phosphoinositide-3 kinase; WPB, Weibel-Palade body. Angiogenesis is dependent on multiple growth factors and receptors and their signalling systems and transcriptional regulators [101] . The process is complex and encompasses the recruitment of mobile ECs and endothelial progenitor cells, the proliferation and apoptosis of these cells, and reorganization of the surroundings [102] . To form stable new blood vessels, the response must be coordinated in time and space, and the Ang/Tie system is involved from beginning to end. To prepare for angiogenesis, Ang-2 destabilizes quiescent endothelium through an internal autocrine loop mechanism [44, 103] . Before vascular sprouting starts, focal adhesion kinase and proteinases such as plasmin and metalloproteinases are excreted [85] . Often, this stage is preceded by activation of innate immunity and inflammation [104] . Apparently, the machinery to clean up after the work has been finished is installed before the work is commenced, again illustrating the close relations among the different processes [104] . Ang-1 maintains and, when required, restores the higher order architecture of growing blood vessels [43,44, 105, 106] . This is achieved by inhibiting apoptosis of ECs by Tie2mediated activation of PI3K/Akt signalling [107] [108] [109] . Ang-1/ Tie2 signalling is involved in angiogenesis induced by cyclic strain and hypoxia [110, 111] . Although its role is less clear, Tie1 might be involved in EC reactions to shear stress [112] . Ang-1 is a chemoattractant for ECs [83] [84] [85] , and both Ang-1 and Ang-2 have proliferative effects on those cells [98, 113] . At the end of a vascular remodelling phase, Ang-2 induces apoptosis of ECs for vessel regression in competition with the survival signal of Ang-1 [106] . This apoptotic process requires macrophages, which are recruited by Ang-2 [70, 114] . ECs require support from surrounding cells such as pericytes, podocytes, and smooth muscle cells [63] . These cells actively control vascular behaviour by producing signalling compounds (for instance, Ang-1 and VEGF) that govern the activity and response of ECs [61] . To attract ECs, Ang-1 secreted by support cells binds to the extracellular matrix. In quiescent ECs, this binding results in Tie2 movement to the site of cell-cell interaction. In mobile ECs, Ang-1 polarizes the cell with Tie2 movement abluminal site [65] . In tumour angiogenesis and in inflammation, Ang-2 recruits Tie2positive monocytes and causes them to release cytokines and adopt a pro-angiogenic phenotype [111] . The Ang/Tie system provides vascular wall stability by inducing EC survival and vascular integrity. However, this stability can be disrupted by Ang-2 injection, which in healthy mice causes oedema [28, 79, 115, 116] that can be blocked by systemic administration of soluble Tie2 [115] . Ang-2 can impair homeostatic capacity by disrupting cell-cell adhesion through E-cadherin discharge and EC contraction [28, 117] . In contrast, through effects on intracellular signalling, the cytoskeleton and junction-related molecules, Ang-1 reduces leakage from inflamed venules by restricting the number and size of gaps that form at endothelial cell junctions [80, 118, 119] . Ang-1 also suppresses expression of tissue factor induced by VEGF and tumour necrosis factor (TNF)-α, as well as expression of vascular cell adhesion molecule-1, intercellular adhesion molecule-1 and E-selectin. As a result, endothelial inflammation is suppressed [120] [121] [122] [123] . In primary human glomerular ECs in vitro, Ang-1 stabilizes the endothelium by inhibiting angiogenesis, and VEGF increases water permeability [124] . Similar observations were made in bovine lung ECs and immortalized HUVECs, in which Ang-1 decreased permeability, adherence of polymorphonuclear leucocytes and interleukin-8 production [123] . Reaction to injury can be seen as an attempt to maintain homeostasis under exceptional conditions. ECs can be affected by several noxious mechanisms. The Ang/Tie system is considered crucial in fine-tuning their reaction to injury and in containing that reaction. Ang-2-deficient mice cannot mount an inflammatory response to peritonitis induced chemically or with Staphylococcus aureus [125] , but they can mount a response to pneumonia, suggesting the existence of inflammatory reactions for which Ang-2 is not mandatory. Ang-2 sensitizes ECs to activation by inflammatory cytokines. In Ang-2-deficient mice, leucocytes do roll on activated endothelium but they are not firmly attached, owing to the lack of Ang-2-dependent upregulation of adhesion molecules and the dominance of Ang-1-regulated suppression of adhesion molecules [120] [121] [122] [123] 125] . In bovine retinal pericytes, hypoxia and VEGF induce Ang-1 and Tie2 gene expression acutely without altering Ang-2 mRNA levels. The opposite occurs in bovine aortic ECs and microvascular ECs, underscoring the heterogeneity of ECs from different microvascular beds [73, 126, 127] . Lipopolysaccharide (LPS) and pro-inflammatory cytokines can shift the Ang/Tie balance, rouse ECs from quiescence and provoke an inflammatory response. In rodents LPS injection induces expression of Ang-2 mRNA and protein and reduces the levels of Ang-1, Tie2 and Tie2 phosphorylation in lung, liver and diaphragm within 24 hours, which may promote or maintain vascular leakage. The initial increase in permeability is probably due to release of Ang-2 stored in WPBs [39, 128] . In a mouse model of LPS-induced lung injury, pulmonary oedema was found to be related to the balance between VEGF, Ang-1 and Ang-4 [129] . In a comparable model, Ang-1-producing transfected cells reduced alveolar inflammation and leakage [130] . In choroidal ECs, TNF induces Ang-2 mRNA and protein before affecting Ang-1 and VEGF levels [131] . In HUVECs, TNF-induced upregulation of Ang-2 is mediated by the NF-κB pathway [132] , and TNF-induced Tie2 expression can be attenuated by both Ang-1 and Ang-2. Without TNF stimulation, only Ang-1 can reduce Tie2 expression [133] . Ang-2 sensitizes ECs to TNF, resulting in enhanced expression of intercellular adhesion molecule-1, vascular cell adhesion molecule-1 and E-selectin [74, 125, 134] . By inhibiting those endothelial adhesion molecules, Ang-1 decreases leucocyte adhesion [122] . Angiopoietins can mediate the synthesis of platelet-activating factor by ECs to stimulate inflammation [90] . Moreover, both Ang-1 and Ang-2 can translocate P-selectin from WPBs to the surface of the EC [135] , and both can also increase neutrophil adhesion and chemotaxis and enhance those processes when they are induced by interleukin-8 [86, 136, 137] . In a rat model of haemorrhagic shock, Ang-1 reduced vascular leakage, and it inhibited microvascular endothelial cell apoptosis in vitro and in vivo [107, 138] . In this model, Ang-1promoted cell survival was partly controlled through integrin adhesion [139] . It has been suggested that EC apoptosis in haemorrhagic shock contributes to endothelial hyperpermeability [140] [141] [142] . Apoptosis is one of the reactions to MODSrelated injury as demonstrated in hypoxia/reperfusion [143] . Ang-1 and Ang-2 are involved in cell-cell and cell-matrix binding [139, [144] [145] [146] . Endothelial permeability is greatly dependent on cell-cell adhesion. The major adherens junction is largely composed of vascular-endothelial cadherin. This complex can be disrupted by VEGF, leading to increased vascular permeability [147, 148] , which can be antagonized by Ang-1 [149, 150] . ECs can also bind to the matrix through the binding of Ang-1 to integrins, which can mediate some of the effects of Ang-1 without Tie2 phosphorylation [146, 151] . At low Ang-1 concentrations, integrin and Tie2 can cooperate to stabilize ECs [151] . Ang-2 might play a role in inflammatory diseases such as vasculitis by disrupting the cell-cell junction and inducing denudation of the basal membrane [152] . Ang-1 can mediate the translocation of Tie2 to endothelial cell-cell contacts and induce Tie2-Tie2 bridges with signal pathway activation, leading to diminished paracellular permeability [65] . In the mature vessel, Ang-1 acts as a paracrine signal to maintain a quiescent status quo, whereas Ang-2 induces or facilitates an autocrine EC response [74, 153] . In general, Ang-1 can be viewed as a stabilizing messenger, causing continuous Tie2 phosphorylation, and Ang-2 as a destabilizing messenger preparing for action [17] . Attempts to unravel the exact molecular mechanisms that control the system are complicated by microenvironment-dependent endothelial phenotypes and reactivity and by flow typedependent reactions to dynamic changes [13, 154, 155] . Hence, the EC must be viewed in the context of its surroundings -the pericyte at the abluminal site, and the blood and its constituents on the luminal site [64] . The Ang/Tie system certainly functions as one of the junctions in signal transduction and plays a key role in multiple cellular processes, many of which have been linked to MODS. A therapy should intervene in the right place and at the right time, with the proper duration of action and without collateral damage [156, 157] . The Ang/Tie system is involved in many processes and lies at the intersection of molecular mechanisms of disease. Thus, interventions targeting this system might have benefits. As in other pleiotropic systems, however, unexpected and unwanted side effects are a serious risk. The absence of redundant systems to take over the function of Ang/Tie2 has the advantage that the effect of therapeutic intervention cannot easily be bypassed by the cell. On the other hand, because the cell has no escape, the effect may become uncontrolled and irreversible. Moreover, the exact function of the Ang/Tie system in the pathological cascade is not fully established. What we see in animal models and in patients is most probably the systemic reflection of a local process. We do not know whether this systemic reflection is just a marker of organ injury or even a mediator of distant organ involvement. Of the three main functions of the Ang/Tie system, it is mainly angiogenesis that has been evaluated as a therapeutic target. So far, the focus of Ang/Tie modulation has been on inhibiting angiogenesis related to malignant and ophthalmological diseases and to complications of diabetes [158, 159] . In peripheral arterial occlusive disease, stimulation of angiogenesis seems a logical strategy to attenuate the consequences of ongoing tissue ischaemia. In a rat model of hind limb ischaemia, combined delivery of Ang-1 and VEGF genes stimulated collateral vessel development to the greatest extent [160, 161] . Thus far, therapy directed at VEGF has reached the clinic, but not therapy directed at Ang/Tie [162] . Targeting homeostasis and repair/inflammation in critically ill patients is an attractive option and has already led to the development of new drugs [45, 158, 163] . From current knowledge, one can speculate about the best options for therapy aimed at the Ang/Tie system. In critical illness, Ang-1 is considered to be the 'good guy' because it can create vascular stability and thus its activity should be supported. In contrast, Ang-2 appears to be a 'bad guy' that induces vascular leakage, so its activity should be inhibited [164] . Production of recombinant Ang-1 is technically challenging as Ang-1 is 'sticky' because of its high affinity for the extracellular matrix [165] . However, stable Ang-1 variants with improved receptor affinity have been engineered. A stable soluble Ang-1 variant has anti-permeability activity [165] . When injected intraperitoneally in mice, human recombinant Ang-1 can prevent LPS-induced lung hyperpermeability [80] . In diabetic mice, a stable Ang-1 derivative attenuated proteinuria and delayed renal failure [166] , and manipulating the Ang-1/Ang-2 ratio changed infarct size [167] . A more profound Ang-1 effect can be achieved by locally stimulating Ang-1 production. In experimental acute respiratory distress syndrome, transfected cells expressing Ang-1 reduced alveolar inflammation and leakage [130] . An adenovirus construct encoding Ang-1 protected mice from death in an LPS model, and Ang-1 gene therapy reduced acute lung injury in a rat model [21, 168, 169] . In hypertensive rats, a plasmid expressing a stable Ang-1 protein reduced blood pressure and end-organ damage [170] . If used in a disease with a limited duration, as critical illness should be, virus/plasmid-driven production of Ang-1 could easily be shut down when it is no longer needed. Manipulating Ang-2 activity is also difficult. Ang-2 stored in WPBs is rapidly released and must be captured immediately to prevent autocrine/paracrine disruption of protective Ang-1/ Tie signalling. Soluble Tie2 or Ang-2 inhibitors should be effective [26, 171] . Neutralizing antibodies against Ang-2 might also be an option. Replenishment of Ang-2 stores could be abolished by small interfering RNA techniques or spiegelmer/aptamer approaches [25, 172, 173] . However, no bad guy is all bad, and no good guy is all good. For example, Ang-1 has been linked to the development of pulmonary hypertension [174] . Also, under certain circumstances Ang-2 can act as a Tie2 agonist and exert effects similar to those of Ang-1 -an unexplained finding that illustrates our limited understanding of the Ang/Tie system [75] . Complete blockade of Ang-2 might also hamper innate immunity and revascularization. Finding the right balance and timing will be the major challenge when developing therapies to target the Ang/Tie system. In the meantime, we might have already used Ang/Tie-directed therapy with the most pleiotropic of all drugs -corticosteroids. In the airways, steroids suppressed Ang-2 and increased Ang-1 expression [26, 171, 175] . Interventions further downstream targeting specific adaptor molecules, signalling pathways, or transcription factors have yet to be explored. In patients with malignant disease, the Ang/Tie system might serve as a tumour or response marker. In patients with multiple myeloma, normalization of the Ang-1/Ang-2 ratio reflects a response to treatment with anti-angiogenesis medication [176] . In patients with non-small-cell lung cancer, Ang-2 is increased in serum and indicates tumour progression [177] . After allogeneic stem cell transplantation in patients with high-risk myeloid malignancies, the serum Ang-2 concentration predicts disease-free survival [178] , possibly reflecting a relation between cancer-driven angiogenesis and Ang-2 serum level. In nonmalignant disease, the levels of Ang/Tie system components correlate with disease severity [28, 29, [34] [35] [36] [37] 39 ]. However, current data are insufficient to justify the use of serum soluble Tie2/Ang levels for diagnostic and prognostic purposes. In critical illness, assessment of the Ang/Tie system in patients with different severities of disease and with involvement of different organ systems might help to define our patient population and allow us to rethink our concepts of MODS. In this way, such work may lead to enhanced diagnosis and prognostication in the future [2] . Accumulating evidence from animal and human studies points to the involvement of the Ang/Tie system in vascular barrier dysfunction during critical illness. Many processes in injury and in repair act through this nonredundant system. Thus far, only preliminary studies in critically ill patients have been reported. Methods to manipulate this system are available but have not been tested in such patients. The response to treatment is difficult to predict because of the pleiotropic functions of the Ang/Tie system, because the balance among its components appears to be more important than the absolute levels, and because the sensitivity of the endothelium to disease-related stimuli varies, depending on the environment and the organ involved. To avoid disappointment, further experimental and translational research must be carried out, and Ang/Tie modulation must not be introduced into the clinic prematurely. Implementing the results of this research in critical care represents an opportunity to show what we have learned [2] . Ang/Tie signalling is a very promising target and must not be allowed to become lost in translation [179] .

Mapping the sequence mutations of the 2009 H1N1 influenza A virus neuraminidase relative to drug and antibody binding sites

In this work, we study the consequences of sequence variations of the "2009 H1N1" (swine or Mexican flu) influenza A virus strain neuraminidase for drug treatment and vaccination. We find that it is phylogenetically more closely related to European H1N1 swine flu and H5N1 avian flu rather than to the H1N1 counterparts in the Americas. Homology-based 3D structure modeling reveals that the novel mutations are preferentially located at the protein surface and do not interfere with the active site. The latter is the binding cavity for 3 currently used neuraminidase inhibitors: oseltamivir (Tamiflu(®)), zanamivir (Relenza(®)) and peramivir; thus, the drugs should remain effective for treatment. However, the antigenic regions of the neuraminidase relevant for vaccine development, serological typing and passive antibody treatment can differ from those of previous strains and already vary among patients. This article was reviewed by Sandor Pongor and L. Aravind.

The recent epidemic of the "2009 H1N1" influenza A virus (also called swine or Mexican flu) has put the world on alert since a new swine flu strain (naturally hosted by pigs) has crossed the species barrier to human and, apparently, acquired the capability for human to human transmission [1, 2] . Given earlier experiences with risks of viral pandemics such as SARS and the avian flu [3] , global control and public health surveillance mechanisms provided sequences of the new flu strain in public sequence databases within weeks of the outbreak. Here, we analyze the protein sequence of its neuraminidase with respect to sim-ilarities and differences to known strains and implications on drug treatment and vaccination. Sequence and residue numbering in this analysis correspond to the neuraminidase [Genbank: ACP41107.1 http://www.ncbi.nlm.nih.gov/protein/227809834] representative for the new strain. Sequence analysis was carried out following an established protocol using the ANNIE resource [4, 5] . The 469 amino acid long neuraminidase (NA) protein ( Figure 1 ) is essential for release of the viral particle from the outer membrane of infected cells by cleaving sialic acid from host glycoproteins that are recognized by the viral hemagglutinin [6] . As a type II transmembrane protein, it is N-terminally attached to the membrane [7] . It consists of a tiny cytoplasmic tail at the N-terminus (residues 1 to 6) [8] followed by the transmembrane region (residues 7 to 34) that is also responsible for translocation of the protein [9] . Next, a presumably unstructured linker region (residues 35 to 82) connects the membrane anchor to the catalytic neuraminidase domain (residues 83 to 469; Figure 1 ). Such unstructured linker regions are rich in small and polar residues and often harbour sites for posttranslational modifications [10, 11] . Probable posttranslational modification sites in the neuraminidase of the new strain are glycosylation motifs involving N88, N146 and N235, which correspond to residues that are also glycosylated in other subtype neuraminidases [12] . However, the minimal and non-specific consensus motif of glycosylation sites (Nx [ST]) is found in total 8 times in the new strain sequence with an apparent clustering (50%) in the unstructured linker region ( Figure 1 ). Interestingly, another putative novel glycosylation site N386, which is unique to the new strain, would be accessible on the surface, as seen in the structural models. Comparing among all strains, the sequence variation is largest in the linker region, including large deleted segments. Nevertheless, this region harbours a cysteine (Figure 2 ) that can be aligned over multiple NA subtypes and is conserved in N1-N5 and N8, but not in N6, N7 and N9. Earlier reports assume that, at least in related viruses, cysteines in the non-globular region could be involved in intermolecular disulfide bridges [13] [14] [15] . Alternatively, by analogy to other influenza proteins such as hemagglutinin [16] and M2 protein [17] , it cannot yet be excluded that cysteine C49 is palmitoylated and that the anchor localizes the protein to lipid rafts [18] . Influenza A virus protein sequences were downloaded from NCBI (as of April 29 th ). Neuraminidases were identified by BLAST (E-value < 0.001) [19] using the representative NA of the new strain as query [Genbank: ACP41107.1 http://www.ncbi.nlm.nih.gov/protein/ 227809834]. Redundancy was removed with cd-hit at a level of maximal 90% sequence identity [20] , the remain-Domain architecture (drawn with http://au.expasy.org/tools/mydomains/) Figure 1 Domain architecture (drawn with http://au.expasy.org/tools/mydomains/). Besides the labelled domains (TM ... transmembrane), grey lollipops indicate known and putative glycosylation sites and the red lollipop marks the conserved cysteine shown in Figure 2 . Representative alignment of the sequence environment of the conserved cysteine C49 that could either serve for intermolecu-lar disulfide bridges or as palmitoylation site Figure 2 Representative alignment of the sequence environment of the conserved cysteine C49 that could either serve for intermolecular disulfide bridges or as palmitoylation site. ing sequences were aligned with MAFFT (using L-INS-I settings [21] ) and the resulting multiple alignment was visualized and annotated in Jalview [22] . A neighbour joining tree with pairwise gap deletion, Poisson correction as distance measure and 500 bootstrap replicates (generated with MEGA [23] ) produces robust groupings consistent with previous studies [24] for the known NA subtypes (clustering of N1, N4, N5+N8 on one side and N2, N3, N6+N7+N9 on the other) and reliably places the new NA with other N1s. Interestingly, inside the N1 cluster, the new NA appeared close to the N1 of H5N1 avian flu viruses. The alignment and corresponding phylogenetic tree are available at http://mendel.bii.a-star.edu.sg/ SEQUENCES/H1N1/. Hence, we repeated the analysis (same protocol as outlined above) for a detailed mapping of only the N1 subtype family with the difference of allowing 95% sequence identity for sequences before 2009 but keeping all new NA sequences (as of April 29 th ). A characteristic clustering emerges ( Figure 3 ) that roughly corresponds to host and geographic distributions, consistent with previous reports [25, 26] . The observed clustering is robust in respect to the method used for tree generation (same for maximum parsimony or neighbour joining trees with JTT distance and gamma-distributed variable rates). The 2009 NA is part of a cluster of avian-like swine flu H1N1 strains predominantly found in European pigs. However, previous examples of human infections from swine flu are also part of the same cluster, for example from 2005 in Thailand [27] . This indicates that, similar to the current outbreak, closely related H1N1 strains have crossed species boundaries on previous occasions as also evidenced by further reports in the literature [28] [29] [30] . Moreover, neuroaminidases of these new H1N1 swine flu examples are more similar to H5N1 avian flu strains than other H1N1 variants found in the Americas or than that of the historic strains such as the 1918 Spanish flu [31] . This is surprising since avian flu strains typically have different hemaglutinin (HA) subtypes (e.g. H5N1). Combinations of HA and NA subtypes need to be fine-tuned to recognize the same type of sialic acid modifications to allow smooth interplay of the two proteins, which is important for the viral cycle [6] . These results support the notion that, also inside the family of N1 subtypes, a clear distinction can be made between avian-like H1N1, such as the one from the current outbreak, and other existing H1N1 strains. The crystal structures of both the historic 1918 NA as well as the avian flu NA are available in complex with currently used drugs. We created a homology model of the new 2009 swine flu NA to map the sequence differences to the three-dimensional structure templates. Using Modeller [32] , the sequence of the new neuraminidase [Genbank Accession: ACP41107.1 http://www.ncbi.nlm.nih.gov/ protein/227809834] was modelled 50 times onto multiple templates (PDB: 2hu4 [33] , 3ckz [34] , 3b7e [35] , 3beq [35] ) and the resulting best model (as judged by DOPE score) further refined with short simulated annealing MD simulations in the presence of a bound inhibitor (zanamivir, oseltamivir or peramivir) as implemented in the Yasara Structure package [36] . The final atom-resolution models are available in PDB format at http://men del.bii.a-star.edu.sg/SEQUENCES/H1N1/. We mapped the level of residue conservation (calculated with the evolutionary trace algorithm [37] ) from the multiple alignment of all NA subtypes to its corresponding position in the structure. The results show the strict conservation close to the neuraminidase catalytic site, which also serves as the drug binding pocket ( Figure 4A ). The remaining conserved patches (for example the sites around N104 or below N146) fit into each other and form the dimerization/tetramerization interfaces [35] . A model of the dimeric version is available in PDB format at http:/ /mendel.bii.a-star.edu.sg/SEQUENCES/H1N1/. Next, we compared the sequences of the new strain with the related H5N1 from avian flu and H1N1 from the Spanish flu ( Figure 5 ). Among 387 residues that were structurally modelled, the "2009 H1N1" neuraminidase differs from the other two in 21 positions. The mapping to the structure ( Figure 4B ) shows that the novel sequence mutations are distributed all around the surface of the molecule leaving the hydrophobic core, but also the catalytic site, essentially untouched. Importantly, none of the new mutations appears sufficiently close to affect the drug binding pocket. For example, all 17 residues within 3 Å of the zanamivir molecule bound to the active site are fully conserved among all three strains. The closest mutation is the conservative V149I substitution at a distance of ~10 Å to zanamivir and ~7 Å to oseltavimir. It has to be noted that indirect effects of the mutations that may alter the binding pocket also from a greater distance are difficult to assess and cannot be excluded. To this extent, we have analysed coevolution patterns in an extensive alignment of more than 6000 non-identical Influenza A neuraminidases to eventually identify connected networks of residues using the SCA algorithm [38] as implemented in [39] , but no network that would connect the surface directly to the core and catalytic site was found (see supplementary material). In fact, all positions of the observed mutations are at the surface and naturally variable, as judged by the conservation and SCA analysis, which would rather indicate that they do not have an effect on the structure of the more distant binding pocket. Thus, we conclude that the drug binding pocket remains unchanged in the new strain and, hence, the binding Phylogenetic tree of neuraminidase protein sequences of the N1 subtype family Figure 3 Phylogenetic tree of neuraminidase protein sequences of the N1 subtype family. behaviour of neuraminidase inhibitors such as oseltamivir (Tamiflu ® ) and zanamivir (Relenza ® ) should be unaffected. Indeed, initial clinical reports suggest that the new virus is susceptible to the two drugs [40] . Our findings support this notion and provide a molecular mechanism. Furthermore, the third currently tested neuraminidase inhibitor, peramivir, should also be effective since it also shares the same binding pocket. Next, we review how the new mutations affect vaccine development through altering antibody interactions as well as antigenic regions. There are 3 crystal structures of related neuraminidases in complex with antibodies [41] [42] [43] . In Figure 5 , we annotate residues that are within 3 Å distance to the respective bound antibody and, hence, crucial for the interaction. Interestingly, residues in sites recognized by both NC41 and NC10 antibodies appear mutated in the new strain. This would suggest that these old antibodies (that were originally directed against N9 neuraminidase) would probably not bind to NA of the new strain. Nevertheless, using the same regions as epitope may be a viable option for novel vaccine development. Additionally, several other known antigenic regions, partially derived from surviving patients of previous flu outbreaks (e.g. H5N1), are reported in the literature [44] [45] [46] and their location is indicated in the alignment ( Figure 5 ). Another extensive source of epitopes and antigens includ-ing neuraminidase of influenza A viruses is the immune epitope database (IEDB) [47] and the complete mapping of epitopes for the new H1N1 NA sequence is available at http://mendel.bii.a-star.edu.sg/SEQUENCES/H1N1/. While several of the new mutations are found in antigenic regions, it is also apparent that they often occur on positions that are hardly conserved among different NA subtypes. Consequently, these regions are evolutionarily more flexible and may mutate fast. This increases the risk of evading antibody responses of human hosts acquired during previous flu infection or from vaccination. After the first wave of new patient sequences arrived, it becomes clear that there are at least two major lineages that are distinguishable by only few mutations. Most notably, N248 has mutated to Aspartate (D248) in the New York infection cluster. All intra-strain mutations available before May 8 th 2009 are indicated in Figures 4B and 5. As expected, they are predominantly found on the surface and in regions that are known to be variable from the conservation analysis. While the drug binding pocket remains unaffected by these most recent mutations, the N248D substitution changes a central part of an antibody recognition site. This has important consequences for vaccine development, forcing to either avoid this epitope or produce combined vaccines to account for the epitope in the first annotation line) is displayed as the respective mutated residue in capital letters if found in multiple patients (e.g. D for the N248D substitution) or lower-case (e.g. "i" for V241I) for single occurrences. In the second annotation row, antigenic regions are labelled as "*". Residues with < 3 Å contact to antibodies are labelled "A" for interactions derived from PDB:1ncb, "B" from both PDB:1ncb and PDB:1nmb, "C" from PDB:1nmb and "D" from PDB:2aep. variation observed in different patient groups. Although the mutation pattern will become less transparent over time it may still serve to delineate chains of transmission, retrospectively. In summary, we provide a sequence analysis and structural modelling of the neuraminidase from the 2009 H1N1 swine flu outbreak. Besides mapping of phylogenetic relationships to other strains, we find that the sequence variation in the new strain does not seem to affect the drug binding site but may very well alter common epitopes. To allow quick analysis of future mutations that could produce drug or vaccine-resistant strains, we provide a tool for 3D visualization of the neuraminidase structure models with mapping of drug and antibody recognition sites on the supplementary webpage http://men del.bii.a-star.edu.sg/SEQUENCES/H1N1/. NA: neuraminidase; HA: hemagglutinin. The authors declare that they have no competing interests. SMS did the alignments and phylogenetic trees. FLS contributed the domain architecture analysis and helped with the phylogenetic analysis. MJ did the structural models and conservation mapping. RLTC contributed the study on antigenic regions and the Jmol visualization on the webpage. SMS and FE wrote the manuscript; all authors approved the final version. Sandor Pongor, International Centre for Genetic Engineering and Biotechnology, Trieste, Italy The results show that this strain is phylogenetically more closely related to European H1N1 swine flu and H5N1 avian flu rather than to the H1N1 counterparts in the Americas. Homology-modeling of the neuraminidase reveals that the novel mutations are not likely to interfere with the active site so the currently used neuraminidase inhibitors (oseltamivir, zanamivir and peramivir) will be effective against the new virus strain. The subject is very timely and the approach is adequate. The authors may want to include analysis of more patient data that were published since the analysis was completed. More and more sequences are being published from all over the world that might be worthwhile to include into this analysis. The authors may consider establishing a periodically updated homepage, if appropriate. In summary, the analysis is careful and carried out in a commendable fashion, and the findings are highly significant. Indeed, there have been additional "2009 H1N1" neuraminidase sequences since the submission of this manuscript for review. Between April 29 th and May 8 th , 45 new sequences became available. Overall, 3 mutations occur in multiple (S95G, V106I and N248D) while 2 mutations are restricted to single patient virus isolates (V83M, V241I), so far. We have included a section about these mutations identified in new patient sequences and mapped them to the structure. As discussed in the main text, the intra-strain variation is typically found on the surface and at positions expected to be variable as judged by the conservation analysis among all NA subtypes. Interestingly, one of the mutations among patients (N248D) is critically affecting one of the antibody binding sites. As the virus will continue to evolve, new mutations will become available and the best way we have found to allow quick mapping and update of new sequence variation in respect to drug and antibody binding sites is to give full access to users/readers via a 3D structure visualization tool at the supplementary webpage http://mendel.bii.astar.edu.sg/SEQUENCES/H1N1/ (instructions to map new mutations are given, including an example). The key finding in the paper is that the binding cavity for neuraminidase inhibitors is unaffected as suggested by molecular modeling. This finding is of significance in the current situation of an outbreak with pandemic potential. However, one issue needs to be highlighted in this regard -mutations far away from the binding site can potentially affect the shape and or binding affinity of the binding pocket. These are not always captured by homology models, especially the issue of affinity. In principle, authors could use a conservation patterns or co-evolution measures (e.g. as in PMID: 10514373) to determine if there are interaction chains that might connect distant residues to the active site. In the least it would be useful to provide the caveat of distant changes affecting affinity in the current paper. We totally agree with the referee that also mutations at a greater distance may affect the binding pocket under spe-cial circumstances and we have added a new paragraph to the manuscript. However, we have to admit that it remains essentially impossible to quantify the influence of non-direct interactions by theoretical means unambiguously. In our experience, co-evolution measures such as the one proposed and several others (PMID: 18056067) produce high rates of false positives and, therefore, the interpretation is difficult. We did the requested analysis with Ranganathan's SCA algorithm to eventually identify connected networks of residues using the webserver implementation at the Gerstein lab, but no network that would connect the surface directly to the core and catalytic site was found (see supplementary webpage for full details). In fact, all positions of the observed mutations are at the surface and naturally variable, as judged by the conservation and SCA analysis, which would rather indicate that they do not have an effect on the structure of the more distant binding pocket. A possible problem why the SCA could not work in this case is the high level of sequence similarity among the neuraminidases which only gives limited numbers of informative correlated mutations. This is a totally different scenario from alignments of highly divergent sequences that still have the same fold, such as the small PDZ domains analyzed by Ranganathan. In the case of a diverse family with shared fold, correlated mutations are indicative of allowed fluctuations also among structurally important residues. However, with the neuraminidases, most variation can be attributed to surface residues, which makes sense given the pressure to avoid immune responses. Additionally, sequence sampling of neuraminidases is not independent but biased by transmission chains and clusters of outbreaks. A possible but also not necessarily more precise alternative to judge indirect effects on drug binding would be to run free energy simulations with the bound drug to judge changes of affinity caused by the mutations but this is a tricky and time-consuming endeavour that would burst the scope of this current manuscript. unspecific consensus motif: change to "non-specific" Response changed. Phylogenetic analysis: While for sequences at this range of similarity Poisson correction may not have negative consequences it is definitely better to repeat the analysis with JTT and variable rates to see if the clustering remains the same or changes drastically. We confirmed that the tree clustering inside the N1 subtype family is robust regarding different tree generation methods and added this also to the text. We also provide the suggested JTT distance tree with variable rates as supplementary at http://mendel.bii.a-star.edu.sg/ SEQUENCES/H1N1/. " This indicates that, similar to the current outbreak, scenarios of breaches in the species barrier between human and pigs have already arisen out of closely related H1N1 strains as also evidenced by further reports in the literature [26, 27] ." The wording of this sentence is somewhat unclear. It appears that the authors wish to state that H1N1 like strains have crossed species boundaries on other occasions, but this is not necessarily clear in the sentence. Response changed to "This indicates that, similar to the current outbreak, closely related H1N1 strains have crossed species boundaries on previous occasions as also evidenced by further reports in the literature". "These results support the notion that, also inside the family of N1 subtypes, a clear distinction can be made to distinguish avian-like H1N1, such as the one from the current outbreak, from other existing H1N1 strains." What would be the explanation for this? A recent recombination between an avian-like H1N1 or has it diverged from other avian like H1N1 with the recombination occurring much earlier. Could this information be superimposed in phylogenetic context on current figure 3 ? Response This is, of course, a very interesting question and difficult to deduce from the phylogenetic tree of a single protein without molecular clock, as in our case. However, this and similar questions have already been analyzed to quite some detail for the existing strains. The current knowledge of the scenario for H1N1 is that around the late 1970s to early 1980s, a human-avian reassortant virus started to be detected in European pigs as host (PMID: 8091678). Then, this avian-like swine flu started to move from the European continent to the UK in the early 90s (PMID: 9049404). More reassortments and emergence of the N2 subtype that quickly spread is also well documented. This phylogenetic analysis plugs the new H1N1 strain into the already known clusterings among the NA subtypes and within the N1 family.

Development of TaqMan(® )MGB fluorescent real-time PCR assay for the detection of anatid herpesvirus 1

BACKGROUND: Anatid herpesvirus 1 (AHV-1) is an alphaherpesvirus associated with latent infection and mortality in ducks and geese and is currently affecting the world-wide waterfowl production severely. Here we describe a fluorescent quantitative real-time PCR (FQ-PCR) method developed for fast measurement of AHV-1 DNA based on TaqMan MGB technology. RESULTS: The detection limit of the assay was 1 × 10(1 )standard DNA copies, with a sensitivity of 2 logs higher than that of the conventional gel-based PCR assay targeting the same gene. The real-time PCR was reproducible, as shown by satisfactory low intra-assay and inter-assay coefficients of variation. CONCLUSION: The high sensitivity, specificity, simplicity and reproducibility of the AHV-1 fluorogenic PCR assay, combined with its wide dynamic range and high throughput, make this method suitable for a broad spectrum of AHV-1 etiologically related application.

China is currently holding the largest waterfowl population in the world and its waterfowl production industry has been characterized by an increasing expansion and rapid development during the past decades [1] . However, infectious diseases represent the biggest obstacle to successful development of this business. Anatid herpesvirus 1 (AHV-1) infection alternatively known as duck virus enteritis (DVE), or duck plague (DP) [2] , is one of the most widespread and devastating diseases of waterfowls in the family Anatidae and has severally affected the waterfowl industry since the early 1900s because relatively high mortality could be observed and a wide host range including domestic [3] and wild ducks [4, 5] , geese and swans of all species as well as other birds like coots are susceptible. Furthermore, serious carcass condemnations and decreased egg production were also observed in affected waterfowls. Like other herpesvirus-induced diseases, AHV-1 infection has latent form and the virus can be persistently shed by birds that recover from the disease [6] . This complicates the control of the disease, particularly under small-holder farming conditions prevalent in China. The causative agent of AHV-1 is grouped in the alphaherpesviridae subfamily of the herpesvirus family [7] and the viral genome is a linear, double-stranded DNA molecule approximately 180 kb in size and its structure is similar to other alphaherpesviruses [8] . The AHV-1 genomic DNA has % G + C content of 64.3, which is the highest reported for any avian herpesvirus in the alphaherpesviridae [9] . Since prevention and early detection are presently the most logical strategies for virus control, various diagnostic procedures including microscopic, immunological and molecular methods have been developed for AHV-1 detection, of which the polymerase chain reaction (PCR) is a powerful tool with exquisite sensitivity for detection of minute amounts of nucleic acids, even against a high background of unrelated nucleic acids. Fluorescent quantitative real-time PCR (FQ-PCR) technique has eliminated the need of sample post-amplification handling required by the conventional PCR assay and has paved the way towards fully automated detection systems now that they usually display very high sensitivity and broad dynamic capacity after optimization [10] [11] [12] . Since virus load and proliferation dynamics serve as indispensable indicators of virus-host interaction, antiviral evaluation, active/ latent infection [13] [14] [15] and guidance for therapeutic intervention, FQ-PCR is therefore of paramount importance by its exquisite virus detection and monitoring ability [16] . The detection of AHV-1 by TaqMan real-time PCR method has only been reported by Yang [17] and with the development of technology, TaqMan Minor Groove Binding (MGB™) probes as an upgrade of the ordinary TaqMan probe has been widely used during the recent years since the following advantages: (1) The TaqMan MGB probe is characterized by the conjugation of minor groove binders which facilitates highly specific of the detection. (2) The TaqMan MGB probe contains a quencher dye that does not emit fluorescence within the detectable wavelength range and results in greater accuracy in the measurement. Therefore a TaqMan MGB-based real-time PCR method for detection and quantitation of AHV-1 is developed to serve as an alternative and improvement of the previously developed ordinary TaqMan real-time PCR method. Following the optimization of FQ-PCR, final concentrations of primers each of 0.3 μmol/L and probe of 0.1 μmol/L were selected. The MgCl 2 concentration was balanced to 6 mM that provided optimal AHV-1 amplification. Therefore the optimized 25-μL real-time PCR reaction system for AHV-1 detection could be summarized as follows: 1 × PCR buffer, 6 mmol/L MgCl 2 , 0.2 mmol/L dNTPs, 0.3 μmol/L each primers, 0.1 μmol/L probe, 1 U Taq and 1 μL DNA template. Following the optimization of conventional PCR, the MgCl 2 concentration was balanced to 2.5 mM and the annealing temperature of 52°C was selected. Therefore the optimized conventional PCR reaction system could be summarized as follows: 1 × PCR buffer, 2.5 mmol/L MgCl 2 , 0.2 mmol/L dNTPs, 0.5 μmol/L each primers, 1.25 U Taq and 1 μL DNA template. The optimized annealing temperature was 52°C. The FQ-PCR amplification curves and the corresponding fluorescent quantitative real-time PCR standard curve ( Figure 1 ) were generated by employing the successively diluted known copy number of pAHV-1 for real-time PCR reaction under the optimized conditions. From the results of correlation coefficient (0.999) and PCR efficiency (86.9%) of the standard curve by the established FQ-PCR, it could be known that the standard curve and the established FQ-PCR are excellent at performance. Different 32 AHV-1 strains kindly provided by the Avian Disease Research Center of Sichuan Agricultural University were examined with the established FQ-PCR method and these specimens all tested positive in the FQ-PCR assay, indicating that this method is sensitive and compatible with wide range of AHV-1 viruses. Ten-fold dilution series of pAHV-1 plasmid standard DNA were tested by the established real-time PCR assay to evaluate the sensitivity of the system and the detection limit was found to be 1.0 × 10 1 copies/reaction. Comparisons were made between conventional PCR and the established FQ-PCR using dilution series to calculate the end point sensitivity of each assay. The results indicate that the established FQ-PCR is around 100 times more sensitive than the conventional PCR method, detecting pAHV-1 down to dilutions of 1.0 × 10 1 , compared to dilutions of only 1.0 × 10 3 for conventional PCR. The test using DNA from several other bacteria and viruses used as template to examine the technique's specificity showed that none of the bacteria, virus (other than AHV-1) and duck embryo fibroblast tested gave any amplification signal and the results demonstrated that the established FQ-PCR assay is of highly specific. The intra-assay and inter-assay CV of this established FQ-PCR was in the range of 1-3% for most of the dynamic range (from 1.0 × 10 9 to 1.0 × 10 2 pAHV-1 plasmid copies/ μL), but increased to more than 6% at viral DNA loads lower than 1.0 × 10 2 pAHV-1 plasmid copies/μL and increased to more than 4% at viral DNA loads more than 1.0 × 10 9 pAHV-1 plasmid copies/μL ( Table 1 ). The results Figure 1 Establishment of the fluorescent quantitative real-time PCR standard curve. Standard curve of the AHV-1 fluorescent quantitative real-time PCR. Ten-fold dilutions of standard DNA ranging from 1.0 × 10 9 to 1.0 × 10 2 copies/μL were used, as indicated in the x-axis, whereas the corresponding cycle threshold (CT) values are presented on the y-axis. Each dot represents the result of triplicate amplification of each dilution. The correlation coefficient and the slope value of the regression curve were calculated and are indicated. demonstrated that the established fluorescent quantitative real-time PCR method was characterized by a wide dynamic range (8 logarithmic decades) of detection from 1.0 × 10 9 to 1.0 × 10 2 pAHV-1 plasmid copies/μL with high precision. However, at lower and higher dilutions quantitation was not always reproducible compared to other properly diluted samples. Therefore the dynamic range of the method was between 1.0 × 10 9 and 1.0 × 10 2 pAHV-1 plasmid copies/μL, which is relatively broad. Viral load quantification through the established AHV-1 FQ-PCR demonstrated that the AHV-1 DNA copy number of each sample could be calculated with the CT value according to the standard curve and 100% of the samples tested were quantifiable (Table 2 ) without the need for further sample dilution or concentration. Conventional etiological, immunohistological and serological methods [18] [19] [20] were routinely used in AHV-1 identification. However, the sensitivity is usually not high enough and the methods were time-consuming since virus propagation in cell cultures is usually required and the onset of virus-induced cytopathic effect (CPE) usually requires at least 2-3 days to develop. Titration of infectious virus in cell cultures is usually achieved by the endpoint dilution method in cell monolayer. Since titration of the virus load is labor-consuming and requires about 5 days for evaluation of virus-induced CPE, distinguishing between virus-induced CPE and non-specific cell alterations may be difficult, the established real-time PCR assay will be particularly suitable in these studies. In addition, an even more important factor is that the virus from tissues of infected birds is usually not readily adapted to cell culture system during the initial several rounds of propagations [21] . The PCR is a rapid, sensitive and specific nucleic acid amplification technique and many conventional qualitative PCR methods for revealing merely the presence or absence of AHV-1 pathogen have been developed and well documented [22] [23] [24] . However, the conventional PCR assays are not sufficient in a variety of clinical situations. They frequently encountered problems including the risk of cross-contamination (leading to false positives) and poor quality of extracts (leading to false negatives). Moreover, the lack of fluorogenic probes in the assay results in relative lower specificity since the amplification and detection of specific PCR products are determined solely by the amplification primers. In this paper, the development of a TaqMan MGB-based real-time PCR by using fluorogenic labels and sensitive signal detection system for detection and quantitation of AHV-1 is described. The optimized FQ-PCR detection system presented in this paper has been designed to address these issues and make it even more applicable for routine diagnostic use with several advantages over conventional PCR. In this assay, the primers and probes have been selected on conserved DNA segments of AHV-1 genome. TaqMan Minor Groove Binding (MGB™) probes as target-specific hydrolysis oligonucleotide employed in this assay are characterized by the conjugation of minor groove binders which increases the Tm of the hybridized probe and facilitates highly specific binding to the targeted sequence [25] . Moreover this probe contains a quencher dye that does not emit fluorescence within the detectable wavelength range and results in greater accuracy in the measurement. This improvement eliminates spectral overlaps with fluorescence emitted by the reporter dye, and results in greater accuracy in the measurement of reporter-specific signals. In view of the great sensitivity of PCR, the occurrence of false negative results is a highly underestimated problem. So an artificial construct generated by cloning of the specific target sequence into a plasmid are often used as internal controls for the amplification step. This internal positive control was incorporated into the reaction system, thus improving diagnostic conclusions, especially negative results, which is most important in the light of quarantine programs. By carrying out direct comparisons between the established FQ-PCR method and the conventional PCR method for AHV-1 detection, the results clearly showed that overall the established FQ-PCR detection method is more sensitive and reliable when compared to conventional gel-based PCR, since it was able to detect as few as 1.0 × 10 1 DNA copies of template. Furthermore, this established AHV-1 FQ-PCR method shows more excellent characteristics such as dynamic range (from 1.0 × 10 9 to 1.0 × 10 2 pAHV-1 plasmid copies/μL, which is approximately 10 3 times broader) and sensitivity (detecting pAHV-1 plasmid down to dilutions of 1.0 × 10 1 copies/μL, which is about 2.3 times more sensitive) than other reported method [17] . The high quality hot start Taq DNA polymerase used in this assay could minimize unspecific amplifications and increase the PCR cycling efficiency. In addition, FQ-PCR reaction and detection is all done in a closed-tube system, the need for post-amplification manipulation is removed since the detection of the PCR products occurs online during real-time PCR amplification, hence greatly reducing the risk of cross-contamination and false positive results. The optimization of the AHV-1 FQ-PCR assay was focused on the concentration of primers and probe and Mg 2+ . When all these different practical refinements are combined, the final result is a molecular diagnostic method that is not only rapid and reliable, but one that is also easy to perform and applicable to use for testing large numbers of samples since the FQ-PCR presented the benefits of increased speed due to reduced cycle time and remove of post-amplification process, offering considerable labor savings and allowing higher throughput analysis than conventional PCR assays and thus is favorable for the transition of this method from research to routine use in laboratories. This method was preliminarily mentioned in a short report [26] but related details of primers and probe sequence, specificity test, sensitivity test, reproducibility analysis, dynamic range and internal control were unavailable. By contrast, great modification and optimization have been made in this paper to improve the quality of this study. The AHV-1 FQ-PCR assay was highly reproducible and linear over a range of eight orders of magnitude from 10 2 to 10 9 copies, allowing a precise calculation of viral DNA load in samples containing a wide range of viral DNA amounts, eliminating the need for sample dilution and minimizing sample handling. The results for intra-and inter-assay precision indicate that both intra-assay and inter-assay CVs were satisfactorily low and the assay is reproducible, even between different batches of reagents used. Probability rather than sample quality variation is the predominant cause of variability at low copy numbers [27] . In conclusion, the FQ-PCR developed in this study is highly specific and sensitive with better parameters than conventional PCR method and is a valuable method for the detection of AHV-1. The method described in this study is especially helpful for high throughput analysis such as evaluating the efficacy of antiviral drugs and experimental vaccines for AHV-1. The research group of authors is currently using this technique to study the AHV-1 distribution characteristics in vaccinated birds and in artificially infected birds. We believe that this method could expedite related AHV-1 research in the AHV-1 viral molecular biology. Duck embryo fibroblast (DEF) monolayer was incubated at 37°C with 5% CO 2 in tissue culture flasks with Minimal Essential Medium (MEM) that contained 10% fetal bovine serum (FBS), 100 U/mL penicillin, and 100 μg/mL streptomycin. Anatid herpesvirus 1 (AHV-1, CHv virulent strain) was obtained from the Avian Disease Research Center of Sichuan Agricultural University (Yaan, Sichuan, China). Virus stock was added onto the surface of the cell layer which was about 90% confluency at time of infection and the maximum virus titers could usually be obtained 48 h postinfection. DNA extraction from AHV-1 infected DEF cells and tissues of AHV-1 infected ducks were performed by using TIANamp Genomic DNA extracting kit (Tiangen Corporation, Beijing, China) according to the manufacture's instructions. The FQ-PCR assay primers and probe (named Real-F, Real-R and Real-P respectively) design was carried out using the Primer Express™ software supplied by Applied Biosystems and their sequences were listed in Table 3 . The forward and reverse primers amplified a 60 bp fragment of AHV-1 DNA polymerase gene as described (GenBank Accession No. AF064639). The fluorogenic probe was labelled at 5' with FAM (6-carboxyfluorescein) dye as reporter and labelled at 3' with TAMRA (tetra-methylcarboxyrhodamine) as quencher and 3'with MGB™ (Minor Groove Binder). The conventional PCR amplification was carried out using primers designed using the Primer Premier™ software according to the sequence as described (GenBank Accession No. AF064639). The forward primer and reverse primer (named Con-F and Con-R respectively) sequences were listed in Table 3 and this primer pair yielded a 498 bp amplicon, in which the 60 bp FQ-PCR fragment was nested. All probes and primers were synthesized by Genecore Corporation (Shanghai, China) and purified by corresponding HPLC system. The real-time PCR was carried out using the ABI AmpliTaq Gold DNA polymerase system with an icycler IQ Realtime PCR Detection System (Bio-Rad Corp., Hercules, CA) according to the manufacturer's instructions. The reaction, data acquisition and analysis were performed using iCycler IQ optical system software. The Real-time PCR was performed in an 25 μL reaction mixture containing 1 × PCR buffer, 0.2 mmol/L dNTPs, 1 U Taq and 1 μL DNA template according to the manufacture's instructions. Autoclaved nanopure water was added to bring the final volume to 25 μL. The two-step PCR cycling condition was as follows: initial denaturation and hot-start Taq The conventional PCR was performed and optimized on a Mycycler™ thermo cycler system (Bio-Rad Corp., Hercules, CA, USA) with a 50 μL PCR reaction system containing 1 × PCR buffer, 0.2 mmol/L dNTPs mixture, 1.25 U rTaq (Takara Bio Inc., Shiga, Japan), 0.5 μmol/L each forward and reverse primers and 1 μL template DNA. All PCR experiments were carried out in 0.2 ml thin-walled tubes with the following cycle parameters: The mixture was subjected to initial denaturation at 95°C for 1 min, followed by 50 cycles of 95°C for 60 s, annealing for 60 s, extension at 72°C for 60 s, and one cycle of final extension at 72°C for 5 min. The amplified 498 bp product then underwent electrophoresis on 1.0% agarose gels. Electrophoresis was carried out at 100 V in a Mini-sub (Bio-Rad Corp., Hercules, CA, USA) gel electrophoresis unit and gels were viewed under a UV transilluminator. The conventional PCR reactions were optimized based on MgCl 2 concentration and annealing temperature selection criteria in a similar way as that of Real-time PCR and the selection was made by the brightness of the amplified 498 bp fragments on the agarose gel under a UV transilluminator. An internal positive control was introduced into the FQ-PCR assay to verify the absence of DNA losses during the extraction step and of PCR inhibitors in the DNA templates. The internal positive control of pGM-T recombinant vector (designed as pB16S) consisting of Bacillus 16S rRNA gene (GenBank Accession No. AJ971894) sequence amplified with primers (IC-F and IC-R) listed in Table 3 was added into the lysis buffer at the concentration of 1.0 × 10 6 copies/μL. Real-time PCR for IC detection was carried out in a separate run, using primers and probe (named IC-F, IC-R and IC-P respectively) listed in Table 3 . The fluorogenic probe was labelled at 5' with FAM as reporter and labelled at 3' with TAMRA. The quantitative real-time PCR protocol was the same as that of AHV-1 detection. From the ratio of the calculated amount of IC to the actual amount of IC, which is shared by the specimen, the normalization could be achieved and the actual amount of AHV-1 in the specimen could be obtained. Actually this internally controlled method has been widely used in other related detection assays [28, 29] . The 498 bp conventional PCR target amplicon band on agarose gel was cut and the DNA was recovered and purified by TIANquick DNA Purification system (Tiangen Corp., Beijing, China) according to the instruction manual of the product. The product was ligated into pGM-T vector (Tiangen Corp., Beijing, China) and transformed into E.coli DH5α competent cells. Recombinant plasmid (designated as pAHV-1) was extracted using TIANprep plasmid extraction kit (Tiangen Corp., Beijing, China). Presence of the target DNA insert was confirmed by PCR amplification and sequencing. The standard curve of the FQ-PCR was generated by successive dilutions of the known copy number of pAHV-1. Recombinant plasmid pAHV-1 concentration was determined by taking the absorbance at 260 nm using a Smartspec 3000 spectrophotometer (Bio-Rad Corp., Hercules, CA) and purity was confirmed using the 260/280 nm ratio. Through its molecular weight, pAHV-1 copy number was then calculated and the purified pAHV-1 plasmid DNA was then serially diluted 10-fold in TE buffer, pH 8.0, from 1.0 × 10 9 to 1.0 × 10 2 plasmid copies/ μL. These dilutions were tested in triplicate and used as quantitation standards to construct the standard curve by plotting the plasmid copy number logarithm against the measured CT values. The Bio-Rad iCycler IQ detection software created the standard curve, calculated the correlation coefficient (R 2 ) of the standard curve, standard deviations of triplicates. Different 32 AHV-1 strains (derived from a wide spectrum of sources, subsequently confirmed through related etiological methods, and then preserved by the Avian Disease Research Center of Sichuan Agricultural University) including virulent and avirulent strains were examined with the established FQ-PCR method to test the sensitivity and compatibility of this method. In addition, the sensitivities of the conventional PCR and FQ-PCR were each determined using triplicates of different concentrations of recombinant plasmid pAHV-1. Template DNA was prepared as follows: plasmids of pAHV-1 were diluted serially in 10-fold steps from 10 10 copies/μL to 10 1 copies/μL using sterile ultra pure water. One microliter from each dilution was used as template and subjected to the conventional PCR and FQ-PCR protocol respectively. The detection limit of the conventional PCR was determined based on the highest dilution that resulted in the presence of clear and distinct amplified fragments (498 bp) on the agarose gel. The detection limit of the FQ-PCR was determined based on the highest dilution that resulted in the presence of CT value in real-time PCR detection. Within-run and between-run reproducibilities of the FQ-PCR assay were assessed by multiple measurements of pAHV-1 samples of different concentrations. The assay was conducted by assessing the agreement between the replicates in five replicates (within-run precision) and in five separate experiments (between-run precision) of the serially diluted pAHV-1 recombinant plasmid samples through performing analysis of the mean coefficient of variation (CV) values of each AHV-1 standard dilution. Dilutions of pAHV-1 recombinant plasmid were used to determine the dynamic ranges of the FQ-PCR assay. The lower and upper limits of quantification were defined by the pAHV-1 recombinant plasmid sample concentrations possessing reasonable precision. AHV-1 infected duck embryo fibroblast culture, allantoid fluid and other specimens including liver, brain, Bursa of Fabricius, thymus, spleen, esophagus, duodenum, ileum, kidney, lung, peripheral blood each collected from AHV-1 infected ducks were employed to assess the ability of the established FQ-PCR to detect AHV-1 in a variety of usually used samples. By this assay viral load quantification was obtained.

The Y271 and I274 Amino Acids in Reverse Transcriptase of Human Immunodeficiency Virus-1 Are Critical to Protein Stability

Reverse transcriptase (RT) of human immunodeficiency virus (HIV)-1 plays a key role in initiating viral replication and is an important target for developing anti-HIV drugs. Our previous study showed that two mutations (Y271A and I274A) in the turn RT (Gln(269)-Arg(277)) abrogated viral replication, but the replication capacity and RT activity was discordant. In this study, we further investigated why alanine substitutions at these two sites would affect viral replication. We found that both RT activity and RT protein were almost undetectable in viral particles of these two mutants, although the Pr160(gag-pol) mutants were properly expressed, transported and incorporated. Using protease inhibition assay, we demonstrated a correlation between the degradation of the RT mutants and the activity of viral protease. Our native gel analysis indicated that the mutations at 271 and 274 amino acids might cause conformational changes, leading to the formation of higher order oligomers instead of dimers, resulting in increased protein instability and susceptibility to viral protease. Thus, residues 271 and 274 are critical to RT stability and resistance to viral protease. The conservation of the two amino acid residues among different strains of HIV-1 lent further support to this conclusion. The knowledge gained here may prove useful in drug design.

Reverse transcriptase (RT) of human immunodeficiency virus (HIV)-1, encoded by pol gene, is a multifunctional enzyme that possesses RNA-and DNA-dependent polymerase activities as well as RNase H activity [1] . RT is indispensable for HIV-1 and it converts the single-stranded viral RNA into double-stranded DNA upon viral entry into host cells. Due to its important role in viral life cycle, RT is one attractive target for antiviral drug design [2] . The biologically active form of HIV-1 RT is a heterodimer consisting of two subunits, p66 (66 kDa) and p51 (51 kDa). The p51 subunit is derived from p66 by proteolytic cleavage of its Cterminal domain [3] . The polymerase domain of p66 and p51, resembling a right hand configuration, consists of four subdomains, which are known as fingers, palm, thumb and connection. The fingers, palm and thumb subdomains of p66 form a nucleic acid binding cleft and the connection subdomains of the two subunits form the floor of the nucleic acid binding site [4] [5] [6] [7] . The thumb subdomain has four a helices. Two antiparallel ahelices of them, a-H (Asn 255 to Ser 268 ) and a-I (Gln 278 to Thr 286 ), are important for holding the primer/template in position during the translocation in polymerization. The primary sequence (Val 254 to Ala 288 ) in the vicinity of these two a helices has been found to share homology with several other nucleic acid polymerases and has been termed the ''helix clamp'' [5, 8] . Extensive studies have been carried out to shed light on the relationship between the ''helix clamp'' and function of RT. The effects of alanine-scanning mutations in a-H and a-I on polymerase activity, primer/template (P/T) binding, fidelity and enzyme kinetics have been determined. While mutations in a-I do not affect P/T binding or fidelity significantly, several a-H mutants exhibit lower binding affinity, processivity and frameshift fidelity [9] [10] [11] [12] . Previous studies have demonstrated that mutations in these two helices can have significant effect on RNase H activity, minus-strand DNA transfer activity and removal of polypurine track primer [13, 14] . Although alanine substitutions at sites 269, 270, 271 and 277 have been investigated in two studies [9, 10] , detailed studies on the functional structure of the ''turn'' (Gln 269 -Arg 277 ) between a-H and a-I were limited. In our previous study on hepatitis B virus RT, conserved residues located at the ''turn'' of helix clamp motif were found important for pregenomic RNA encapsidation during the assembly of nucleocapsids [15, 16] . Since this homologous helix clamp motif is also present in HIV-1 RT, we hypothesized that residues in the turn may play important roles in viral life cycle. Our recent study showed that alanine substitutions at 271 and 274 of HIV-1 RT drastically affected viral replication, but discordance between viral replication and RT activity was observed [17] . In this study, we confirmed our previous observations and further investigated why the two mutations abrogated viral replication. Our study demonstrated that these two mutations lead to rapid degradation of RT in viral particles, indicating that the residues of 271 and 274 are critical for maintaining the stability of HIV RT. The parental HIV-1 proviral plasmids, pLAI.2, pNL4-3-DE-EGFP and pHEF-VSV-G, were obtained through the NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH [18] [19] [20] . The pNL4-3-DE-EGFP based mutants (Y271A, G273A, I274A, K275A, V276A, R277A) and pLAI2 based mutants (Y271A and I274A) were constructed by sitedirected mutagenesis using QuikChange II XL Site-Directed Mutagenesis Kit (Stratagene, USA) according to manufacturer's instruction. The open reading frame of RT p66 subunit was amplified from wild type or mutant pNL4-3-DE-EGFP using a pair of primers and cloned into pET-28b vector (Novagen, Shanghai) to obtain expression plasmid pET-p66 with the 66 His tag at 39 terminus. The paired primers used for site-directed mutagenesis and construction of plasmids are listed in Table 1 . Cell lines 293FT, HeLa and U373-MAGI-CXCR4 CEM [21] were maintained in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% fetal bovine serum and antibiotics (Invitrogen, USA). MT2 cells [22, 23] were maintained in RPMI 1640 supplemented with 10% fetal bovine serum and antibiotics. To prepare pseudoviruses and live viruses, 293FT cells were cotransfected with 7.5 mg wild-type or mutant pNL4-3-DE-EGFP, and 2.5 mg pHEF-VSV-G, or 10 mg wild-type or mutant pLAI.2 using Lipofectamine 2000 (Invitrogen, USA) according to the manufacturer's instructions. Supernatants were collected 48 h after transfection and stored at 280uC after spinning down and filtering to remove cell debris. The cells were rinsed with ice-cold phosphate-buffered saline (PBS), scraped from each plate and lyzed in cell lysis buffer (Boehringer, Germany) and 16 protease inhibitor cocktail (Roche, USA) on ice for 30 min. Cell lysates were stored at 220uC after centrifugation at 13,0006g at 4uC for 15 min to remove cell debris. One cycle infection assay was carried out using normalized pseudoviruses as described previously [20] . Briefly, Jurkat cells (0.5610 6 ) were infected with viral supernatants containing 250 ng p24, which was measured by Vironostika HIV-1 antigen MicroELISA kit (Biomerieux bv Boxtel, Netherlands). The virus and cells mixture was spun at 1,800 g at 30uC for 2 h. After the 2h spin infection, Jurkat cells were washed with 2 ml culture medium twice, then cultured in 24-well plates at 37uC for 48 h. The cells were collected and washed twice with PBS. After the cells were fixed with 1% paraformaldehyde in PBS for 30 min on ice, the infected cells, as determined by the expression of GFP, were measured using a FACSCalibur instrument (Becton Dickinson, USA) and analyzed with Cell Quest software (Becton Dickinson, USA) as described previously [20] . Infection with live viruses was conducted in U373-MAGI-CXCR4 CEM and MT-2 cells. U373-MAGI-CXCR4 CEM cells were infected with normalized wild type or mutant live viruses in 24-well plate as described previously [21] . Briefly, triplicate wells (6610 4 cells/well) were infected with live viruses (60 pg p24/well) which were diluted in DMEM containing 20 mg/ml of DEAEdextran (Amersham Biosciences). After cultured at 37uC in 5% CO 2 incubator for 48 h, the cells were fixed in 1% formaldehyde-0.2% glutaraldehyde in PBS for 5 min. After washing twice with PBS, the cells were stained with 400 mg/ml of X-Gal (5-bromo-4chloro-3-indolyl-b-D-galactopyranoside), 4 mM MgCl 2 , 4 mM potassium ferrocyanide, and 4 mM potassium ferricyanide in PBS for 2 h at 37uC. The plate was washed twice with PBS and blue foci were observed under microscope. Infection with live viruses was also carried out in MT-2 cells as described previously [22, 23] . In brief, MT-2 cells (1610 4 cells/well) in 96-well plate were infected with live viruses (20 pg p24/well). After cultured for 6 days, the virus-specific cytopathic effect (CPE) was observed under microscope. The RT activity in pseudoviruses was measured by a RT assay using colorimetric kit (Roche, USA). Briefly, viral supernatants containing 2 mg p24 were centrifuged at 4uC for 2 h at 40,0006g and the viral pellets were resuspended in 50 ml lysis buffer. Lyzed viral pellets were 10 fold serially diluted and the subsequent procedures were carried out according to the manufacturer's instructions. RT activities of mutants were calculated and compared to that of wild type pseudovirus. Expression and subcellular localization of precursor Gal-Pol polyprotein were detected by immunofluorescence microscopy as described previously with some modifications [24] . Briefly, HeLa cells cultured on coverslip in 24-well culture plate were transfected with 600 ng pNL4-3-DE-EGFP and 200 ng pHEF-VSV-G. Two days post-transfection, the cells were fixed with 500 ml 4% paraformaldehyde (PFA) at room temperature for 15 min. After washing 3 times with PBS, 500 ml of 50 mM ammonium chloride was incubated with the cells for 10 min to neutralize residual PFA. The cells were washed 3 times with PBS and treated with 0.05% Triton X-100 for 3 min. After washing 3 times with PBS, 500 ml 10% normal rabbit serum (NRS) in PBS was added to block the slip overnight at 4uC. Primary antibody (mouse monoclonal antiintegrase, 1:100, Santa Cruz) and secondary antibody (Texas Red dye-conjugated Rabbit anti-mouse IgG, 1:100, Jackson Immu-noResearch) were incubated with samples in dark at room temperature. Following each incubation, samples were subjected to 3 washes with 1% NRS in PBS. With 3 additional PBS washes, the coverslip was mounted using fluorescence mounting medium (Dako) and observed under LSM510 Meta confocal microscope (Carl Zeiss). Gal-Pol polyprotein and its products were tested by Western blotting as described previously with some modifications [25] . Briefly, viral proteins were separated by 10% SDS-polyacrylamide gel electrophoresis (SDS-PAGE) and electro-blotted onto Hybond-P PVDF membrane (GE Healthcare, Bio-sciences). After blocking with 5% skim milk for 1 h at room temperature, the membrane was incubated with primary antibodies (Rabbit polyclonal anti-RT, 1:3000; mouse monoclonal anti-RT, antiintergrase, anti-protease or anti-capsid, 1:500) for 1 h. Membrane was washed three times with PBS containing 0.1% Tween 20 (PBS-T) and then incubated with horseradish peroxidase (HRP)-conjugated secondary antibody (goat anti-rabbit IgG or goat anti-mouse IgG, 1:4000) for 1 h. After the blots were rewashed three times in 0.1% PBST, signals were visualized using ECL Western blotting substrate reagents (Amersham Biosciences, USA) and KODAK BioMax Scientific Imaging Film (Eastman Kodak). Wild type and mutant pET-p66 expression vectors were respectively transformed into E. coli BL21(DE3) and the expression was induced with 0.3 mM isopropyl b-D-1-thiogalactopyranoside (IPTG) when E. coli grew up to an OD600 of 0.7,1.0. E. coli expressing p66 was spun down and re-suspended in the binding buffer (50 mM NaH 2 PO4, 300 mM NaCl, 10 mM imidazole). After the bacteria were disrupted by ultrasonication and centrifuged at 15 000 g for 30 min at 4uC, the wild type and mutant p66 in supernatants were purified using Ni-NTA magnetic agarose beads (Qiagen, Germany). The purified p66 proteins were subjected to NativePage Novex Bis-Tris gel analysis and Western blotting according to the protocol recommended by Invitrogen, USA. RT sequences of 1083 HIV-1 strains obtained from the HIV complete sequence database (http://www.hiv.lanl.gov/content/ sequence/NEWALIGN/align.html) were aligned and compared. Based on the X-ray crystal structures of HIV-1 RT (1RTH) from the Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB), structural models of wild type and mutant RTs were analyzed using Swiss-PdbViewer software (http://spdbv.vital-it.ch/). Statistical analysis of RT activities was performed by Student's t test using Stata statistical software. Results were considered significant at P#0.05. The effect of six RT mutations at the helix clamp turn in HIV-1 RT on viral replication was tested with pseudoviruses. When Jurkat cells were infected with the wild type and mutant pseudoviruses (250 ng p24 virus/5610 5 cells), viral replication was almost completely inhibited in the mutants Y271A and I274A, while replication of other mutants did not show significant difference as compared to that of the wild type (Fig. 1A) . A similar result was also obtained when the cells were infected with lower amount (150 ng p24) of pseudoviruses (data not shown). To confirm the above results, replication of wild type, Y271A and I274A live viruses was further monitored in U373-MAGI-CXCR4 CEM and MT2 cells. Mutants Y271A and I274A were undetectable in U373-MAGI-CXCR4 CEM cells (Fig. 1B) and no virus-specific CPE was found in MT-2 cells (Fig. 1C) . Since reverse transcription is the first step for HIV replication after viral entry into host cells, our results suggested that the two mutations might have affected the RT activity. RT activity and RT proteins were undetectable in pseudoviral particles of mutants Y271A and I274A We thus interrogated the pseudovirus mutants for RT activity. The RT activity recovered from mutants Y271A and I274A were almost unnoticeable, as compared with that of wild type ( Fig. 2A) . This result was further confirmed using wild type, Y271A and I274A live viruses (data not shown). Next, we asked if the viral particles contain dysfunctional RT or the RT was not incorporated into the viral particles. By Western blotting, RT protein was basically undetectable in viral particles of these two mutants (Fig. 2B) . The results suggested that the mutations could affect the incorporation of RT into the viral particles, leading to a defect in reverse transcription which initiates viral replication. Mutations Y271A and I274A did not affect Pr160 gag-pol expression and transportation We next investigated whether the loss of RT in viral particles of the mutants was attributed to altered expression and transportation of the precursor protein, polyprotein Pr160 gag-pol , during virus assembly. Pr160 gag-pol in lysates of cells transfected with the wild type and mutants was detected by Western blotting. As shown in Fig. 3A , similar levels of Pr160 gag-pol , Gag protein (Pr55 Gag ) and capsid protein p24 (CA p24) were found in cells transfected with the wild type or mutant constructs, indicating that the expression and stability of RT precursor protein were not affected by the mutations. Immunofluorescence staining further demonstrated a normal subcellular localization of mutant Gag-Pol polyproteins as compared to that of the wild type, suggesting that the transportation of Gag-Pol was unlikely to be affected by alanine substitutions (Fig. 3B) . The Gag-Pol polyprotein was incorporated into mutant virions of Y271A and I274A, but the RT was degraded by viral protease To investigate whether the precursor Gag-Pol polyprotein was indeed incorporated into pseudoviral particles of mutants Y271A and I274A, products of the Gag-Pol polyprotein, integrase (IN), protease (PR) and p24, in wild type and mutant pseudoviral particles were examined by Western blotting. The results showed that, except for RT (Fig. 2B) , all products of the Gag-Pol polyprotein, IN p32, PR p11 and CA p24, were detected in the mutants Y271A and I274A at a level similar to that of the wild type pseudoviral particles (Fig. 4A) . Thus, Pr160 gag-pol was indeed incorporated into the virions and processed properly. To study if the RTs in the viral particles of Y271A and I274A mutants were degraded by proteolysis that made them undetectable, pseudoviruses of wild type and mutants were generated in the presence or absence of indinavir, a highly specific inhibitor of HIV-1 protease. Indinavir treatment was effective, because the treatment resulted in a dose-dependent inhibition of Pr55 Gag processing into p24 for wild type and mutant viruses (Fig. 4B ). As shown in Fig. 4C , in the absence of indinavir, both RT p66 and p51 were readily detected in the wild type pseudovirus, but not in the mutant viruses. In the presence of indinavir, however, both RT p66 and p55 were markedly reduced in the wild type virus but became detectable in mutant Y271A viral particles, while RT p66 was also detectable in mutant I274A virus. We also detected the RT subunits in the viral supernatants collected at earlier time points (12, 24 and 36 hours post-transfection), but the mutant RTs were still undetectable (data not shown). These results suggested that the Y271A and I274A RTs were degraded after incorporation of Gag-Pol polyprotein into the virions, which might be attributed to the activity of viral protease. Since RT dimer is the stable form which remains resistant to the proteolysis, we tested whether treatment of dimerization enhancer could reduce the mutant RT proteolysis. Wild type and mutant pseudoviruses were generated in the presence or absence of Efavirenz (EFV), the most potent dimerization enhancer [26] . The results showed that the RT mutants were still undetectable in the presence of EFV (Fig. 5A) . The conformation of the wild type and mutant RT p66 was further analyzed by native gel electrophoresis followed by Western blotting (Fig. 5B) . Under native conditions, wild type p66 subunit formed homodimer. However, the formation of homodimer in mutant Y271A was markedly reduced and a higher order oligomer appeared, which might be tetramer according to its molecular weight, while mutant I274A existed as at least two higher order oligomers, which were likely tetramer and octamer based on their molecular weights. However, the exact nature of the oligomers formed by mutant p66 subunits remained to be clarified. Furthermore, the treatment with b-mercaptoethanol could partially disrupt the higher order oligomers and improve dimer formation. This result implicated that dimer formation might be necessary for the stability of RT and its resistance to proteolysis after incorporation of Gag-Pol polyprotein into the viral particles. Amino acids at 271 and 274 are relatively conserved and big side chains may be important in maintaining RT stability and resistance to proteolysis By comparing 1083 complete sequences of HIV-1, it was found that the amino acids at 271 and 274 were relatively conserved. As shown in Table 2 , tyrosine (Y) is the predominant naturally existent amino acid at 271 residue (99.19%), while phenylalanine (F), histidine (H) and cysteine (C) occur rarely (,1%). Similarly, isoleucine (I) is the predominant naturally existent amino acid at 274, which accounts for 98.43%, while in less than 2% of all cases, valine (V) and leucine (L) are found at this site. This finding suggested that these amino acids might probably play important role in maintaining the active conformation of RT. Consistently, structural analysis suggested that amino acids at these two sites are buried in the thumb region of RT (Fig. 6A) . It was further revealed that all wild type RTs have relatively big side chains, but they are absent from the mutants 271A and 274A (Fig. 6B & 6C) . Thus, the loss of the side chains in the RT mutants plausibly leads to conformational change of RT, leading to aberrance in dimer formation and susceptibility to proteolysis by HIV-1 protease. HIV RT plays a key role in viral replication and is an important target for development of anti-HIV drugs. However, the turn between helices H and I of HIV-1 RT thumb region (amino acid residues 269 to 277) has not been well characterized yet. To understand the structure-function relationship of this turn in details, we investigated whether alanine substitutions in this region would affect RT activity. Except for residues 269 and 270, which have been reported to have no significant influence on RT activity [9, 10] , and for residue 272, which is originally an alanine, six mutant pseudoviruses (residues 271 and 273 to 277) were constructed and viral replication was compared with that of the wild type virus. The replication of two mutants, Y271A and I271A, were found to be almost completely abolished (Fig. 1A) . This result was further confirmed in live virus system (Fig. 1B & 1C) . Since it has been reported that bacterially expressed Y271A mutant has only 1% activity of wild type enzyme [9] , we asked whether the mutants would similarly affect RT activity in the viral particles. The results showed that the RT activity was basically undetectable in viral particles of these two mutants ( Fig. 2A) . It was also found that the loss of RT activity in the viral particles might be attributed to the absence of RT in the viral particles rather than the incorporation of dysfunctional RT into the virions (Fig. 2B) . After viral entry into the cells, the first step of HIV-1 replication is reverse transcription. Our results thus suggested that replication of mutant viruses was abrogated, because the mutant RT did not exist in the virion. Since HIV-1 RT is incorporated into the virion in the form of Pr160 gag-pol , which is transported to cell membrane where the virus is packaged, through interaction with Pr55 gag [27] [28] [29] [30] , we thus investigated whether the mutations would affect Pr160 gag-pol expression and transportation. Our results showed that Pr160 gagpol was properly expressed and transported in cells transfected with mutant constructs, Y271A and I271A (Fig. 3) , indicating that these mutations did not affect either the production of the precursor or the interaction between Pr160 gag-pol and Pr55 gag . It was further found that the Pr160 gag-pol was indeed incorporated into the virions of these two mutants, because except for RT, all other products of Pr160 gag-pol including p24, protease and integrase [31, 32] , could be detected in the viral particles of mutants Y271A and I271A (Fig. 4A) . Although it has been reported that the domain between residues 183 and 305 of RT is likely responsible for RT incorporation [33] , our study ruled out the involvement of RT was detected in wild type and mutant pseudoviruses generated in the presence and absence of EFV, the most potent dimerization enhancer, by Western blotting using mouse monoclonal anti-RT and anti-CA p24 antibodies, respectively. (B) Purified wild type and mutant RT subunit p66 were analyzed by native gel electrophoresis followed by Western blotting using mouse monoclonal anti-RT antibody in the presence or absence of b-mercaptoethanol (b-ME). Compared with wild type p66, which basically existed as homodimer, p66 subunit of the 271A and 274A mutants formed higher order oligomers, suggestive of conformational change. doi:10.1371/journal.pone.0006108.g005 the turn (Gln 269 -Arg 277 ). Another report has suggested that two mutations in RT, L234D and W239A, led to premature cleavage of the Gag-Pol precursor and reduced levels of viral enzymes in the virions [34] . Our results also excluded that the mutations of Y271A and I271A affected the cleavage of the Gag-Pol precursor. By a viral protease inhibition assay using indinavir, we finally demonstrated that the RT mutants were degraded after incorporation of the precursor polyprotein into the virions and the degradation was associated with the activity of viral protease (Fig. 4C) . Our results indicated that the mutations in residues 271 and 274 would reduce the stability of RT inside the viral particles, rendering it susceptible to viral protease. Post-incorporation degradation of RT has been reported previously. Wapling et al. [35] have reported that mutations of W401L and W401A in RT can inhibit RT dimerization, resulting Fig. 6A . A big side chain (blue) is found in naturally occurring residues but not in 271A. (C) Structural models of naturally existing 274I, 274V and 274L residues as well as the lethal 274A mutation of p51 subunit. The same region was highlighted above in the right panel of Fig. 6A . A big side chain (pink) is found in naturally occurring residues I, V and L, but not in 274A. doi:10.1371/journal.pone.0006108.g006 [36] . Another mutation, L289K in p66 subunit, has also been reported to be able to abrogate dimerization [37] . Another group, when they tried to generate infectious molecular clones of Simian Immunodeficiency Virus, found that glutamic acid replacement at position 287 affected the stability of RT [38] . The biologically active and stable form of HIV-1 RT is a heterodimer consisting of p66 and p51, while the immediate precursor of p66/p51 heterodimer is the p66 homodimer. Proteolytic removal of the RNase H domain in one of the p66 homodimer subunit by HIV-1 protease leads to the formation of stable heterodimer p66/p51 [39] . RT exists as an equilibrium mixture of monomers and dimers that include the p66/p51 heterodimer and the p66/p66 and p51/p51 homodimers, among which the heterodimer is the most stable and the p51/p51 homodimer is the most unstable [40] [41] [42] [43] . Plausibly, the degradation of RT as reported in previous studies may be ascribed to the inhibition of RT p66 dimerization by the mutations. The mechanism of RT degradation in the virions observed in this study, however, may be different. Our results showed that RT mutants were still undetectable in the viruses generated in the presence of a potent dimerization enhancer (Fig. 5A ). Our native gel analysis showed that the dimer form of RT mutant 271A was detected, although it was markedly reduced as compared to the wild type, while mutant 274A could not form dimer. Instead, higher order oligomers of RT were detected in both mutants (Fig. 5B ). It has been reported that higher order oligomerization may occur in HIV-1 RT [41, 44, 45] . Our results have also indicated that Y271A and I274A substitutions may change the conformation of RT, leading to oligomerization. The higher order oligomers formed in these 2 mutants, perhaps tetramer and octamer, may be associated with instability of RT mutants and their susceptibility to viral protease. Furthermore, according to the locations of residues 271 and 274 (Fig. 6A) , it is highly probable that alanine substitutions at 271 and/or 274 can affect the conformation of p51 thumb region and subsequently its interaction with RNase H domain of p66, resulting in the exposure of the 7-amino-acid p51-RNas H cleavage sequence in the p66 subunit [7, 46] and consequent proteolytic degradation by HIV-1 protease. The relative conservation of RT 271 and 274 residues and the loss of big side chains at these two sites in the mutants as shown in the structural models (Fig. 6B & 6C ) also support their important roles. We tried to demonstrate this in vitro, by treating wild type and mutant RTs with recombinant protease, but wild type RT, as well as mutant RTs, could not be digested in vitro (data not shown). This result is consistent with that reported by Abram and Praniak [36] . This could be explained by different proteolytic stability of RT in vitro and in virions. Taken together, our study showed that alanine substitutions at residue 271 or 274 of HIV-1 RT could cause conformational changes, rendering RT unstable and susceptible to viral protease. Thus, it is demonstrated that the ''turn'' (Gln 269 -Arg 277 ) between two helices may be important in maintaining protein stability and formation of bioactive dimer. Mutations in residues 271 and 274 of RT may inactivate the virus, which may enter cells but can not replicate. These findings may prove useful in anti-HIV drug design and vaccine development. Considering the emergence of resistance to the current RT inhibitors, this turn, especially amino acid residues 271 and 274, may be an ideal new target for antiviral drug design. The agents targeting this region may be able to affect the stability or dimerization of HIV-1 RT and subsequently inhibit viral replication. Moreover, the potential drug resistant mutations in this region, especially at residues 271 and 274, may probably inactivate the virus, which prevents the appearance of drugresistant virus.

CVTree update: a newly designed phylogenetic study platform using composition vectors and whole genomes

The CVTree web server (http://tlife.fudan.edu.cn/cvtree) presented here is a new implementation of the whole genome-based, alignment-free composition vector (CV) method for phylogenetic analysis. It is more efficient and user-friendly than the previously published version in the 2004 web server issue of Nucleic Acids Research. The development of whole genome-based alignment-free CV method has provided an independent verification to the traditional phylogenetic analysis based on a single gene or a few genes. This new implementation attempts to meet the challenge of ever increasing amount of genome data and includes in its database more than 850 prokaryotic genomes which will be updated monthly from NCBI, and more than 80 fungal genomes collected manually from several sequencing centers. This new CVTree web server provides a faster and stable research platform. Users can upload their own sequences to find their phylogenetic position among genomes selected from the server's; inbuilt database. All sequence data used in a session may be downloaded as a compressed file. In addition to standard phylogenetic trees, users can also choose to output trees whose monophyletic branches are collapsed to various taxonomic levels. This feature is particularly useful for comparing phylogeny with taxonomy when dealing with thousands of genomes.

Traditional molecular phylogeny makes use of small subunit ribosomal RNA (SSU rRNA) sequences or a few orthologous proteins. Some more recent phylogenomic studies are based on concatenation of a larger number of proteins. The ever burgeoning genome sequencing projects worldwide have prompted several whole-genome phylogenetic approaches. However, most-if not all-rely on sequence alignment at some stage and therefore depend on many parameters, such as the use of scoring matrices. As modern prokaryotic and fungal taxonomy depends more and more on the traditional phylogeny, there is an urgent need to develop alternative approaches. CVTree provides such an alignment-free and parameterfree phylogenetic tool using composition vectors (CVs) inferred from whole genome data (1) . As a web server, it was first introduced in 2004 (2). The CV method has been effectively apaplied to phylogenetic study of viruses (3, 4) , chloroplasts (5) , prokaryotes (1, 6, 7) and fungi (submitted for publication). So far, the CV method has been cited in more than 70 papers not of our own, including some reviews (8, 9) . Since a CV consists of 20 K (for proteins) or 4 K (for DNA sequences) components for each organism, the calculation is simple but CPU time and memory consuming. In order to catch up with the increasing amount of genomic data, we have redesigned the data processing strategy and implemented a new user-friendly web interface to improve the new CVTree server in several aspects: (1) the inbuilt database has been enlarged and is now updated monthly from the NCBI FTP site (10). (2) Users may upload sequences of their own and carry out phylogenetic study together with genomes selected from the inbuilt database. (3) Many kinds of tree files are provided to facilitate comparison with taxonomy. Some tree files are directly uploadable to MEGA (11) or the Interactive Tree Of Life (iTOL) project (12) in order to display the results in different ways. (4) The efficiency of CVTree has been significantly enhanced to meet the requirement of treating thousands of genomes in a single run. All these improvements make the CVTree server a useful complement to various phylogenetic projects *To whom correspondence should be addressed. Tel/Fax: +86 21 6565 2305; Email: xuzh.fdu@gmail.com, xuzh@fudan.edu.cn ß 2009 The Author(s) This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/ by-nc/2.0/uk/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. such as AToL (Assembling the Tree of Life, http:// atol.sdsc.edu) or AFTOL (Assembling the Fungal Tree of Life, http://aftol.org) by providing independent verification and support to the SSU rRNA and few gene-based phylogenies (13) (14) (15) (16) (17) . Since the algorithm used in CVTree has been described previously (1, 2, 6) , we only give a brief account here. One collects all protein products in a genome and counts the number of (overlapping) K-tuples to form a raw CV with 20 K or 4 K components, depending on whether protein or coding DNA sequences are used (both options are allowed in CVTree, but protein sequences are recommended). Furthermore, one predicts the number of K-tuples from that of K À 1-mers and K À 2-mers by using a simple Markovian assumption. The differences between the prediction and the actual counts are taken as new components of a 'renormalized' CV. One may consult (1, 2, 6) or the online user's manual (available from the CVTree home page or http://tlife. fudan.edu.cn/cvtree/help/help.pdf) for more detailed description. The key improvement to accelerate CVTree's speed consists in avoiding repeated calculations among all jobs submitted after a major update of the database. All intermediate results of raw and renormalized CVs are kept until a major change taking place in the database. The response to a new submission may be deceptively fast if one's genome list coincides largely with that of a previous job. In the CVTree web server, the processing is carried out in two steps ( Figure 1) . First, the CV for each organism is calculated. CV files containing high dimensional vectors for all organisms are dumped to the hard disk. This strategy ensures that the CV is calculated only once for each organism. If the sequences of one organism have not been changed during the monthly update, the corresponding CV file will be kept. Second, the pairwise distances between the 'renormalized' CVs are calculated to generate a dissimilarity matrix. After the dissimilarity matrix has been produced, the standard neighbor-joining method (18) generates the tree files. The program neighbor is borrowed from the PHYLIP package (19) . CVTree reads amino acid or nucleotide sequences in FASTA format. It permits two kinds of input data: selected genomes from the inbuilt database and user's uploaded data. The inbuilt database consists of prokaryotic genomes downloaded from the NCBI FTP site (ftp://ftp.ncbi.nlm. nih.gov/genomes/Bacteria/) and fungal genomes collected manually. Only the compressed faa, rpt and gbk files are downloaded. The ffn files are locally extracted from the gbk files by the program extractfeat in the EMBOSS package (20) . Judging by the DEFINITION line in the gbk files, files that represent plasmids, mitochondria, phages or extrachromosomal elements are not fetched. Only chromosomal sequences are used. The NCBI Taxonomy ID was extracted from the rpt file of each organism. As of 1 April 2009, there were 799 bacteria and 56 archaea genomes. More than 80 fungal genomes have been collected manually from various sequencing projects. These fungal genomes together with their origin are listed in the online user's manual. The NCBI Taxonomy ID of fungi is assigned manually. (Currently some manually collected genomes including some fungi do not contain the ffn files, therefore could not be used to perform the DNA type calculation.) We have also included a few eukaryote genomes frequently used as outgroup in previous publications to bring the total number of built-in genomes to 941. This number will grow with monthly updates. By the way, the convention of using abbreviations for prokaryotic names has been given up, as it becomes inconvenient when organism number gets enormous. The binomina with full strain specification are used instead. Users may upload their own genomic sequences to the CVTree web server. All sequences of one and the same organism should be included in one FASTA file. The file name (without extension) will be displayed as the organism name in the trees. For user's convenience, sequences for a number of organisms may be wrapped Figure 1 . Two-step implementation of CVTree. CVTree implements a two-step strategy to produce phylogenetic trees. In the first step, CVTree reads in each genome sequence and counts the frequency of all K-tuples. Then the CV of each organism is calculated and dumped to the hard disk (CV files). In the second step, CVTree calculates the dissimilarity matrix from the correlation between CVs. Finally, the tree files are generated by the neighbor-joining program in PHYLIP package. into one compressed file. Many types of compressed file are accepted, such as GZIP(*.gz), BZIP2(*.bz2), ZIP(*.zip), TAR(*.tar) and RAR(*.rar). Due to disk limitation, up to 100 M disk space can be used for a user's uncompressed sequences in a project. Uploading files are restricted to 20 M at a time. These restrictions will be weakened in the future. In the inbuilt genome page, one can use the keyword filter to pick up the species of interest. For example, for the time being entering 'Archaea' as a keyword would bring up all the 56 arachaea names, while a keyword 'Streptococcus' would show up all the 38 species/strains in this genus. A user can click on the 'Check All/Uncheck All' button to select/unselect all organisms in one click. By combined use of the keyword filter and the taxonomy selector, it is convenient to make user-specific dataset for study. An overview of the new CVTree web server is given in Figure 2 . Once connected to CVTree's interface, a user may alternate between six pages shown in the figure, depending on how the job is being submitted and processed. We describe these pages one by one. The CVTree home page contains a link to an online user's manual and provides several ways to get to the project page. A first-time user may choose to create a new project or load an example project for a quick start. The difference consists in that the new project will start with an empty project space and the example project will bring up a preselected list of bacteria and archaea names from the inbuilt database. Users may recall a previous project by entering its project number. The project number also enables one to share the results with others. We suggest users always to create a new project to do their analysis, in this way the CVTree server does the garbage collection more efficiently. Any of the 'Create a new project', 'Example project' and 'Reload project' actions will redirect the user to the project page. Parameters are set in the project page. In the CV approach, length K of the oligopeptides or oligonucleotides controls the resolution of the method and is important for getting good results. Our previous studies have shown that better results may be achieved by setting K to 5 for virus (3, 4) , 5 or 6 for prokaryotes (1,7) and 6 or 7 for fungi. Our further study on how to choose K will make the subject of a separate publication. The sequence type (DNA or protein) and email to receive the results are to be entered in the project page. For DNA sequences K may be chosen from 6 to 18 with increment 3. For protein sequences, which are recommended, K = 3-7. If an email is entered, the web page may be safely closed after the project gets running. Using the project page, users can upload/delete their own sequences. To upload, first click the 'Browse' button (or 'Choose File' button in Google Chrome web browser) to find a sequence file locally, and then press 'Upload this file' to transfer. To delete, first select the sequences to be deleted and then press the 'Delete selected files' button. With user's files uploading or deleting the table in this page may stretch or shrink. User can select inbuilt genomes from the inbuilt genome page by clicking the 'See details' button. After that, the 'Download selected genomes' button will be activated. Clicking on this button, the user will be brought to the download page. This page shows the organism list of all inbuilt genomes. The default view shows organism name, proteome size (or cDNA length for DNA sequences) in MB, accession number for chromosome sequences and the superkingdom label extracted from NCBI Taxonomy Browser. Full taxonomy information can be shown by putting the mouse on the organism name. Clicking on the organism name will redirect the user to the NCBI Taxonomy Browser. This table is sortable by clicking at one of the header items. This is useful, for example, when a user wants to select a few smallest or largest genomes to study. Users can see organisms from designated taxa by using the taxonomy selector. The selected taxonomy will be shown in the last column of the organism list table. By choosing the taxonomy label and typing the appropriate keywords, the user can pick up the species of interest quickly. After filtering the organism list, the user can tick the box in the table header to select or deselect all organisms in the current list. When the selection is finished, the status filter can be used to review and check the list. When the inbuilt genomes have been selected, the 'Download selected genomes' button in the project page will be enabled. By clicking on this button, the user will be asked to wait while the selected sequence files are being prepared for downloading. Then the user will see a link appearing in download page. This link remains available as long as the project has not been destroyed or the user does not choose some other genomes to download again. The CVTree result page shows the run-time information and displays the final results when calculation ends. The CVTree web server returns three kinds of result files: (1) a dissimilarity matrix file matrix.txt: this file can be used to construct the phylogenetic trees by calling different programs of the user's choice. (2) Two Newick tree format files: NJtree.nwk for a full tree and Genus NJtree.nwk for a tree 'collapsed' to genus level. These files can be viewed in MEGA and in some other phylogenetic programs. (3) Two ASCII tree files: NJtree.txt for a full tree and Genus_NJtree.txt for a tree collapsed to genus level. These files can be displayed directly in any text editor with monospace font. The result page appears with the five file names listed in the upper part and the NJtree.txt displayed as default in the lower window. By clicking at a file name any of the five files may be displayed. Users get to the tree page by following the 'Show collapsed trees' link in the result page. In this page, we provide trees partially 'collapsed' to certain taxonomic level according to the NCBI taxonomy. The necessity of so doing requires some explanation. At present, the progress of prokaryotic and fungal phylogeny has made detailed comparison with taxonomy feasible. However, it is not easy to comprehend a tree with hundreds or more leaves. To simplify the job, we collapse an original genome tree to different taxonomic levels taking monophyleticity of branches as a guiding principle. For example, at the phylum level the 36 species/strains classified as Cyanobacteria do form a monophyletic branch. This branch is replaced by a single node labeled by Cyanobacteriaf36g. The reduction can only be partial, as, for example, the phylum Proteobacteria does not appear as a monophyletic group in a tree. However, three out of the five classes in this phylum do form monophyletic branches. Therefore, Alphaproteobacteriaf104g, Betaproteobacteriaf62g and Epsilonproteobacteriaf24g nodes appear when the tree is collapsed to class level. In this way, the number of leaves in a collapsed tree may be greatly reduced. The collapsing process requires the knowledge of organism lineage. The NCBI Taxonomic Browser, though disclaimed to be a taxonomic reference, is, in fact, more dynamic and up-to-date as compared to the Taxonomic Outline of Bacteria and Archaea (TOBA) (17) or the Bergey's Manual (21) . That is why we download taxonomic information from NCBI. Since the Genus_NJtree in the result page is generated according to the genus part of an organism's binomen, it might be different from the Genus tree given in the tree page. For example, according to NCBI taxonomy the genus Aliivibrio contains also the species Vibrio fischeri, which is classified under genus Vibrio in TOBA. Therefore, in the genus tree in the tree page we see both Aliivibriof3g and Vibriof7g, however there is only Aliivibriof1g but no Vibriof9g in the Genus_NJtree in the result page. The neighbor-joining program or other treeing software does produce branch lengths from the dissimilarity matrix generated by the CVTree method. However, as the calibration of branch length in CVTree is a subject of current research, we recommend users pay more attention on the tree topology than branch lengths. This is especially true for the collapsed trees as the collapsing is carried out on the NJtree files directly without redefining distances. Although the tree page appears only as a table of file names, the files can be displayed online by clicking at their names. Some tree files, listed at the lower part of the tree page, are directly uploadable to a user's iTOL personal account in order to be displayed in a different manner. In particular, the NCBI taxonomy information may be seen on the branches in the iTOL tree. All the files in the result page and tree page are sent to the user if an email is given in the project page. More examples of output trees can be found in the online user's manual. The new CVTree web server comes with a greater, monthly auto-updated inbuilt database, with a more user-friendly and intuitive interface and a faster data processing pipeline. A phylogenetic tree of more than 900 genomes will be calculated in several hours if the job runs from scratch. Subsequent calculations take much less time if the genome list coincides largely with a previous job. CVTree also provides a useful tool to find the phylogenetic position of the user's-specific genome data. However, there are still many eukaryote genomes not included in the new CVTree web server. These genomes will be put online when the CV method has been fully tested on these data. We will further improve the implementation of CVTree to meet the need of efficiently processing thousands of genomes. Suggestions and comments are welcome.

Remission of Invasive, Cancer Stem-Like Glioblastoma Xenografts Using Lentiviral Vector-Mediated Suicide Gene Therapy

BACKGROUND: Glioblastoma is the most frequent and most malignant primary brain tumor with a poor prognosis. The translation of therapeutic strategies for glioblastoma from the experimental phase into the clinic has been limited by insufficient animal models, which lack important features of human tumors. Lentiviral gene therapy is an attractive therapeutic option for human glioblastoma, which we validated in a clinically relevant animal model. METHODOLOGY/PRINCIPAL FINDINGS: We used a rodent xenograft model that recapitulates the invasive and angiogenic features of human glioblastoma to analyze the transduction pattern and therapeutic efficacy of lentiviral pseudotyped vectors. Both, lymphocytic choriomeningitis virus glycoprotein (LCMV-GP) and vesicular stomatitis virus glycoprotein (VSV-G) pseudotyped lentiviral vectors very efficiently transduced human glioblastoma cells in vitro and in vivo. In contrast, pseudotyped gammaretroviral vectors, similar to those evaluated for clinical therapy of glioblastoma, showed inefficient gene transfer in vitro and in vivo. Both pseudotyped lentiviral vectors transduced cancer stem-like cells characterized by their CD133-, nestin- and SOX2-expression, the ability to form spheroids in neural stem cell medium and to express astrocytic and neuronal differentiation markers under serum conditions. In a therapeutic approach using the suicide gene herpes simplex virus thymidine kinase (HSV-1-tk) fused to eGFP, both lentiviral vectors mediated a complete remission of solid tumors as seen on MRI resulting in a highly significant survival benefit (p<0.001) compared to control groups. In all recurrent tumors, surviving eGFP-positive tumor cells were found, advocating prodrug application for several cycles to even enhance and prolong the therapeutic effect. CONCLUSIONS/SIGNIFICANCE: In conclusion, lentiviral pseudotyped vectors are promising candidates for gene therapy of glioma in patients. The inefficient gene delivery by gammaretroviral vectors is in line with the results obtained in clinical therapy for GBM and thus confirms the high reproducibility of the invasive glioma animal model for translational research.

Glioblastoma is the most frequent and most malignant primary brain tumor. Despite advances in neurosurgery, radiation and chemotherapy, the prognosis of patients remains poor with a median survival of 14 months [1] . A major drawback in translational brain cancer research has been the lack of suitable animal models. Syngeneic-or xenograft tumors based on glioblastoma cell lines cultured as monolayers grow as circumscribed and highly angiogenic lesions in vivo [2] , lacking the invasive tumor cells, which represent an important feature of human glioblastoma. The invasive cells migrate away from the initial tumor mass and can cause recurrent tumors in different regions of the brain. Thus, these cells represent a major therapeutic target. A recently established model in which glioblastoma biopsybased spheroids are serially passaged in the brains of nude rats shows highly invasive and angiogenic features [3] . Therefore, this model is well suited for the study of new therapeutic strategies. Still, reports using this or other clinically relevant models for experimental therapy are scarce. Recently, we analyzed the therapeutic potential of the HSV-1-based oncolytic Herpes vector G207 in the biopsy spheroid-based GBM model. The tumor volume in treated animals was reduced compared to control groups, but there was no significant survival advantage [4] . In contrast, the same therapy was more effective in a cell line-based animal model [5] and as a result is currently investigated in a phase I/II clinical study [6] . In the present investigation we used the invasive xenograft model to evaluate transduction and therapeutic efficacy of lentiviral pseudotyped vectors. Gammaretroviral vectors derived from the Moloney murine leukemia virus (MMLV) have been the most frequently used retroviruses for gene therapy of brain tumors [7] [8] [9] [10] . However, despite promising results in animal models, clinical trials using retroviral vector supernatants or retroviral packaging cells have failed [11] [12] [13] . One major drawback of gammaretroviral vectors is the exclusive transduction of dividing cells, since in human gliomas, the majority of tumor cells do not divide within a given treatment window. Therefore, lentiviral vectors with their ability to also transduce non-dividing cells are attractive candidates for the treatment of brain cancer. In previous studies, we have developed gammaretroviral and lentiviral vectors pseudotyped with the glycoproteins (GP) of the lymphocytic choriomeningitis virus (LCMV) [14, 15] . These vectors have a broad host range and can be concentrated by ultracentrifugation for in vivo applications. In addition, LCMV-GP is not cytotoxic, and stable recombinant packaging cell lines can be established [16, 17] . Recently, we showed that lentiviral LCMV-GP pseudotypes efficiently delivered transgenes to rat glioma cells in vivo, while resident neurons were not transduced [18] . Furthermore, we showed a significant therapeutic effect of LCMV-GP pseudotyped lentiviral vectors in the cell-line based 9L rat glioma model using the suicide gene HSV-1-tk. VSV-G lentiviral pseudotypes also showed a significant efficacy, similar to that of LCMV pseudotypes, which was mainly mediated by a bystander effect of transduced normal brain cells [19] . In the presented work, we showed that both, VSV-G and LCMV-GP pseudotyped lentiviruses efficiently transduced human glioma cells in vitro and in vivo, whereas gammaretroviral transduction was inefficient. The gene transfer to glioma cells was efficient for both lentiviral pseudotyped vector types. However, it was more specific using LCMV-GP pseudotyped vectors, as VSV-G pseudotypes also transduced host brain cells in invasive areas. Analysis of transduced tumor cells revealed that both lentiviral vectors targeted CD133-positive as well as CD133negative cancer cells. Furthermore, transduced glioblastoma cells expressed the stem cell markers nestin and SOX2. Importantly, when evaluated for therapeutic application using HSV-1-tk as a transgene, both lentiviral vectors mediated complete tumor regression on MRI, resulting in a highly significant survival benefit (p,0.001) compared to the control groups. The collection of human biopsy tissue was approved by the regional ethical committee. The handling of the animals and the surgical procedures were done in accordance with the Norwegian Animal Act and the local ethical committee approved the protocol. The human embryonic kidney cell line 293T (ATCC number CRL-11268) and the TE671 cell line were obtained from the American Type Culture Collection (ATCC, Manassas, VA) and maintained in Dulbbeco's modified eagle medium (DMEM) supplemented with 10% fetal calf serum (FCS) and 1% glutamine. All cell lines were grown at 37uC in a humidified atmosphere of 5% CO 2 . Tumor fragments from glioblastoma multiforme patients were obtained during surgery. Tissue specimens were taken from viable tumor areas that corresponded to regions with contrast enhancement on preoperative MRI-scans. The specimens were transferred to test tubes containing complete growth medium, and spheroids were prepared as previously described [20] . The same method was applied for tumor material passaged in nude rats. Briefly, tissue samples were minced into ,0.5 mm fragments and placed into 80 cm 2 tissue culture flasks (Nunc, Roskilde, Denmark) base-coated with 0.75% agar (Difco, Detroit, MI). The spheroids were maintained in a standard tissue culture incubator with 5% CO 2 and 100% relative humidity at 37uC. The medium was changed once a week. Spheroids with diameters between 400 and 600 mm were selected for in vitro experiments and for intracerebral implantation. Tumors were dissociated using the Neuronal Dissociation Kit (Miltenyi, Bergisch-Gladbach, Germany) according to the manufacturer's protocol. Cells were analyzed and sorted on a MoFlo cell sorter (Beckman Coulter, USA; former Cytomation, USA), equipped with a Coherent Enterprise 621 argon-ion laser tuned to 488 nm (used at 180 mW), and 635 nm Diode (25 mW). Two mg/ml propidium iodide -PI (Molecular Probes) were added to the samples before flow sorting to facilitate dead cell discrimination. The GFP and PI were excited at 488 nm and fluorescence was measured through 530/40 BP and 613/20 BP optical filters (all filters from Omega Optical, Brattleboro, VT, USA), respectively. Doublets were discriminated using a forward light scatter (FSC) versus pulse width. FL3 channel (in logarithmic mode) with FSC were used to display and gate out PI positive/dead cells. FSC and side light scatter (SSC) signals were detected and shown in linear mode. GFP+ cells were defined on SSC versus FL1 (in logarithmic mode) dot plot after exclusion of dead cells and debris as described above. For analysis of CD133 expression cells were stained with allophycocyanin (APC) conjugated monoclonal CD133/1 (AC133) antibodies (Miltenyi, Bergisch-Gladbach, Germany), according to the manufacturer's general protocol for immunofluorescent staining (for 10 min in the dark at 4uC). CD133-APC was excited at 635 nm, the fluorescence was measured through 670/ 30 BP optical filter, and alive CD133+ cells were defined on SSC versus FL6 (in logarithmic mode) dot plot. Non-stained cell suspension was used as a control. GFP+ tumor cells were sorted in ''purify 1'' mode into polypropylene round-bottom Falcon tubes (Becton Dickinson Labware Europe, France) containing culture media, that were placed on ice. Aliquots from some samples at the end of the sorting were removed and reanalyzed for control of the sort purity that was greater than 98%. Sorted cells were either cultured in neurobasal medium (Invitrogen, Carlsbad, CA) with B27 supplement (20 ml/ml; Invitrogen), Glutamax (10 ml/ml; Invitrogen), fibroblast growth factor 2 (20 ng/ml; Peprotech, Rocky Hill, NJ), epidermal growth factor (20 ng ml; Peprotech) or transferred to DMEM supplemented with 10% fetal calf serum (FCS) and 1% glutamine and grown on cover slips in 24 well plates. Immunofluorescence staining of spheroids/adherent cells Spheroids/adherent cells were stained with human-specific mouse-anti-nestin antibodies (Millipore, Billerica, MA), goat-anti-SOX2 antibodies (R&D), mouse-anti-b2tubulinIII (Millipore) antibodies and mouse-anti-GFAP antibodies (Millipore). Primary antibodies were incubated overnight at 4uC. Alexa-Fluor647-goatanti-mouse und Alexa-Fluor647-donkey-anti-goat antibodies (Dianova, Hamburg, Germany) were used as secondary antibodies over night at 4uC (for spheroids) or for 2h at room temperature (for adherent cells). Spheroids were examined under a fluorescence microscope (Nikon, Tokyo, Japan) and adherent cells were analyzed by confocal scanning laser microscopy (Zeiss, Jena, Germany). The lentiviral vector plasmid pRRL.sinCMVeGFPpre was published by Naldini et al. [21] . The construction of the lentiviral vector pRRL.sinCMV-TK/eGFPpre has been described previously. The retroviral vector pMP71-eGFP-pre has been described previously [22] . The 293T cell line was used for transient lentiviral vector production. The lentiviral vector plasmid pRRL.sinCMV-TK/ eGFPpre (5 mg) or pRRL.sinCMVeGFPpre (5 mg), the HIV gagpol-REV expression plasmid pCMV-dR8.91(12.5 mg) [21] and 2 mg of the envelope expression plasmid pHCMV-LCMV-GP [14] or pCMV-G [23] were cotransfected into 293T cells and concentrated as described previously [18] . For the production of retroviral vectors, 293T cells were transfected with 7.5 mg of pMP71-eGFP-pre, 12.5 mg of pSV-Mo-MLVgagpol, and 2 mg of the envelope expression plasmid pHCMV-LCMV-GP [14] or pCMV-G [23] . Vectors were harvested and concentrated as described previously [24] . Vectors were titered on TE671 cells as described previously [18] . Intracranial implantation of glioblastoma spheroids was done as described previously [25] . Three weeks to one month after implantation, the animals were anesthetized and prepared for vector injection. The skin was withdrawn to reveal the location of the craniotomy. 2 times 10 mL of vector stocks were delivered into the centre of the tumors using a glass syringe (model 701, Hamilton, Bonaduz, Switzerland) secured in a microprocessor-controlled infusion pump (UMP 2-1, World Precision Instruments, Aston, Stevenage, UK). The injection coordinates were estimated after analyzing MRI images for each individual lesion. Vector infusion was done by convection enhanced delivery in the course of 25 min (200 nl/min for 10 min, followed by 400 nl/min for 10 min, and finally 800 nl/ min for 5 min). After infusion, the needle was left in place for 5 min to avoid vector reflux. The needle was slowly retracted and the skinfolds were closed with polyamide surgical thread. Following surgery, rats were allowed to recover in an incubator set at 35uC before returning them to the cages. Rats bearing glioblastoma xenografts were treated by daily i.p. injections of 50 mg/kg ganciclovir (GCV, Roche, Basel, Switzerland). Animals were euthanized and perfused with sterile saline and thereafter with 4% paraformaldehyde. Brains were removed, suspended in 30% sucrose for three days, and then snap frozen in isopentane chilled with dry ice. Coronal sections (12 mm) were prepared on a cryostat. For immunofluorescence analysis, sections were stained with human-specific anti-nestin antibodies (Millipore) for human glioblastoma cells, mouse-anti-NeuN (Millipore) antibodies for neurons, rat specific mouse-anti-nestin antibodies (Millipore) for astrocytes and progenitor cells. Primary antibodies (dilution 1:200) were incubated overnight at 4uC. Biotinylated goat-anti-mouse and goat-anti-rabbit (Vector Laboratories, Burlinghame, CA) were used as secondary antibodies (dilution 1:100) for 2 h at room temperature. Sections were incubated with Extravidin-Cy3 (Sigma, St. Louis, MO) as fluorochrome (dilution 1:200) for 1 h at room temperature. The sections were examined under a fluorescence microscope (Nikon) and analyzed by confocal scanning laser microscopy (Zeiss, Jena, Germany). For analysis of transduction efficacy, consecutive sections (every 20.-30.) throughout the tumors were examined under a fluorescence microscope (Nikon) with an automated stage using 106magnification. The transduction volume was calculated using Nikon Lucia imaging software. Paraffin embedded formalin-fixed tissue sections from rat brain and patient material were placed in xylene bath for 263 minutes, absolute ethanol 263 minutes, 96% ethanol 262 minutes and finally in distilled water for 30 seconds for removal of paraffin and rehydration. Epitope retrieval was performed by heating the sections at 99uC for 20 minutes in 10 mM citrate buffer at pH 6.0. The sections were incubated with a monoclonal human-specific anti-nestin antibody 1:200 in TBS/1%BSA over night at 4uC. A biotinylated goat-anti-mouse antibody (Vector Laboratories) was used as secondary antibody (dilution 1:100) for 1 h at room temperature followed by ABC-complex incubation for 30 min. Sections were developed with with 393-diaminobenzidine (DAKO Cytomation), following the manufacturer's instructions. Using a Bruker Pharmascan 7 Tesla MR scanner (Bruker Biospin, Billerica, MA), axial T2-weighted RARE sequences were acquired (repetition time, 4,200 ms; echo time, 36 ms; slice thickness, 1 mm; field of view, 3.2 cm; matrix size, 2566256; 20 slices). During scanning, the animals were kept under anesthesia with 1.5% isofluorane (Schering-Plough, Kenilworth, NJ) mixed with 50% air and 50% O2. Survival was analyzed by a log-rank test based on the Kaplan-Meier test using SPSS software. Differences between pairs of groups were determined by the Student's t-test. P values,0.05 were considered significant. Human glioblastoma spheroids derived either directly from the patient (low generation) or from serial in vivo passages in the brains of nude rats (high generation), were infected with lentiviral LCMV-GP (5610 4 in 10 ml) or VSV-G pseudotyped lentiviral vectors (both 5610 4 particles in 10 ml) or with retroviral MLV-based vectors pseudotyped with LCMV-GP (1610 5 particles in 10 ml). The vectors were prepared in the same way for in vitro and in vivo experiments (see materials and methods). Both lentiviral vectors transduced patient spheroids and high generation spheroids very efficiently ( Figure 1 ). In contrast, retroviral vectors transduced only a few single cells in high generation spheroids ( Figure 1 ) and failed to transduce patient spheroids (data not shown). In conclusion, both lentiviral vectors are much more efficient in transducing human glioblastoma spheroids in vitro than retroviral vectors. To compare the transduction efficiency of lentiviral and gammaretroviral vectors in vivo, we used a xenograft model that reflects the angiogenic and invasive features of human glioblastoma in situ. The xenograft also expresses the neural progenitor marker nestin and closely recapitulates the histology of the patient tumor ( Figure 2 ). The vectors were injected into the center of progressively growing lesions using microprocessor-controlled stereotactic infusion. The injection coordinates were estimated after analyzing MRI images for each individual lesion. The injection volume applied was 2610 ml and the vector titre 1610 7 / ml for all vectors. Transduction efficiency was evaluated 7 days after vector injection. Both lentiviral pseudotyped vectors showed very efficient transduction of the tumors ( Figure 3A,D) . When analyzed at higher magnification, both LCMV-GP and VSV-G pseudotyped lentiviral vectors showed efficient transgene delivery to nestin-positive tumor cells in solid ( Figure 3B ,E) and invasive tumor areas ( Figure 3C ,F). In contrast, the retroviral vector only transduced a few scattered tumor cells near the injection site (Figure 3 G,H) . For a quantitative comparison of transduction efficiency between the two lentiviral pseudotyped vectors, the GFP-positive areas were measured on histological slides (see material and methods). The total volume of transduced tumor tissue was 7.0563.51 mm 3 for LCMV-GP pseudotyped vectors and 4.0562.04 mm 3 for VSV-G pseudotyped vectors ( Figure 3I) . Although there was a difference in the mean, it was not statistically significant (p = 0.269) due to high interindividual differences (standard deviations). To analyze transduction specificity, histological sections of invasive tumor areas were stained for rat specific markers NeuN (for neurons) and nestin (for astrocytes and progenitor cells). LCMV-GP pseudotyped lentiviral vectors exclusively transduced tumor cells in all invasive areas ( Figure 4A ), while normal brain cells were not transduced ( Figure 4B,C) . Also VSV-G pseudotyped vectors showed specific transduction of tumor cells in some invasive areas ( Figure 4D,E) , however, they also transduced a few host brain cells at other sites ( Figure 4G,H) . To analyze the potential of both lentiviral vectors to infect cancer stem-like cells, transduced tumors were enzymatically dissociated and CD133 expression was measured by flow cytometry. Both vectors transduced CD133-positive and CD133-negative cells ( Figure 5A ). Although there were high interindividual differences in the fraction of total CD133-positive tumor cells in the different xenografts, both vectors showed similar efficiency in transducing CD133-positive cells, which was slightly higher than the overall fraction of CD133-positive cells (table 1) . The GFP-positive cells from tumors transduced with LCMV-GP or VSV-G pseudotyped lentiviral vectors were sorted and cultured in neural basal medium supplemented with EGF and bFGF. Transduced tumor cells from both vectors were able to form spheroids ( Figure 5B,E) . Spheroids expressed the stem cell markers nestin and SOX2 ( Figure 5C,D,F,G) . Sorted cells were also plated under serum conditions. The cells continued to show significant expression of the stem cell markers nestin and SOX2, but also of the differentiation markers GFAP and b-tubulinIII ( Figure 5H-O) . To evaluate the therapeutic efficacy of both lentiviral pseudotyoped vectors in the invasive xenograft model, vectors expressing the suicide gene HSV1-tk fused to eGFP were injected into established tumors when visible on MRI using the same method as described for the in vivo tropism study. The animals were treated daily with 50 mg/kg ganciclovir for 30 days starting 7 days post vector injection. Tumor growth was measured every 7-14 days by MRI. After 4 weeks of treatment, 7 out of 8 animals in both the LCMV-and the VSV-pseudotype treated groups had a complete remission on MRI ( Figure 6 ). One animal in each group had a stable disease until the end of GC treatment. All animals in the control groups developed large tumors during the treatment period of 30 days ( Figure 6 ). Both, LCMV-and VSV-pseudotype treated animals had a highly significant survival advantage (p,0.001) compared to the control groups ( Figure 7A ). There was no statistically significant difference in survival between the two treatment groups. Upon cessation of prodrug administration, all animals developed recurrent tumors, which could be classified into three different groups ( Figure 7B-E, table 2) . The animals showed either 1) local recurrences or 2) local and/or contralateral recurrences or 3) recurrences in other distant brain areas. There was no clear difference in the recurrence pattern between the two vector types, but LCMV-pseudotyped vector-treated animals had more contralateral recurrences, whereas VSV-G pseudotyped vector-treated animals had more local recurrences ( Figure 7B -E, table 2). Histopathological and confocal microscopic analysis of the lesions revealed GFP-positive cells in all recurrent tumors demonstrating that not all transduced glioma cells were killed by ganciclovir treatment (Figure 7F-M) . In VSV-G pseudotype-treated animals, the GFP-positive surviving cells were found in invasive areas ( Figure 7F ,G), the corpus callosum region ( Figure 7H ) and also in some regions of distant recurrences. One animal also showed a focus of transduced normal brain cells at the tumor border that survived GC treatment ( Figure 7I ). In the LCMV group, most GFP-positive cells were found in the ipsilateral hemisphere, in solid and invasive tumor areas ( Figure 7J ,K,L), with only a few cells seen in the contralateral hemisphere ( Figure 7M ). Future success of glioma gene therapy will depend on more potent vector systems that show higher transduction efficiency than the systems that are available today. In addition, the application of representative animal models that recapitulate both, the invasive and angiogenic features of patient tumors, is vital in order to minimize the huge discrepancies between the experimental results and clinical outcomes previously observed for gene therapeutic strategies for brain cancer. To this end, we applied one of the most clinically relevant animal models for glioblastoma known. This model was established several years ago [20] and its growth pattern as well as geno-and phenotypic similarity to glioblastoma in patients has been extensively characterized [3] . A striking difference of our model compared to other cell-line based models is the highly invasive behaviour of the lesions, similar to glioblastoma in patients. Our model is based on spheroids derived from patient biopsies that are passaged serially in the brains of nude rats. First generation tumors are highly invasive and grow without signs of angiogenesis. Late generation tumors show an angiogenic phenotype, but are still invasive. Our in vitro experiments revealed that both lentiviral vectors transduced spheroids derived from both low and high generation tumors very efficiently. In contrast, retroviral vectors transduced only high generation spheroids and displayed a much lower efficiency of gene transfer than both lentiviral vectors. This difference can only be attributed to the vector backbone, as the glycoprotein which is responsible for virus entry into the cell was the same for the retroviral and one of the lentivral vectors applied (LCMV-GP). The most important feature that distinguishes lentiviral from retroviral vectors is their ability to infect non-dividing cells. It is known that glioma spheroids, especially primary biopsy spheroids, contain a significant fraction of non-dividing cells, which cannot be transduced by retroviral vectors. It has previously been shown that the cultured biopsy spheroids show a similar cell proliferation as seen in glioblastoma patients [20] . Previous studies have demonstrated that retroviral vectors can very efficiently transduce highly proliferative monolayer cultures of glioma cell lines [14] . However, monolayer cultures change their geno-and phenotypic characteristics under long term culture and thus are not a suitable model to answer clinically relevant questions [26] . For in vivo experiments, we selected a high generation xenograft that showed both angiogenic and invasive features and recapitulated the histology of the patient lesion. Furthermore, this xenograft showed a high level of nestin expression similar to the patient material. In translational research it is crucial to assure that the experimental tumors truly reflect the corresponding patient's tumor properties to avoid using non-relevant animal models. However, this strategy is not common practice yet, based on the simplicity of using established cell lines for in vivo experiments. We showed a high transduction efficiency of lentiviral vectors for glioma cells in vivo, whereas retroviral vectors transduced a few scattered tumor cells near the injection track. This is in contrast to in vivo studies by others where retroviral vectors were very efficient, but as mentioned above, the models applied were non-invasive, based on cell lines cultured as monolayers [27] . Of note, the results of clinical studies using retroviral vectors showed the same low transduction efficiency as observed in our model system [12] . This finding provides further evidence that the glioma model used here has a higher predictive value for the performance of a novel therapeutic approach in the clinic than previous animal models. The tropism for glioma cells was more specific with LCMV-GP lentiviral pseudotyped vectors, as VSV-G pseudotyped lentiviral vectors also transduced few normal brain cells in invasive areas. In previous studies using a rat glioma model, we also showed a more specific transduction of glioma cells by LCMV-GP pseudotyped vectors compared to VSV-G pseudotyped vectors [18, 19] . In fact, in these studies, the VSV-G pseudotyped vectors transduced normal brain cells at a much higher frequency than in this study. This can be explained by the mode of vector delivery, because in the previous studies, we injected the vectors both into the tumor core as well as into tumor border areas. In the present study, we injected the vectors into the tumor core only using convection enhanced delivery. We used this method because it results in a high distribution volume of drug and vector and is currently used as a delivery method in clinical studies [28] . In addition, the tumor model we apply here is highly invasive and lacks a sharply demarcated brain tumor/normal brain interface, present in the rat glioma model. Another explanation could be the species difference as we used human glioma cells in this study in contrast to rat glioma cells in the previous work. VSV-G pseudotyped vectors might have a higher tropism for human glioma cells than for rat normal host cells. However, as the receptor for VSV-G is unknown [29] , this remains a hypothesis. The targeting of cancer stem cells or cancer stem-like cells in human tumors including glioblastoma has recently evolved as a major aim in cancer therapy. These stem cells are suggested to initiate cancer and might be resistant to therapy, thus being responsible for tumor recurrence [30] [31] [32] . Yet, recent studies have initiated a controversial discussion whether cancer stem cells really exist [33, 34] . Therefore, we use the term ''cancer stem-like cells'' in our study to designate cells which have certain stem-like properties described previously. We showed that both lentiviral vectors transduced CD133-positive and CD133-negative cells. CD133-positive cells have been identified as cancer initiating cells and cancer stem cells in many different cancers including glioblastoma [30, 35] . However, more recent reports questioned these findings showing that CD133-negative popopulations can include cancer initiating cells as well [36, 37] . The efficient targeting of CD133-positive and CD133-negative glioma cells has also been described for adenoviral vectors [38, 39] . We further demonstrated that sorted cells from tumors transduced either by lentiviral LCMV-GP or VSV-G pseudotyped lentiviral vectors had the ability to form spheroids upon culturing in neural basal medium supplemented with EGF and bFGF. Spheroids from both sorted cell populations expressed the neural stem cell markers nestin and SOX2 and showed the ability to express the differentiation markers GFAP and b-tubulinIII under serum conditions. These properties have been described for neural progenitor cells [40] as well as for cancer stem-like cells in human glioblastoma [26] . Thus, the cell populations transduced by LCMV or VSV pseudotyped lentiviral vectors showed features of cancer stem-like cells which might be an important target for therapy. In a therapeutic approach using the suicide gene HSV-1-tk fused to eGFP, we demonstrated a highly significant therapeutic effect for both lentiviral vectors compared to control groups. Using MRI to follow tumor growth, we detected complete remission in 7 out of 8 animals for LCMV-GP and VSV-G pseudotyped vectors after 30 days of ganciclovir treatment. HSV-1-tk has been reported to be an effective therapeutic gene in previous studies [19, 41] . The limited success in clinical studies has been a result of inefficient gene delivery systems rather than lack of efficacy of the suicide mechanism [11] . However, there is still space for improvement of the prodrug delivery such as application length, treatment intervals and route of delivery. Tai et al. demonstrated that multiple cycles of prodrug application are superior over one cycle of prodrug [42] . In our study, we still detected GFP-positive tumor cells after one cycle (30 days) of ganciclovir administration in treated animals indicating that application in cycles might also improve the therapeutic effect in this setting. Further, these results demonstrate that vector-transduced tumor cells retain the ability to invade brain tissue and migrate even to distant brain regions. In conclusion, the present study demonstrates an efficient transduction and therapy of experimental human glioblastoma by lentiviral vectors. The inefficient gene transfer by gammaretroviral vectors is in line with the results obtained in clinical trials and thus confirms the high relevance of the spheroid-based glioma animal model for translational research.

A Microarray Based Approach for the Identification of Common Foodborne Viruses

An oligonucleotide array (microarray) incorporating 13,000 elements representing selected strains of hepatitis A virus (HAV), human coxsackieviruses A and B (CVA and CVB), genogroups I and II of Norovirus (NV), and human rotavirus (RV) gene segments 3,4,10, and 11 was designed based on the principle of tiling. Each oligonucleotide was 29 bases long, starting at every 5th base of every sequence, resulting in an overlap of 24 bases in two consecutive oligonucleotides. The applicability of the array for virus identification was examined using PCR amplified products from multiple HAV and CV strains. PCR products labeled with biotin were hybridized to the array, and the biotin was detected using a brief reaction with Cy3-labeled streptavidin, the array subjected to laser scanning, and the hybridization data plotted as fluorescence intensity against each oligonucleotide in the array. The combined signal intensities of all probes representing a particular strain of virus were calculated and plotted against all virus strains identified on a linear representation of the array. The profile of the total signal intensity identified the strain that is most likely represented in the amplified cDNA target. The results obtained with HAV and CV indicated that the hybridization profile thus generated can be used to identify closely related viral strains. This represents a significant improvement over current methods for virus identification using PCR amplification and amplicon sequencing.

Polymerase chain reaction (PCR) coupled to reverse transcription (RT) represents the most significant improvement in the area of RNA virus detection over classical cell culture based methods. In the classical culture based method, the principal mode of virus identification uses growth of the virus in permissive cells and observation of the morphological changes brought about by virus replication in the host cell [1] . Although it is possible to differentiate between cytopathic and non-cytopathic hepatitis A virus (HAV) strains due to a difference in the morphology of infected cells [2] , in practice such morphological identification is of limited value because the morphological effects are cell-line specific, and many viruses in the same genus (e.g. Enterovirus) produce rapid and similar cytopathic changes in many of the celllines normally used for virus detection. Moreover, using multiple cell-lines for virus detection is also labor intensive and time consuming, and further confirmation and identification often requires the use of additional techniques such as serotyping [1] . Molecular methods based on viral RNA amplification by RT-PCR have evolved as rapid alternatives to cell culture for the detection and identification of viral strains [3] . For *Address correspondence to this author at the Division of Molecular Biology, Office of Applied Research and Safety Assessment (OARSA), Food and Drug Administration, 8301 Muirkirk Road, HFS-025, Laurel, MD, 20708, USA; Tel: +1-301-210-7812; E-mail: biswendu.goswami@fda.hhs.gov example, the differential identification of strains within a species is possible based on the difference in the size of the amplified PCR product (amplicon) detectable by gel electrophoresis ( [4] or single-strand conformational polymorphism (SSCP) [5] . Indeed, we have utilized agarose gel electrophoresis following RT-PCR using primer pairs straddling a 14 base insertion at the non-coding region of some HAV genomes to identify specific cytopathic strains from noncytopathic strains of HAV [4, 6] . We also reported the use of SSCP analysis following Alu 1 or Hinf 1 digestion of amplicons generated from the 3' end of the viral genome to provide differential identification of multiple HAV strains [5] . However, SSCP is a multi-step procedure involving radiolabeling of restriction fragments prior to electrophoretic separation of individual DNA strands. Consequently, this procedure works best when the restriction fragments are small enough to provide sufficient single-stranded DNA separation for effective strain identification. For genetically wellconserved viruses such as HAV, the region to be amplified for SSCP analysis has to be carefully chosen in order to represent areas of reasonable diversity [7, 8] . Due to these considerations, it has been preferable to sequence the PCR amplified DNA fragment in order to specifically identify the genotypes or strains of the viruses. While sequencing amplified PCR products is considered a precise technique for identification, PCR amplification of a mixed population of target sequences may be biased in favor of a dominant (by copy number) target such that subsequent sequence analysis may not reveal the presence of other closely related target sequences in starting populations. Putative mixed virus populations (e.g. of the same or different species) can exist in isolates obtained from environmental and infected-host samples particularly those resulting from RNA virus replication that is known to generate a sub-population of "quasi-species" [9] . Therefore, a threshold number of RNA molecules must have the same specific mutation in order to be unambiguously detectable by RT-PCR and sequencing, due to possible inhibition of amplification of a less abundant template by template competition [10] . Conversely, the dominant mutation present in a population may be preferentially amplified, and therefore, sequence analysis would represent the dominant mutant [11] . Therefore, while sequencing remains a "goldstandard" for target sequence identification, the identification of multiple viral species or tracking species mutations necessitated the development and application of a broader approach to identification prior to undertaking sequence analysis. As an alternative to sequencing, Proudnikov et al. [12] applied a hybridization-based technique to the detection of genetic variants of poliovirus within a virus population or among viral strains. Oligonucleotide probes are synthesized and then immobilized on a solid surface. A target consisting of amplified viral complementary DNA (cDNA) then labeled and hybridized to the immobilized probes and the hybridization to the individual probes detected [12] . The presence of a change in the nucleotide sequence in the target is detected by the absence or the reduction of hybridization to the wild type probes around the change, or by the ratio of the signals generated by a mutant against a reference strain. Modifications of the above technique including the use of amplified viral complementary RNA (cRNA) were used to identify genetic variations arising during cultivation of a vaccine strain of poliovirus and the emergence of vaccine derived poliovirus in immunized patients showing signs of vaccine associated paralytic poliomyelitis [13, 14] . Application of this procedure was restricted, however, to identifying known mutations in specific virus strains. Advances in microarray technology have allowed the identification of genetic variability over very long stretches of DNA in bacterial genomes [15] . These newly developed high density microarrays contain thousands to hundreds of thousands of oligonucleotide probes, instead of a few dozen, in a single array thereby expanding the power of identification [15] [16] [17] . In the current investigation we report the design and use of a high density oligonucleotide microarray for the identification of HAV and coxsackievirus (CV), both foodborne human pathogens. Our results indicate that the microarray hybridization technique can be applied to the identification of viruses of differing genus and species present in a sample and detect single nucleotide polymorphisms (SNP) to identify closely related viral strains belonging to the same species. Viruses and Plasmids. Hepatitis A virus strains HM175/clone 1 and 18f, and coxsackievirus (CV) serotypes B1, A3 and A5 strains used in this study were obtained from ATCC (Manassas, VA) and further grown in FRhK4 cells [18] . The plasmid pHAV/7 contains a full length cDNA copy of wild type HAV strain HM175 cloned into the vector pGEM-1 [19] that was grown and purified as previously described [20] . HM175 clone1 and 18f are culture-adapted strains derived from continuous culture passage of the wildtype strain HAV HM175 [21] . Oligonucleotide Arrays. All microarrays used in this study were manufactured by NimbleGen Systems Inc. (Madison, WI) using a maskless array synthesis (MAS) technology for in situ synthesis of DNA oligonucleotides directly onto glass microscopy slides [16, 17] . Oligonucleotide design was based on available complete viral genome sequences obtained from GenBank for CV (n=25), HAV (n=23), Norovirus genogroup I (n=4), Norovirus genogroup II (n=21), rotavirus (various species) segments 3 (n=11), 4 (n=19), 8 (n=11), and 11 (n=12) where n equals the number sequences obtained for each virus group. All genomic sequences within a virus group were aligned using CLUSTALX [22] , and dendrograms were generated and consensus sequences constructed based on these analyses. Examples of these dendrograms are shown for HAV and CV (Figs. 1, 2 , respectively). For the purpose of generating representative viral genomic sequences on which to base subsequent oligonucleotide designs, the HAV strains were clustered into 5 groups whose viral genome sequences were constructed as follows: i) a consensus sequence based on the seven genotype Ib (i.e. genotype I, subgenotype b) strains that clustered into group 1 which includes the HAV HM175/wt strain (M14707), ii) a sequence derived from M20273 based on the pairing of M20273 and AF314208 (genotype Ib sequences in group 2), iii) a sequence derived from the single HAV genotype II sequence (IIb) available (AY032861) and assigned as group 3, iv) two consensus sequences based on either cluster group 4 or 5 derived from fourteen genotype Ia sequences that were clustered into either of these two groups. The three consensus sequences representing cluster groups 1, 4 and 5 were obtained following a group sequence alignment and the assignment of the most frequently occurring nucleotide at positions containing nucleotide differences. The clustering of either one or two sequences within a group (as in groups 3 and 2, respectively) resulted in the selection of a single sequence representing that group. Due to the highly diverse (genetic) nature of the CV genome sequences, clustering of strains for generating a group consensus sequence was only done for serotype strains B1 and B3 (groups 1 and 2, respectively). Four additional unique strain sequences were selected as representative sequences for broadly clustered strains identified as groups 3-6. Viral genomic sequences (approximately 3000 bases) from either the 3' end of the HAV genome group sequences or the 5' end of the CV genome group sequences were submitted for design of a tiling oligonucleotide array consisting of oligonucleotides of length 29, starting at every 5 th base in every sequence, resulting in an overlap of 24 bases in two consecutive oligonucleotides [15] . Similar methods were applied to the development and tiling of oligonucleotides as probes for norovirus and rotavirus sequences on the array. The resulting array contained approximately 13,000 viral probes. All reverse transcription (RT) reactions were completed using RNA templates obtained from linearized plasmid pHAV/7 transcribed in vitro with SP6 polymerase, total cellular RNA (1 g) isolated from virus infected cells using the RNA AqueousKit (Ambion, Austin, TX), or viral genomic RNAs (equivalent to 5 x 10 6 infectious particles) isolated directly from clarified tissue culture supernatants using the Fig. (1) . Dendrogram showing the grouping of HAV strains based on their genetic relatedness for developing viral probe sequences to be used for oligonucleotide design. HAV strains are identified by their GenBank accession number and their respective complete sequences were used to generate the dendrogram. Brackets encompass strain sequences selected to derive a group consensus sequence while arrows identify group individual (i.e. non-consensus) strain sequences for probe set development. Group sequences are designated by numbers 1-5 followed by the probe set identifier (within parentheses), and the genotype (I or II) and subgenotype (a or b) designation. RNeasy Micro Kit (Qiagen, Valencia, CA); a mixture of oligo(dT 15 ) and random hexamers (pdN6) as primers; and AMV reverse transcriptase (Promega, Madison, WI) as previously described [4, 20] . In vitro transcribed and infected cell RNA templates represent in vitro and in vivo replication, respectively. PCR amplification with HAV or CV specific primers was carried out in 50 l reactions using 5 l of each RT reaction as template or 5ng of pHAV/7 plasmid DNA as previously described [20] . PCR products (5 l) were analyzed by agarose gel electrophoresis to confirm authenticity of product formation (data not shown). PCR Primers. Two primers, 3399 -3423 (forward) and 7084 -7105 (reverse), were used to amplify an approximately 3.7 kb region of the HAV genome [4, 6, 20] . Tables 1 and 2 show the sequences around the primer binding sites of selected HAV strains represented on the array. Tables 3 and 4 contain the sequence alignments at the forward and reverse Dendrogram showing the grouping of CV serotype strains based on their genetic relatedness for developing viral probe sequences to be used for oligonucleotide design. CV strains are identified by their GenBank accession number and their respective complete sequences were used to generate the dendrogram. CV serotypes are given in parenthesis following the accession numbers. Brackets encompass strain sequences selected to derive a group consensus sequence while arrows identify group individual (i.e. non-consensus) strain sequences for probe set development. Group sequences are designated by the numbers 1-8 followed by the probe set identifier (within parentheses). Brackets (outer right) indicate which human enterovirus species (HEA, HEB and HEC) are represented by the CV serotype strains used to develop the dendrogram. primer binding sites for selected CV strains. The reverse primer for CV is degenerate owing to sequence differences among strains in this region [23] . These primers amplify a 746 bp fragment from several B and A strains (data not shown). Labeling of PCR Products and Hybridization. PCR products were purified using a spin column procedure [Qiagen or Stratagene, (La Jolla, CA)]. One g of each purified PCR product was labeled with biotin-dUTP in a primer extension reaction using random hexamers and Klenow po- forward primer GAATGATGAGAAATGGACAGAAATG a Strains are identified by their GenBank accession number. b HAV sequence alignment is presented as the positive (sense) genomic strand in 5' to 3' orientation. The primer sequence and nucleotide identity is based on the genomic sequence and nucleotide numbering of HM175 strain 18f (M59808) at nucleotide positions 3399 to 3423. reverse primer CCGAAACTGGTTTCAGCTGAGG a HAV strains are identified by their GenBank accession number. b HAV sequence alignment is presented as the reverse complement (antisense) of the genomic strand in 5' to 3' orientation. The primer sequence and nucleotide identity is based on the genomic sequence and nucleotide numbering of HM175 strain 18f (M59808) at nucleotide positions 7105 to 7084. lymerase (Exo -). Labeled products were purified by spin column chromatography, and concentrated by centrifugation through Microcon ® (Millipore, Billerica, MA) filters. Biotinlabeled DNA was denatured in a total volume of 20 l of hybridization solution containing 5XSSC, 0.1%SDS, 5 g poly A, and 5 g human Cot-1 DNA and 6 l used per hybridization reaction per well of a 12 well sample pod (Nim-bleGen Systems, Inc.). The microarray slide (NimbleGen) was laid on top (oligonucleotide side down) of the sample pod and held in place in a metal cassette provided by the manufacturer. Hybridization was carried out for 12h at 42 °C. The slides were washed sequentially with 2XSSC/0.1%SDS, and 0.1XSSC/0.1%SDS at 42 o C then distilled-deionized water at room temperature. The slides were then stained with a Cy3-streptavidin conjugate (Amersham Biosciences, Piscataway, NJ) as described in Jackson et al. [15] . Data Extraction and Analysis. Hybridized, Cy3-stained microarrays were scanned using an Axon GenePix ® 4200A scanner at 5 m resolution using a 532 nm laser. Fluorescence intensities of each feature (oligonucleotide probe) were extracted utilizing NimbleScan TM software (NimbleGen Systems Inc), and all subsequent data analyses were performed using MS Excel. Data were analyzed independent of comparison to a reference strain assuming that each virus strain is unique. Following normalization for background fluorescence, the fluorescent intensity of each probe (normalized probe intensity) was plotted against the genome position of each probe to generate a hybridization profile for each viral strain [15, 17] . To generate the average probe intensity for each probe set per hybridized virus strain, the sum of all normalized probe intensities for individual probes within a probe set (i.e. set of probes derived from an individual strain or group sequence) was divided by the number of probes within that set [15] . As members of the positive-stranded RNA virus family Picornaviridae, Hepatitis A virus and coxsackievirus belong to the genera Hepatovirus and Enterovirus, respectively. At the nucleotide level there is substantial genetic diversity between these two groups with a greater within group diversity observed for coxsackieviruses than for hepatitis virus strains. Indeed, the coxsackievirus genomes are much less conserved and mutations are distributed throughout the genome [23] . Sequence analyses of small segments of different strains of HAV have led to the recognition of six genotypes (I to VI) of this virus [24] . Genotypes I, II and III have been further subdivided into two subgenotypes, a and b [24, 25] . Within each genotype the strains have greater than 85% sequence homology, whereas subgenotypes may differ from each other in up to 7.5% of nucleotide positions. Genotype I is the most prevalent HAV genotype world wide [25] . For the present investigation, complete HAV genome sequences available at the time of chip design and construction belonging to genotypes I and II were clustered into five groups (Fig. 1) . Two of these groups contain subgenotype Ia (groups 4 and 5) while sequences belonging to subgenotype Ib were clustered into two groups (1 and 2) and a single sequence of genotype II was designated group 3 ( Fig. 1) . As members of the genus Enterovirus, the Human enterovirus A and Human enterovirus C virus species comprise most of the species assigned to CV serotype A strains with the exception of serotype A9 [23] . Coxsackievirus serotypes B1-6, coxsackievirus serotype A9 (CVA9), enteroviruses 69 and 73 and the majority of echoviruses are classified within the Human enterovirus B species that is one of the largest species group in the family Picornaviridae [26] . Fig. (2) shows the CLUSTALX analysis of the different coxsackievirus serotype strains used for designing the tiling array. For B1 and B3 strains a consensus sequence was developed based on the sequences available for individual members of these groups, whereas single sequences were used for the strains B4, B5, A16 and A21, due to extensive sequence variations between individual members of these groups [23, 26] . Since sequence diversity among the CV strains was distributed over the entire length of the genome, the first 3000 nucleotides were determined to contain sufficient sequence diversity to identify a strain without ambiguity. Identification of HAV Genotype by Microarray Hybridization. Fig. (3) shows the hybridization profile obtained with a target synthesized by PCR amplification of the plasmid pHAV/7. This plasmid contains a copy of the entire HAV sequence of wild-type HM175 strain HM175 [19, 27] that originated from an Australian outbreak, and was designated as genotype Ib by subsequent sequence analysis [24, 28] . The hybridization signals (normalized probe intensities) produced a profile indicating areas of intense hybridization at the position where the HAV sequences are clustered in the array. However, variations in the intensity of hybridization can be observed within these sequences, where the target hybridization intensities against group 1 probes (hav1Cb) differ from probes derived from groups 2 through 5 (hav2b, hav3b, hav4Cb and hav5Cb) sequences. This is more clearly observed in Fig. (4) , where the normalized probe intensities for individual probes within each group sequence present in the array were converted to average probe intensities and plotted for the target. The plot reveals that the HAV genotype 1b (HM175 wild-type) target hybridized most efficiently to probes from genotype Ib, group 1 consensus sequence (hav1cb). These results are consistent with the fact that the viral genome sequence for HAV HM 175 wt strain (14707) is a member of, and therefore most closely related to, group 1 derived probe sequences. Probes representing a closely related HAV Ib strain from group 2 (hav2b) hybridize the target about two thirds as efficiently, while probes from the more genetically distant genotype II virus (hav3b) hybridize with the least intensity. The other probe groups (hav4cb and 5cb) both representing genotype Ia consensus sequences (Fig. 1 , cluster groups 4 and 5) hybridize less efficiently than genotype Ib. Given the readily observable differences in both the normalized and average signal intensities among the genotype group sequence probes (groups 1-5) following genotype Ib target hybridization, and the fact that viruses belonging to different subgenotypes can differ by as much as 7.5% in sequence [21, 24, 28] , the data in Fig. (4) indicate that it is possible to identify HAV strains at the level of both genotype and subgenotype with this type of array. To further explore genotype/subgenotype differentiation, different HAV strains belonging to the same subgenotype Ib sequence (Fig. 1, group 1) were hybridized to the array. As shown in Fig. (5) , HAV strains HM175 wt, clone 1 and 18f hybridize most efficiently to genotype Ib (consensus group 1) probes (hav1Cb). Lower efficiencies of hybridization are observed for all three targets against all other probe sets. These results reflect a greater target specificity for the probe set that contains target member sequences than for the other genotype Ib derived probe set (hav2b) that does not contain target member sequences (group 2 in Fig. 1) . For all target strains, the remaining probe sets yielded signal intensities equivalent to or less than intensities for probe set hav2b. Thus, in support of the interpretation of results of Fig. (4) , this array has the potential to discriminate viral targets at the level of both their genotype and subgenotype. It is important to note that despite the variation in average probe intensities for the individual strains against probe set hav1Cb (Fig. 5) , the information as presented cannot be used to identify actual target nucleotide Fig. (4) . Hybridization profile of wild-type HAV HM175 strain target: average probe intensity. The hybridization data (normalized probe intensities) for HAV HM175 from Fig. (3) was converted to average probe intensities [15] and plotted vs each individual probe set. A given probe set represents all probes derived from their respective sequence group. differences. For example, differences in signal height could be attributed to differing hybridization efficiencies between two different experiments. Indeed, a target derived from in vitro synthesized RNA from pHAV/7 (representing in vitro replication of the viral genome) was indistinguishable from plasmid derived target, or the virus following several rounds of replication in culture except for the peak height (data not shown). We pursued, therefore, an alternative method of analysis because the tiling array design offers the potential to distinguish between these closely related strains following hybridization by i) determining the normalized probe intensities for each target, and ii) plotting the change in signal intensity of hybridization by each target to the same probe set as the ratio (fold-change in probe intensity) vs the individual probes. As discussed by Jackson et al. [15] , this method of analysis can reveal distinct peaks with defined slopes (above background/signal noise) where changes in signal strength would occur with probes tiled further up or down stream of the nucleotide change. The presence of a mutation in the genome causes a destabilization of a number of probes around the mutation, which can be identified by the appearance of well defined peaks. Therefore, this method of analysis offers the potential to differentiate closely related strains of virus belonging to subgenotype Ib at the level of individual nucleotide differences, thereby producing data that can be used to tell them apart. In order to complete this analysis, the two different HM175 strains designated clone 1 and the cytopathic 18f strain were again subjected to hybridization and the total normalized intensities of all probes belonging to the different HAV probe groups were plotted as in Fig. (3) . Again, we found no overall differences in the hybridization profile but rather found peaks of hybridization intensities with the strongest hybridization intensities for the group 1 (HAV1Cb) consensus sequence following calculation of average probe intensity (data not shown). The fold-change in intensity between clone 1 and 18f targets was calculated for each probe in the probe set HAV1Cb. As shown in Fig. (6) , ten well defined peaks were observed over the range of the HAV1Cb probe set and the probe number that corresponds to each peak was identified. It is important to note that due to the initial size of the graphical analysis output, it was necessary to compress the scale of the x-axis (HAV1Cb probe number) in order to fit all data points within a smaller graph. As a result, analysis of the hybridization (signal) values revealed two features not readily discernable on the graph; i) a probable single peak at probe 109 rather than what appears as two adjacent (overlapping) peaks, and ii) a possible second overlapping peak adjacent to probe 441. Since the HAV1Cb probe set (group 1) is a consensus sequence developed from the alignment of seven strains assigned to this group (Fig. 1) , there are nucleotide differences between each group member and the consensus sequence. Plotting the fold-change in intensity between clone 1 and 18f would potentially identify nucleotide sequences in a probe that are identical to clone 1 but not identical to 18f. Indeed, upon comparative analysis of clone1 and 18f amplified target sequences with HAV1Cb probes set sequence synonymous with the target sequences, one would predict a total of 11 peaks to occur by this method of analysis. We then sought to determine whether the "peak" probes contained nucleotide differences that could be mapped to nucleotide differences [e.g. single-nucleotide po- Fig. (5) . Comparison of hybridization profiles for three HAV genotype Ib strain targets. Average signal probe intensities were calculated and plotted following hybridization of targets generated as PCR products from reverse transcription of RNA derived from either in vitro transcribed pHAV/7 (black bar), HAV HM175 clone 1 infected cells (white bar) or clarified supernatant from HAV 18f infected cells (grey bar). lymorphisms (SNPs), deletions, or insertions] that exist between clone 1 and 18f (and the probe set). As shown in Table 5, we were able to conservatively detect 10 out of 11 predicted nucleotide changes in the 18f genome identifiable by this method of analysis. It is important to note that these nucleotide changes represent mutations arising in the 18f virus during its emergence as a cytopathic strain from the HM175 noncytopathic strain which were identified by direct sequencing [21] . These results demonstrate a strong correlation between results obtained by direct sequencing and array hybridization and strongly suggest that tiling arrays can be used to detect nucleotide changes instead of sequencing amplified PCR products over a much longer span of the genome in a single experiment. Identification of CV Serotype by Microarray Hybridization. Unlike HAV strains, there is tremendous genetic diversity between CV strains, even within the same species as observed, for example, among serotype B strains although they are all members of HEV species [23, 26] . We sought, therefore, to determine whether this array hybridization technique could be used to identify a CV serotype strain target. A typical hybridization profile with a 746 bp segment amplified from CV strains is shown for CVB1 in Fig. (7, panel A) where the data is presented as average probe intensity for all probes derived from the same group sequence, i.e. probe set. Similar to the results obtained following hybridi-zation with HAV targets, CVB1 targets hybridized very efficiently and with greatest intensity to probes (coxB1Ca) derived from a consensus sequence based on its own sequence, i.e. serotype B1 strains (Fig. 2, group 2) . As indicated by the significantly lower probe intensities, minimal hybridization was observed among the remaining 7 CV probe sets indicating a lower efficiency of hybridization to non-CVB1 sequences represented on the array. In fact, hybridization to probes representing all other (non-CV) viruses was essentially at background signal intensity. The results are consistent with the extensive sequence heterogeneity that exists between the CV serotype A and B virus strains, the members within a serotype (A or B), as well as the probe sets derived from these strains. Importantly, these results demonstrate that even with highly (genetically) diverse viruses, such as coxsackieviruses, this array design can discriminate between strains of the same (or different) virus species. We next sought to determine whether discrimination between virus strains or species was possible when the viral target contains sequences not represented by either an individual or a consensus probe set on the array. To complete this experiment, a 746 bp targets derived from coxsackievirus serotype A3 and A5 strains were hybridized to the array. CVA3 and CVA5 serotype strains are both members of HEA species, however the probes' sequence (group 7, coxA16a) for the species was derived from CVA16 (Fig. 2) . Analysis of normalized probe intensities reveal a striking reduction in the overall level of Fig. (6) . Detection of nucleotide differences between two genetically related HAV strains. Average signal probe intensities were calculated following hybridization of targets generated as PCR products from reverse transcription of RNA derived from HAV HM175 clone 1 infected cells or clarified supernatant from HAV 18f infected cells. The amplified targets (3.7 kb) derived from both clone 1 and 18f contain nucleotide sequences synonymous with the first 2.7 kb (probes 1-543) of the 3.1 kb group 1 (HAV1Cb; Fig. 1 ) consensus sequence used to develop the HAV1Cb probe set that is comprised of 608 probes. The individual points on the graph represent specific probe numbers; however, due to graphical compression of the original data, there are iterative probes not represented by individual points on the graph. Arrows identify the probe number having the peak intensity difference between clone 1 and 18f where a nucleotide(s) present in the consensus sequence is identical to nucleotide(s) in clone 1 but not identical to nucleotide(s) in 18f. probe hybridization (normalized) intensities for CVA3 and CVA5 derived targets compared to those values obtained following hybridization with a CVB1 target (data not shown). As shown in Fig. (7, panels B and C) , this is also observed following conversion to average probe intensity. The peak average probe intensity for these hybridizations is approximately 2750 units and 900 units with CVA3 and CVA5 targets, respectively. The results indicate that in the absence of matching probe sets on the array the sequence heterogeneity between these CV targets and the existing probe sets precludes the establishment of any strong or efficient hybridization to a single probe set. It is important to note, however, that neither of these targets hybridizes with any significance to non-CV probe sets suggesting that the genetic diversity between CV targets and probe sets does not prevent or obscure virus target group (i.e. CV) identification. In addition, these hybridization profiles are not only distinct from B1 (Fig. 7) but also from one another suggesting the possibility that unique hybridization profile patterns (calculated as normalized and/or averaged probe intensity) could be used for CV serotype target identification. The results from Fig. (7) also suggest that in a single experiment it is possible to identify whether a virus belongs to group A or group B. Indeed, identification of coxsackieviruses at the level of serotype strain may be possible without single nucleotide polymorphism (SNP) analysis and limited only by the number of probe sequences/sets present on the array. Currently, RT-PCR is the most widely used molecular method for the detection and identification of viruses in biological and environmental sources [27, 28] . Identification of genotypes of virus strains are based on the amplification of specific regions of the viral genome using gene specific primers followed by sequencing of the amplicon by standard procedures [28] . In some instances a preliminary identification is possible using the techniques of single strand conformational polymorphisms (SSCP) or restriction fragment length polymorphisms (RFLP) [5] . Multiplex PCR allows the detection of more than one species of virus in a single analyte [29] . However, these techniques have limitations on sensitivity and versatility, and require extensive prior knowledge of the sequences to be amplified. The requirement of size differences in the amplicons to be analyzed by gel electrophoresis following amplification by multiplex PCR also limits its utility. Different strains of HAV and many enteric viruses show variable sequence diversity [23, 24, 26] . This allows easy identification of a virus at the genotype level by sequencing discrete segments of the viral genome amplified by RT-PCR. Ideally, sequencing should be done on amplicons that are known to have multiple nucleotide differences between strains. However, designing PCR primers that will capture a significant number of members of that group requires significant sequence homology, and therefore, a relatively variable region flanked by conserved regions is needed for sequence based identification. While for some virus groups such as HAV it is relatively easy to find PCR primers that can capture many members, it is much more difficult with CV genomes due to extreme sequence diversity. The length of the amplified region is another constraint for sequence based identification. Sequencing an amplicon larger that 500 bp generally will require designing multiple primers for sequence walking. Although automated sequencing techniques currently available can be used for rapid sequencing of a moderate sized amplicon, the process is still too time consuming to be used on a routine basis where a quick identification is needed. a Detection of (putative) nucleotide differences by array hybridization between clone 1 and 18f was initially based on the hybridization profile in Fig. (6) . b Probe numbers are from Fig. (6) and represent the oligonucleotide probes whose sequence contains nucleotide change(s) between clone 1 and 18f when the clone 1 target sequence is identical to the probe sequence. Nucleotide changes were indentified (grey boxed) based on comparison of the clone 1 and 18f GenBank sequences (accession numbers in Fig. (1) ) used to develop the HAV1Cb group 1 consensus probe set (Fig. 1) . The probe nucleotide range numbering is defined by the 29-mer probe and corrected to 18f nucleotide numbering from Lemon et al. [21] . c The nucleotide change and position between clone 1 and 18f as reported by Lemon et al. [21] . d Probe 120 defines a 26 nucleotide base region of 18f due to the three-base GAT deletion. e This probe identified as a potential peak overlapping with the peak at probe 441. Fig. (7) . Comparison of hybridization profiles of three CV strain targets: average probe intensity. Viral genomic RNAs isolated directly from clarified tissue culture supernatants of infected cells were used for RT followed by PCR amplification and labeling prior to array hybridization as described in Materials and Methods. The hybridization data (normalized probe intensities) were converted to average probe intensities [15] and plotted vs each individual probe set following hybridization with either CVB1 (panel A), CVA3 (panel B) or CVA5 (panel C). The underlined identifies the human enterovirus species (HEA, HEB and HEC) represented by a CV probe set. We investigated whether hybridization of fluorescently labeled amplified DNA (target) to a microarray containing many oligonucleotide probes representing many different viral genomes can identify a virus without sequencing. Unlike sequencing, these arrays can interrogate thousands of bases of a viral genome in a single experiment [15] [16] [17] . We determined the feasibility of this approach by using labeled targets amplified from either the DNA (i.e. as recombinant plasmid) or RNA from several strains of HAV and CV. As shown in Figs. (4-6) , a single hybridization experiment using a multi-well array with different samples loaded in different wells of a 12-well sample pod can identify HAV and CVB by the unique profile generated with no ambiguity or crosshybridization to oligonucleotides representing an unrelated virus. Within the broad genus of hepatovirus of which HAV is the only species member, different genotypes which differ from each other by 5% to 8% of base positions (Fig. 1) can be identified (Figs. 4-6) . Within the same subgenotye Ib, strains such as wild type HM175 and the cell culture adapted variants including the cytopathic 18f strain differ by only 0.5% of base positions. We have shown that differentiation of these strains is possible by analyzing the ratio of the signal probe intensities generated by the isolates when hybridized to the probe sets present on this tiling array ( Fig. 6 and Table 5 ). A sequence based identification of the same 3.7 kb amplicon would require several sequencing reactions with multiple primers in order to identify nucleotide differences. In addition, mutations accumulating in the HM175 genome during its evolution into the cytopathic 18f strain can be identified by ratio analysis (Fig. 6 and Table 5 ). Thus, the present array design is suitable for identification of species (e.g. CV and HAV) and HAV subgenotypes since in the latter case the nucleotide differences are very few. We have also demonstrated that it is possible to distinguish between CVB and CVA strains by virtue of their hybridization profiles. In addition, individual members of A and B groups show distinct and characteristic hybridization patterns. Thus, members of CVA strains such as A3 and A5 can be easily identified not only as belonging to group A CV, but also a genotype A3 or A5. More virus strains need to be examined to determine if the method is applicable to other members of this group. The more closely the target sequence matches the probe set, the stronger the hybridization signal. Diversity between and among probe sets representing virus strains within a group such as CV increases the power of discrimination (due to heterogeneity) particularly when the target is highly similar to one of the probe sets. This enables discrimination even at least at the strain level (e.g. among strains within the same serotype group such as CV group B serotypes). This level of discrimination is lost when a target whose sequence is not represented by a probe set is not present on the array. Again, this has been shown to be problematic with highly (genetically) diverse viruses such as CV. However, despite the loss of serotype discrimination, the diverse nature of such viruses does still enable the differentiation between virus groups as shown between CVA and HAV, NV, and rotavirus. Our results show that an oligonucleotide array incorporating thousands of probes representing genomes of multiple foodborne RNA viruses including multiple hepatitis A virus genotype strains and multiple coxsackievirus serotype (A and B) strains can be used to differentiate between virus members of either genus to identify the genotype/serotype of these viruses by array hybridization assay. Because the large number of probes can bind and detect labeled targets over a much larger area of the viral genomes, producing distinctive signal patterns for each genotype/serotype, the need for large scale sequencing is eliminated for this level of discrimination.

Is the risk of multiple sclerosis related to the ‘biography’ of the immune system?

Multiple sclerosis (MS) with onset in childhood offers a unique opportunity to study the infectious background of this disease but the immune reactions against infectious agents in such children have only recently been investigated. These and other epidemiological studies strongly implicate involvement of one or more infectious agents in the aetiology of MS, with Epstein-Barr virus (EBV) being the prime candidate. Rather than being the actual cause of MS, it is more probable that these agents are involved in the development of immunoregulatory pathways. These pathways, if disturbed by hygiene-related factors including an altered sequence of infections, may generate and maintain a deficit within the immunological network that facilitates, to particular early events in the development of MS, preceding the onset of MS disease by years or a decade. A framework that can serve as a guide for further epidemiological, immunologic and molecular biologic investigations is formulated. This approach may shed light on the complex natural history of MS and may lead to rational preventive and therapeutic strategies. It is possible that, in the future, MS could be prevented by vaccination against EBV in early childhood; the framework is of relevance to the design of an appropriate type of vaccine.

The pathogenesis of multiple sclerosis (MS) is certainly complex and heterogeneous in nature [1] , involving an interplay between innate and environmental factors [2] [3] [4] and genetic factors, notably the polymorphism of HLA [5] . The epidemiology of this disease strongly indicates that it has emerged as a major neurological disorder in industrially developed nations over the last 150 years and is likely to be affected by hygiene-related factors [3, 6] . According to the so-called hygiene hypothesis [7] , modern living conditions in the industrialised nations isolate infants and children from many infectious challenges that are required for the development of appropriate immunoregulatory networks. This hypothesis has been advanced to explain the rise in incidence of disease associated with immune dysregulation, including asthma, allergy, autoimmunity and at least some forms of cancer, in the industrially developed world [7] . In addition, the hypothesis suggests strategies for interventions for the prevention of such diseases, as illustrated by studies on the epidemiology of malignant melanoma which indicate that certain vaccinations, BCG, vaccinia and yellow fever, can substitute for the significant protective effect of natural B. Krone infections [8 -11] . As epidemiological investigations on MS strongly indicate that this disease is likewise affected by hygiene-related factors and by a history of infections [3, 6] , the challenge is to determine whether there is an underlying distortion of immune responses in this complex disease that could be prevented or corrected by therapeutic intervention. Evidence from epidemiological studies of a role for any of the many common infections in the aetiology of MS is largely inconsistent [3, 6] . This could be due to the prolonged silent stage of the disease, but recent studies on MS commencing in childhood might be able to shed light on this issue [6, [12] [13] [14] . Although there is no convincing evidence that any specific active infection directly causes MS, at least 14 specified infections have demonstrable associations with this disease on the basis of serological characteristics [2, 3, 6, [12] [13] [14] [15] , with Epstein-Barr virus (EBV) emerging as the most likely single candidate for a leading aetiological role [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] . This infection is unique in that the prevalence of specific antibody is consistently higher in children with MS as compared with healthy agematched controls [12] [13] [14] , whereas at the typical age of onset of MS (late 20s to early 30s) nearly all patients and control subjects have already experienced EBV infection. The two key points to emerge from these studies in relation to EBV are that all children with MS had already been infected with EBV [12] [13] [14] , and that the levels of antibody against the EBV EBNA1 antigen were very significantly higher in both adults and children with MS as compared with controls [13, 14, 16, [23] [24] [25] . In this context, infections with EBV usually occur in infancy in the developing nations where MS is rare while in the industrialised nations infection usually occurs later in life, often being delayed until adolescence or early adulthood. At the time of diagnosis of MS in children, the earlier EBV infection had become latent and there was no serological sign of reactivation (IgM or anti-early antigen titres C80 as measured in an indirect immuno fluorescent assay) [12, 13] . Moreover, in adult patients, the MS-associated serological EBV pattern was probably established many years before the onset of clinically evident MS as prior infectious mononucleosis is a strong risk factor for MS manifesting 2-20 years (mean around 10-12 years) later [18, 25] . Taken together, these studies strongly indicate that latent EBV infection is an essential predisposing condition for the development of MS but other genetic, environmental and hygiene-related conditions appear necessary for the actual expression of the disease and the unifying condition might be a dysregulated immune response. A distinct possibility is that other infections synergise with EBV in the aetiology of the disease and the timing of infections might be important. In this context, Chlamydia pneumoniae (CP) infection may play a special role [26] , as in a recent study on children with MS, CP-specific IgM antibody points to a high frequency of fresh, recent or reactivated infection with this pathogen at the onset of the disease [12] . A final key point is the demonstration that slight, though statistically significant, elevated levels of antibody to certain common infectious agents other than EBV and CP occur in children and adults with MS compared with age matched control subjects [12, 15, 27] . This finding does not, however, imply a direct causative relation of any infection to MS as it could reflect a more general dysregulation of immune function as a cause or consequence of the development of the disease. It is indeed likely that the elevated antibody concentrations have no significance per se for the development of MS but reflect a shift in patterns of immune reactivity away from a protection and towards enhancement of the risk of disease. Nevertheless, studies on these MS-associated infectious agents could lead to the identification of specific antigenic determinants involved in the generation and maintenance of immune dysregulation. A hypothesis which relates MS to the 'biography' of the immune system On the basis of the available epidemiological evidence it may be postulated that MS is dependent on an infection with EBV which, owing to hygiene-related factors, occurs later than usual in life. Under such circumstances, the EBV infection might encounter patterns of immune responsiveness generated by prior infections with certain other microorganisms. In this paper we suggest a scenario in which the sequence of certain common infections results in immune dysregulation favouring the onset of MS and in the following sections this hypothesis is elaborated. Recent studies have shown that the immune system contains a very complex network of regulatory pathways. To some extent these pathways are genetically determined, but there is growing evidence that they are critically determined by the nature and timing of infections and other immune challenges that an individual experiences earlier in life. This could be termed the 'biography' of the immune system. The immunoregulatory pathways are based on populations of lymphocytes, termed regulatory T cells (T reg s), in which there is currently considerable interest. In the case of infectious disease, such populations may lead to rapid resolution, the establishment of latent or persistent infection or to tissue damage by autoimmune processes [28] . Accordingly, T reg s have been termed 'a dangerous necessity' [29] . This term implies that T reg s are neither 'good' or 'bad' per se but may, according to the overall pattern of responsiveness, participate in appropriate immune reactions leading to resolution of disease or in inappropriate ones resulting in immunopathology. The temporal sequence of infections, especially initial and early ones, is crucial to the development of patterns of immune reactivity as prior contacts with other antigens may have induced cross-reactive T-helper cells competing with T reg s. As a consequence T reg s normally induced by the second pathogen may be marginalized or even eclipsed. The latter phenomenon, also known as lateral inhibition, has many parallels in biology, particularly in neurology. The locking of an immune response into an eclipsed state seems to involve an active deletion of clones of T-cells occurring as a result of reinfections or reactivations [28] . In the case of MS, infections such as those with HHV-6 [30, 31] and, possibly, with CP [12, 26] occurring before or at the time of initial or reactivated EBV infection could have such an effect. Thus these prior contacts could have induced populations of T reg s that have a crucial role in protection against MS but also cross-react with an epitope on EBV. A subsequent infection by EBV could therefore generate a dominant population of cross-reactive T-helper-cells which could suppress or delete the T reg s. Under normal physiological conditions, these T reg s could either suppress other populations of T-cells which would otherwise be able to induce autoimmune processes, including those involved in the pathogenesis of MS, or they could cause the expansion of a population of specific CD8 ? T-cells which would have an immune repair function. The nature of the epitope or epitopes involved in this postulated process remains unknown but, by analogy with the parallel studies on melanoma mentioned above [10, 11] , it is suggested that key host epitopes involved are coded for by certain human endogenous retroviruses (HERVs) as activation of these has been extensively documented in MS [32, 33] . Accordingly, a challenge in MS research is to delineate patterns of MS-related immune responses [34, 35] , and the T reg s involved, that affect the risk of MS both beneficially and detrimentally and the likely targets of these responses. The IgG-anti-EBNA1 antibody concentrations are particularly elevated in patients with MS [13, 14, 16, 18, [23] [24] [25] , and systematic studies on the T-helper cell epitopes in the EBNA1 protein revealed that CD4 ? T-cells from healthy EBV carriers matched for MS-associated HLA-DR alleles recognised several epitopes within the central region of the C-terminal domain of this protein but not other EBVencoded proteins [36, 37] . In contrast, those from MS patients recognised many more epitopes spread over the entire domain [36] . Concomitantly, the number of memory T-cells directed against this domain is increased about tenfold in MS and have been shown to be T-helper 1 cell precursors and polarised effector memory cells [36, 37] , containing a subfraction of regulatory T-cells (T reg s) [38, 39] . T reg s were suppressed in MS [40] , and high level of CD8 ? T-cell activation against EBV but not cytomegalovirus was demonstrated early in the course of MS [41] . A search for the possible origin of competing T-helper cells was undertaken with the BLAST analytical program [42, 43] . Sequence homologies were evident over the entire expanded EBNA1 epitope with proteins from CP and HHV-6 ( Table 1 ). This possible involvement of HHV-6 and CP in T cell competition is supported by the observation that the targets of T-and B-cells which have been found to be MS-associated by systematic studies [34, 35] also have homologies in HHV-6 and CP ( Table 2) . The central EBNA1 epitope marginalized in MS (amino acids FENIAEGLRALLARSHVER) could well have different functions in health and in MS and is therefore a major putative candidate for generation of T reg s which control the relevant immune processes. A specific or functional deficiency of T reg s in MS has only recently been recognised, and the need for a large cohort of T reg s for the resolution of experimental autoimmune encephalomyelitis has been demonstrated [44] . For the purpose of studying the potential infectious and immunological background of MS, it is relevant to search for homologies to this 'epitope No. 1' in the MS-associated pathogens [42, 43] . Interestingly, homologies to the putative epitope were found in all these pathogens (Table 3) . It is likely that a diverse range of MS-associated infectious agents other than EBV and, possibly, CP, is involved in the generation and maintainance of the postulated immune responses associated, beneficially or detrimentally, with MS. By generating populations of T reg s or of competing T-helper cells, such infectious agents would play a role in the generation of various immunological networks based on T reg s which, in turn, would facilitate the expansion or suppression of populations of effector T cells including epitope-specific CD8 ? T-cells. The epitope recognised by these T-cells should be common to the MS-associated infectious pathogens and to one or more cellular gene products. The latter was preferentially sought in HERVs since patients with MS expressed HERV-W env at higher copy numbers as compared with controls (P = 0.00003) [32, 45] and a HERV-K18 env genotype was described as a risk factor for MS [46] . A putative target epitope for effector T-cells in the processes suggested above ('epitope No. 2') was identified in a short peptide from the HERV-W env gene complex: MPVPSAPST. It is predicted that this peptide is presented, though only subdominantly, by diverse HLA class I molecules including Ld, A*0201, B*0702, B*5101, as determined by reference to the SYFPEIHI database for MHC ligands and peptide motifs [47] . Only three pathogens bearing the two homologies on the same protein, which is postulated to be optimal for a co-operation of the corresponding T-cells, have been identified, namely, herpes simplex (1 and 2) virus (tegument protein), measles (nucleoprotein), and varicella (tegument). The MS-association of the serology of these pathogens (higher specific antibody concentration) [16] was confirmed in a recent study: herpes simplex-2, P \ 0.0001; measles, P \ 0.0001 and varicella, P \ 0.0001 [12] . The epidemiological evidence for the MS-association of the majority of the other pathogens in Table 3 is only weak and inconsistent. Moreover, a simultaneous occurrence of homologies to epitopes 1 and 2 was also found in a diverse range of pathogens causing respiratory and gastrointestinal infections and which have also been associated, beneficially and Specific T-cells directed against this region were found to be expanded in MS patients as compared to control individuals [36] ; the region with homologies in HHV-6 and CP proteins extends to EBNA1 amino acid position 640 * for identical amino acid; ? = conserved amino acid exchange; / = missing amino acid; arabic numbers for additional amino acids: 1 = DKK; 2 = PF; 3 = D detrimentally, with the risk of MS [48] [49] [50] . Likewise, unknown parasitic infections have recently been found to be associated with a reduced risk of MS [51] , and some parasitic organisms share the two homologies (Table 3 , footnote). In addition, preliminary studies indicate that parasite infections in MS patients lead to fewer exacerbations and this has been linked to the emergence of T reg s [52] . The relatively widespread occurrence of these two homologies explains, at least in part, why the infectious background of MS has proved so complex and difficult to define. The MPVPSAPST-peptide is the amino-terminal part of a small hypothetical protein of 29 amino acids encoded by the complementary DNA strand of the HERV-WE1 env gene which is conserved in all homologous HERV-W sequences in the human genome. Gene transcripts of 21 of 25 open reading frames with an initiating start codon have been found in association with MS according to genetic data bank entries. Moreover, several other HERVs (-H, -K, -L) exhibit this homology. As these HERV peptides are all self-antigens, they could serve as targets, but not as inducers, of the postulated MS-protective immune response. The main HLA class I molecule A*0201 for the presentation of the peptide, with a frequency of about 30% in a European population, was shown to be associated with a significantly reduced MS risk (OR = 0.52, P = 0.0015) [53] . Immune dysregulation in MS is likely to be an early event [18, 22, 25, 41] preceding the onset of MS disease by many years or a decade [18, 22, 25] . It should thus be emphasized that the epidemiologic observations on the possible infectious background of MS partly, if not predominantly, reflect the earliest stage in the natural history of MS. A situation similar to that postulated here has been described in malignant melanoma in which, as discussed above, cross-reactive protective immune responses are induced by homologous epitopes in BCG, vaccinia and yellow fever vaccines given at least 10 years before the onset of disease [8] [9] [10] [11] . It was suggested that melanocytes in the early stages of malignant transformation, may be eliminated or repaired by CD8 ? T-cells which recognise cells expressing a HERV peptide. This immune reaction seems to suppress the genetic activity of HERV proviruses (env genes) associated with malignant transformation [10, 54, 55] . The HERV env proteins probably impair the redox Table 2 Homologies in proteins of sero-epidemiologic MS-associated pathogens to MS associated EBV-epitopes [34, 35] Enhanced immune reactivity in MS patients in comparison with healthy control subjects as identified by systematic studies [34, 35] ; consensus in other proteins to the EBV sequence given by * for identical amino acid, ? for similar amino acid (conservative exchange), and / for missing amino acid; one additional homology in vaccinia virus regulation within the cells via reduced levels of glutathione peroxidase [10] . In MS there is still another environmental risk factor, namely, suboptimal levels of bio-active vitamin D [4] , which, as demonstrated in rat astrocytes, may impair via c-glutamyl transpeptidase intra-cellular glutathione levels [56] . Table 3 MS-associated pathogens and homologies in proteins thereof to candidate epitopes No. 1 (EBV EBNA1, putative T reg ) and No. 2 (HERV-W) * for identical amino acid; ? = conserved amino acid exchange; / = missing amino acid; arabic numbers for additional amino acids: 1 = TEE; 2 = AG; 3 = QKE; 4 = YCK; 5 = AT; 6 = V; 7 = V; 8 = LAT Pathogens against which higher antibody concentrations were observed as compared with control individuals. Homologies to both epitopes were also found in the following pathogens causing respiratory and gastrointestinal diseases: Bordetella parapertussis, Corynebacterium diphtheriae, cytomegalo virus, Haemophilus influenzae, human corona virus, human rota virus, Mycobacterium tuberculosis (also M. bovis, strain BCG), Salmonella enterica, Pseudomonas aeruginosa, Serratia marcescens, A and in two parasites causing gastrointestinal infections: Entamoeba histolytica, Giardia lamblia Proteins with homologies to epitope 1/epitope 2, respectively: adeno virus: DNA stabilization protein/protein V; Chlamydia pneumoniae: hypothetical protein cpB U609/hypothetical protein; Epstein-Barr virus: EBNA1/BBLF2/BBLF3; herpes simplex-1: tegument/tegument; herpes simplex-2: tegument/tegument; HHV-6: tegument/major capsid; measles virus: nucleoprotein/nucleoprotein; Mycoplasma pneumoniae: phosphate import ATP-binding protein pstB/enolase; parainfluenza-2: large protein/nucleocapsid; RSV: fusion protein precursor/glycoprotein; rubella: RNA-directed RNA polymerase/E1; vaccinia: 14K membrane protein/putative DNA-binding core; varicella virus: tegument/tegument In cell culture experiments, gangliosides of the neolacto series, such as LM1, were identified as possible relevant mediators as they suppress HERV genes [10] and induce lytic replication cycles in cells latently infected with EBV [57, 58] , thereby releasing a viral antigen which would be readily recognisable by the specific immune system. The failure of such protective mechanisms would facilitate an uncontrolled establishment of the observed extensive EBV infection of brain lymphatic tissue in MS, albeit only at a low level of virus activity [59] [60] [61] [62] . The specificity of the process for the brain may be associated with the high content of gangliosides in nervous tissue, as other gangliosides may antagonize the activity of those that possibly protect against MS. Multiple sclerosis is certainly a complex disease entity and the pathogenic process involves more than just increased neuro-degeneration. In particular, reduced remyelination contributes to disease progression. Thus, relevant targets in addition to the HERV peptide might exist, possibly host proteins with a homology to the HERV peptide. One such candidate is neuron-specific ankyrin-2 which, owing to complementary binding sites, forms a complex with spectrin. In an animal model of remyelination of axons the latter is the target of an immune repair mechanism mediated by a monoclonal antibody [63] . Based on epidemiological considerations, it is postulated that the relatively high risk of MS in the industrialised nations is due to hygiene-related changes in the sequence of common infections, resulting in the emergence of patterns of immune reactivity that either cause, or fail to protect against, the development of MS. Further epidemiological studies are required to determine whether the timing of EBV infection is related to the risk of MS [20] . Based on studies of altered immune responses to certain infectious agents and evidence for the expression of HERVs in MS, epitopes that have a putative role, depending on the appropriate or inappropriate activity of immunoregulatory networks, in protection against or predisposition to MS have been delineated. Studies based on the above considerations should focus on processes that initiate MS years or a decade before manifestation of the disease. Subsequently, a better understanding of the highly complex immunopathology of MS will, hopefully, lead to the eventual development of vaccination strategies for the prevention or correction of anomalies in immune function. Such vaccines could well work by preventing or correcting hygiene-related immune dysregulation [64] . It has previously been postulated that a vaccine based on EBV should afford a high level of protection against MS [16] , but the exact nature of an efficient and safe vaccine would depend on the nature of the relationship between EBV and MS. If the framework presented in this study is in principle confirmed, the appropriate vaccine is most likely to be a living attenuated one that induces a strong T cell response to EBV epitopes but is not expressing EBV EBNA1 in latency.

Evidence of Recombination and Genetic Diversity in Human Rhinoviruses in Children with Acute Respiratory Infection

BACKGROUND: Human rhinoviruses (HRVs) are a highly prevalent cause of acute respiratory infection in children. They are classified into at least three species, HRV-A, HRV-B and HRV-C, which are characterized by sequencing the 5′ untranslated region (UTR) or the VP4/VP2 region of the genome. Given the increased interest for novel HRV strain identification and their worldwide distribution, we have carried out clinical and molecular diagnosis of HRV strains in a 2-year study of children with acute respiratory infection visiting one district hospital in Shanghai. METHODOLOGY/FINDINGS: We cloned and sequenced a 924-nt fragment that covered part of the 5′UTR and the VP4/VP2 capsid genes. Sixty-four HRV-infected outpatients were diagnosed amongst 827 children with acute low respiratory tract infection. Two samples were co-infected with HRV-A and HRV-B or HRV-C. By comparative analysis of the VP4/VP2 sequences of the 66 HRVs, we showed a high diversity of strains in HRV-A and HRV-B species, and a prevalence of 51.5% of strains that belonged to the recently identified HRV-C species. When analyzing a fragment of the 5′ UTR, we characterized at least two subspecies of HRV-C: HRV-Cc, which clustered differently from HRV-A and HRV-B, and HRV-Ca, which resulted from previous recombination in this region with sequences related to HRV-A. The full-length sequence of one strain of each HRV-Ca and HRV-Cc subspecies was obtained for comparative analysis. We confirmed the close relationship of their structural proteins but showed apparent additional recombination events in the 2A gene and 3′UTR of the HRV-Ca strain. Double or triple infections with HRV-C and respiratory syncytial virus and/or bocavirus were diagnosed in 33.3% of the HRV-infected patients, but no correlation with severity of clinical outcome was observed. CONCLUSION: Our study showed a high diversity of HRV strains that cause bronchitis and pneumonia in children. A predominance of HRV-C over HRV-A and HRV-B was observed, and two subspecies of HRV-C were identified, the diversity of which seemed to be related to recombination with former HRV-A strains. None of the HRV-C strains appeared to have a higher clinical impact than HRV-A or HRV-B on respiratory compromise.

Human rhinoviruses (HRVs) are a highly p revalent cause of the acute respiratory infection (ARI) defined as the common cold [1, 2, 3] , which is frequently associated in children with bronchitis, bronchiolitis, wheezing, pneumonia, asthma and otitis [4, 5, 6, 7, 8, 9] . HRVs are classified in genus Enterovirus (HEVs) in family Picornaviridae [10] . HRVs are non-enveloped, single-stranded, positive-sense RNA viruses of approximately 7200 nt, composed of a 59 untranslated region (UTR), followed by a long open reading frame coding for capsid proteins VP4, VP2, VP3 and VP1, and seven non-structural proteins 2A, 2B, 2C, 3A, 3B, 3C and 3D, and terminated by a short 39UTR and poly A tract. More than 100 serotypes of HRV are known, which have been classified into two species, HRV-A and HRV-B, according to comparative alignment of nucleotide fragments of VP1 [11, 12] , VP4/VP2 [13] and 59UTR [14, 15] , and more recently, on complete genome nucleotide sequences [16] . Moreover, some genomic sequences have been found not to cluster with HRV-A and HRV-B species, which suggests the existence of other species (HRV-C and HRV-D) [16] . A new species of HRV-C was recently identified worldwide by comparative analysis of VP4 or VP4/VP2 genes [7, 17, 18, 19, 20, 21] and 59UTR [14, 15] . However, discrepancies have appeared in the classification of some of the new HRV-A or HRV-C strains, depending on the size and location of the nucleotide sequence in the viral genome and on the phylogenetic methods used for direct analysis of HRV sequences [3, 7, 14, 15, 17, 18, 20, 21, 22, 23] . These recent data underline the lack of knowledge about the biodiversity of HRV strains and their worldwide distribution [14, 17] . Moreover, little is known about the characteristics and diversity of HRVs circulating in a given area in a short period of time. In the present study, we looked for HRVs in a 2-year collection of nasopharyngeal swabs (NPSs) of children with ARI visiting a district hospital in Shanghai, and compared sequences in two regions previously defined for genetic classification of HRV serotypes [13, 15] . Our study showed a high diversity of HRV species and genotypes, and a prevalence of the novel HRV-C species in NPSs of children with bronchitis and pneumonia. This biodiversity appeared to result partly from recombination events in the 59UTR, between HRV-C strains and those close or similar to HRV-A species, which led to the suggested classification of HRV-C into at least two subspecies. Eight hundred and twenty-seven samples were collected from a group of children consulting the Shanghai Nanxiang Hospital during a 2-year period, and tested for 17 respiratory viruses using a multiplex RT-PCR (mRT-PCR). Sixty-four samples (7.7%) were positive for HRV, according to the length of the amplified fragment in the VP4/VP2 region visualized on agarose gel (data not shown). A larger fragment of 924 nt, including part of the 59UTR (starting at nt 163) and the VP4/VP2 genes (ending at nt 1086), was amplified and cloned into plasmid vectors for genetic analysis (Table 1) . Only one sample, N1, could not be amplified and was amplified in two steps in the 59UTR (nt 163-552) and in the VP4/VP2 region (nt 528-1086), respectively. Analyses of different clones for each sample allowed the identification of multiple infections: sample N16 was shown to contain one HRV-A and one HRV-B (N16A and N16B, respectively), while N58 contained one HRV-A and one HRV-C (N58A and N58C, respectively) (see below). In order to further characterize the 66 HRVs identified in the 64 samples, the nucleotide sequences located in the 59UTR (285 nt) and the VP4/VP2 genes (420 nt) were chosen to allow comparative alignment with sequences of reference serotypes and field strains available in GenBank. We first compared and classified Shanghai strains according to their VP4/VP2 sequences. The pairwise nucleotide divergence in the VP4/VP2 region ranged from 0 to 72.2%. Twenty-seven HRVs (40.9%) showed .81% nucleotide identity with the closest HRV-A clusters, and five HRVs (7.6%) showed .88.8% nucleotide identity with HRV-B clusters ( Table 2 ). The remaining 34 strains (51.5%) diverged from HRV-A and HRV-B species by .47.3% in their VP4/VP2 nucleotide sequence ( Table 2) . These strains showed from 68.3 to 100% nucleotide identity with each other and were related to the recently described HRV-C strains, NAT001 and NAT045, isolated in California, USA [18] , C024, C025 and C026 in Hong Kong [20] , and QPM in Australia [7, 24] ( Fig. 1 ). Strains N34, N35 and N68 were closely related to the recently identified strains C025 and NTA001 with .95.9% nucleotide identity, whereas the 31 remaining HRV-C strains showed only 74.4-86.4% nucleotide identity with six other recent strains (QPM, NAT001, NAT045, C024, C025 and C026; Table 2 ), which were classified tentatively as HRV-C species [7, 18, 20, 24] . Classification of the strains into three different species was also demonstrated by construction of a phylogenetic tree using aligned VP4/VP2 sequences (Fig. 1) . To characterize and classify further the Chinese HRV strains, 59UTR sequences were considered (Table 3 , Fig. 2 ). They were compared to the 59 UTR of all 101 reference HRVs, to those of 26 new strains identified in children with respiratory illness in Wisconsin (indicated as W) [15] , and to those of other recently identified HRV-C strains [14] . Pairwise nucleotide divergence between the three HRV species was 0.7-64.3%, and a limit of ,9% divergence between genotype pairs was chosen for similar genotype assignment in one species [15] . New genotypes were identified when they had 9-30% pairwise nucleotide divergence from the nearest serotype in the same species (Table 3) . Fifty-five HRVs shared .94.4% nucleotide identity with strains already identified, and 11 HRVs showed 9.5-20.9% nucleotide divergence with the nearest known HRVs. They may represent newly discovered genotypes. These strains clustered with HRV-A (N6) or HRV-C species (N4, N8, N21, N62, N63, N67, N24, N25, N28 and N32) ( Table 3) . Most surprisingly, 20 of the 34 strains classified as HRV-C by comparative analysis of VP4/P2 sequences ( Table 2) were related more closely to HRV-A strains when their 59UTRs were analyzed, and showed incongruent clustering in phylogenetic trees (Figs. 1 and 2). The nucleotide sequences in the 59UTR of these strains were related closely to those of formerly identified QPM, NAT001, NAT045, C024, C025 and C026 HRV-C strains, and to some of the W strains recently identified as HRV-A [15] (Fig. 2) . However, our 20 strains clustered together with these six strains in two major branches in the phylogenetic tree constituted a subspecies of HRV-C called HRV-Ca (Table 3 , Figs. 2 and 3) . The fourteen other HRV-C strains formed another unique branch of HRV-C subspecies, called HRV-Cc, which clustered differently from other species of HRV-A and HRV-B, and from subspecies HRV-Ca in the phylogenetic tree based on 59UTR sequences (Table 3 ; Figs. 1 and 2). These strains were related closely to some W strains of HRV that were classified as HRV-C [15] (Fig. 2 , Table 3 ) In order to characterize more precisely the differences observed in the 59UTR between HRV-Ca and HRV-Cc subspecies, and to localize possible recombination sites in the 59UTR of the genome of HRV-Ca subspecies, bootscanning and similarity plot analyses were conducted in the gene fragment of 868 nt that included the 59UTR and adjacent capsid genes. HRV-Ca nucleotide sequences were scanned against sequences of N10, R16 and R52, which are considered as representative strains of HRV Cc subspecies and HRV-A and HRV-B species, respectively. Stretches of nucleotide sequences that were closer to HRV-A (R16) than to HRV-Cc (N10), flanked by sequences related to HRV-C could be detected in the 59UTR of HRV-Ca strains, as exemplified with HRV-Ca N25 strain ( Fig. 3a and b) . These HRV-A-related nucleotide stretches were thus flanked by putative recombination sites. These sites were located differently among the HRV-Ca strains, delimiting HRV-A-related stretches that ranged from 150 to 400 nt in length (Fig. 3c ). While variable among the strains, the identified recombination sites were all located inside the 59 UTR and none of them was identified in the downstream VP4/VP2 coding sequence. HRV-A-related nucleotide sequences and putative recombination sites were also found in the 59UTR of the previously described HRV-C strains C024, C025, C026, NAT001, NAT045 and QPM (Fig. 3) . The results corroborated the clustering observed in the phylogenetic tree based on 59UTR sequences (Fig. 2) , since strains gathered in the same HRV-Ca subcluster (for example N24, N25, N28 and N32, or N4, N7, N8, N21, N36 and N46) displayed the same recombination pattern. These subclusters revealed different recombinant lineages, each of which originated from independent recombination events. In order to further characterize the genome of the HRV-Cc subspecies, for which no full-length sequence was yet available, we sequenced the remaining genes that covered the whole coding sequence and 39UTR of N10 strain, which was chosen as the representative of this subspecies (Tables 2 and 3 ). The full-length N10 genome sequence was compared to those of the HRV-Ca subspecies strains C024, C025 and C026, and to that of N4 strain, which was sequenced as the representative of the HRV-Ca subspecies. The genome sequences were also compared to those of the HRV-A strains N13 and R44, and to the HRV-B strains R14 and R52 (Table 4 ). The full-length nucleotide sequence of N10 strain contained 7111 nt, excluding the poly(A) tract, which was shorter than sequences from HRV-A and HRV-B strains, but similar to those of HRV-Ca strain N4 and other related strains (C024, C025 and C026). The 2144 aa lengths of the polyprotein and of each of the individual proteins of N10 were slightly different from those of HRV-A and HRV-B species, but similar to those of other HRV-C strains. The most divergent amino acid length between HRV was observed for the VP1 protein that was shorter in HRV-C species ( Table 4 ). The unique putative cleavage (M/S) site between VP4 and VP2 protein identified previously for QPM, C024, C025 and C026 strains [20] was also observed for N10 and N4 strains. It was different from those of the HRV-A strains N13 and R44 (Q/S), and from those of the HRV-B strains R14 and R52 (N/S) (data not shown). Alignment of the VP1 amino acid sequence of HRV-Cc strain N10 with those of other HRV-A and HRV-B species and HRV-Ca subspecies, designated in Table 4 , showed structural features typical of HRV-C species [16, 20, 25] (data not shown). In particular, footprints including deletions in the BC, DE and HI loops and conserved amino acids potentially involved in Inter-Cellular Adhesion Molecule 1 (ICAM-1) receptor binding [7, 11, 20, 24] were conserved within the HRV-C species (data not shown). Bootscanning and similarity plot analysis conducted on the genomic sequences of N4 (HRV-Ca), N10 (HRV-Cc), R16 (HRV-A) and R52 (HRV-B) confirmed that N4 featured a 59UTR sequence that was related to the R16 sequence (stretch I), followed by a capsidic sequence related to the N10 sequence. N4 nonstructural sequence (2A to 39UTR) was related more closely to N10 than to R16 and R52 sequences. However, in stretch II (nt 3300-3500 according to N4 numbering), N4 strain (HRV-Ca) was closer to R16 (HRV-A) than to R52 (HRV-B) or N10 (HRV-Cc), which resulted in high bootstrap values between N4 and R16 2A sequences (Fig. 4) . This may have been the result of a recombination event that occurred in the 39 part of the 2A-encoding sequence of the parental strain N4, and which involved an HRV-A strain. Nevertheless, the HRV-A parental strain or ancestral strain could not be identified since the closest HRV-A 2A nucleotide sequence available was ,80% identical to that of N4 in this region. In contrast from nt 6,550 to the 39 end (stretch III in Fig. 4) , the N10 strain genome was found to be closely related to that of N4, with nucleotide identity .98%. This result is corroborated in Figure 5 , which shows a phylogenetic analysis of the 39UTR sequences of N10 and N4 compared to those of HRV-Ca subspecies and HRV-A and HRV-B species. This suggests that N4 and N10 strains share a common recent ancestor through recombination. 1 and 2 ) and on local recombination in 59UTR (see Fig. 3 ). b Strains closely related with more than 93% identity. c These strains clustered differently when based on VP4/VP2 sequences (see Table 2 and Figs. 1 and 2) . doi:10.1371/journal.pone.0006355.t003 Clinical outcome from HRV strains isolated from pediatric outpatients Among the pediatric patients, 46 were males and 18 females, and their age ranged from 5 months to 14 years. The majority of HRV infections were diagnosed between 2 and 6 years of age (84.6%). Bronchitis (73.4%) and pneumonia (26.6%) were highly prevalent in children with comparable incidence in HRV-A and HRV-C infections (Table 5) . Moreover, the ratio of pneumonia over bronchitis (36.2%) was comparable to that in the whole cohort of 827 children (40.7%). Only one child among the 64 HRV-positive patients had asthma and was co-infected with HRV-C, influenza A virus (IAV) and respiratory syncytial virus (RSV) ( Table 5) , whereas 100 of the 827 patient were diagnosed with asthma. HRVs were isolated throughout the 2 years, with a predominance of HRV-C viruses in the cold season (Table 5) . Interestingly, different HRV genotypes were detected within the same period (for example, N1 and N4, N9 and N11/N12, N44/N48 and N51, and N55/N56 and N62/N67), with a larger diversity and distribution of individual or paired HRV-A genotypes compared to HRV-C strains, which clustered in closely-related genotypes (Fig. 2, Tables 2 and 3) . Conversely, N4 and N21 strains of samples collected at 10 months interval showed 99.8% identity (Fig. 2) . Single HRV infection was diagnosed in 42 children and coinfections were identified in 22 patients (Table 5) , with 17 double and five triple infections. The viruses most often identified in HRV co-infection were RSV (six cases) and human bocavirus (HBoV; four cases), and two patients were co-infected with HRV, HBoV and RSV (Table 5 ). There was no difference between HRV-Ca or HRV-Cc subspecies and any of the clinical or epidemiological data (data not shown). In this report, we looked for HRVs in a 2-year collection of NPSs from children with ARI visiting a district hospital in Shanghai, and found a high diversity of HRV strains that belonged to different species and genotypes. We characterized by RT-PCR and sequenced 66 HRVs, among them 27 HRV-A, five HRV-B, and 34 HRV-C strains. When sequencing the VP4/VP2 region of the HRV genome, several recent studies have identified new strains of viruses from children and adults with ARI, asthma, or otitis, which are clustered differently from HRV-A and HRV-B, and have been classified into a novel HRV-C species [7, 8, 17, 18, 19, 20, 21, 25, 26, 27, 28] . Other groups have also identified novel HRV-C strains by sequencing the VP1 gene [29] or the 59UTR [14, 15] . The different sizes and locations of the regions amplified in the HRV genomes renders difficult comparative genetic analysis. Recently, Palmenberg et al. (2009) have finalized the full-length genome sequences of all HRV-A and HRV-B reference strains, and identified structural features of these two species and the novel HRV-C species [16] . In our study, we identified 34 HRVs (51.5%) that clustered differently from HRV-A and HRV-B in a phylogenetic tree that was established on the basis of VP4/VP2 sequences, which were related to recent strains classified in the novel HRV-C species (Fig. 1, Table 2 ). Fourteen HRV-C strains (41.2%) segregated from the other 20 strains (58.8%) that were closely related to HRV-A in their 59UTR sequence (Fig. 2) . This led us to propose a classification of two HRV-C subspecies, HRV-Cc and HRV-Ca. In previous studies [15] . These strains clustered with our field strains within the HRV-Cc subspecies. Moreover, 17 strains that clustered with HRV-A, and had 12-35% pairwise nucleotide divergence from the nearest reference serotype [15] , clustered within the two major branches of HRV-A and HRV-Ca strains (Fig. 3) . Therefore, we cannot ensure that some of the 17 strains were HRV-A or HRV-Ca strains. Kiang et al. (2008) have identified five novel HRVs out of 24 clinical samples (20.8%), which segregated from HRV-A and HRV-B, and were classified as HRV-C, and three additional strains (12.5%) that also clustered with QPM, C024, C025, C026, NAT001 and NAT045 [14] (Fig. 2) . However, the field HRV strains of these previous studies were sequenced using a 59UTR that did not match fully our sequence and that of Lee et al. (2007) [15] , and could not be included in the present study for comparative analysis. Interestingly, the five strains identified in California in 2007 [14] and N42 and N45 from our study were closely related to strain W37 isolated in Wisconsin in the late 1990s [15] , and to NAT001 isolated in the winter of 2004 in California [18] , which confirms that similar genotypes of HRV-Ca are widespread [17] . The strains of HRV-C species identified in the present study were characterized by analyzing the 59UTR, VP4, and part of VP2 (Fig. 3) . This approach showed the advantages of covering only 59NCR, VP4/VP2, VP1 or 3D genome fragments. Analyzing sequences that covered the 59UTR and the downstream VP4/ VP2 capsid region allowed identification of co-infections when several clones were sequenced, and helped to locate the recombination sites in strains of the HRV-Ca subspecies. Thus, this region of the genome may be useful for building a database of the novel strains that are circulating worldwide. The genome of HEVs is subject to frequent recombination [30, 31, 32, 33, 34, 35, 36] , with interspecies exchanges observed in the 59UTR [37] . Palmenberg et al. (2009) have observed intraspecies recombination in three HRV-A, with structural characteristics and phylogenetic evidence that suggests a novel clade D classification [16] . Tapparel et al. (2009) observed phylogenetic incongruities in 59 UTR, VP1 and 3CD sequences of two clinical isolates of HRV-A related to recombination [38] . We observed incongruent clustering of N12, N44 and N48 strains of HRV-A species in the phylogenetic trees based on the 59UTR or the VP4/VP2 regions of their genomes (Figs. 1 and 2, Table 3 ), which suggests intraspecies recombination in the 59UTR. We observed one co-infection with HRV-A and HRV-B (N16A and N16B), one with HRV-A and HRV-C (N58A and N58C), and three co-infections of HRV-A and HRV-B with HEVs that may favor recombination events. Previous comparison of genome sequences between 34 HRVs showed only limited recombination events and a pattern of genetic diversity lower than that observed with other picornaviruses [25] . The presence in HRV-C subspecies of sequences that share 90.5-98.6% identity with HRV-A strains (Table 3) suggests that recombination events occurred between HRV-C and HRV-A. Bootscanning of the 59UTR of HRV-C strains also showed different sites and lengths of recombination (Fig. 3c) , which suggested that there were several independent events that led to several groups of HRV-Ca genotypes, which formed clusters in the phylogenetic tree (Fig. 2) . Comparative analysis of the full-length nucleotide sequences of two field strains of different HRV-C subspecies (N4 and N10) with those of other HRV species suggested that multiple interspecies recombination events occurred in the 59UTR and in the NS2A protein gene, and that recombination also occurred in the 39UTR between N4 and a strain close to N10. These findings are in agreement with those observed for other HEVs, for which recombination events in the capsid-encoding sequence are very rare, probably because of structural constraints that restrict the functioning of chimeric capsids [31] . This result appeals for the full-length genome sequencing of the major representatives of the HRV-C species, in order to establish a clear understanding of the evolution and classification of the novel virus into subspecies. Comparison of the coding sequences of N10 HRV-Cc with other strains of HRV-Ca subspecies [20] , including our field strain N4, showed high similarities in the lengths of the 11 proteins, their cleavage sites, and the structural features of VP1. These characteristics and the absence of growth in cell culture, noted in our laboratory and by others (data not shown), support the classification of the novel strains into a unique HRV-C subspecies. Our clinical specimens all originated from NPSs from pediatric outpatients. The remarkable outcome of the study is the large diversity of genotypes that has circulated in a relatively small group of people in a district of Shanghai during a 2-year observation. Although some clusters of similar genotypes in a limited period of time were observed, co-circulation of different genotypes and HRV species and subspecies, and co-infections with two HRV species were observed. The prevalence of the novel HRV-C in our specimens (4.1%) differed from previous studies that associated the prevalence of the novel variant with severe disease outcomes, which ranged from influenza-like illness or infection of the low respiratory tract [17, 28] to asthma exacerbation, bronchiolitis, and febrile wheeze [7, 8, 15, 18, 20, 21, 29, 39] . All our patients showed bronchitis or pneumonia, with no etiological correlation with any of the species or subspecies of HRV. Only one patient co-infected with HRV-C, IAV and RSV was diagnosed with asthma among the HRV-positive patients (1.6%), whereas 100 of the 827 children had asthma (12%). The difference observed with previous studies, 44.6% [8] and 12% [29] asthma in HRV-positive patients, may be related to the criteria for enrolment. Moreover, none of the patients in our study were hospitalized, which makes comparison with hospitalized children difficult [20, 24] . Another criterion to consider in the trend to correlate clinical symptoms with HRV infection is the presence of co-infecting pathogens. In our study, four strains of HBoV and six strains of RSV (17.6%) were identified in association with HRV-C (11.7%). HBoV and RSV are common viruses diagnosed in ARI, which are often associated with HRV [40, 41] , and HBoV was identified in .50% of children co-infected with HRV [20] . Nevertheless, the incidence of HBoV in ARI and in severe outcomes remains elusive [42] . More studies need to be carried out on large numbers of samples from severe and mild diseases, to identify any obvious role of HRV sequence diversity and association with other pathogens in disease severity. Since a large diversity of recombination in HRVs has become obvious, we must be aware of the occurrence of novel HRVs that may become highly virulent. This study was approved by the ethical committee of Shanghai Nanxiang Hospital and written informed consent was obtained from the parents of the children. Clinical specimens (n = 827) from NPSs were collected from children under 14 years old, who experienced a lower respiratory tract infection, and who were consulting the pediatric department of Shanghai Nanxiang Hospital during the period October 2006 to October 2008. Total RNA was extracted from NPS specimens using QIAamp viral RNA Mini Kit (Qiagen, Hilden, Germany), and stored at 280uC. RNA was amplified using the Qiagen One Step RT-PCR Kit. A five-tube mRT-PCR was used for virus identification as previously described [43, 44] . Tube 1 targeted IAV, influenza B virus, RSV, and human metapneumovirus; tube 2, parainfluenza viruses 1 to 4; tube 3, HRV and influenza C virus; tube 4, human coronaviruses (HCoVs) 229E-HCoV, OC43-HCoV, NL63-HCoV and HKU1-HCoV; and tube 5, adenovirus and HBoV. Amplified products were analyzed in 0.5 mg/ml ethidium bromide/2% agarose gel. Samples that showed positive results for HRV were amplified again using specific primers P1-1F and VP4/2R, located in the 59UTR and VP2 gene, respectively (Table 1) . One strain of HRV-C (N1) could not be amplified using the P1 and VP2 extreme primers and was amplified using primers in 59UTR and VP4/ VP2, respectively (Table 1 ). In brief, 2.5 ml of extracted RNA was mixed with 56 buffer and 0.4 mM dNTPs, 0.2 mM of each of the primers, and 1 ml of enzyme mix, and diethylpyrocarbonatetreated ultrapure water was added to a final volume of 25 ml. Amplification programs included reverse transcription at 50uC for 30 min, inactivation at 95uC for 15 min, followed by 40 cycles at 94uC for 30 s, 50uC for 30 s, 72uC for 70 s, and final extension at 72uC for 10 min. The amplified DNA products were detected by ethidium bromide-agarose gel electrophoresis. The lengths of P1-VP2, VP4-VP2 and P1-P3 amplicons were 924, 559 and 390 nt, respectively. DNA products were extracted from agarose gels by using QIAquick Gel Extraction Kit (Qiagen), and were ligated into pMD20-T vector (Takara Biotechnology, Dalian, China), and at least two recombinant plasmids were sequenced in Biosune Sequence Company and Life Biotechnology in Shanghai, China. Sequences of different clones of N16 and N58 showed identities for either HRV-A or HRV-B strains. More plasmids were sequenced for these strains to confirm that the two patients were originally coinfected with two different HRV species. Sequences of three complete genomes of HRV were obtained for strains N4 (reference R3061207002 collected on December 7, 2006) , N10 (R3070614001 collected on June 14, 2007) and N13 (R3070719007 collected on July 19 2007). Primers used for the amplification of viral genomes were designed after multiple alignments of sequences from the genomes of different HRVs available in GenBank (Table 1) . Overlapping amplified DNA products were obtained after PCR of cDNA that was obtained using oligodT and a Transcriptor High Fidelity cDNA Synthesis Kit (Roche, Mannheim, Germany), following the manufacturer's protocols. Briefly, 10.4 ml viral RNA was mixed with 1 ml oligodT and heated at 65uC for 10 min, and then kept on ice for 2 min. After addition of 4 ml 56buffer, 0.5 ml Protector RNase Inhibitor, 2 ml dNTPs, 1 ml DTT, and 1 ml RT enzyme, the reaction was incubated at 50uC for 1 h, inactivated at 85uC for 5 min, and stored at 220uC. Amplification of a 3D region of N4, N10 and N13 HRV strains was carried out by nested-PCR using Takara EXTaq (Takara Biotechnology) and specific primers (Table 1) , for 35 cycles of 30 s at 94uC, 30 s at 55uC, and 70 s at 72uC. To amplify VP1 (upstream of 2A) sequences of N4 and N13 strains, nested PCR was carried out using Takara LATaq with GC buffer I(Takara Biotechnology), specific primers (Table 1) , and incubation for 35 cycles of 30 s at 94uC, 30 s at 60uC, and 4 min at 68uC. The fragment VP2-3D of N10 HRV strain was obtained by seminested PCR and specific primers VP2 F, and 3D inner and outer reverse primers (Table 1) , using Takara LATaq with GC buffer II, for 35 cycles of 30 s at 94uC, 30 s at 60uC, and 6 min at 68uC. The terminal part of the whole genome was obtained by rapid amplification of cDNA ends using 59/39 rapid amplification of the cDNA kit, following the manufacturer's protocol (Roche). To perform 59 terminal RACE, 4 ml of 56cDNA buffer, 2 ml dNTPs, 1.25 ml specific primer 1 (10 mM), 9.2 ml RNA, 1 ml control primer neo1/rev (12.5 mM), 1 ml control RNA, 1 ml RT enzyme, and 0.6 ml RNase inhibitor (Roche) were mixed and incubated for 55uC for 60 min, followed by inactivation at 85uC for 5 min, and stored on ice. The product was purified using the Qiagen PCR Purification Kit and eluted with 30 ml deionized distilled water. A polyA tail was added to the cDNA, by mixing 9.5 ml DNA with 1.25 ml 106 reaction buffer, 1.25 ml (2 mM) dATP, and after incubation at 95uC for 3 min, the reaction was chilled on ice for 2 min. After addition of 0.5 ml terminal transferase, the reaction was incubated at 37uC for 30 min, inactivated at 70uC for 10 min, and kept on ice. Nested PCR was performed by using the Expend High Fidelity PCR kit (Roche). A mixture of 2.5 ml poly-dA-tailed cDNA, 0.5 ml oligodT-anchor primer 37.5 mM, 0.62 ml SP2 primers (10 mM) (Table 1) , 0.5 ml control neo2/rev primer (12.5 mM), 0.5 ml dNTP, 0.35 ml enzyme, 2.5 ml 106 buffer, and 18 ml ddH 2 O was incubated for 40 cycles of 30 s at 94uC, 30 s at 60uC, and 30 s at 72uC. To perform 39 terminal RACE, the method was similar to normal two-step RT-PCR using 3D inner and outer F primers (Table 1) . Sequence alignment, phylogenetic analyses and recombination analysis DNA sequences used for P1-P2 gene analysis were based on HRV-16 nt 178-462 and those used for VP4/VP2 gene analysis were based on HRV-16 nt 626-1045. Multiple sequences were aligned using Clustal X [45] . The multiple-sequence alignment was subjected to phylogenetic analyses using programs in the PHYLIP package (v3.6). Bootstrap analysis was performed using SEQBOOT, with a replicate number of 1000. Then, DNADIST and NEIGHBOR were used to obtain distance matrices with the F84 parameter, and a transition/transversion ratio of 4. Consensus trees were computed by CONSENSE, and then re-rooted with RETREE. The final tree was visualized and edited with MEGA version 4 [46] . Recombination analysis was carried out by using Recombination Detection Program v.3.22. Manual bootscanning was performed by using the Juke-Cantor algorithm and the neighbor-joining method [47] , with a window size of 200 nt, a step size of 20 nt and 100 replicates. Pairwise identities between sequences were determined with SimPlot software method [48] ,with a window size of 200 nt and a step size of 20 nt. Pairwise homology matrices were obtained by using CLC Combined Workbench 3.0 software (CLC bio, Aarhus, Denmark). The original P1-VP2 sequences described in this study were deposited in GenBank under accession nos. GQ223119 to GQ223136. The VP4-VP2 sequences were deposited under the nos GQ223137 to GQ223181, and the P1-P2 sequences under the nos GQ223182 to GQ223226. The full length genomes sequences of N4, N10 and N13 strains were deposited under the nos. GQ223227, GQ223228, GQ223229, respectively.

Cytomegalovirus infection in critically ill patients: a systematic review

INTRODUCTION: The precise role of cytomegalovirus (CMV) infection in contributing to outcomes in critically ill immunocompetent patients has not been fully defined. METHODS: Studies in which critically ill immunocompetent adults were monitored for CMV infection in the intensive care unit (ICU) were reviewed. RESULTS: CMV infection occurs in 0 to 36% of critically ill patients, mostly between 4 and 12 days after ICU admission. Potential risk factors for CMV infection include sepsis, requirement of mechanical ventilation, and transfusions. Prolonged mechanical ventilation (21 to 39 days vs. 13 to 24 days) and duration of ICU stay (33 to 69 days vs. 22 to 48 days) correlated significantly with a higher risk of CMV infection. Mortality rates in patients with CMV infection were higher in some but not all studies. Whether CMV produces febrile syndrome or end-organ disease directly in these patients is not known. CONCLUSIONS: CMV infection frequently occurs in critically ill immunocompetent patients and may be associated with poor outcomes. Further studies are warranted to identify subsets of patients who are likely to develop CMV infection and to determine the impact of antiviral agents on clinically meaningful outcomes in these patients.

Cytomegalovirus (CMV) is a major  herpes virus and a significant human pathogen. Infection is common with seroprevalence rates increasing steadily from 65% among 40 to 49 year olds to 91% in those aged 80 years or over [1] . After primary infection, CMV, like other  herpes viruses, establishes lifelong latency. In immunocompetent individuals, asymptomatic viral shedding may be detectable in saliva or urine; however, cell-mediated host immune responses prevent the development of overt CMV disease. In contrast, CMV infection has been shown to lead to significant disease in immunocompromised hosts such as those with HIV infection or transplant recipients. End-stage HIVinfected patients with a CD4 lymphocyte count of less than 50 cells/mm 3 are at the highest risk of developing CMV retinitis [2] . In transplant recipients, CMV disease occurs in 11 to 72% of patients especially in the first three months after transplant while the patients are receiving maximum immunosuppression [3] . In addition to febrile syndrome and end-organ disease directly as a result of viral replication, immunomodulatory characteristics of CMV may contribute to opportunistic infections, allograft rejection, and higher mortality in transplant recipients. Clinical trials have shown that preventive approaches utilizing antiviral agents have lead to a reduction in the rates of CMV infection and disease, and indirect sequelae associated with CMV [3] [4] [5] . Currently, prophylaxis or periodic monitoring and antiviral therapy targeted towards patients with viral replication are routinely employed at many transplant centers. APACHE : Acute Physiology and Chronic Health Evaluation; ARDS: acute respiratory distress syndrome; BAL: bronchoalveolar lavage; CI: confidence interval; CMV: cytomegalovirus; HHV: human herpesvirus; HSV: herpes simplex virus; ICU: intensive care unit; IL: interleukin; MeSH: medical subject headings; NF-B: nuclear factor-B; OR: odds ratio; PCR: polymerase chain reaction; SAPS: simplified acute physiology score; SOFA: sepsis-related organ failure assessment, TNF: tumor necrosis factor; VAP: ventilator-associated pneumonia. It has increasingly come to be recognized that critically ill patients who are traditionally considered immunocompetent may also be at risk for CMV infection. For example, septic insult as a result of bacterial or fungal infections has the potential to promote the release of immunomodulatory cytokines and lead to reactivation of CMV [6, 7] . Reactivation from the latency rather than primary infection is believed to be the cause of CMV infection because none of the critically ill CMV seronegative patients developed CMV infection as opposed to 13 to 56% of seropositive patients [8, 9] . Several observational studies have shown an association between CMV infection in critically ill patients and poor clinical outcomes [8, 10, 11] . However, available data are limited by relatively small sample sizes, diversity in patient populations studied, difference in methodological assays employed for CMV, and variability in reported outcomes that preclude generalizability of the results of the individual reports. The objectives of this review are to summarize the frequency and predictors of CMV infection, and outcomes in critically ill immunocompetent patients with CMV infection. Additionally, we discuss the pathophysiologic basis of CMV reactivation and the implications of these data for optimizing outcomes in critically ill patients. English-language reports of published studies on CMV infection in critically ill immunocompetent patients were identified through November 2008 by cross-referencing the following medical subject headings (MeSH) keywords and text words: cytomegalovirus, immunocompetence, critical illness, intensive care units, intensive care, reactivation, sepsis, and shock. Databases searched included PubMed, EMBASE, Cochrane Database of Systematic Reviews, and Cochrane Central Register of Controlled Trials. Bibliographies of original articles were manually reviewed for additional articles. Non-Englishlanguage reports were also identified in PubMed using the same keywords in order to supplement our search. We included studies in which: critically ill immunocompetent adults were monitored either retrospectively or prospectively for the development of CMV infection in the ICU and; the rate of CMV infection was explicitly reported. CMV infection was defined as evidence of positive viral cultures, antigenemia, and/or DNAemia by PCR from blood or a clinical specimen. Patients were considered to be immunocompetent if they were not solid organ or hematopoietic stem cell transplant recipients, not infected with HIV, did not have primary immunodeficiencies, and were not recipients of immunosuppressive agents such as calcineurin-inhibitors, anti-TNF- drugs, antilymphocyte antibodies, or chemotherapeutic agents for treating cancer. We excluded studies in which an increase in CMV serologic titers in the absence of viremia was the sole evidence for CMV infection. Two of the authors independently searched articles and extracted the following data for analyses: study design, inclusion criteria, type and frequency of CMV assays, rate of CMV infection, rate of CMV IgG positivity, the time elapsed from ICU admission to CMV infection, risk factors for CMV infection, and outcomes (i.e. mortality, duration of ICU stay). Any discrepancies were resolved by review and discussion. Authors of published studies were contacted if reported data required further clarification. Additional mortality data was provided in one study [12] . The initial database search identified 524 English-language and 77 non-English-language studies. After review of the title and abstract and manual search for bibliographies of the potentially relevant articles, 26 studies were selected for fulltext review . Ten studies were excluded after full-text review because: CMV infection was diagnosed based on an increase in titers [19] [20] [21] [22] [23] [24] ; the study was not explicitly conducted in the ICU [25] ; or the rate of CMV infection was not reported [26, 27] . Data from one institution with overlapping study cohorts in two articles were analyzed only once to avoid duplication in the results [16, 28] . We found three studies in which CMV disease was sought as etiology of acute respiratory distress syndrome (ARDS) or ventilator-associated pneumonia (VAP) [29] [30] [31] . Considering that CMV disease (organspecific symptoms or signs plus the detection of CMV in organ biopsy samples by histopathology) is a distinct entity with worse outcomes than CMV infection, the data from these studies were summarized separately. Thus, we identified a total of 13 studies that have described CMV infection in immunocompetent critically ill patients with sample sizes ranging from 23 to 237 (Table 1) [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] . These included nine prospective observational studies [7, 8, [10] [11] [12] [13] [14] [15] [16] , three retrospective studies [9, 17, 18] , and one study where insufficient details were available to determine the type of the study [6] . CMV IgG was positive in all the subjects in four studies [10, 11, 14, 16] , not measured in four studies [12, 15, 17, 18] , and positive in 28 to 94% of the subjects in five studies [6] [7] [8] [9] 13] . Other inclusion criteria were sepsis in four studies [6, 7, 16, 17] , mediastinitis in one study [13] , simplified acute physiology score (SAPS) II of 41 or higher in one study [10] , prolonged ICU stay in three studies [8, 9, 16] , fever for more than 72 hours in one study [18] , and shock or organ failure in two studies [12, 16] . There were five studies in which corticosteroid use was documented [8, 12, [16] [17] [18] . In one study, eight of 48 patients were recipients of long-term corticosteroid therapy or had a malignancy [12] . In a case-control study, 22 of 40 of the cases and 13 of 40 of the controls were recipients of short-term corticosteroids ( 3 months) [18] . The details of corticosteroid use were unavailable in three studies [8, 16, 17] . Based on nine prospective studies that assessed CMV infection in all study subjects, the rate of CMV infection ranged from 0 to 36% with the median rate of 25% [7, 8, [10] [11] [12] [13] [14] [15] [16] . The specimens used to assess CMV infection included blood, urine, and respiratory secretions. All the studies used blood to assess CMV infection. Blood was used solely in four studies [7, 11, 12, 15] , whereas urine and respiratory specimen in addition to blood samples were used in two [13, 16] and four studies [8, 10, 14, 16] , respectively. The rate of CMV infection was reported separately based on the type of specimens in these studies (Table 1 ). Considering only studies that assessed for viremia, the CMV infection rate ranged from 0 to 33% with the median rate of 20%. Thereafter, we reported the rate of CMV infection diagnosed based on the presence of viremia. Next, studies were categorized based on the frequency of virologic monitoring because it can influence the rate of CMV infection. In five studies in which CMV infection was monitored at least weekly, the rate of CMV infection ranged from 6 to 33% with the median of 32% [7, 8, 10, 11, 16] . In contrast, CMV infection rate was 0.8 to 2.1% in two studies in which assessment for CMV infection was performed only once, 1.8 to 4 days after ICU admission [12, 15] . There are three methods of diagnosing CMV infection: viral cultures, antigenemia, and PCR assays [32] . Culture-based assays (conventional and shell-vial cultures) are considered obsolete because of their low sensitivity and time-consuming nature. The antigenemia assay is based on direct detection of the CMV protein pp65 using monoclonal antibodies. It is sensitive and quantitative, although it requires sufficient leukocytes in peripheral blood and is more labor-intensive than the PCR assays. Finally, the PCR assays have been considered gold standard given their high sensitivity and rapid turnover time, although these are not fully standardized [33] . In our review, the assays used to assess CMV infection were viral cultures in five studies [8, 10, 13, 14, 16] , antigenemia in three studies [7, 12, 16] , and PCR in six studies [7, [10] [11] [12] 14, 15] . CMV infection rate was 0 to 20% (median 4%), 0 to 32% [ (median 18%), and 0 to 33% (median 17%) by viral cultures, antigenemia, and PCR, respectively in these indices. PCR assay and antigenemia were performed in all patients in two studies [7, 12] . Twelve of 82 patients developed CMV infection detected by the PCR assay compared with six of the 82 patients by antigenemia. No cases of CMV infection were diagnosed solely by antigenemia. Lastly, we focused on three studies in which the PCR assay was used at least weekly because this method is currently widely utilized to monitor CMV infection; the rate of CMV infection was 32 to 33% in these studies [7, 10, 11] . Based on five studies where CMV infection was assessed at least weekly, the mean (or median) time to onset of CMV infection ranged from 4 to 28 days [7, 8, 10, 11, 16] . Among studies where PCR was used, the mean (or median) time ranged from 4 to 12 days [7, 10, 11] . A PCR assay seems to help diagnose CMV infection earlier than the antigenemia assay. In a study that used both methods for the diagnosis of CMV infection, the median time between onset of sepsis and CMV infection determined by PCR and antigenemia were four days (range 1 to 23 days) and 11 days (range 1 to 23 days), respectively [7] . Candidate variables assessed as predictors for CMV infection varied for different studies and primarily included demographic and clinical characteristics, and severity of illness markers. Age did not appear to be a risk factor for CMV infection [7, 8, 10, 11, 13, 16, 18] and the association between CMV infection and gender was inconsistent [7, 8, 10, 11, 13, 16, 18] . Other risk factors identified included mechanical ventilation at admission (odds ratio [OR] 8.5, 95% confidence interval [CI] 1.1 to 66.5 for high-grade CMV viremia, i.e. CMV PCR > 1000 copies/ml) [11] , bacterial pneumonia [8] , and sepsis (OR 4.62, P = 0.02) [10] . Corticosteroid use was a risk factor in one study [8] . In a retrospective case-control study where variables used in a multivariate logistic regression model were not clearly documented, neither corticosteroid use nor sepsis was a risk factor for CMV infection [18] . Transfusion within 24 hours of admission was associated with high-grade viremia, i.e. CMV viral load greater than 1000 copies/ml (OR 6.7, 95% CI 1.1 to 42.7) [11] ; however, no association between CMV infection and transfusion during hospitalization was documented in two other studies [18] . The mean number of packed red blood cell transfusion was larger in patients with CMV infection than in those without it (22.3 units vs. 11.2 units, P = 0.002); however, this difference was not statistically significant after controlling for other risk factors [8] . Malignancy was not associated with CMV infection [9, 10, 18] . None of the disease severity scores including Acute Physiology and Chronic Health Evaluation (APACHE) II score [7, 8, 11, 13] , sepsisrelated organ failure assessment (SOFA) score [16] , or SAPS II [10, 18] correlated with a risk of CMV infection. Organ dysfunction Organ dysfunction was reported in three studies [8, 13, 18] . One study showed a higher incidence of hepatic dysfunction (international normalized ratio > 1.5 or total bilirubin > 2.5 mg/ dL) in CMV infection group (70% vs. 36%, P < 0.047) [8] . In two studies, renal failure was observed more frequently in those with CMV infection (55 to 58% vs. 32 to 33%, P < 0.05 in each study) [13, 18] . Based on six studies where the duration of ICU stay was assessed, the mean (or median) duration of ICU stay ranged from 33 to 69 days in patients with CMV infection as compared with 22 to 48 days among those without (P < 0.05 in each study) [8] [9] [10] 13, 16, 18] . The mean (or median) duration of mechanical ventilation ranged from 21 to 39 days in patients with CMV infection compared with 13 to 24 days in those without CMV infection in four studies (P < 0.05 in each study) [8, 9, 16, 18] . Nosocomial infection was more frequently observed in patients with CMV infection as compared with those without CMV infection in one study (75% vs. 50%, P = 0.04) [18] . A higher level of CMV viremia was associated with death or continued ICU hospitalization at 30 days in one study [11] . An OR of combined outcome of death or continued ICU hospitalization at 30 days was 1.7 (95% CI 1.2 to 2.4) for each logarithmic increase in maximum CMV viral load measured [11] . Mortality rate in critically ill patients with CMV infection was 29 to 100% as compared with 11 to 74% in those without CMV infection. Except for two retrospective studies [9, 18] , no other studies showed a significant difference in the mortality rates between those with and without CMV infection. CMV viremia at any level was associated with death or continued ICU hospitalization at 30 days (OR 5.7, 95% CI 2.1 to 15.6) [11] . Human herpesvirus (HHV) -6 and -7 have been associated with a greater risk of developing CMV disease in transplant recipients [34] [35] [36] [37] . Furthermore, HHV-6 and HHV-7 have been shown to be significant contributors to morbidity and poor outcomes, particularly when concurrent infection with CMV exists [35, [38] [39] [40] . Thus, it is of interest to investigate whether the association between CMV and other herpes viruses exists in critically ill patients. HHV-6 infection has been frequently observed in critically ill patients [12, 15] . In one study, HHV-6 infection occurred in 53.5% of all patients (54/101) requiring hospitalization in the ICU as compared with none of the healthy volunteers [15] . In another study, HHV-6 infection occurred in 54% of all ICU patients with at least two organ failures as opposed to 15% of those with less than two organ failures [12] . HHV-6 viremia was assessed by the PCR assay at a mean of day 4 and day 1.8, respectively [12, 15] . Potential association between CMV and HHV-6 infection could not be assessed because only one patient developed CMV infection in each study. Both studies failed to show an association between HHV-6 infection and mortality. HHV-7 infection was observed more commonly in healthy volunteers (18/50, 36%) as compared with ICU patients (14/101, 14%; P = 0.002) in one study [15] . Three studies evaluated herpes simplex virus (HSV) infection in conjunction with CMV infection. In one study, HSV was isolated in the bronchial aspirate of 8 of 25 patients (32%) consisting of six of eight patients with CMV reactivation and 2 of 17 patients without CMV infection (P = 0.004) [16] . In other studies, none or only one patient developed co-infection of CMV and HSV, therefore, the association between CMV and HSV could not be evaluated [8, 17] . CMV disease and acute respiratory distress syndrome or ventilator-associated pneumonia CMV pneumonia was diagnosed by open-lung biopsy while investigating the etiology of pulmonary disease in 29 to 50% of critically ill patients with ARDS or VAP (Table 2 ) [29] [30] [31] . In a study in which open-lung biopsy led to the diagnosis of CMV pneumonia in 30% of critically ill patients with ARDS, lung tissue culture was positive only in 10% of these patients [29] . The sensitivity/specificity of bronchoalveolar lavage (BAL), blood, and urine culture for the diagnosis of histologically proven ventilator-associated CMV pneumonia was 53%/92%, 20%/83%, and 13%/62%, respectively [31] . Clinical outcome data were largely lacking; however, there was no difference in duration of ICU stay in one study [31] . Our review demonstrates that depending on the methodological assay used and the patient populations studied, CMV infection occurs in 0 to 36% of the critically ill otherwise immunocompetent hosts in the ICU. Among the most frequently studied inciting event for CMV infection in these patients is sepsis [6, 7, 13, 16, 17] . The risk of CMV infection was five-fold higher in patients with sepsis even when controlled for age and the initial severity of illness [10] . In a murine model of CMV infection, cecal ligation and puncture resembling post-surgical intraabdominal sepsis led to reactivation of latent CMV in the lungs and ultimately pulmonary fibrosis [41, 42] . The propensity of sepsis to promote CMV infection may result from its pleiotropic effects on the host immune system. Pro-inflammatory cytokine production such as TNF- and IL-1 in the early phase of sepsis has the potential to activate NF-B and other transcription factors that are key in the reactivation of CMV from latency [43, 44] . The later phase of sepsis, characterized by the generation of immunosuppressive cytokines such as IL-10 and IL-4 is often referred to as compensatory anti-inflammatory response syndrome [45, 46] . Once latent virus is reactivated, these cytokines may further enhance CMV replication. Indeed, in lung transplant recipients, elevated levels of IL-10 in the BAL and/or plasma were associated with delayed CMV clearance [47] . A sustained high level of IL-10 in patients with sepsis has been associated with poor outcomes, presumably due to excessive anti-inflammatory effects [48] . Transfusion within 24 hours of admission was identified as a risk factor for high-grade CMV viremia in critically ill patients [11] . This association may be explained by potential transmission of CMV by blood products, but more likely by the immunomodulatory effect of transfusion per se. Previous studies have shown that allogeneic blood transfusion resulted in a reduction in T-helper cells, induction of suppressor T cells, and suppression of natural killer cell activity [49] . The transfusionrelated immunosuppression has been associated with clinically important sequelae such as improvement of renal allograft survival, increased risk of tumor recurrence, and postoperative infections [50] [51] [52] . The risk of CMV transmission by leukocyte depleted blood products is at least as low as by CMV seronegative blood products [53, 54] , supporting the hypothesis that transfusion-related immunomodulatory effect plays a major role in CMV infection in critically ill patients if transfusion were truly a risk factor. However, these data were not available for most studies. For example, only one in three studies that evaluated transfusion as potential risk factors for CMV infection reported use of leukocyte depleted blood products explicitly [18] . A body of literature based largely on serologic assays for the diagnosis of CMV suggests that severe burn injuries are a major risk factor for CMV infection [55, 56] . At least a four-fold rise in serologic titers suggestive of CMV reactivation has been documented in 45 to 56% of the burn patients [19, 21, 22] . Recently, in a study where patients with severe burn injuries comprised a subset of critically ill patients, CMV viremia using PCR was observed in 55% (11/20) of the burn patients [11] . Burn injuries are associated with profound changes in cell-mediated immunity and a predominant Thelper 1 cell response that may facilitate CMV infection [57] [58] [59] . Susceptibility to sepsis due to the loss of skin integrity in these patients may also contribute to the risk of CMV infection. Attempts to utilize the severity of 'critical illness' to predict CMV infection have not shown a correlation between scoring systems such as APACHE II or SAPS II and the risk of CMV [7, 8, 10, 11, 13, 16, 18] . Severity of illness scores have typically (page number not for citation purposes) been assessed in the first 24 hours after ICU admission whereas CMV infection does not usually occur until late in the ICU stay. Additionally, these scores are based on age, physiologic parameters, basic laboratory values, and chronic medical conditions and may not be necessarily representative of host immunologic deficits that lead to CMV infection. CMV infection rate was 0.8 to 2.1% in two studies where the PCR assay was performed only once at a mean of 1.8 and 4 days following the onset of illness requiring ICU admission [12, 15] . The median time to first detectable CMV viremia was 12 days (range 3 to 57 days) in a study where the PCR assay was performed thrice weekly [11] . Thus, it appears that CMV infection is a rare event very early in the course of critically ill patients and that most infections develop between 4 and 12 days after the onset of illness requiring ICU stay, which could lead to a hypothesis that CMV infection may coincide with the development of compensatory anti-inflammatory response syndrome, and not with the initial surge of pro-inflammatory cytokines. A key question is whether CMV infection adversely affects outcomes in critically ill patients. Virtually all studies have documented that CMV infection was related to prolonged mechanical ventilation and duration of ICU stay in patients with CMV infection. CMV infection has also been associated with organ system failure and at least two studies have documented significantly higher mortality rates in patients with CMV infection compared with those without it [9, 18] . Thus, although these data do not prove a causal association as CMV infection may have been more likely to develop or diagnosed in sicker patients, existing evidence suggests that CMV infection is associated with poor outcomes even in immunocompetent critically ill patients. We believe that a causal association between these can only be assessed by carefully conducted clinical trials designed to show whether suppression of CMV has a mitigating effect on the severity of illness. Another major unresolved issue is whether CMV infection is associated with overt disease or clinical manifestations directly attributable to this virus in critically ill patients. CMV infection in immunocompetent patients generally presents with mononucleosis-like symptoms including fever and malaise with liver enzyme abnormalities [60, 61] , which are typically benign. However, 31 to 42% of the hospitalized patients with CMV infection may have organ involvement [60, 62] and rarely life-threatening CMV infection has also been reported [63, 64] . In critically ill patients, 10% (2/20) of those with CMV infection eventually developed severe CMV disease (pneumonitis, neurologic disease) in one study [10] . CMV pneumonia has also been diagnosed in 29 to 50% of patients with ARDS or VAP [29] [30] [31] ; however, this does not necessarily mean that CMV is the cause of ARDS or VAP. Critical illness due to serious pulmonary disease may predispose these patients to CMV infection in the lungs. In a cohort study in the ICU, 17% of critically ill patients who experienced fever for three or more days had CMV infection [18] . Current guidelines for the evaluation of new fever in critically ill adult patients list transfusion-asso- ciated CMV mononucleosis as a cause of fever [65] . However, it remains to be determined whether and how often CMV produces febrile syndrome and whether coexistent infection with HHV-6 is a contributor to this entity as shown in the transplant setting [34, [36] [37] [38] [39] [40] . Experimental studies have shown that ganciclovir prevented murine CMV reactivation and the development of pulmonary fibrosis in immunocompetent mice with sepsis [42] . Two retrospective studies where small subsets of ICU patients received antiviral agents for CMV infection have yielded inconclusive results and data on the utility and efficacy of antiviral therapy for CMV in critically ill immunocompetent patients are largely lacking [17, 18] . Employment of potent antiviral therapy in all critically ill patients may be impractical, logistically infeasible, and potentially harmful given a large number of ICU patients and potential adverse effects of ganciclovir such as bone marrow suppression or teratogenicity. A more prudent approach may be to identify subgroups of patients at high risk for developing CMV infection and targeting antiviral prophylaxis towards these patients. These subgroups may include patients with sepsis, persistent fever, or those receiving transfusion. An alternative approach is to employ antiviral therapy only in those with CMV viremia. Regardless, carefully conducted clinical trials are warranted to discern the impact of antiviral agents on clinically meaningful outcomes before employing antiviral therapy or even considering routine monitoring of CMV in critically ill patients. Several limitations of our study deserve to be acknowledged. We found considerable heterogeneity in the methodology used to assess CMV infection and in patient characteristics. As noted in the Results, the frequency and type of CMV monitoring influenced the rate of CMV infection. Although all the studies in this review were conducted in the ICU, the overall mortality of the studied patients ranged from 5 to 71% [7, 15] , suggesting that study populations were significantly diverse. Furthermore, these studies were published over a period spanning nearly two decades in different regions with diverse clinical practices. Considering the heterogeneity of available data, quantitative analyses such as meta-analysis can be misleading [66, 67] and our results are therefore presented in a descriptive fashion only. Second, while we excluded the recipients of iatrogenic immunosuppressive agents that enhance the risk of CMV reactivation, critically ill patients in whom corticosteroids were employed were included. Controversy abounds whether corticosteroids alone without other immunosuppressive agents lead to reactivation of CMV from latency or merely promote the replication of activated virus [68] [69] [70] . Corticosteroids were employed in a subset of patients in 5 of 12 studies in this review. Given that corticosteroid use is a common practice in the ICU [71] , these studies reflect clinical scenarios encountered by care providers and therefore their inclusion in this review was deemed appropriate. In summary, accumulating data suggest that CMV infection is a frequent occurrence in critically ill patients. Considering a large number of patients requiring ICU level of care, the scope of impact of CMV infection in these patients may be equally or potentially wider than in other immunocompromised hosts traditionally recognized to be at risk for CMV infection. For example in the USA, an estimated 383,000 cases with sepsis require ICU admission as opposed to 28,360 solid organ transplant cases yearly [72, 73] . Mortality rate in patients with sepsis ranges from 20 to 50% and approaches 70% in those with multiple organ failure [74] . Furthermore, the incidence of sepsis and the number of sepsis-related deaths appear to be increasing [73, 74] . Precise identification of the role of CMV as a contributor to outcomes in these patients may therefore have far reaching implications. Subsets of critically ill patients who are at risk for developing CMV infection and for poor outcomes remains to be determined. These studies have significant implications for future investigations to determine the potential benefits and for guiding the study design to evaluate the impact of antiviral agents on clinically meaningful outcomes in critically ill patients. The authors declare that they have no competing interests. RO participated in the study design, literature search, data acquisition, interpretation of the data, and the drafting of the manuscript. NS participated in the study design, literature search, data acquisition, interpretation of the data, and the revision and editing of the manuscript. Both authors read and approved the final manuscript. • CMV infection occurs in 0 to 36% (median 25%) of critically ill patients between 4 and 12 days after ICU admission, especially those with sepsis, requiring mechanical ventilation, and receiving transfusion. • CMV infection is associated with poor outcomes; however, it is not known whether the causal association exists, that is, CMV is truly a pathogen or CMV infection is just an indicator of immunosuppression. • It remains to be determined whether CMV produces febrile syndrome or end-organ disease directly in critically ill patients. • Further studies are warranted to identify subsets of patients who are at high risk of developing CMV infection and to determine the role of antiviral agents on clinically important outcomes in critically ill patients.

Heliox reduces respiratory system resistance in respiratory syncytial virus induced respiratory failure

INTRODUCTION: Respiratory syncytial virus (RSV) lower respiratory tract disease is characterised by narrowing of the airways resulting in increased airway resistance, air-trapping and respiratory acidosis. These problems might be overcome using helium-oxygen gas mixture. However, the effect of mechanical ventilation with heliox in these patients is unclear. The objective of this prospective cross-over study was to determine the effects of mechanical ventilation with heliox 60/40 versus conventional gas on respiratory system resistance, air-trapping and CO2 removal. METHODS: Mechanically ventilated, sedated and paralyzed infants with proven RSV were enrolled within 24 hours after paediatric intensive care unit (PICU)admission. At T = 0, respiratory system mechanics including respiratory system compliance and resistance, and peak expiratory flow rate were measured with the AVEA ventilator. The measurements were repeated at each interval (after 30 minutes of ventilation with heliox, after 30 minutes of ventilation with nitrox and again after 30 minutes of ventilation with heliox). Indices of gas exchange (ventilation and oxygenation index) were calculated at each interval. Air-trapping (defined by relative change in end-expiratory lung volume) was determined by electrical impedance tomography (EIT) at each interval. RESULTS: Thirteen infants were enrolled. In nine, EIT measurements were performed. Mechanical ventilation with heliox significantly decreased respiratory system resistance. This was not accompanied by an improved CO2 elimination, decreased peak expiratory flow rate or decreased end-expiratory lung volume. Importantly, oxygenation remained unaltered throughout the experimental protocol. CONCLUSIONS: Respiratory system resistance is significantly decreased by mechanical ventilation with heliox (ISCRTN98152468).

Respiratory syncytial virus (RSV) is the most important causative agent of lower respiratory tract disease (LRTD) in infancy [1] . Approximately 100,000 infants are annually admitted with RSV-induced bronchiolitis in the USA, and the number of hospitalizations is increasing [2] . Because of this, RSV-associated disease imposes a major burden on health care resources [3] . There is no effective therapy against RSV available, prevention can only be achieved through passive immunisation using monoclonal antibodies [4] . RSV LRTD is pathophysiologically characterized by sloughed necrotic epithelium, excessive mucus secretion, bronchial mucosal oedema and peribronchial inflammation that contributes to airway obstruction resulting in increased airway resistance with subsequent air-ANOVA: analysis of variance; ARDS: acute respiratory distress syndrome; CO 2 : carbon dioxide; Cstat: static compliance; EELV: end-expiratory lung volume; EIT: electrical impedance tomography; ELISA: enzyme-linked immunosorbent assay; ET-CO 2 : end-tidal carbon dioxide; FiO 2 : fraction of inspired oxygen; LRTD: lower respiratory tract disease; MAP: mean airway pressure; MV: mechanical ventilation; OI: oxygenation index; PaCO 2 : partial pressure of arterial carbon dioxide; PaO 2 : partial pressure of arterial oxygen; PEEP: positive end-expiratory pressure; PEFR: peak expiratory flow rate; PICU: paediatric intensive care unit; PIP: positive inspiratory pressure; Ptrach: intratracheal pressure; relative Δ EELV : relative change in end-expiratory lung volume; Rlung: lung resistance; Rrs: respiratory system resistance; RSV: respiratory syncytial virus; SPO 2 : oxygen saturation; V D : dead space; VI: ventilation index; Vte: expiratory tidal volume. trapping and respiratory acidosis [5, 6] . Although the majority of infections run a mild disease course, mechanical ventilation (MV) for up to 10 days is necessitated in approximately 2% to 16% of previously healthy hospitalised infants due to severe lower respiratory tract infection including bronchiolitis or pneumonia [1, 7, 8] . Helium is an inert gas with a density that is one-seventh that of air. In addition, carbon dioxide (CO 2 ) diffuses more easily through helium than through air [9] . With helium, a more laminar flow is preserved in narrowed airways, resulting in lower resistance to gas flow allowing for increased bulk flow [10] . Based on these properties, MV with heliox could be considered in mechanically ventilated infants with RSV LRTD. Its use in these patients has been studied once but with inconclusive results [11] . We hypothesized that the use of heliox in mechanically ventilated infants with RSV LRTD would result in decreased respiratory system resistance (R rs ). In addition, MV with heliox would result in less air-trapping defined by the relative change in end-expiratory lung volume (EELV), and improved CO 2 clearance. The objective of our study was to test this hypothesis in a prospective, double cross-over intervention trial comparing heliox 60/40 with conventional gas (nitrox) using lung function testing and electrical impedance tomography (EIT) measurements. The study protocol (ISCRTN98152468) was approved by the hospital's Institutional Review Board and written informed consent was obtained from patients before enrollment. Eligible for inclusion were infants younger than 12 months of age with a virologically confirmed clinical diagnosis of RSV LRTD (either a positive direct immunofluorescent assay or ELISA) who were admitted to the nine-bed paediatric intensive care unit (PICU) facility of the VU university medical center for MV during the RSV seasons (autumn and winter) between 2005 and 2007. Infants were excluded if no informed consent was obtained, fraction of inspired oxygen (FiO 2 ) was more than 0.4, corticosteroids were used prior to admission, they were on high-frequency oscillatory ventilation or a haemodynamically significant congenital heart defect (i.e. significant left-to-right shunting with or without pulmonary hypertension) was present. Patients were in supine position, intubated with an uncuffed endotracheal tube size 3.5 or 4.0 mm, and put on a timecycled, pressure-limited ventilation mode (Pressure Control, AVEA ventilator, Cardinal Health, Yorba Linda, CA, USA). Aims of ventilation were transcutaneously measured oxygen saturation (SpO 2 ) 88 to 92%, and partial pressure of arterial carbon dioxide (PaCO 2 ) 45 to 65 mmHg (if pH >7.25). Inspir-atory times were fixed at 0.5 seconds, positive end-expiratory pressure (PEEP) was set 1 to 2 cmH 2 O below total PEEP (i.e. extrinsic PEEP + intrinsic PEEP). The flow-time curve was observed thoroughly throughout the study period in each patient to examine if expiration was complete in order to prevent dynamic hyperinflation. Patients were sedated with midazolam and morphine, paralysis was achieved using intravenous rocuronium. Endotracheal suctioning was performed 30 minutes prior to the start of, but not during, the experimental protocol. Bronchodilators (either nebulized or intravenous) or ketamine were not used before or during the study period. Arterial blood samples were drawn from an arterial line to determine PaCO 2 and partial pressure of arterial oxygen (PaO 2 ). End-tidal carbon dioxide (ET-CO 2 ) concentration, and expiratory tidal volume (V Te ) were measured at the airway opening. ET-CO 2 was measured using a side-stream Microstream (Philips Medical Systems, Best, The Netherlands) and V Te was measured with a proximal flow sensor connected to the AVEA ventilator (Cardinal Health, Yorba Linda, CA, USA). The ventilator is designed to detect which gas is used and adjusts its pneumotachograph automatically in order to measure the correct V Te . A chest radiograph was obtained and evaluated by one pediatric radiologist in each patient prior to the start of the experimental protocol to evaluate the presence of hyperinflation (defined by a depression of the diaphragm below the sixth anterior rib) or an infiltrate (described as opacities with irregular markings without loss of volume) [12] . The experimental protocol started within 24 hours of PICU admission and lasted for 90 minutes. At four intervals (T = 0 (baseline), T = 30, T = 60, and T = 90 minutes) data were collected and respiratory variables measured. At T = 0 and T = 60, patients were ventilated with nitrox. At T = 30 and T = 90, patients were ventilated with heliox (helium 60%, oxygen 40%). Ventilator settings were kept constant throughout the experimental protocol. Positive inspiratory pressure (PIP), intratracheal pressure (Ptrach), mean airway pressure (MAP), PEEP, SpO 2 , ET-CO 2 , respiratory rate and V Te were measured. Ptrach was measured with a pressure transducer placed at the distal end of the endotracheal tube. Blood samples were drawn for the determination of the PaO 2 , PaCO 2 and pH. Static compliance (C stat ), R rs and peak expiratory flow rate (PEFR) were measured using the AVEA ventilator (Cardinal Health, Yorba Linda, CA, USA) according to the manufacturer's manual. In summary, R rs was defined by the ratio of the airway pressure differential to the inspiratory flow 12 ms prior to the end of inspiration. Lung resistance (R lung ) was defined by the ratio of the tracheal pressure differential to the inspiratory flow 12 ms prior to the end of inspiration. At each interval, EIT measurements were made using the Göttingen Goe-MF II EIT system (Cardinal Health, Yorba Linda, CA, USA). Sixteen electrodes (Blue Sensor BR-50-K, Ambu, Denmark) were applied circumferentially around the infant's chest at the mammary line. A 30 second reference measurement at 13 Hz scan rate was recorded. All further measurements were referenced to this measurement. All other measurements were made at a scan rate of 44 Hz for 180 seconds. A 5 mA peak-to-peak, 50 kHz electrical current was injected at each adjacent electrode pair, and the resultant potential differences were measured at the remaining adjacent electrode pairs. Subsequently, all adjacent electrode pairs were used for current injection, thus completing one data cycle. The impedance map was built using the back-projection image reconstruction algorithm [13] . It calculates the relative impedance ΔZ, defined by (Z inst -Z ref )/Z ref (where Z inst is the instantaneous local impedance and Z ref the reference impedance, determined from each cycle of current injections and voltage measurements in each pixel). Both the respiratory and cardiac components of the EIT signal were identified in the frequency spectra generated from all EIT measurements (Fourier transformation). The EIT data was lowpass filtered with a cut-off frequency of 2 Hz to eliminate small impedance changes synchronous with the heart beat [14] . The calculations performed on the sums of values from all pixels of the 32 × 32 pixel matrix EIT image were described as 'global'. In addition, sums of values from the left and right lung regions were described separately, and the entire EIT image was divided into 64 regions-of-interest (32 left and 32 right lung) from anterior to posterior as previously described by Frerichs and colleagues [15] . Ventilation-induced tidal volume (ΔZ VT ) was quantified by measuring the relative ΔZ from the highest point at end inspiration to the lowest point at end expiration, and an average ΔZ was calculated from multiple breaths. Changes in ΔZ VT were calibrated to volume using the known V T . The relative change in end-expiratory lung volume (relative ΔZ EELV ) was determined by measuring the median impedance from the lowest point at expiration during the sampling time (Z EELV ) [16] . The relative ΔZ EELV was normalized to volume (relative Δ EELV in ml) by multiplying the median impedance with the ratio V T /ΔZ VT . The oxygenation index (OI) was calculated as follows: (FiO 2 × 100 × MAP in cmH 2 O)/PaO 2 in mmHg. The ventilation index (VI) was calculated as follows: (PaCO 2 in mmHg × respiratory rate × (PIP -PEEP in cmH 2 O))/1000. VI is used as determinant for CO 2 elimination because the respiratory rate, PIP, and PEEP were kept constant throughout the study period [17] . Dead space (V D ) was calculated according to the Bohr-Enghoff equation: V D = V Te × (1 -(P ET-CO2 /PaCO 2 )) [18] . As no data on relative Δ EELV in mechanically ventilated infants with RSV LRTD were available, we performed a power analysis after inclusion of all patients using the paired t-test. The data were analyzed with one-way repeated measures analysis-of-variance (ANOVA) with Tukey post-hoc testing between T = 0 versus T = 30, T = 30 versus T = 60, and T = 60 versus T = 90. P < 0.05 was accepted as being statistically significant. Data are expressed as mean ± standard deviation unless stated otherwise. Statistical analysis was performed using SPSS version 15.0 (Chicago, IL, USA). Thirteen patients were included in 11 EIT studies; good-quality EIT signals were obtained from nine patients. Descriptive data, ventilator settings and baseline data of respiratory system mechanics and gas exchange are summarized in Table 1 . Although three patients were born prematurely (one at 32 weeks and two at 36 weeks' gestation), none of the patients had chronic lung disease. Hyperinflation was present in 10 patients, four of these patients also had infiltrates. Ten patients had hypercapnia (PaCO 2 >45 mmHg) and seven infants had PaO 2 /FiO 2 less than 200 at baseline (T = 0). Tidal volume remained constant throughout the experiment (Figure 1 ). Leakage around the uncuffed endotracheal tube was less than 5% in all patients. Mechanical ventilation with heliox had an overall significant effect on R rs (P < 0.001; Figure 2 ). R rs decreased from 69.1 ± 6.9 cmH 2 O/L/sec at T = 0 to 50.2 ± 6.0 cmH 2 O/L/sec (P = 0.020) after 30 minutes of ventilation with heliox. After reintroduction of nitrox, R rs increased significantly to 70.7 ± 7.2 cmH 2 O/L/sec (P = 0.016) but decreased again to 42.9 ± 3.3 cmH 2 O/L/sec (P = 0.001) when heliox was reintroduced. Course of tidal volume Course of tidal volume. (page number not for citation purposes) R lung was not significantly influenced by MV with heliox ( Figure 3 ). PEFR was not significantly improved by MV with heliox compared with nitrox (P = 0.520; Figure 4 ). C stat was 1.9 ± 0.4 L/ cmH 2 O at T = 0 and not significantly different throughout the study (P = 0.214; Figure 5 ). The mean relative Δ EELV ± standard deviation at T = 0 was 76.6 ± 15.1 ml. With an estimated reduction of 25% with heliox, nine patients were needed to recruit in order to detect a statistically significant difference with α 0.05 and β 0.90. The degree of airtrapping as defined by the relative Δ EELV in ml was overall not significantly reduced by heliox (P = 0.493; Figure 6 ). This was due to differences in response to MV with heliox. Five patients showed a reduction in relative Δ EELV when heliox was introduced, and when conventional gas was reintroduced relative Δ EELV increased in only three patients (Table 2 ). There were also patients who had an increase of relative Δ EELV with heliox that was either reversed or increased when conventional gas was reintroduced. To investigate if a time-dependent effect of heliox could be found, the change in relative Δ EELV was correlated with the change in R rs for T = 30 to T = 0 (R 2 0.068, P = NS), T = 60 to T = 30 (R 2 0.110, P = NS) and T = 90 to T = 60 (R 2 0.498, P = 0.01). Fractional ventilation (i.e. the distribution between left and right lung), as well as the center of ventilation of the left and right lung, also remained constant throughout the study period (Table 3) . Table 4 summarizes the effect of mechanical ventilation with heliox on indices of gas exchange and V D /V T . Elimination of CO 2 defined by the VI (P = 0.661), as well as a reduction in V D /V T (P = 0.929) was not positively influenced by MV with heliox. Importantly, oxygenation as defined by the OI (P = C stat = static compliance; EIT = electrical impedance tomography; N/A = not available; PaCO 2 = partial pressure of arterial carbon dioxide; PaO 2 = partial pressure of arterial oxygen; PEEP = positive end-expiratory pressure; PEFR = peak expiratory flow rate; PIP = positive inspiratory pressure; Pt = patient; R rs = respiratory system resistance. 0.477) and alveolo-arterial oxygen gradient (Aa-DO 2 ) remained unaltered throughout the study period. The major finding of our study is that MV of infants with RSV LRTD with heliox 60/40 resulted in a significant reduction of the respiratory system resistance. Increased R rs resulting from airway narrowing due to sludging, excessive mucus secretion, edema, and possible bronchoconstriction has been described in mechanically ventilated infants with RSV LRTD [19] [20] [21] [22] [23] . Measures to alleviate increased R rs such as nebulisation of bronchodilators or nitric oxide have yielded inconclusive results [20, 22, 24, 25] . However, these studies are methodologically different compared with ours. For instance, we excluded patients with chronic lung disease or congenital heart disease. The decrease in R rs led not to an improved CO 2 clearance as defined by the VI or a reduction in PEFR. Some explanations for this may be proposed. First, it is uncertain how much of the observed reduction in R rs could be partitioned to the ventilator circuit or the endotracheal tube because no endotracheal suctioning was performed during the study. Increased mucus production during RSV LRTD is common, and may further obstruct the airways [26] . As the AVEA ventilator is able to calculate the R lung , we also studied if MV with heliox resulted in a reduction in R lung , but were unable to demonstrate this. This could mean that MV with heliox does not affect the resistance of the small airways of the infants; it cannot be ruled out, how- Effect of mechanical ventilation with heliox on respiratory system resist-ance Effect of mechanical ventilation with heliox on respiratory system resistance. Data are expressed as mean ± standard deviation. * P < 0.05 T = 30 vs T = 0; ** P < 0.05 T = 60 vs T = 30; *** P < 0.05 T = 90 vs T = 60. Effect of mechanical ventilation with heliox on lung resistance Effect of mechanical ventilation with heliox on lung resistance. Data are expressed as mean ± standard deviation. Effect of mechanical ventilation with heliox on peak expiratory flow rate Effect of mechanical ventilation with heliox on peak expiratory flow rate. Data are expressed as mean ± standard deviation. Effect of mechanical ventilation with heliox on static compliance Effect of mechanical ventilation with heliox on static compliance. Data are expressed as mean ± standard deviation. ever, that the resolution of the AVEA's signal of R lung (1 decimal) might not be sufficient enough to detect true differences in R lung in small children with little tidal volume. Second, the measured R rs in our patients is lower than previously reported in mechanically ventilated infants with RSV LRTD designated to have an obstructive disease phenotype [20, 22, 27] . This could indicate that our patients had mild-to-moderate airway obstruction, although hyperinflation suggesting airway obstruction on chest radiograph was present in all but one patient. Unfortunately, there is no gold standard for the radiological definition of hyperinflation especially in mechanically ventilated infants. Furthermore, the degree of air-trapping might vary between patients, indicating that severe RSV LRTD necessitating MV is a heterogeneous disease in which patients express to a varying degree both restrictive and obstructive disease characteristics explaining why some patients had a PaO 2 /FiO 2 ratio of less than 200 or a C stat less than 0.3 ml/cmH 2 O/kg in our study. This assumption opposes the previously proposed dichotomization of RSV LRTD by Hammer and colleagues, who have observed that mechanically ventilated infants with RSV LRTD showed either a disease pattern compatible with acute respiratory distress syndrome (ARDS) or a disease pattern characterized by increased airway resistance [27] . Although our study was not designed to investigate differences in clinical phenotype, we would dare to challenge this dichotomy in clinical phenotype for several reasons. Hammer and colleagues included prematurely born infants with chronic lung disease and infants with congenital heart disease [27] . C rs is significantly lower in these patients compared with healthy infants [28] [29] [30] . In addition, the term 'bronchiolitis' to describe RSV LRTD is strictly speaking a histopathologic diagnosis and hampered by universal differences in its clinical interpretation [31] . Controversy exists about whether differences in parameters for gas exchange correlate with clinical phenotype [32, 33] . The lack of improved CO 2 clearance in our study is compatible with the observations by Gross and colleagues [11] . They were unable to demonstrate a beneficial effect on PaCO 2 of various heliox mixtures (ranging from 50%/50% to 70%/30%) compared with T = 0 (PaCO 2 45 ± 10 mmHg) in 10 mechanically ventilated infants with moderate severe RSV LRTD. It should be mentioned, however, that our study population was probably more ill than theirs based on a higher T = 0 PaCO 2 and lower PaO 2 /FiO 2 ratio. Previously, we did observe a beneficial effect of heliox in a small infant with obstructive airway Effect of mechanical ventilation with heliox on relative change in end-expiratory lung volume Effect of mechanical ventilation with heliox on relative change in endexpiratory lung volume. Data are expressed as mean ± standard deviation. Negative values indicate a decrease in relative change in end-expiratory lung volume (relative Δ EELV ). EIT = electrical impedance tomography. disease [34] . This disparity in results cannot easily be explained except for the fact that this particular patient had severe respiratory acidosis. EIT is a non-invasive bedside technique to assess global and regional lung volumes that has primarily been used in acute lung injury or ARDS [35] . Hinz and colleagues have shown that compared with the validated nitrogen-washout method it is an appropriate tool to study EELV in critically ill patients [16] . To our knowledge, the use of EIT in the determination of the dynamic process of air-trapping in patients with small airway disease has not been used before, although its use in this disease condition can be rationalised. In our study, MV with heliox did not result in a universal reduction of air-trapping as defined by the relative Δ EELV . However, there were some patients who seemed to benefit from MV with heliox as they did show a reduction in relative Δ EELV that was reversed by MV with conventional gas. Several explanations for the non-universal reduction in relative Δ EELV may be proposed. First, not all alveoli have the same degree of hyperinflation due to the difference in time constants throughout the lung, indicating that hyperinflation is a regional phenomenon rather than a global problem [36] . This would implicate that the technique of EIT may be insufficient to detect regional differences in viralinduced small airway disease due to heterogeneity of the disease, a problem that can be overcome by increasing the reso-lution of the EIT signal. In favor of EIT, however, is the study by Adler and colleagues showing that with EIT dynamic hyperinflation could be adequately monitored [37] . Second, during the study no endotracheal suctioning was performed. Increased mucus production could obstruct the airways, resulting in the collapse of alveoli that is reflected by a decrease in EELV. As tidal volume remained constant throughout the experiment, we think that not performing endotracheal suctioning did not influence our results ( Figure 5 ). Third, if there is a difference in expression of clinical phenotype of RSV LRTD a universal response in relative Δ EELV would not be expected. Some patients responded with a decrease in relative Δ EELV whereas others did not in our study. Also, redistribution of ventilation within each lung or between the left and right lung was not significantly influenced by MV with heliox. This is in line with a heterogeneous clinical phenotype of RSV LRTD. There are some limitations to our study that should be mentioned. First, the small sample size of our study. This sample size does not allow discrimination between responders and non-responders nor a categorization of clinical phenotype based on chest radiographs, but this should be the subject of further research. Second, patients were paralyzed throughout the study, thus prohibiting spontaneous breathing and mucus clearance by the patient itself. We choose to do so to eliminate any confounding effect of spontaneous breathing on the Data are expressed as mean ± standard deviations. Aa-DO2 = alveolo-arterial oxygen gradient; OI = oxygenation index; VI = ventilation index; V D /V T = dead-space/tidal volume ratio. (page number not for citation purposes) degree of dynamic hyperinflation in order to truly assess the effect of MV with heliox. However, our findings require re-evaluation in spontaneously breathing mechanically ventilated infants. Supportive therapy maintaining spontaneous breathing could very well be a key element while awaiting therapeutic modalities for mechanically ventilated infants with RSV LRTD [38] . Third, the measurements of our study were not blinded because connection of the heliox and the measurements were conducted by one investigator (MK). However, this might have introduced measurement bias. Fourth, ventilation with heliox may have influenced the tidal volume measurements of the AVEA ventilator. The AVEA is equipped with the Bicore CP100™ pulmonary mechanics monitor that has been validated previously [36, 39] . Finally, the AVEA performs in a similar way with respect to tidal volume measurement when heliox is used [40, 41] . MV with heliox significantly reduced R rs in mechanically ventilated infants with RSV LRTD with a heterogenous effect on the degree of hyperinflation and CO 2 elimination. These findings warrant further study in order to identify a subgroup of mechanically ventilated infants with RSV LRTD who might benefit from MV with heliox. • MV with heliox decreases respiratory system resistance in RSV LRTD. • MV with heliox does not reduce air-trapping in RSV LRTD. • MV with heliox does not improve gas exchange in RSV LRTD. • RSV LRTD may actually be a heterogeneous disease.

Regulation of the apoptosis-inducing kinase DRAK2 by cyclooxygenase-2 in colorectal cancer

BACKGROUND: Cyclooxygenase-2 (COX-2) is over-expressed in colorectal cancer (CRC), rendering tumour cells resistant to apoptosis. Selective COX-2 inhibition is effective in CRC prevention, although having adverse cardiovascular effects, thus focus has shifted to downstream pathways. METHODS: Microarray experiments identified genes regulated by COX-2 in HCA7 CRC cells. In vitro and in vivo regulation of DRAK2 (DAP kinase-related apoptosis-inducing kinase 2 or STK17β, an apoptosis-inducing kinase) by COX-2 was validated by qRT-PCR. RESULTS: Inhibition of COX-2 induced apoptosis and enhanced DRAK2 expression in HCA7 cells (4.4-fold increase at 4 h by qRT-PCR, P=0.001), an effect prevented by co-administration of PGE(2). DRAK2 levels were suppressed in a panel of human colorectal tumours (n=10) compared to normal mucosa, and showed inverse correlation with COX-2 expression (R=−0.68, R(2)=0.46, P=0.03). Administration of the selective COX-2 inhibitor rofecoxib to patients with CRC (n=5) induced DRAK2 expression in tumours (2.5-fold increase, P=0.01). In vitro silencing of DRAK2 by RNAi enhanced CRC cell survival following COX-2 inhibitor treatment. CONCLUSION: DRAK2 is a serine–threonine kinase implicated in the regulation of apoptosis and is negatively regulated by COX-2 in vitro and in vivo, suggesting a novel mechanism for the effect of COX-2 on cancer cell survival.

Colorectal cancer (CRC) remains a leading cause of cancer death, with worldwide one million new cases each year and as many as half a million cancer deaths annually (Boyle and Leon, 2002) . Cyclooxygenase-2 (COX-2) expression is increased in the majority of colorectal tumours (Eberhart et al, 1994) and this induction is associated with advanced tumour stage and correlates with poor clinical outcomes (Sheehan et al, 1999) . Nonsteroidal anti-inflammatory drugs (NSAIDs), which inhibit COX activity, show anti-neoplastic effects in vitro (Richter et al, 2001; Sheng et al, 1997) and human studies have demonstrated their use to be associated with a reduced incidence of colorectal neoplasia Sandler et al, 2003) . Although more recent studies have confirmed the chemopreventive activity of COX-2selective NSAIDs (Arber et al, 2006; Baron et al, 2006; Bertagnolli et al, 2006) , it is also clear that their long-term use is associated with an unacceptable increase in the risk of cardiovascular events (Bertagnolli et al, 2006; Bresalier et al, 2005) . The anti-neoplastic properties of these agents result from the inhibition of prostaglandin generation, particularly that of prostaglandin E 2 (PGE 2 ), the most abundant in vivo product of COX-2 activity in CRC cells (Pugh and Thomas, 1994; Rigas et al, 1993) . Although it appears that PGE 2 modulates various processes that are fundamental to tumour cell survival, such as altered proliferation and susceptibility to apoptosis (Sheng et al, 1997 (Sheng et al, , 1998 (Sheng et al, , 2001 Tang et al, 2002) , the precise molecular mechanisms remain unclear. A strong rationale exists therefore to generate a more complete understanding of the downstream targets of COX-2 activity (Doherty and Murray, 2009 ). This may lead to the development of more refined therapies, with side-effect profiles that allow their generalised use. Previous studies examining gene regulation by COX-2 in CRC cells have focused on long time points and have used relatively high doses of NSAIDs (Zhang and DuBois, 2001) . With this in mind, we set out to explore early changes in gene expression in CRC cells resulting from low-dose treatment with a selective COX-2 inhibitor, to improve our understanding of the early signalling events downstream of prostaglandin production. One candidate gene that we have identified, DRAK2 (DAP kinase-related apoptosis-inducing kinase 2 or STK17b), is one of a family of serine threonine kinases that share the ability to induce apoptosis (Sanjo et al, 1998) . The aim of this study was to explore the relationship between COX-2 and DRAK2 as a potential downstream regulator of cell survival in CRC. All cells were grown in culture at 371C in a humidified 5% CO 2 incubator. HCA7 cells were kindly donated by Susan Kirkland (ICRF, London, UK). HCA7 cells were cultured in DMEM with 10% FBS, supplemented with 1 mM sodium pyruvate and 100 mg ml À1 kanamycin, to approximately 90% confluence before treatment. HT29 cells were purchased from the ATCC (Rockville, MD, USA) and maintained in McCoy's 5A medium containing 1.5 mM L-glutamine, 10% FBS, penicillin 100 U ml À1 and streptomycin 100 mg ml À1 . SC236, a selective COX-2 inhibitor, was a gift from Dr Peter Isakson (Searle, Skokie, IL, USA). PGE 2 was purchased from Cayman (St Louis, MO, USA). Staurosporine was purchased from Calbiochem (San Diego, CA, USA). A validated siRNA against a target sequence in exon 3 of the DRAK2 (STK17b) gene was purchased from Ambion Inc (Austin, TX, USA). siRNA to scrambled DRAK2 sequence target (scrambled/negative control) was in vitro transcribed from oligonucleotide template using Silencer siRNA construction kit (Ambion Inc) . pEGFP-N1 vector was purchased from BD Clontech (San Jose, CA, USA). Total RNA was isolated from cells and tissue following homogenisation in RNA lysis buffer (Qiagen GmbH, Hilden, Germany) supplemented with 1% b-mercaptoethanol. Extraction was performed using RNeasy Midi kits (Qiagen GmbH). RNA quality was determined by agarose gel analysis and RNA concentration was determined by spectrophotometry (GeneQuant pro; Amersham Biotech, Bucks, UK). Total RNA was isolated as outlined above from HCA7 cells treated for 4, 6 and 8 h with SC236 (5 mM) or vehicle. cDNA was synthesised using the Custom SuperScript ds-cDNA synthesis kit (Invitrogen, Carlsbad, CA, USA). Samples from various time points were pooled and underwent in vitro transcription using the ENZO IVT kit (Affymetrix, Santa Clara, CA, USA) to form biotin-labelled cRNA. This was fragmented and three independent biological samples for both control and treated conditions were then hybridised to Affymetrix U95Av2 GeneChips according to Affymetrix protocols. A detailed description of microarray analysis is included as Supplementary methods. Total RNA (1 mg) was reverse transcribed using Moloney murine leukaemia virus reverse transcriptase (Promega, Madison, WI, USA) according to the manufacturer's instructions. DRAK2, COX-2 and vascular endothelial growth factor (VEGF) were quantified by RT -PCR using SYBR Green Universal Master Mix (Roche Diagnostics Corp., Indianapolis, IN, USA) . Reactions were carried out in a 96-well format in the ABI 7700 Sequence Detector (PerkinElmer/Applied Biosystems, Warrington, Cheshire, UK). Results were then normalised to 18S rRNA amplified from the same cDNA mix and expressed as fold induction compared with the controls. cDNAs were amplified using the following primer pairs: DRAK2, AAAATAGGGCATGCGTGTGAA and TATTATACC AATATTCCACATATCTGTTGCT; COX-2, TTGTACCCGGACAGG ATTCTATG and TGTTTGGAGTGGGTTTCAGAAATA; VEGF, CA TGCAGATTATGCGGATCAA and TTTGTTGTGCTGTAGGAAGCT CAT. Bax and Bcl-xl quantification was performed using TaqMan pre-developed assay reagents according to the manufacturer's protocol (Applied Biosystems, Foster City, CA, USA). HCA7 cells were grown on glass chamber slides and treated with SC236 (5 mM) or vehicle control for 24 h. Slides were air-dried, fixed (methanol) and permeabilised (cold acetone). Slides were then blocked with 1% BSA in PBS (2 h) and incubated with a 1 : 50 dilution of anti-DRAK2 antibody (Santa Cruz Biotech, Santa Cruz, CA, USA) overnight at 41C. After serial washes, the slides were incubated with a 1 : 250 dilution of fluorescently labelled donkey anti-goat Alexa Fluor 488 (Molecular Probes, Leiden, the Netherlands) and subsequently propidium iodide (Molecular Probes) diluted at 1 mg ml À1 in PBS as a nuclear counterstain for 5 min. Slides were mounted with fluorescent mounting media (Dako, Carpinteria, CA, USA) and imaged with an LSM510 Zeiss Axioplan-2 upright confocal microscope (Carl Zeiss, Göttingen, Germany). HCA7 cells were seeded at a density of 0.5 Â 10 6 cells per well in six-well plates and were treated with SC236 (5 mM) or vehicle control for 24 h. Cells were detached using Versene for 5 min at 371C and floating and adherent cells were then pelleted by centrifugation and washed in a solution of 2% BSA suspended in PBS. We performed FITC -Annexin V/propidium iodide labelling was performed using a TACS Annexin V -FITC Apoptosis detection kit (R&D Systems, Abingdon, UK). Briefly, cells were re-suspended in a 100 ml working stock of binding buffer, propidium iodide and Annexin V -FITC conjugate and incubated in the dark for 15 min. Cells were re-suspended in binding buffer before analysis in a FACScalibur flow cytometer (Becton Dickinson, Oxford, UK) with measurement of fluorescence emission at 530 nm (FL1 channel) and at 4575 nm (FL3 channel). Four quadrant analyses using CellQuest (Becton Dickinson, Oxford, UK) software allowed quantification of cell populations according to labelling characteristics. The results were verified by staining/morphology by confocal microscopy. HT-29 cells were seeded at a density of 0.5 Â 10 6 cells per well in six-well plates and allowed to adhere overnight, then cotransfected with 3 mg of DRAK2 siRNA and 1 mg of pEGFP-N1 (as a marker to select transfected cells) DNA per well using Fugene 6.0 (Roche Diagnostics Corp.) with a 2 mg:1 ml ratio of RNA to transfection reagent. Transfections were performed in OptiMEM, 1 ml per well (final concentration of siRNA, approximately 200 nM). A negative control siRNA to a scrambled target in the DRAK2 gene (without sequence homology to any other known transcript as verified by BLAST analysis) was used in the mock transfection controls. Populations of both mock-and positively transfected cells were subsequently treated with either SC236 (5 mM), staurosporine (1 mM) or vehicle control for 24 h. Adherent and floating cells were then pelleted, washed and then re-suspended in 250 ml of PBS with 7-aminoactinomycin D (7-ADD; Molecular Probes) at a final concentration of 20 mg ml À1 and incubated on ice for 20 min in the dark. Samples were analysed by flow cytometry using an established method for assessment of cell viability in transfected cells selected according to GFP fluorescence (Vezina et al, 2001 ). The protocol was approved by the ethics (medical research) committee of Beaumont Hospital, Dublin and all patients provided written, informed consent. Colon cancer tissue and matched normal colonic mucosa were obtained from patients at the time of surgery and were immediately placed in RNAlater solution (Qiagen GmbH). Total RNA was extracted from the tissue samples as above. Additional patients with newly diagnosed distal CRCs were recruited under a separate protocol, approved by the Irish Medicines Board and Beaumont Hospital Ethics Committee, for treatment for 5 -7 days with the selective COX-2 inhibitor rofecoxib at a dose of 25 mg daily. Tumour was sampled endoscopically at day 0 and again on completion of therapy and was placed in RNAlater. Compliance with drug therapy was monitored by measurement of whole-blood monocyte COX-2 activity as previously described (Panara et al, 1995) . Analysis of tumour/normal differences were performed using twotailed Student's t-test. Differences in gene expression across time points were analysed by ANOVA with Bonferroni multiple comparisons test. Linear correlation was assessed using Pearson's test. t-Test, ANOVA and correlation statistics were performed using In-Stat, version 3.0 (GraphPad Software, La Jolla, CA, USA). We examined the regulation of apoptosis by COX-2 in a human CRC cell line. HCA7 cells have previously been reported to express high levels of COX-2 with abundant PGE 2 generation (Sheng et al, 1997) . Apoptosis was quantified by fluorescence labelling with FITC-conjugated antibody to Annexin V and propidium iodide assessed by flow cytometry (Figure 1 ). Following a 24 h treatment with the selective COX-2 inhibitor SC236 (5 mM), a significant decrease (P ¼ 0.01) in HCA7 cells viability was observed ( Figure 1B ), mirrored by a doubling of the proportion of the cell population gated to the early apoptosis phase (Annexin V positive, propidium iodide negative) ( Figure 1B ). The observed increase in apoptosis at low micromolar doses of SC236 was prevented by coadministration of exogenous PGE 2 (1 mM), whereas more marked increases in apoptosis seen with higher doses of SC236 were not rescued. A similar effect was noted with regard to effects on cell proliferation (see Supplementary Figure S1 ). The ability of PGE 2 to promote cell survival confirmed that at low micromolar doses the effects of SC236 are largely dependant on its ability to inhibit COX-2. High-density oligonucleotide arrays identify SC236mediated changes in gene expression in HCA7 cells Having confirmed that COX-2 inhibition causes an apoptotic phenotype in HCA7, we used high-density oligonucleotide microarrays (HDONAs) to examine early global changes in gene expression associated with the abolition of prostaglandin production. Total RNA from cells treated with SC236 (5 mM) or vehicle control for 4, 6 or 8 h was pooled and used to probe Affymetrix HGU95Av2 GeneChips that feature probe sets for over 12 000 different human transcripts. Three independent biological replicates were assayed for each condition (i.e. vehicle control or COX-2 inhibitor). Robust Multichip Average (RMA)-based analysis was performed to compare expression measures. Magnitude of changes in expression of individual genes was small in most cases, reflecting the early time points used. Correspondence analysis (CoA) of RMA-based measures, however, highlighted significant global differences in gene expression between control and treated samples ( Figure 2A ). This technique, similar to principal components analysis, represents expression (hybridisation) as points projected in three-dimensional space and demonstrates a consistent alteration in the transcriptome of cells treated with COX-2 inhibitor. Expression of DRAK2 (STK17b) consistently showed strong induction with COX-2 inhibition. This novel serine -threonine kinase is a member of the death-associated protein (DAP) kinase family and like other family members has been implicated in the control of apoptosis (Sanjo et al, 1998) . However, a potential function in the regulation of cell death and viability in cancer has not been explored to date. A comparison of our expression data set with another recently published HDONA experiment (Levitt et al, 2004 ) (evaluating selective COX-2 inhibition in a breast cell line) revealed a cluster of genes regulated by COX-2 in both systems (see Supplementary data; Table S1) and differential expression of DRAK2 once again featured as a strong signal. It was thus chosen for detailed validation and further analysis. Differential DRAK2 expression was confirmed by quantitative RT-PCR (qRT-PCR) on template from HCA7 cells. Induction of DRAK2 expression was observed as early as 4 h after treatment with SC236 and was inhibited by co-administration of PGE 2 ( Figure 2B) . A 4.4-fold induction of DRAK2 expression (relative to control) was observed with SC236 (5 mM) after 4 h (P ¼ 0.001), a response that was attenuated on co-incubation with prostaglandin (SC236 vs SC236 þ PGE 2 , P ¼ 0.01), suggesting that the changes in DRAK2 expression were dependant on variations in prostaglandin generation. Indirect immunofluorescent staining for DRAK2 in HCA7 cells showed a corresponding increase in levels of DRAK2 protein with COX-2 inhibitor treatment, with development of intense nuclear staining for DRAK2 at 24 h post-treatment with SC236 ( Figure 2C ). We examined the expression of both DRAK2 and COX-2 in a bank of colorectal tumour samples using total RNA extracted from tumour and normal mucosa sampled from the freshly resected tumours of 10 patients with CRC (not taking aspirin or NSAIDs). The levels of COX-2 transcript (normalised to 18S rRNA) were elevated in the majority of tumour samples relative to normal mucosa ( Figure 3B , mean 2.4-fold increase, P ¼ 0.006). DRAK2 expression showed an opposite pattern, with relative suppression of DRAK2 expression in tumour compared to normal from the same patient ( Figure 3A , mean decrease approximately 50%, P ¼ 0.003). A negative correlation between the ratio of DRAK2 and COX-2 expression (in tumour relative to normal) was noted ( Figure 3C , R ¼ À0.68, R 2 ¼ 0.46, P ¼ 0.03) reflecting an inverse relationship between the tumour/normal differences for the two genes across the patients sampled. Our observations suggested a suppression of DRAK2 expression, either directly or indirectly, by the actions of COX-2-derived prostaglandins in tumour samples. To test this hypothesis further, five patients with newly diagnosed CRC (not taking aspirin or NSAIDS) were recruited to a pilot study. Endoscopic biopsies of tumour tissue were obtained before and after a short course (5 -7 days) of treatment with the COX-2-selective inhibitor rofecoxib at a dose of 25 mg daily, a dose that has been demonstrated to selectively inhibit COX-2 in vivo (Ehrich et al, 1999) . The study was conducted before the withdrawal of rofecoxib by the manufacturer. qRT-PCR demonstrated an induction in DRAK2 expression ( Figure 4A ; mean 2.5-fold increase, P ¼ 0.01) in tumour from each of the patients treated with rofecoxib, reflective of the changes seen with COX-2 inhibition in vitro. The expression of a panel of additional genes was also evaluated for comparative purposes ( Figure 4B ). Although no significant change in several genes that modulate apoptosis (Bax, Bcl-xl) was observed, a significant decrease in VEGF expression was noted, with a mean reduction of close to 40% (P ¼ 0.01). This pro-angiogenic factor has previously been linked to COX-2 activity and reduced VEGF expression has been observed in animal models following treatment with rofecoxib (Yao et al, 2003) . This finding provided additional validation of the suppression of COX-2 activity in the tumours of these patients. Finally we evaluated the phenotypic activity of DRAK2 in cell systems. We suspected DRAK2 expression was already significantly suppressed in HCA7 cells, given high levels of COX-2 expression whereas HT-29 cells showed lower levels of COX-2 with higher levels of DRAK2 expression. Following transient transfection of HT-29 cells with a commercially available sequence validated siRNA to DRAK2, significant repression of DRAK2 transcription was seen ( Figure 5A ). To counteract the effects of low transfection efficiency and maximise detection of a phenotype associated with DRAK2 silencing, a GFP expression vector was cotransfected to 'gate'-transfected or 'silenced' cells in subsequent assays to quantify cell survival using an established method (Vezina et al, 2001) . The percentage of viable cells (without 7-ADD uptake) and non-viable cells (with 7-ADD uptake) was calculated in mock-transfected cells and in DRAK2 siRNA-transfected cells selected (gated) by GFP fluorescence. Loss of cell viability was observed in mock-transfected HT-29 cells following treatment with both SC236 and staurosporine ( Figure 5B ). In contrast, the siRNA-transfected (GFP-gated) cells showed no loss of cell viability following treatment with the COX-2 inhibitor SC236. They did however still show a significant loss of viability with staurosporine, suggesting a specific effect on the pathway involved in the action of SC236. Traditional NSAIDs have anti-neoplastic properties, but their prolonged use is limited by their association with gastrointestinal side effects, particularly the risk of gastrointestinal haemorrhage. The new-generation coxibs have also shown promise for chemoprevention of CRC but are now considered by many to carry an unacceptable risk of thrombotic events. Our aim was to identify novel 'effector' pathways operating downstream of COX-2 in tumours, so that this knowledge might allow future selection of more refined therapeutic targets. There are a number of ways in which COX-2 may promote cancer progression, either through an effect on cancer cells, the associated immune response or angiogenesis in response to the tumour. There is evidence that COX-2 expression protects cells from apoptosis and conversely that treatment of CRC cells with selective COX-2 inhibitors causes cell-cycle arrest, growth inhibition and induction of apoptosis (Richter et al, 2001; Sheng et al, 1997 Sheng et al, , 1998 . Evidence from both animal models (Hansen-Petrik et al, 2002) and human studies (Sinicrope et al, 2004) suggests that modulation of cell survival by altered susceptibility to apoptosis is equally important in vivo. However, the precise mechanisms by which COX-2 exerts these effects in tumour cells are unclear. Enhanced COX-2 expression in cells alters susceptibility to tumour necrosis factor-related apoptosis-inducing ligand by reduction in membrane death receptors (Tang et al, 2002) and may also alter the threshold for intrinsic pathway activation (Tang et al, 2002) . Over-expression of COX-2 leads to the generation of prostanoids, particularly PGE 2 , and signalling through the various EP receptors, several of which have been implicated in carcinogenesis (Watanabe et al, 1999; Sonoshita et al, 2001; Mutoh et al, 2002) . PGE 2 stimulates TCF-b-catenin-mediated gene transcription (Castellone et al, 2005; Shao et al, 2005; Eisinger et al, 2007) , possibly by activation of EP2 and EP4 receptors (Fujino et al, 2002) . Some of the important effects of PGE 2 are probably also related to its ability to transactivate the epidermal growth factor receptor through a number of distinct mechanisms (Pai et al, 2002; Shao et al, 2003; Al-Salihi et al, 2007) . It seems likely therefore that prostaglandins generated by CRC cells do not exert their activity by a single mechanism but have a range of downstream signalling targets. We made use of HDONAs to identify 'early' target genes downstream of COX-2. Earlier studies in other cell systems have demonstrated the ability of COX-2 to regulate gene expression, such as the expression of MDR1 (Patel et al, 2002) . Specific studies of differential gene expression in colon cancer cells (using suppressive subtractive hybridisation and differential screening) have previously demonstrated altered expression of genes involved in cell proliferation and viability (Zhang and DuBois, 2001) , however high NSAID doses and long time points make a direct comparison with our expression studies difficult. We chose to look at early time points following treatment with the highly selective COX-2 inhibitor SC236 to identify crucial early responses. COX-2 independent effects of SC236 have been described (He et al, 2006) . However, we chose to use SC236 at a dose which was associated with only modest changes in rates of apoptosis but where we observed rescue by co-incubation with PGE 2 , in the belief that study of a more subtle but specific phenotypic change would yield more significant findings. We identified DRAK2 as showing consistent changes in expression. This predominantly nuclear serine -threonine kinase has similarity to the DAP kinases and is involved (at least in certain cell types) in initiation of apoptosis (Sanjo et al, 1998) . Unsupervised hierarchical clustering of our expression data (Supplementary Figure S2 ) selected a cluster of genes showing a similar pattern of upregulation with COX-2 inhibition. This cluster of genes have either known nuclear localisation within the cell or have a bipartite nuclear localisation signal, which suggests their ability to traffic to the nucleus. Many of these are known to be involved in transcriptional regulation or in the induction or potentiation of apoptosis. Prior expression analyses have previously identified nuclear proteins as significantly over-represented in genes induced by aspirin treatment in colon cancer cells (Hardwick et al, 2004 ) and a mechanism regulating the nuclear trafficking of DRAK2 has recently been suggested (Kuwahara et al, 2008) . We have confirmed that DRAK2 expression in colon cancer cells in vitro is regulated by COX-2 ( Figure 2 ) and also demonstrated an in vivo relationship between COX-2 and DRAK2 expression in human CRC (Figure 3 ). Both COX-2 and DRAK2 have been implicated in the modulation of T-lymphocyte function, with opposing phenotypic consequences. Activated T cells in patients with SLE markedly upregulate and sustain COX-2 expression and resist inactivation and cell death (Xu et al, 2004) . Indeed, COX-2 expression has been shown to mediate lethal T-cell activation (Brewer et al, 2003) . By contrast, over-expression of DRAK2 in activated T cells enhances apoptosis in the presence of IL-2 (Mao et al, 2006) , and studies in mice with targeted disruption of DRAK2 have demonstrated that this protein is an important negative regulator of T-cell activation, raising the threshold for activation through the T-cell receptor (McGargill et al, 2004; Friedrich et al, 2005) and acting as an important regulator of T cells response and survival (Schaumburg et al, 2007; McGargill et al, 2008) . Although it is clear that DRAK2-mediated cell death is cell-type and context dependent, there is a growing body of evidence to indicate that its function is directly related to the regulation of cell survival. In addition to in vivo evidence from DRAK2 transgenic mice (Mao et al, 2006) , induction of DRAK2 in vitro in a variety of contexts and cell types augments apoptosis (Sanjo et al, 1998; Kuwahara et al, 2006; Mao et al, 2006) . We also show that silencing DRAK2 expression in HT29 cells promotes cell survival and abrogates the effects of a COX-2 inhibitor in vitro. These findings confirm published observations on the effect of DRAK2 silencing by RNAi on susceptibility of rat colon cancer cells to UV-induced apoptosis (Kuwahara et al, 2006) . We also recently observed that DRAK2 may be important in the ability of COX-2 to protect cardiomyocytes from doxorubicin-induced apoptosis (Neilan et al, 2006) , suggesting that the modulation of DRAK2 expression may not be confined to cancer but may also be important in the positive effects of COX-2 on cell viability and tissue healing in other systems. For the moment, the precise mechanisms by which DRAK2 impacts on cell survival are unclear. Recent observations suggest that the ribosomal kinase p70S6, involved in cell-cycle regulation, is a substrate for DRAK2 (Mao et al, 2009) , highlighting a possible involvement in regulation of cell-cycle dynamics. In an analysis of publicly available expression data (Whitfield et al, 2002) , we have noted that DRAK2 forms part of a cluster of genes showing similar cyclical changes in expression with cell-cycle periodicity in HeLa cells (Supplementary Figure S3 ; Supplementary Table 2 ). There is an interesting degree of overlap (summarised concisely in Supplementary Figure 4) between this panel and COX-2-regulated genes identified by our gene expression analysis and a number of other gene discovery studies that examined the effects of COX inhibitors (Iizaka et al, 2002; Levitt et al, 2004) . How does the identification of the regulation of DRAK2 by COX-2 further our understanding of the role of COX-2 in cancer? One of the most consistent observations about COX-2 in cancer is its contribution in vivo to the risk of systemic metastases (Tomozawa et al, 2000; Yao et al, 2004) . This fact likely explains the impaired survival observed in individuals whose tumours express COX-2 (Sheehan et al, 1999) . Previous observations about the effects of DAP kinase (the best-characterised member of the kinase family to which DRAK2 belongs) in cancer cells may provide a unifying explanation, linking the ability of this DAP kinase to regulate apoptosis to a suppression of metastatic potential (Inbal et al, 1997) . The most plausible explanation is that the enhanced cell survival conferred by DRAK2 silencing (by COX-2) provide a resistance to various forms of cell death and possibly crucially to anoikis (matrix detachment-induced cell death) (Hofmann et al, 2007) , the form of cell death that may be the most important in deciding the viability of circulating tumour cells and their ability to form distant metastases.

Rooting human parechovirus evolution in time

BACKGROUND: The Picornaviridae family contains a number of important pathogenic viruses, among which the recently reclassified human parechoviruses (HPeVs). These viruses are widespread and can be grouped in several types. Understanding the evolutionary history of HPeV could answer questions such as how long the circulating lineages last shared a common ancestor and how the evolution of this viral species is shaped by its population dynamics. Using both strict and relaxed clock Bayesian phylogenetics we investigated 1) the substitutions rates of the structural P1 and capsid VP1 regions and 2) evolutionary timescale of currently circulating HPeV lineages. RESULTS: Our estimates reveal that human parechoviruses exhibit high substitution rates for both structural P1 and capsid VP1 regions, respectively 2.21 × 10(-3 )(0.48 – 4.21 × 10(-3)) and 2.79 × 10(-3 )(2.05 – 3.66 × 10(-3)) substitutions per site per year. These are within the range estimated for other picornaviruses. By employing a constant population size coalescent prior, the date of the most recent common ancestor was estimated to be at around 1600 (1427–1733). In addition, by looking at the frequency of synonymous and non-synonymous substitutions within the VP1 gene we show that purifying selection constitutes the dominating evolutionary force leading to strong amino acid conservation. CONCLUSION: In conclusion, our estimates provide a timescale for the evolution of HPeVs and suggest that genetic diversity of current circulating HPeV types has arisen about 400 years ago.

Parechoviruses belong to the Picornaviridae family which includes other pathogenic viruses such as foot-and-mouth disease virus (FMDV), hepatitis A virus, enteroviruses and rhinoviruses [1, 2] . The Parechovirus genus includes two species: human parechoviruses (HPeV) and the zoonotic Ljungan virus. HPeV are non-enveloped pathogens with a single-stranded genomic RNA of positive polarity with around 7.400 nucleotides organized into a single long open reading frame in between a 5'UTR and 3'UTR. The open reading frame can be divided into three main regions: P1 (encoding capsid proteins VP0, VP3, VP1), P2 (nonstructural proteins) and P3 (nonstructural proteins, including the viral RNA polymerase) [2] [3] [4] . In HPeV, out of the three capsid proteins that constitute the monomeric units of the viral icosahedric-shaped capsid [3] , VP1 protein plays a crucial role in cell entry via interaction of an Arg-Gly-Asp (RGD) triplet with integrins on the cell surface [5] . However, some HPeVs (among which the type 3 strains) lack the RGD motif in VP1, and their mode of cell recognition and entry is less clear [6] . Typing of HPeV is based on the VP1 sequence providing a reliable locus to type all the identified HPeV strains as described for enteroviruses by Oberste et al [7] . As a result, the majority of HPeV available nucleotide data concerns the VP1 gene. In general, HPeV is transmitted by the oral-fecal route causing in most cases relatively mild respiratory and gastrointestinal symptoms [2, 8] , though conditions such as bronchiolitis [9] and severe neonatal infections [10, 11] have also been reported. HPeV1 and HPeV2 were first isolated in 1956 and classified by serotyping as enteroviruses, respectively echovirus types 22 and 23 [3, 12] . HPeV3 was first described in 2004 [6] and is associated with more severe conditions related to CNS symptoms [10, 11, 13] . Subsequently, improvements in HPeV-specific screening tools allowed a successful identification of HPeV4 and HPeV6 throughout North America, Japan and Europe [14] [15] [16] [17] [18] . Moreover, an HPeV variant originally classified as HPeV2-Connecticut was reclassified as HPeV5 [14] . Currently, sequences have become available for two novel types that were recently isolated in Pakistan and Brazil [19, 20] and in the Netherlands one more novel type was identified (HPeV14, [21] ). Unfortunately sequences of the HPeV types 9 to 13 are not available for analysis yet. Of all types, HPeV1 and HPeV3 are the most prevalent strains [11, 22] . Understanding the mechanisms underlining pathogenicity and persistence of pathogens in human populations is an important aspect of disease epidemiology and control. Fixation of mutations into nucleotide substitutions, a key principle behind phylogenetic signatures, is shaped by major evolutionary forces such as selection (molecular adaptation deriving from an increasing fitness of a corresponding phenotypic trait) and genetic drift (stochastic gene sampling process at reproduction) [23, 24] . A useful tool to detect and measure selection in viral gene sequences is the ratio between synonymous (dS) and nonsynonymous (dN) substitutions. Whereas a ratio above 1.0 is an indicator of positive selection operating at the amino acid sequence level [25] , significantly lower values are generally referred to as purifying selection and refer to preservation of the phenotypic trait. RNA viruses yield the highest mutation rates of all groups of pathogens which is approximately six orders of magnitude higher than in most DNA organisms [23, 26] . In the context of viral population genetics, substitution or evolutionary rates can be defined as the number of fixed mutational changes that accumulate in the population per nucleotide site per unit of time [27] . This rate is driven by the short-generation times of viruses and their error-prone RNA polymerase proteins lacking proofreading activity. Combined with their small genomes, these characteristics make RNA virus ideal models for evolutionary research [23, 28, 29] . In addition, recombination events may also play a role in RNA virus evolution [23] . While lacking a fossil record, evolutionary histories of RNA viruses can be calibrated because they represent 'measurably evolving populations', in which genetic diversity accumulates over a timescale of human observation [30] . Their evolutionary history and population dynamics can be reconstructed by means of genealogy-based coalescent approaches using nucleotide sequences sampled over an epidemiological time frame in order to estimate timed viral ancestry as well as the rates of genetic change [27, 29] . The most advanced methods operating on time-stamped sequence data use Bayesian Metropolis-Hastings Markov Chain Monte-Carlo (MCMC) algorithms that accommodate for the uncertainty of phylogenies rooted in time. Here, we estimated the substitution rates for the P1 and VP1 regions of HPeV with such a Bayesian approach, which provides a statistical framework for evolutionary analysis [31] . The identification of several novel types within the last few years may be conceived as a relatively recent introduction of HPeV into the human population, but this is not necessarily the case. By reconstructing the evolutionary history of HPeV we shed light on this issue. We investigated when current HPeV diversity emerged by determining the time of divergence from the most recent common ancestor (TMRCA). Dataset 1 comprised 29 nucleotide sequences from the P1 structural region (2291 nt) from different HPeV isolates (12 sequences of HPeV1, 1 sequence of HPeV2, 4 sequences of HPeV3, 5 sequences of HPeV4, 2 sequences of HPeV5, 3 sequences of HPeV6, 1 sequence for HPeV7 and 1 sequence for HPeV8). Dataset 2 comprised 199 nucleotide sequences of the VP1 capsid region (647 nt) (117 sequences of HPeV1, 2 sequences of HPeV2, 40 sequences of HPeV3, 18 sequence of HPeV4, 9 sequence of HPeV5, 10 sequences of HPeV6, 1 sequence of HPeV7, 1 sequence of HPeV8 and 1 sequence of HPeV14). To date, sequences of HPeV9-13 have not been made available [32] . The accession numbers of the sequences from both data sets are available in Additional file 1. Sampling date (year) for dataset 1 (1956-2007) and for dataset 2 (1975-2007) was either collected directly from Genbank record or following direct contact with the relevant authors. Multiple alignments of the P1 and VP1 regions of HPeV were conducted in ClustalW [33] and sequences were edited manually with Se-Al v2.0 [34] . Overall evolutionary rates for P1 and VP1 regions were measured as the number of nucleotide substitution per site per year (s/s/y). Relevant parameters were summarized as the median of posterior distributions by Bayesian coalescent Markov chain Monte Carlo algorithm implement in the Bayesian Evolutionary Sampling Trees (BEAST) software package version 1.4.8 [31] . To identify the optimal substitution model we performed a maximum likelihood analysis using the Modelgenerator package [35] . The model that best fit both sequence datasets was General Time Reversible (GTR) model with a discretised γ-distribution (GTR+Γ), allowing for nucleotide rates to vary among sites within the protein coding sequence alignments. Codon partitions (1+2)+3 were applied to both alignments, keeping first and second positions (mostly to non-synonymous changes) in one partition and the third position (related to increase redundancy and prone to synonymous changes) in a separate partition [36] . Relative rate parameters were estimated in separate for each partition, in order to accommodate rate variation at the third codon position. We employed both strict and relaxed lognormal molecular clocks, the latter allowing rate variation among branches [37] . The coefficient of variation (σ r ) was used as a quantification of the rate variation among branches (σ r > 0.2 was considered as significant rate variation among branches) ( Table 1) . A constant size demographic model was used as coalescent prior. Each alignment of both data sets was analyzed using Markov Chain Monte-Carlo (MCMC) computations run over a sufficient time to achieve convergence of the chains, which was analyzed by inspection of the MCMC samples using TRACER 1.4 [38] . The 95% highest posterior density (HPD) interval is the shortest credible interval that contains 95% of the samples values. Statistical uncertainties of the substitution rates and the TMRCA were summarized as the lower 95%, median, and upper 95% values of the HPD. Out of the tested models (GTR + Γ, both with strict and relaxed lognormal molecular clocks), the clock model that performed better was the lognormal molecular clock, which yielded the highest marginal likelihood. Clock models were also compared in terms of Bayes Factors (BF, Table 2 ). The relaxed model clock following a lognormal distribution was also supported by the highest log 10 BF as suggested [39] . The fact that a relaxed lognormal molecular clock fits best to our data was consistent with an estimated coefficient of variation of 0.29 and 0.41 (respectively, for dataset comprising P1 and VP1 regions) that reflected significant rate heterogeneity, thus rejecting a strict molecular clock. The resulting trees for each run were summarized using Tree-Annotator and the maximum clade credibility tree was visualized with FigTree v1.1.2 [34] . BEAST xml files are available as additional files 2, 3, 4 and 5. Overall selective pressures acting on VP1 antigenic region were estimated by using the CODEML program in the PAML package [40] . We used site models M7 (with a discrete distribution of 10 categories and accounting for sites not allowed to be positively selected) and M8 (estimates dN/dS for an extra class (p11) of sites, accounting for positively selected sites with dN/dS>1). Models were compared by means of likelihood ratio test and statistical support was taken from the Bayes-Empirical-Bayes output (BEB, see additional file 6: Log-likelihood and parameter estimates for PAML analysis) [40] . To detect adaptative molecular evolution, we used the complete dataset 2. We first identified the best-fitting substitution model for the HPeV sequences using the Modelgenerator package (GTR + Γ) [35] , and tested whether the evolution of the P1 and VP1 genetic regions was better described by a strict or relaxed lognormal molecular clock. A relaxed lognormal molecular clock provided a better fit to both datasets according to Bayes Factor (BF) analyses (P1: log 10 BF = 7.03 and VP1: log 10 BF = 27.8, Table 2 ). This is in accordance with significant rate variation among the branches of the inferred phylogeny as measured by a non-zero coefficient of variation (σ r ) obtained with the relaxed molecular clock analysis (P1: σ r = 0.29; VP1: σ r = 0.41) (see Methods for details). Using the available P1 and VP1 dated sequences of HPeV, our analysis inferred a similar rate of nucleotide substitution for both regions (P1 median: 2.21 × 10 -3 s/s/y, 95% HPD [0.48 × 10 -3 , 4.21 × 10 -3 ]; VP1 Table 2 ). The higher rate indicated for the VP1 region is possibly related to its antigenic properties, perhaps reflecting a difference in the level of gene expression or mirroring the involvement of the VP1 capsid protein in the viral entry mediated by cellular integrins. Despite our study focused on the available sequences of HPeV, more accurate estimates could probably be obtained with broader and more homogenous sampling timescale, preferably for all types. Yet, this may be a daunting task because it is difficult to obtain older samples and some of the HPeV types e.g. HPeV2, HPeV4, HPeV5 and HPeV6 appear to be relatively rare (see e.g [17, 18] ). Moreover, a common pitfall on estimating evolutionary rates is its underestimation due to mutational saturation of synonymous sites [41] [42] [43] [44] . By using a gamma distributed substitution model, we assured that rate variation among sites was allowed. Therefore, the effect of possible saturation of synonymous sites was alleviated by permitting a proportion of these sites to change at a higher rate [43] . In addition, we used partitioning in codon positions that allows different codon positions to have different substitution rates (and different amount of rate heterogeneity) (see Methods for details) [31, 45] thus further accommodating rate variation among synonymous and non-synonymous positions. The high rates of evolutionary change obtained in this study are in accordance with the evolutionary rates of other RNA viruses [32, 33] . Consistently, HPeV replication mechanism relies on an RNA-dependent RNA polymerase that lacks proofreading capacity. This increases the number of mutations incorporated in viral genomes over time and settles the ground for a relatively rapid genetic diversification [46] . The evolutionary rates of a few members of the Picornaviridae family have been studied. Despite the fact that most of the studies used different evolutionary frameworks, the rate of evolutionary change estimated in this study for the capsid region of HPeV VP1 is 1) faster than the rate of Hepatitis A virus [47] , 2) resem-bles the rate estimated for the antigenic region of Echovirus 71 [41, 48, 49] and finally 3) it is nearly one order of magnitude lower than the rates of poliovirus (2.09 × 10 -2 s/s/y) [48] or FMDV (2.7 × 10 -2 s/s/y) [50] . RNA viruses are the most suitable object of study for rates of change and divergence times. This is due in large part to the rapid rate at which they evolve allowing genetic diversity to accumulate within a timescale approximately the same as mutations are fixed in viral populations [29] . Yet, a deeper understanding of the replication machinery of HPeV (e.g. generation times, fidelity of RNA polymerase) may deliver insights on the molecular basis of these high rates of evolutionary change [44] . According to our analysis based on 199 HPeV VP1 available sequences, these viral species diverged from their most recent common ancestor (MRCA) at the year 1600 (95% HPD [1427-1733]) ( Figure 1 , Table 2 ). Moreover, and focusing on the two most recently isolated types (HPeV7 [19] , HPeV8 [20] . Taken together, we suggest that the genetic diversity of the currently circulating HPeV types has arisen around 400 years ago (Figure 1 ). The wider 95% Bayesian credible intervals obtained for the estimates using dataset 1 composed by the total of 29 available P1 sequences to date (Table 2) probably reflect a less heterochronous sequence data. Yet, an identical timescale was obtained when performing the MCMC approach with the dataset comprising the P1 region (1603, 95% HPD [940-1883]) ( Table 2 ). Despite holding new pieces to solve the puzzle of HPeV origins, the evolutionary rates and the timescales for the most recent common ancestor and type lineage-splitting events, may be better framed once a larger number of sequences are available [51] . However, the overlapping of the 95% Bayesian credible intervals obtained in our analysis for both genomic regions indicates that our estimates on the TMCRA of the HPeV lineages are robust ( Table 2) . One facet of fast evolving RNA viruses that induce acute infections (as the case of HPeV) is that they are likely candidates for jumps between species boundaries [29] . While the latter appears to be clearly established for e.g. SARS-CoV or influenza H5N1, a zoonotic link remains to be elucidated for HPeV. Because Ljungan virus shares a close phylogenetic proximity with HPeV virus, it is likely that both species have had a common ancestor [52] . Moreover, the reservoir host for Ljungan virus is Myodes glareolus, a widely distributed rodent commonly named as bank vole [4] . Despite the connection of Ljungan virus infection and human disease still remains to be clarified, bank voles are recognized as the reservoirs of other infectious agents, e.g. Puumala Hantavirus [53] and have been linked to a significant number of outbreaks over Europe [54] [55] [56] . Bayesian time-scaled phylogeny of HPeV based on VP1 sequence analysis Figure 1 Bayesian time-scaled phylogeny of HPeV based on VP1 sequence analysis. Maximum clade credibility tree obtained with BEAST with a constant size coalescent prior showing lineage splitting events (nodes A-F) since the most recent common ancestor to the presently circulating HPeV types. The divergence times correspond to the mean posterior estimate of their ages (in years). For the TMRCA, the correspondent 95% Bayesian credible intervals are shown (median 1600). Time axis is shown in years and ranges from the TMRCA to the present year. Deeper and some subtype nodes with posterior probability of higher than 0.8 are pointed out. Each colour corresponds to a specific HPeV, as indicated in the box on the right. The dashed grey circle depicts the extent of genetic diversity of the sampled HPeV strains. HPeV-1-"Harris-like" strains (*) clustered separately from the contemporary HPeV-1. In search for the driving force that shapes the evolution of the HPeVs, we looked at the ratio of non-synonymous-tosynonymous substitutions (the dN/dS ratio) [24] . For most codons in the VP1 region the ratio is < 0.1 (Figure 2 ). We noticed a few sites that tend to escape from purifying selection displaying dN/dS values > 0.3 (position Q61, A119, G203), or even > 1.0 (position N202 of our alignment, see additional file 7), however with statistically poor support (see additional file 6, Log-likelihood and parameter estimates for PAML analysis). Also other studies have found an overall low dN/dS ratio for the HPeVs [1, 57] . Our analysis confirms on a codon level that throughout the structural region strong purifying selection is dominant, leading to the conservation at the level of the amino acid sequence. Future analysis may shed lights not only in a unified framework of evolution for this viral species but also help preventing major burdens associated with HPeV pathogenicity. The HPeV are highly prevalent human RNA viruses and thus far no studies have addressed the evolutionary history of these pathogens. The Bayesian analysis presented here first indicates that the structural P1 and the capsid VP1 region of this viral species evolve at a high rate of evolutionary change (~10 -3 substitutions per site per year). Additional genomic and epidemiological data will help to reveal the relation between such rates and the widespread of this viral species. We also show that the currently circu-lating HPeV types have shared a common ancestor around four centuries ago. Since then, HPeV evolved into different lineages that have spread widely. Overall, a strong tendency for phenotypic conservation could be observed, suggesting that genetic drift plays an important role in the generation of the diversity within the regions under investigation. In summary, by delivering insights into the evolutionary mechanisms of HPeV, this study provides the foundations for a unified understanding of HPeV evolution.

HIV/AIDS prevention in China: A challenge for the new millennium

China’s first HIV infection was officially reported in 1985 and by the end of 1996, there may have been up to 200,000 people affected nationwide. In 2001, this figure probably exceeded 600,000. By 2003, the predicted number of HIV cases had reached 1.5 million. At least 80,000 individuals now have fullblown AIDS. China may soon have the largest HIV-infected population in the world, possibly 6 million cases by 2005. With infection rates rising at about 30% per year, it is feared this figure might exceed 10 million by 2010. Although the Chinese government was initially slow to accept the problem, in the late 1990s definite changes began occurring. In 2003 Premier Wen Jiabao publicly shook the hand of an AIDS patient and his government promised to introduce a range of free HIV-related services. Large preventive education campaigns are now underway. Unfortunately, there will still be many obstacles in controlling the epidemic and preventing further spread of this disease. Without doubt, China faces a serious predicament in the new millennium, and one which will pose numerous challenges for preventive medicine.

China's first AIDS case was officially reported in 1985 in Yunnan Province, adjacent to the border with Burma (1). By the first half of 1995, a total of 1774 patients had been identified (2) . In 1996 it was estimated that there were probably between 150,000 and 200,000 people affected nationwide (3) . According to the Chinese Ministry of Health however, there were only 8303 cases officially recorded by October 1997 (3) . By September 2001, this official number had risen to 28,133, although estimates suggest that it most likely exceeded 600,000 (4) . Until 2002, the Chinese government maintained that there were only about 30,000 HIV cases nationwide, although their official count was later revised 25-fold upwards to 840,000. By 2003 the United Nations had estimated that the number of cases was closer to 1.5 million (5) . In the first half of 2001 alone, China's HIV incidence rate increased by 67% (6) , and at least 80,000 people now have full-blown AIDS (7) . The Chinese Academy of Preventive Medicine finally acknowledged in 2001 that they may soon have the largest HIV infected population in the world, possibly 6 million cases by 2005 (8) . With infection rates now rising at about 30% per year, the United Nations estimates that China may have more than 10 million HIV positive patients by 2010 (9) . Given this alarming situation, we considered it necessary to conduct an in-depth literature review on the topic. We conducted a detailed literature search using standard internetbased medical search engines. We also retrieved some older documents from library archives and identified other material from the reference lists of published papers. An up-to-date reference list of HIV/AIDS in contemporary China appears at the end of our manuscript. The three phases Most epidemiologists agree that HIV spread through China in three distinct phases (10) . The first phase (the entry phase) occurred between 1985 and 1988, involving a small number of imported cases, and was officially viewed as a 'foreign' problem (4) . In 1987, China's Health Minister suggested the disease might be curtailed because sexual promiscuity and homosexuality were limited. HIV infected foreigners were subsequently banned from living in China, and it was also noted that it was illegal for Chinese citizens to have sex with foreigners (10) . The second phase (the spreading phase) occurred between 1989 and 1993, with the detection of HIV cases among minority groups from Western province communities that bordered Myanmar, Laos and Vietnam (4) . The third phase (the expansion phase) seems to have begun around 1994 (4), although several aspects of the 'expansion phase' and the 'spreading phase' were probably overlapping. In 1995, an article appeared in the scientific literature describing a mother and her two daughters who had become HIV infected after donating plasma (11) . This was followed by a Chinese newspaper story in 2000 titled 'Strange disease in a Henan village shocks top officials,' and one which later appeared on the US Embassy web site (12) . Blood selling in Henan Province probably created China's first significant HIV reservoir. Traditional Chinese taboos against blood donation made blood-buying a profitable venture, with entrepreneurs visiting many rural areas and collecting blood from poor villagers and townsfolk (8) . The first case of HIV acquired through plasma donation was reported in 1995 (11) . In the 1990s it was estimated that 70% to 90% of Chinese blood reserves were obtained from people who sold their blood commercially (2) . Approximately 1 million poor farmers trying to supplement their incomes regularly donated blood in Henan province during this period (6) . Unsanitary conditions and limited or non-existent infection control procedures contributed to the early spread of HIV. However, it was the practice of donating plasma only that enabled multiple donations from the same individual, who was presumably encouraged by the notion of multiple payments. In this manner, blood was centrifuged and the plasma removed, before pooling red blood cells from multiple donors and then reinjecting the mixture back into the original donor (10). Plasma selling not only disseminated HIV infection more widely, but it also ensured high viral titres in recipients of contaminated blood. It is estimated that up to 250,000 blood donors became HIV positive from blood donation during the 1990s (13) . In Donghu village for example, over 600 villagers sold their plasma to blood companies, among whom 231 became HIV positive (13) . In Wenlou village Henan, over 60% of the population are now HIV positive (6) . The overall HIV prevalence among former blood donors now appears to range from 1% to 40%, depending on the individual village (13) . Although the Chinese government initially responded to the Henan outbreak by banning journalists and international aid workers (13) , it eventually addressed the issue of unsafe blood donation and officially banned the practice of pooled donations. Their countermeasures were reasonably successful, with few blood-donation-related HIV cases appearing after 1996 (8) . Nonetheless, the spreading phase probably never ended as such; rather it was superseded by the expansion phase as HIV positive individuals began unknowingly infecting others. This situation was facilitated by a number of interrelated and complex social factors that China faced as it rapidly industrialised and social freedoms increased. HIV now has a firm foothold in Chinese society, with HIV-1 subtypes now including A, B, Thai B, C, D, E, F and G, as well as BC and BB recombinants (4). Intravenous drug users represent one of the most important social groups for HIV transmission in contemporary China, with approximately 60% to 70% of reported HIV infections occurring within this demographic (4) . Most appear to be men aged between 20 and 29 years of age (8) . Interestingly, some investigations have revealed that the HIV prevalence rate may be higher among female drug users than males (1) . Either way, their overall number has certainly increased rapidly in recent years. For example, there were officially 170,000 drug users nationwide in 1993, a figure which increased to 400,000 in 1994 (2) . By 2001, the official population of drug users was put at 900,000 (4). The prevalence of HIV among drug users, although variable from study to study, is known to be uniformly high. In Yunnan province, bordering Burma, almost half of all Heroin users are infected. In Xinjiang province, some studies have shown a HIV prevalence rate of 85% among drug users (8) . It has been demonstrated that after 1 year of injecting drug use, the prevalence of HIV infection reaches 68% (1) . HIV infection also occurs in conjunction with Hepatitis C Virus (HCV) and other blood-borne diseases. A recent study of drug users in Sichuan province showed the prevalence of HCV and HIV to be 71% and 11% respectively (14) . HIV has now spread to all 30 provinces throughout China (6) . At present, Yunnan has the highest number of HIV infected persons, representing approximately one-quarter of all reported cases. In Dehong for example, more than 1% of all pregnant women now suffer from HIV (15) . Although Yunnan is one of China's poorest provinces, increased foreign trade in recent years has paradoxically improved the economy while simultaneously facilitating the spread of HIV (16) . HIV in China is also being spread via sexual contact and heterosexual transmission now accounts for about 7% of all cases (4). There are a few reasons for this trend. Firstly, an increasing proportion of the country's youth are no longer waiting until marriage before they have sex (6) . This large demographic of 15 to 24 year old persons is estimated at approximately 210 million (6) . A history of multiple sex partners is also increasing among young Chinese, especially in the more urban areas. Traditional Chinese conservatism has made it difficult to talk openly about sexual matters, particularly in the older generation. Lack of public health knowledge regarding HIV/AIDS also represents a problem, with many contemporary Chinese knowing very little about sexually transmitted diseases and their prevention (5) . A recent survey showed that less than 3% of people were aware that condoms minimized the risk of HIV infection and only 21% knew they could be infected through sexual intercourse. Furthermore, almost 20% of the respondents had never even heard of HIV or AIDS (5) . Prostitutes represent another important social group for HIV transmission, and similar to injecting drug users, their numbers have been steadily increasing in recent years. According to the Ministry of Public Security, 137,000 prostitutes and their clients were arrested in 1990, 240,000 in 1992 and 400,000 in 1993 (2) . This number had risen to 700,000 by the late 1990s (1). The overall number of prostitutes in modern China is now believed to exceed 3 million. Among them, high levels of HIV and other sexually transmitted disease have been revealed. Research in Yunnan and Guangxi demonstrated the HIV prevalence among sex workers to be about 5% and 11%, respectively (6) . Concurrent gynaecological infections may also facilitate sexual transmission of HIV, particularly in rural communities (6) . In this regard, prostitute arrest records have revealed that the majority of commercial sex workers have at least one active, sexually-transmitted disease (17) . Prophylactic usage is not widely practiced among Chinese sex workers. Estimates of condom usage among prostitutes ranges from 1% to 30% (1). Condom usage usually depends on the customer. Some studies have shown that most prostitutes have never had formal sex education (17) . Many of them are also intravenous drug users, while sexually transmitted diseases are commonly treated with over-the-counter medicines. As prostitution is officially illegal in China, many infected sex workers are understandably reluctant to seek help or make themselves available for HIV testing and counselling. Their customer demographic is also concerning. A recent study of prostitution users showed that their main clients were middle-class men under the age of 35. Men with the highest incomes were considerably more likely to have bought sex than those less well-off (6) . In this manner, commercial sex workers may transmit HIV to their married clients, who will then transfer it to their families. Another factor compounding prostitution usage is the traditional social preference for male children, which became particularly acute following China's one-child policy (10) . There is now an increasing demographic trend for males to outnumber females, with the rate currently believed to be about 120 males for every 100 females (18) . This situation has resulted in less available women for sexually active males, which in turn, encourages or even necessitates prostitution usage. The problem may not be temporary either. From current figures it is anticipated that, over the next decade, around 10 to 15 million young Chinese men will have zero future prospects for marriage. This large demographic of unmarried men may be destined to find no sexual outlet during their lives other than commercial prostitution (10). Homosexual transmission is another probable HIV vector in China, with the number of gay men currently estimated at between 5% and 7% of the total male population (18) . Men who have sex with men probably represent a group of between 2 and 8 million people (19) . A survey of Chinese college students in the 1990s suggested that up to 8% may have homosexual tendencies (2) . Interestingly, around 80% of gay Chinese men are currently married (6) . Approximately 2% of married Chinese in rural areas engage in homosexual activities, a higher rate than for urban areas (4). Although homosexuality was removed from the official list of mental disorders in 2001, it remains illegal and highly stigmatized throughout Chinese society (1) . The problem may worsen in future years if China's homosexual remains highly marginalised. In the year 2000 for example, gay men comprised one-third of all AIDS patients in two dedicated AIDS care hospitals in Beijing (6) . Furthermore, safer sex appears not to be widely practiced among China's homosexual community. For example, a recent study showed that condom usage, either with heterosexual or homosexual partners, occurred less than 10% of the time (6) . Another investigation revealed that homosexual Chinese men may often engage in bisexual behaviour and/or paid sex with commercial sex workers, thereby expanding their range of potential contacts (20) . This latter point is particularly important, as HIV prevalence is probably increasing among homosexual communities within China. A study of homosexuals recruited in Beijing for example, demonstrated that the HIV prevalence was about 3% and was independently associated with being older than 39 years and having more than 20 male sexual partners (19) . Another of China's major HIV carriers appears to be the increasing number of itinerant workers who are for the most part, young, poorly educated labourers, in a sexually active period of their life (10) . During the Maoist era, a strict system of household registration kept most people confined to the immediate area in which they lived. With China's increasing economic prosperity and personal freedoms, such limitations were lifted and people were freely able to travel around. Rapid modernisation and industrial development also contributed to this demographic (4). The situation was compounded in poorer rural areas, where many young people left their hometown in search of a more prosperous life in urban centres (18) . Additionally, there were around 40 million workers who became unemployed when previously state-owned enterprises failed due to free-market economic reform. Itinerant workers now number at least 120 million (6) . Most of them are between 15 and 45 years of age, and by the nature of their demographic, represent a floating population that is difficult to educate, monitor and treat. They typically travel between their rural hometowns and cities seeking itinerant work, thereby providing an ideal vector for the extensive dissemination of HIV infection to remote areas (10) . Although prostitution is illegal in China as previously noted, commercial sex workers are readily available to itinerant workers and other travelling subpopulations such as longdistance truck drivers (21) . Again as previously noted, condom usage is not widely practiced among prostitutes, HIV knowledge and preventive measures are not well known; and like their nomadic clients, commercial sex workers themselves are also a very mobile subgroup. China's itinerant workforce does not represent the only mobile vectors however. Newfound economic prosperity and individual freedoms have made international travel possible for many, with almost 100 million people visiting China in 2002 and around 17 million Chinese travelling abroad (21) . These individuals may also provide additional vectors for HIV dissemination and importation. Although China's first HIV patient was reported almost 20 years ago, official government reaction was initially unrealistic and suboptimal. This situation was also compounded by a general community belief that HIV was simply a 'foreigners' disease that must be kept out (3) . Most early programs targeting HIV seem to reflect such notions. Even early predictions of HIV prevalence were hopefully optimistic, suggesting that it might only affect between 50,000 and 250,000 by the year 2000 (2) . As previously noted however, the real figure was probably double or triple these early government estimates. In many ways, Chinese officials were reluctant to acknowledge that their country was beginning to suffer from a disease, which was at the time, often associated with sexual promiscuity or moral degeneration (8) . The HIV issue appears to have been essentially ignored until the summer of 2002 (7), by which time the national HIV reservoir had probably reached a critical figure of 1 million. The problem was not solely due to central government inaction however, with local government officials in Henan province (one of the worst hit areas) often detaining aid workers, journalists and other health workers in futile attempts to cover up their own regional epidemics (7) . Although the gallant efforts of various individuals undoubtedly occurred behind closed doors for many years during China's growing HIV epidemic, official and visible government response was essentially delayed until the late 1990s. In January 1997, the Chinese government announced that it would take positive steps to keep HIV infections below 1.5 million by 2010 (3). In November 2001, an unprecedented international meeting on HIV/AIDS was held in Beijing, combining the talent of over 2700 health care experts (22) . Two important AIDS symposiums were also held in Beijing in November 2003, where guest speaker Bill Clinton warned that HIV might hinder China's continued economic progress (9) . However, it was the unprecedented and highly symbolic handshake between Premier Wen Jiabao and an AIDS patient in late 2003 that reverberated strongly across the country (10). On World AIDS Day, December 1, virtually every major media outlet published the handshake photo acknowledging this formerly taboo subject. Premier Wen even told the AIDS patients: 'You must have confidence. All of society cares about you' (5) . In a country steeped in tradition and heavily influenced by symbolism, the message was clear. It was also the first time a high ranking official from the ruling party finally confronted the type of discrimination that HIV positive Chinese are burdened with on a daily basis (10) . On World AIDS Day 2003, China's health minister finally acknowledged that their government was probably not doing enough to fight the emerging HIV epidemic. He offered to expand four important services: free HIV testing, free antiretro-viral drugs, free care for HIV-infected mothers and free education for AIDS orphans (5) . Most importantly the testing and treatment would be free for those who could not afford to pay for it themselves (23) . Aggressive public health education and prevention campaigns are now underway in China, although their effectiveness may be limited and are yet to be adequately demonstrated (24) . Nevertheless, in Yunnan Province, the government recently issued Regulation 121 which focuses on aggressive HIV education. Hotels with prostitutes must now offer condoms (15) . A new HIV vaccine is also being developed at the Chinese Academy of Medical Sciences, and one which uses distinctly Chinese primates as the research instrument (25) . China's health care system Although positive steps have now been taken by the Chinese government in combating HIV/AIDS, numerous challenges remain. One issue is the structural inadequacy of China's current health care system that was graphically revealed during the SARS outbreak (26) . Tuberculosis offers another pertinent example of how these shortcomings may compound the situation. China now has twice the level of multi-drug resistant tuberculosis of other countries (7) . Providing affordable and accessible health care to the general public has also become increasingly difficult following the introduction of free-market reforms in 1978. In this manner, the rural health cooperatives that once provided health insurance and affordable health care have now been effectively abolished (7) . Furthermore, the 'middle tier' community hospitals which formerly dispensed a great proportion of total health care have also disappeared (27) . Such facilities will be essential to the success of any future HIV treatment and prevention programs. Social inequalities are also problematic, as increasing numbers of Chinese hospitals now charge fees for services, thereby excluding those who cannot afford to pay. Drug users, prostitutes and itinerant workers represent important HIV-affected sub-populations that have limited capacity to pay for medical care and drug treatments. Although antiretroviral medications (such as indinavir, efavirenz and nevirapine) are now available from Chinese hospitals, most HIV positive individuals still rely on traditional Chinese medicine (1). Many patients also alternate between traditional Chinese medicine and standardised pharmaceutical regimes, which in turn, creates problems with drug compliance and resistance. In one Beijing study for example, 20% to 30% of the subjects had become resistant to nevirapine after 9 months (24) . The costs of these drugs are also prohibitive for many Chinese HIV patients, although the price has fallen considerably in recent years. In 2001 for example, the average annual cost was US$10,000 (22) . By 2004, this had fallen to US$200 or US$300 (7) . Unfortunately, the drop in price was not uniform among all drug types and from all drug companies. For example, the main drug of choice for treating HIV in Africa and Thailand still remains prohibitively expensive in China (24) . Although China has recently offered to provide free HIV drug treatment for those who cannot afford it themselves, it remains to be seen whether the country can keep such a bold promise. Nevertheless, it is estimated that these costs may be minimised to about US$30 million per year, a figure which is probably affordable in the current economic climate (7) . The accurate identification of HIV positive individuals represents another barrier to effective public health interventions in China's HIV/AIDS epidemic. For example, although an estimated 80,000 Chinese HIV patients now have symptomatic AIDS, less than 5% have been identified (24) . General knowledge regarding the biological mechanisms and public health implications of HIV also remains problematic. There appears to be a false sense of security among ordinary Chinese that HIV is a foreigners' disease which only affects marginal communities (26) . A previous community survey demonstrated that only 3% of Chinese people were aware that HIV was transmitted through an exchange of bodily fluids. However, in the same study, 54% of participants thought that sharing chopsticks could spread the disease (8) . Lack of preventive knowledge is another serious issue hampered by limited HIV-related knowledge. Aside for hampering the ultimate success of public health intervention programs, this type of biological ignorance also stigmatizes people already carrying the virus, forcing them even deeper underground and away from potential sources of help. HIV/AIDS clearly represents one of the greatest challenges for preventive medicine in modern China. There are at least 1 million HIV positive individuals at present, and this figure may reach 10 million by the year 2010. However, the situation is not hopeless. In particular, the SARS epidemic provided an important learning experience and a key to the future battle against HIV/AIDS in China. Firstly, it revealed some important deficiencies in the medical and political infrastructure. Secondly, and perhaps crucially, it captured the attention of the international community and China's top leadership (28) . This may indicate tentative acceptance of the comprehensive response needed to combat the nation's greatest public health challenge. In some ways, Thailand can be seen as a role-model for HIV prevention within developing countries. Thailand has been particularly successful in reducing new HIV infections among prostitutes and military recruits with their 100% condom program (6) , and this approach may be appropriate for China. Either way, there will be many obstacles in controlling the epidemic, not the least of which will be to change public perceptions of HIV and raise awareness of preventive measures. Whether the Chinese government can muster the resources needed to keep its bold promises is also unknown. What is certain is that China now faces a serious epidemic in the new millennium, and one which will pose numerous challenges for public health and preventive medicine.

Emerging zoonotic diseases: An opportunity to apply the concepts of nidality and one-medicine

The use of animals as sentinels of human disease revolves around the concept of nidality. That is, an agent of disease occupies a particular ecologic niche and alterations in that niche will change the function of that agent relative to traditional host-agent-environment relationships. Nidality is a derivation of the root word nidus. Nidus is defined as a nest or breeding place, particularly a place where microbes such as bacteria, fungi, viruses, as well as other organisms and larger parasites, are located and multiply. Application of the concept of nidality and development of prevention strategies has most frequently been associated with military campaigns and interruption of tick-borne infections. Modern usage of the phrase “one-medicine” was popularized in the United States and Europe by Calvin Schwabe and the concept is attributed to Rudolph Virchow. It is applied today to the study of zoonotic disease and interventions in rural agricultural communities that share close living arrangements between people and their families, their pastoral work environment, and the animals for which they care. Integration of the two concepts of one-medicine and nidality provides an opportunity to apply a systems approach (i.e. general systems theory) to dealing with emerging zoonotic diseases in today’s global agricultural and industrial settings.

The use of animals as sentinels of human disease revolves around the concept of nidality. That is, an agent of disease occupies a particular ecologic niche and alterations in that niche will change the function of that agent relative to traditional host-agent-environment relationships (1) . Nidality is a derivation of the root word nidus. Nidus is defined as a nest or breeding place, particularly a place where microbes such as bacteria, fungi, viruses, as well as other organisms and larger parasites, are located and multiply. Application of the concept of nidality and development of prevention strategies has most frequently been associated with military campaigns and the interruption of tick-borne infections (2) . Modern usage of the phrase "one-medicine" was popularized in the United States and Europe by Calvin Schwabe and the concept is attributed to Rudolph Virchow (3). It is applied today to the study of zoonotic disease and interventions in rural agricultural communities that share close living arrangements between people and their families, their pastoral work environment, and the animals for which they care. A recent movement in the United States, the "Medicine and Public Health" initiative (http://www.sph.uth.tmc.edu/mph/) recognized that human clinical medicine and public health communities had developed parallel but separate enterprises and there was a need to reassess this division. Medicine (focusing primarily on the health of individuals) and Public Health (with a population perspective) needed to find innovative and joint solutions to improve the health of the people of the United States. Similar schisms are evident between veterinary medicine and the agricultural industry in the United States and initiatives are underway to bridge these gaps. Integration of the two concepts of one-medicine and nidality provides an opportunity to apply a systems approach to dealing with emerging zoonotic diseases in today's global agricultural and industrial settings. In the late 1990s, our School of Public Health embarked on a series of Bioterrorism Awareness Seminars for health care students and primary care providers, i.e. family practitioners, pediatricians, general internists, nurse clinicians, physician assistants, and others. I struggled with the development of a communications approach that would capture the interest of the audience, especially prior to the events of September 11, 2001. The presentation of information needed to raise a clinician's awareness for the early diagnosis of a bioterrorism incident was considered by many to be esoteric, an unlikely scenario, and not relevant to the daily practice of medicine in the United States. As a public health veterinarian, I knew that many of the biological organisms identified as potential agents of bioterrorism occur naturally in North America, specifically Texas which is situated on the southern border of the United States of America. Texas has enzootic foci of anthrax, tularemia, and plague. Mosquito-borne encephalitides such as Saint Louis Encephalitis, Eastern Equine Encephalitis, Western Equine Encephalitis, and most recently West Nile Virus occur in our geographic region. We have had cases of hemorrhagic fever (Ebola alice) in primate colonies in Texas. Dengue is endemic in countries immediately to our south. Vibrio species infections are identified in association with our seafood and shellfish industry in the Gulf of Mexico. It is not unusual for sporadic human cases of infection with these organisms to be diagnosed by clinicians in Texas. The teaching challenge was to accomplish our goal of "heightened awareness" for agents of bioterrorism and also make the seminars relevant to the daily practice of medicine (4). There is an overall lack of knowledge regarding the natural history and ecology of these zoonotic organisms (5) . There is an abysmal lack of understanding among the general public and media of the natural occurrence of many biological organisms in the agricultural and animal industry setting as well as in free-ranging animal populations. The natural history, normal ecology, and epidemiology of potential agents of bioterrorism provide scenarios for explaining risk to the human population at large. Understanding the natural history of these organisms and their role in agriculture and free-ranging animal populations provides a context for health care providers to make appropriate risk assessment decisions. When first confronted with the emergent threat of bioterrorism I used a mental paradigm based on the principles of nidality. Instead of considering the biological niche of the organism in nature, however, I chose to define the agent's nidality from an epidemiological perspective. Every organism that causes disease in human populations has an expected epidemiologic presentation. The epidemiologic pattern provides clues as to whether the clinical presentation is "expected" or "un-expected" in that physician's range of practice experience. A corollary to the expected/un-expected clinical paradigm is the potential route of exposure to the organism in a particular geographic and/or occupational setting (6) . Routes of exposure can be categorized as "natural" or "un-natural" and may in fact influence the constellation of signs and symptoms observed by the clinician. I now employ these concepts in teaching and instructional activities to prepare health care providers in biodefense preparation and response. The paradigm that I employ is as follows: ‫ع‬ Expected and natural clinical presentation and epidemiologic circumstances ‫ع‬ Unexpected but natural clinical presentation and epidemiologic circumstances ‫ع‬ Unexpected and unnatural clinical presentation and epidemiologic circumstances Let me illustrate by providing examples of the three clinical presentations of anthrax in humans: cutaneous; gastrointestinal; and inhalational. In South Texas it would be usual and expected to diagnose a sporadic case of human anthrax. The presentation would normally be a single cutaneous eschar on a visible surface such as the hand or arm. The epidemiologic history would most likely involve contact with a known diseased animal. The latest example was an adult male who cut himself while removing the hide (skinning) from a buffalo (American Bison) that had died at a game ranch. The geographic area had recently experienced a severe die-off of white tailed deer from anthrax. This clinical presentation and epidemiologic history would be usual and normal in South Texas. It is rare to diagnose gastrointestinal anthrax in humans in Texas. Animal protein is abundant. Cattle raised for human consumption are inspected. The hunting culture is one of harvesting healthy and active game animals, not culling old and/or impaired specimens. When anthrax epizootics occur in free-ranging populations, they are in remote areas with minimal human interaction. However, one could envision a scenario of ranch workers and/or unwitting hunters harvesting anthrax infected cattle or deer for the meat. Diagnosis of a cluster of gastrointestinal anthrax cases in a family or several acquaintances with a history of hunting their own food would be unexpected, but given the epidemiologic evidence, natural. It is difficult to develop a scenario for the state of Texas where any occurrence of primary respiratory infection with the anthrax organism would be expected. However, Laredo, Texas is the largest inland port in North America. Ten thousand tractor trailer trucks cross the Mexico-United States border every day. If a trailer contained a cargo of anthrax infected animal hides, it is feasible that the truck driver and or illegal passengers riding in the trailer could be exposed through the respiratory route. This fictional example would certainly represent unexpected and unnatural circumstances. All three of the preceding scenarios for cutaneous, gastrointestinal, and primary respiratory infection with the anthrax organism share several epidemiologic characteristics. Bacillus anthracis occupies a natural ecologic niche in the southwestern United States and parts of Mexico. The number of individuals involved in each scenario was small. The clinical presentations followed classical expectations. The epidemiologic picture matched the clinical presentation. Thus, the information presented to my clinical audience was relevant to their practice of medicine today in the state of Texas and hopefully raised their level of knowledge regarding the natural history of one of the potential agents of bioterrorism. I conclude with "clues" to understanding how the epidemiologic picture may differ if the anthrax exposures were man-made and intentional rather than a part of what is expected in our part of the world: the epidemiology would be wrong; there would be large numbers of human cases; there may be concomitant deaths of other animal species; and, there may be unusual strains of the organism. The concepts of nidality and one-medicine provide the opportunity to explore the natural history of agents of disease; the inter-relatedness of human activity, animal industry, and free-ranging animal populations. Application of the principles espoused by the Medicine and Public Health initiative provides a framework for human medical clinicians, public health practitioners, and the veterinary medicine community to interact and address the threat of bioterrorism in a reasoned and scientific manner based on effective risk assessment.

Disease ecology and the global emergence of zoonotic pathogens

The incidence and frequency of epidemic transmission of zoonotic diseases, both known and newly recognized, has increased dramatically in the past 30 years. It is thought that this dramatic disease emergence is primarily the result of the social, demographic, and environmental transformation that has occurred globally since World War II. However, the causal linkages have not been elucidated. Investigating emerging zoonotic pathogens as an ecological phenomenon can provide significant insights as to why some of these pathogens have jumped species and caused major epidemics in humans. A review of concepts and theory from biological ecology and of causal factors in disease emergence previously described suggests a general model of global zoonotic disease emergence. The model links demographic and societal factors to land use and land cover change whose associated ecological factors help explain disease emergence. The scale and magnitude of these changes are more significant than those associated with climate change, the effects of which are largely not yet understood. Unfortunately, the complex character and non-linear behavior of the human-natural systems in which host-pathogen systems are embedded makes specific incidences of disease emergence or epidemics inherently difficult to predict. Employing a complex systems analytical approach, however, may show how a few key ecological variables and system properties, including the adaptive capacity of institutions, explains the emergence of infectious diseases and how an integrated, multi-level approach to zoonotic disease control can reduce risk.

The growing problem of globally emerging infectious diseases (EIDs) has prompted a substantial effort by the biomedical research establishment to identify the causes and recommend action. As reported in the most recent of a series of volumes (1), a main finding is that the current episode of global infectious disease emergence is the result of a convergence of factors involving complex interactions among numerous variables. This includes genetic, biological, social, economic, political, ecological, and physical environmental factors, and calls for an interdisciplinary research agenda. It is also concluded that "human development and large scale social phenomena are closely associated to infectious disease threats at a global level," which points to the need for research focused on "social and ecological factors affecting infectious disease emergence" (1) . The phenomenon of globally emerging infectious diseases requires understanding biological systems in the broadest sense and dealing with their extraordinary complexity. This includes processes operating at the level of transmission and evolution of a pathogen within and among host species and humans. It extends to and includes processes involving ecosystems and regional environmental change occurring on a global scale (2) . In fact the scale and magnitude of anthropogenic activity has reached a point of virtual co-dominance with natural processes of energy and material flows globally (3) . Understanding these kinds of processes traditionally has been the domain of classical ecology, or natural history, plus systems ecology (4, 5) . Adding to this are recent ecological perspectives and models applied at the molecular, cellular level, and organismal levels (6) , and others addressing the complexity, multiple variables, cross-scale influences, and dynamic behavior at the level of natural ecosystems (7) , and social-ecological systems (8) . Along with the research at the organismal level and below, that aimed at the level of social-ecological (coupled humannatural systems) is critical to the development of the comprehensive scientific framework necessary for understanding zoonotic infectious disease emergence in particular. Not only does this new area address the dynamic behavior of complex, large scale systems, but also bridges theory from the traditionally separate biological and social science disciplines, thus contributing to the interdisciplinary research agenda also called for in the above reports. The purpose of this paper is to consider how regional and global zoonotic disease emergence trends might be explained on the basis of current thought in biological ecology including the very recent developments new to the field of infectious disease ecology. Here we draw on ecological science as broadly defined as a basis for identifying causal mechanisms of zoonotic disease emergence, the ultimate goal being to enhance disease prevention and control programs. Several authors have categorized causal factors of infectious disease emergence, including explicitly citing 'ecological' ones involving land use change (9) (10) (11) (12) or 'land use drivers' (13) , human movement (10, 12) , encroachment and wildlife translocation (10, 11) , rapid transport (9, 10) and climate change (11, 12) . Most recently, the Institute of Medicine (1) described these along with others (13 categories of factors in all) and a model stating the major categories of factors have historically converged to bring about the current global emerging infectious disease crisis. Ecological factors are described as one of four major categories of factors that have converged with social, political, and economic factors; genetic and biological factors; and physical environmental factors. We take a different approach to understanding the interaction of the above factors and their causal relationships by focusing on disease emergence as an ecological-evolutionary phenomenon influenced by human factors. Our interest is in how human factors interact with natural processes and, in particular, how mechanisms operating at levels meaningful to understanding pathogen transmission and evolution can result in regional and ultimately global phenomena (i.e., regional endemism, epidemics, or global pandemics). As a first step we distinguish between the two broad categories of human factors, 'demographic and societal' and 'disease intervention and policy' suggested previously by Gubler (10) ( Table 1) . This categorization differentiates between factors associated with specific kinds of environments or ecosystems and those involving biological and policy factors not so associated. However, both can be described as part of a single ecological framework involving interaction of systems of essentially natural versus human design, respectively. Our focus is on the first category from the standpoint of how disease emergence is explained by ecological concepts and principles. This includes some relatively new models and theory not previously used to explain the current trend in increasing emerging infectious diseases. We present a general model of zoonotic disease emergence on this basis. We also discuss recent explanations based on complexity theory for how human behavior and ecosystems interact to contribute to disease emergence. Classical ecology, or natural history, has been the basis and mainstay of infectious disease research since its origins with 'Koch's Postulates' and subsequent development of microbiology and zoonotic disease epidemiology during the 19th and 20th Centuries (14, 15) . Throughout much of its early history, zoonotic disease research involved this descriptive, empiricallybased ecology: identifying the life cycle, transmission, incidental and natural hosts of pathogens, along with demographic, life history, dispersal, and habitat attributes of reservoirs and vectors. A substantial theoretical dimension has developed, beginning with Ross's mathematical analysis of malaria transmission (16) and extending to Anderson and May's (17) recent synthesis Infectious Diseases of Humans. Although essential to designing effective prevention and control programs, empirical field based, disease ecology has been neglected in recent years (18) . Fortunately, theoretical disease ecology, stimulated largely by the work of Anderson, May and others has flourished and led to a significant syntheses involving application of ecological-evolutionary biology to the study of infectious diseases (19, 20) . Parallel to this, systems ecology has begun to extend its domain by applying complexity theory to emerging infections with at least initial suggestions of its implications (6) . This development in particular, along with observations from several decades of applications of systems ecology to natural resources and economic development (8, (21) (22) (23) , have resulted in important insights of significant potential in understanding zoonotic disease emergence as a cross-scale process. This area uses complex systems theory applied to coupled, human-natural systems to explain how processes such as local phenomena can result in a cascade of effects ultimately reaching global proportions. The finding suggests this crossscale behavior is controlled by relatively few variables, and is mitigated by social and ecological resilience. The loss of this resilience in ecological systems is observed to lead inevitably to unpredictable events or the 'surprise' characteristic of complex systems generally. This combination of socialecological systems and resilience theory helps explain the unpredictability of disease emergence events. It represents another potentially useful area of application to understanding emerging infectious diseases along with those areas generally considered within the domain of ecology mentioned above: population ecology and genetics, community ecology, and systems ecology. Population ecology, genetics and disease emergence Of particular relevance to disease emergence is the explanation provided by theoretical population biology, already mentioned, of how host (including human) population size determines whether or not a pathogen can persist in a population. The accumulated findings demonstrate thresholds exist, depending on the type of pathogen and host population, below which a pathogen cannot be sustained. Considered in light of the exponentially increasing size of human and domestic host and vector populations in the world, the breaching of thresholds of pathogen persistence can explain much of the increase in emerging infectious diseases. This can be explained as follows. Although zoonotic disease emergence is not entirely a tropical phenomenon, it is mostly associated with tropical developing regions undergoing the most rapid population growth and ecological changes. Prior to the post-WWII economic era, most regional ecosystems in the tropics consisted of relatively scattered human settlements, and only a few large cities (>500,000) (24) . These were separated by large expanses of cropland and pastureland and relatively undisturbed forest. Since then, in what has been the most rapid period of large scale ecological transformation in human history, the pattern has essentially reversed (25) . The once scattered settlements and few large cities have coalesced into expansive megacities and surrounding periurban settlements with only remnant patches of undisturbed forest remaining in a sea of cropland, scrub, and ecologically degraded lands. Dacca, Bangladesh, which grew from a population size of 200,000 to 13 million from 1970 to 2003, is one of many examples. Thus the existence of population density-dependent thresholds for disease emergence is particularly relevant (26, 27) . This explains the abrupt transitions of urban diseases between non-persistence to endemic and endemic to epidemic behavior as population densities of susceptible humans, hosts, and vectors reach critical densities. The classic illustration is that of measles which, given its particular transmission rate, requires human settlements with population sizes in excess (>250,000) of what historically existed in most pre-industrial states and geographically isolated populations even today (28). Thus, for example, many infectious diseases endemic on continents have not become established on islands despite their occasional introduction and the occurrence of local outbreaks. The same mathematical ecology that explains why measles and virtually all diseases have threshold densities, explains the much lower thresholds existing for vector-borne diseases such as arboviruses (29) . Particularly noteworthy is the theoretical demonstration that the pathogen 'reproductive rate' increases with the square of vector population density. This indicates threshold densities can fast be breached as domestic and peridomestic hosts and vectors expand (or re-expand) their geographic ranges (once they are introduced or re-introduced) and increase their densities. This helps explain the explosive re-emergence of dengue and dengue hemorrhagic fever in the American Tropics as vector populations responded to relaxed controls and new breeding habitats associated with urbanization (30) . This phenomenon can be likened to the gradual build-up of 'dead and down' wood across a forested landscape with a history of fire suppression. The build-up of fuel, like that of host or vector populations, becomes an 'accident waiting to happen' when a single ignition event in one locality, similar to a single infection event, spreads to the entire region. Another consequence of the dramatically increased densities of humans, host reservoirs, and vectors is the increased number of pathogen genomes. The resulting increased levels of genetic variability can accelerate microbial adaptation, including evolution of pathogenesis, and antimicrobial resistance. Genetic variability increases with population size and density through a variety mechanisms including mutation. The probability of producing more virulent variants not only increases with host population size but also with crowding and co-mingling of different host species (31) . In general, parasite (pathogen)-host relations naturally constitute a co-adaptive/ evolutionary 'dance' along the pathogenicity threshold, which is likely to be crossed with greater frequency due to unnatural anthropogenic disturbances (32) independent of increasing population sizes and pathogen genetic diversity. The study of ecological communities and the 'community ecology' theory it has yielded includes a number of principles and mechanisms that describe how human disturbances as well as natural environmental variation can contribute to any of the above population level factors (33) . There are a number of implications to zoonotic disease emergence, although most have not yet been described in terms of disease emergence or in the medical, public health, or zoonotic disease literature. Of critical significance from this area of ecology is the general principle of community assembly. Research has demonstrated that communities of arranged predicably in terms of 'assembly rules' (34) . This order, in terms of the spatial distribution, composition and the abundance of each species in an ecological community, is affected by interspecific interactions (predation, competition, and parasitism). Density independent factors (e.g., weather, natural catastrophes) play an important, but a more ephemeral role in most ecosystems. The process of community assembly (and disassembly) has been particularly well demonstrated through studies of insular ecosystems, the so-called 'species area relationship,' and the phenomenon of faunal collapse (35) (36) (37) . A principal outcome and the ecological significance of this body of research first described in terms of the process of 'habitat fragmentation', has been identified as the principal mechanism by which human land and resource use contributes to species extinction (38) . Although initially debated, a significant amount of evidence has since accumulated resulting in its general acceptance and applicability, particularly to tropical forest ecosystems (36, 39, 40) . The critical significance of habitat fragmentation and related human disturbances to disease emergence stems from it contribution to the disassembly of orderly natural communities. For example, human activities such deforestation, use of pesticides, and various forms of pollution often result in the loss of predators. In fact, carnivorous mammals typically are the first species to disappear following forest fragmentation. Their local extinction represents the loss of 'top down' natural control in ecological communities. This can in turn result is an increase in abundance, even 'hyper-abundance' (41) , of species such as rodents and biting insects. Community disassembly and the resulting loss of natural population control mechanisms for such species generally is associated with the conversion of natural landscapes to urban and agriculture landscapes. Broad spectrum pesticide use and habitat simplification, along with habitat fragmentation, are important contributing factors. The reduction in species diversity can contribute to the phenomenon of 'ecological release' in remaining species, whose predators, competitor and parasites are reduced in numbers or eliminated. Some of these may be already serving as zoonotic reservoirs or vectors. If so, ecological release may result in their proliferation. This helps explain why many emerging zoonotic diseases occur in areas where settlement and agricultural expansion recently have encroached into tropical forests. A similar phenomenon, associated with the regrowth of forests in developed regions, can lead to zoonotic disease emergence. The pattern of reforestation in Eastern North America during the past half-century resulted in increased habitat favored by forest edge adapted species such as whitetail deer and white-footed mice. White-tailed deer, the principle host of the adult tick Ixodes scapulari, reinvaded the area and with few predators and competitors the population exploded. This pattern of ecological change arguably has been a major factor in the emergence of Lyme disease in this region. An extension of the concept of 'ecological release' has been suggested to explain the invasive species phenomenon in which super-successful introduced (alien) species have escaped from their natural complement of parasites (42) , as well as predators and competitors. This may explain in large part our most successful invasive domestic species, Rattus and Aedes species, which are hosts and vectors of some of our current and historically most problematic urban zoonotic diseases. In sum, zoonotic infectious disease emergence can be explained in part as a consequence of the disruption of natural ecological communities, and the breakdown of naturally existing predator-prey, competitive, and host-parasite relationships that tend regulate and stabilize species' abundance. This can occur through the use of non-selective pesticides through changes in land use and land cover that affect the distribution and abundance of species. It should be noted that the disassembly of natural ecological communities that results from habitat fragmentation is a protracted process compared to exponential decay. Depending on the extent of habitat loss and other factors such as the sizes, shapes and spatial relationships of remaining fragments, the process may require decades, centuries or longer as communities 'relax' toward a new equilibrium (36, 43, 44) . It follows that the frequency of disease emergence can be expected to follow a similar path: highest rates at first, followed by gradual decline as an ecosystem 'stabilizes'. The application of systems thinking to ecological communities centers on the ecosystem concept and has been the basis of significant research activity in ecological science since the 1960's (4) . Much of the focus has been on describing and understanding energy and nutrient fluxes across different kinds of ecosystems, and more recently emphasizing the human dimensions of global environmental change (45) . Although global scale ecosystem change has been considered in reference to human and infectious diseases (13, 46) , no systematic framework has yet been presented in this regard. The evidence presented has been anecdotal. However, the development and application of systems theory independently of and largely outside the realm of mainstream systems ecology (47) shows significant promise in providing a basis for more systematic interpretation of the role of ecosystem change in disease emergence. Recent thinking draws in particular on the application of the theory of 'complex adaptive systems' to ecological systems in which ecosystems are represented as scaled, self-organizing, far from equilibrium, evolutionary, and non-linear (6) . Organization, diversity, and resilience are important 'emerging' properties of complex systems-often equated with a 'healthy' state. 'Surprise', or qualitative disagreements of system behavior with a priori expectations, is another property (48) . Their association with loss of system resilience and increased vulnerability applies both to increasing inflexible social systems 'managing' and attempting to 'control' natural variables (i.e., vectors, pathogens) and that of the ecosystems whose component variables (e.g., vector and pathogen populations) are targeted for management or control. Alteration of natural disturbance regimes (via control of natural variation via flood control projects for example) reduces resilience, while secondary disturbance events (wildfires, storms, floods, and earthquakes) precipitate events caused by crossscale influences (e.g., a thunderstorm igniting a fire locally that spreads regionally as a result of fuel build-up from years of artificial fire suppression). Regional environmental change such as that associated with population growth development often does not accommodate the need to maintain resilience (21) . Such ecosystems have been described as 'over-connected' or 'brittle' and, as stated above, 'accidents waiting to happen' (8, 22) . Floods, often associated with waterborne disease outbreaks, occur more frequently and with greater severity as a result of lost resilience in natural systems due to attempts to 'control' (in contrast to manage) natural variation. Conversion of upland forests to plantations or cropping systems for more 'efficient' natural resource production is another, as is agriculture generally. Both result in a reduction of heterogeneity in ecosystems that tend to 'buffer' against disease outbreaks, which can spread more readily across the more uniform landscape including immunologically vulnerable domesticated species. Gunderson et al. (8, 23) describe this as general pattern of institutional behavior. It begins with the targeting of a natural variable for control, followed by increasing institutional efficiency and inflexibility in the control methods used. As the target variable and ecosystem changes and initial signs of failure are ignored, the ultimate result is a crisis. This 'pathology of regional resource and ecosystem management' and is apparently applicable to infectious disease management as well. The above body of ecological theory and observations involving specific emerging infectious disease cases suggests a causal schema that links ecological phenomena on the scale of pathogen transmission and evolution to regional and global transformations. This is represented by Figure 1 , focusing on the role of demographic and societal factors in disease emergence (shown in Table 1 ). This schema is constructed using the generally adopted view of human-environment interaction in which the impact of human population and technology is taken as the driving force, or ultimate cause. Clark et al. (25) elaborate on this in their seminal treatment on global and regional change. Here, the combination of population, technological capacity, and sociocultural organization act as the system drivers of regional environmental change. These and 'mitigating' forces are in turn influenced by 'human behavior', referring to patterns of actions and the rationales giving rise to them. Broadly speaking, these forces and their affects on ecosystems represent the 'ecological factors' in social-ecological systems, while human behavior represents the 'social factors' for the purposes of discussion. This schema represents zoonotic disease emergence from the perspective of the ecosystems within a regional environment, the large scale processes involved, and the associated ecological effects and processes involved in disease emergence. Thus, referring to Table 1 , the factors under 'Demographic and Societal Changes' can be identified in reference to particular ecosystems and processes (i.e., urban and urbanization, agriculture and agricultural intensification, forest and forest alteration), and associated factors operating on regional and global scales. These are unprecedented population growth, unplanned and uncontrolled urbanization, non-biodegradable packaging of consumer goods, and modern transportation, all contributing to conditions that promote pathogen transmission and persistence. For zoonotic diseases, which by definition involve the jumping of a pathogen from its natural host to humans, and in some instances extension of its host cycle to include humans, conditions can be described for this simple schema as follows. The likelihood or frequency of transmission events change when the natural host or pathogen changes, humans change, or the ecosystem supporting both changes. Thus, fundamentally the processes influencing transmission of zoonotic pathogens can be described as a consequence of one or a combination of three possible kinds of change: expansion of the habitat or geographic range of a host, of a pathogen or both; expansion of human's habitat or geographic range; or change in the habitat or ecosystem occupied by both humans and the natural host. Evolutionary adaptation of pathogens is omitted from the figure. However, it can be assumed any factor contributing to increased likelihood of transmission, as well as increased population size of hosts and pathogens will contribute to the potential that new, increasingly pathogenic, infectious, antimicrobial resistant variants will emerge. Anthropogenic climate change, while not incorporated in the diagram, can be considered to potentially contribute to disease emergence through its contribution to habitat alteration. Our review of ecological theory and the resulting complex model described here demonstrates the limitation of classical disease ecology and natural history. For example, without incorporation of population genetic theory in ecology as 'evolutionary ecology', along with the concepts of pathogen spillover and cross-scale ecosystem dynamics, classical disease ecology cannot explain recent emergence events like those involving HIV/AIDS, SARS, avian influenza, E. coli 0:157, dengue, malaria, West Nile virus, and Nipah virus, among others. The resurgence of vector populations, having acquired pesticide resistance and lost predators and competitors as natural controls after a regimen of inappropriate pesticide use, similarly is explained by modern ecology and evolutionary biology. This has direct application to control and intervention policies and practices. For example, it points to the need to adopt control strategies that integrate landscape and habitat management with careful rotation of targeted chemical and biological agents (30) . Developing such strategies requires detailed field studies, built on traditional natural history and classical disease ecology, but supplemented with advanced molecular techniques, as well as ecological research that takes into account the community and systems level dimensions of a particular pathogen-host complex. The continual expansion of human populations since prehistoric times, and particularly since the advent of settled agriculture with its associated domesticated animals exposing humans to their pathogens as well as those spilling over from wild animal populations, has incrementally added to the pathogen load through successive invasions by different organisms over time (49, 50) . Well established principles of population ecology, applied via mathematical epidemiology as Anderson and May and others so aptly have done, readily explain why, ceteris paribus, infectious disease incidence should generally be increasing with human population size, as it has in the world's poorest and most populous regions. Yet in spite of what are in general commonly understood principles, and warning signs that went unheeded in the 1970's and 1980's (10, 30), biomedical and public health institutions were unprepared for the recent surge in emerging infectious diseases (1, 9, 10, 51) . Not until the 1990's, prompted in part by the HIV/AIDS pandemic and the failure of 'quick-fix' solutions based on drugs, vaccines, and pesticides (18) , did the biomedical science and public health communities begin to launch a significant response. In light of the complexity of the factors involved, this lack of preparedness should perhaps not be surprising. As explained by the infectious disease ecology described here, zoonotic disease emergence involves biological processes operating on the scale of molecules and cells to that of coupled, regional scale human-natural systems. The latter involve political economic factors and policies driving regional environmental change, spreading geographically across the globe. It is this process of globalization-its ecological underpinnings and consequences-to which the current EID global trend of zoonotic disease emergance can be largely attributed. Lack of awareness of the ecological implications of Fig. 1 Diagram depicting infectious disease ecology causal schema. The aggregate variable at the top of the diagram, representing population and consumption, along with mitigating socio-cultural attributes, is the driver or 'forcing function' responsible for land use and land cover change characteristic of a particular region. The result is varying degrees (and types) of urbanization, agricultural intensification (including food production and distribution), and alteration of natural habitat. These changes at the level of landscapes and habitat produce conditions influencing the ecological and evolutionary dynamics of vector and host species and vector/host-pathogen dynamics. In turn, these conditions facilitate the geographic expansion, novel appearance, or increased epidemic activity of a disease. regional environmental transformations, and of their synergy with the ecological and evolutionary consequences of inadequate or inappropriate policies or methods of vector control and disease prevention, have unwittingly promoted disease emergence. The 'ecological' changes taking place as a result of public health agencies' actions (or inaction) involve pathogen biology and are small scale in time and space: selection for vector pesticide and antimicrobial resistance. However, the cumulative effect of micro-scale processes involving pathogen adaptation and host range expansion (or re-expanding) can ultimately produce regional and even global consequences. However, policy action and inaction outside the domain of biomedical or public health agencies has produced ecological changes at a historically unprecedented rate and scale. Urbanization, agricultural and food production intensification, alteration of natural habitat, along with concomitant loss of ecosystem functions, have transformed entire regional ecosystems during the past 50 years. The role of urbanization cannot be underestimated. The direct land use changes associated with urbanization effectively concentrate human, animal reservoir and vector populations at unprecedented densities. But urbanization also is strongly tied to socio-economic and cultural factors, along with human migration dynamics, affecting agricultural and food production intensification, as well as rural and natural ecosystems. In the case of food production and distribution, dramatically increased contact with and transport of wildlife (e.g., bush meat trade in Africa), and increased ruralurban transport and concentration of wild species for the exotic food market (e.g., civet cats in Guangdong, China), is another contributor to increasing disease emergence. It can be assumed these impacts on wildlife populations also contribute to changes in the composition of ecological communities on a regional scale (52) , and often result in a hyper-abundance of small mammal species low in the food chain, which are likely to serve as human disease reservoirs. Similar effects of ecological disruption appear to generally apply to invertebrate commu- Whether a disease 'jumps' to the regional population and the global human population scale is determined by quite different processes. These depend on demographic and transport patterns related to processes, like urbanization, influencing pathogen transmission and behavior operating on the regional and global scales. Physical environmental processes such as climate variability (on the right side of the figure), which are episodic by nature, include short term and small scale variations in the form of seasonal storms (e.g., monsoons), for example, as well as larger time and space scale variations. These include, for example, decadal and regional scale changes in weather patterns such as El Niño. These climate changes and weather events can precipitate floods and droughts. These act as cross-scale mediators that directly affect disease reservoir and vector populations, or pathogens (e.g, dispersal via flood waters). They cause the disease to jump from a smaller demographic unit to a larger one (e.g., from a single village to a district). Human Factors such as land use and land cover changes (Table 1 and Fig. 1 ) produce ecosystem changes contributing or magnifying the cross-scale processes. For example, urbanization and deforestation increase the magnitude of floods and droughts resulting from natural or anthropogenic climate variation. nities (53) , the kinds of organisms most commonly serving as human disease vectors. The estimated acceleration of natural species extinction rates by orders of magnitude over nonanthropogenic extinction rates (54) gives an indication of the scale and magnitude of change in natural communities, particularly in the tropics. In these regional ecosystems, the original extent of tropical forest has declined to a fraction of its original area with concomitant affects on biodiversity (38) . Large scale agricultural expansion (a common proximate cause of tropical deforestation) has decelerated and all but ended in many regions. This suggests that further contributions from what historically has been a main factor in natural habitat alteration, ecosystem disturbance and biodiversity loss, will decline. Yet agricultural intensification has replaced expansion with technologies that further impact native biodiversity (55) . The reduction in plant species richness that accompanies agricultural intensification leads to changes in the community composition of the pest complex-herbivorous insects, their natural enemies (predators and parasites), and the microbial community attacking crops (56) . The schema described here demonstrates how regional environmental change, involving ecological dynamics such as demographic or landscape transformations on the scale of decades, interact with change on the scale of host-parasite/ pathogen dynamics. How these cross-scale mechanisms produce regional or global scale disease emergence patterns is beyond the realm of conventional epidemiology, or analytical approaches generally. This may explain why such changes either are not apparent or their implications in terms of disease emergence are a low priority to biomedical and public health agencies. Such cross-scale processes are however characteristic of ecosystems, whose dynamics involve interaction of variables and operate on vastly different time and space scales, involving natural processes that are discontinuously distributed as shown in Figure 2 . Change in such systems is difficult to understand employing conventional analytic approaches alone, but require new thinking and analytical methods drawn from complex systems theory. The discontinuous character of the processes or variables makes their interaction intuitively improbable except for the mediating effect of 'cross-scale influences' as illustrated in Figure 2 . Examples include the eruption of Hantavirus in the American Southwest, thought be associated with weather events precipitated by El Nino Southern Oscillation. Storms events are similarly episodic, and represent a discontinuous form of variation transporting pathogens via flood waters in domestic and natural environments. On the other hand, massive environmental events, like the recent Southeast Asian tsunami, are potentially capable of producing epidemics across an entire region, and may episodically extinguish some zoonotic diseases by temporarily destroying reservoir and vector populations and habitat. The roles of resilience and vulnerability in determining the severity of such events, both in terms of social systems and ecological systems is another critical aspect of social-ecological systems behavior (8) . The recent worldwide upsurge of zoonotic infectious diseases, involving the resurgence of a growing number of diseases previously believed under control or the emergence of newly recognized diseases, has been attributed to a list of global factors characterized in terms ranging from microbial adaptation and land use to changing ecosystems, breakdown of public health, and poverty (1, 57) . The categorization provided by Gubler (10) , elaborated here based on an expanded view of infectious disease ecology, provides the basis for a schema describing the causal relations involving factors previously identified, along with hypothesized mechanisms. The complicated nature of this problem, which obviously entails numerous interacting variables operating on different time and space scales pose a significant challenge to biomedical science and epidemiological research as well as public health intervention. However, the current trend of increasing global emerging infectious diseases is linked with another issue of 'global governance', sustainable development, with which disease control and prevent strategies must be integrated. Here too, the problems of politically stability, population growth, unplanned urbanization and economic development, income disparity, and environmental degradation, are all integral to the solution. This interdisciplinary imperative challenges what historically has been an increasing disciplinarily focus in infectious disease research. Greater investment in research has succeeded in revealing more detail about diseases within specific disciplines. Yet this arguably has been at the cost of greater disintegration rather than integration among disciplines. The division is widest between the genetic and biological aspects of disease and health on the one hand and the social, economic, political, and physical environmental factors on the other. The disciplinary distinctions within infectious disease research of course belie the true transdisciplinary nature of this and most problems in the global heath and environment arena (59). The above disciplinary divisions largely reflect the gap between disciplines focusing on systems at or below and above the level of the organism or species. The former involve disciplines addressing the genetics, pathogenesis, or immune response within particular organisms (humans, vectors of pathogens), for example, or disease as a statistical phenomena at the species population level (the mainstay of epidemiology). The latter deal with social or ecological phenomena, essentially representing higher levels of biological organization, within which the organism-level processes operate but that also involve many other variables or processes operating on a wide range of biological scales (from genes to the biosphere) characterized by cross-scale influences, and interactions at multiple societal and ecological levels. It follows that for intervention to be globally effective, in addition to rebuilding public health infrastructure based on the comprehensive view of infectious disease ecology presented here, at least three elements are required: 1. a multilevel ecosystem approach, involving cross-scale integration. 2. incorporated ecological theory and data for the specific disease system, 3. local scale intervention using a participatory approach that matches pathogen management with sustainable development across ecosystem and institutional scales.

Globalization and emerging governance modalities

This paper explores the possibilities for global governance effectively dealing with the international transmission of disease. First, zoonotic regulation and control pose a special case for public health agencies, and this paper proposes a propositional model for an effective public health stance. Second, globalization dynamics are briefly reviewed in terms of an emerging consensus on the need for global governance in public health. Third, a brief examination of global governance modalities suggests that a strong global governance case has distinct limitations (despite its possible justification); an exploration of contemporary directions in global governance follows. Finally, the paper examines the phenomenon of contemporary zoonotic control within the conditions of an effective regulatory regime.

The global governance of zoonoses presents a special case for global public health. The structural framework and dynamics of public health define the parameters of how emerging mechanisms of global governance are likely to influence the control of zoonoses. I begin this paper with a review of some of the endemic tensions that exist within efforts to extend public health into a global framework. I then discuss how the dynamics of contemporary globalization influence the transformation of both social and governmental behavior. I review the notion of "regimes" of governance that embrace trans-national and supra national entities. Finally, I inquire about the conditions for an effective regime for zoonoses control via global governance. Efforts to control infectious disease lie within the writ of state authority and are bound by the framework of public health. Wherever society has legitimized private capital as to some degree separate from that of the state, a structural tension exists between the state's interest in defining health on behalf of its subject population, and the resulting regulation of private inter-ests. When the state does not permit the existence of private capital within a framework of civil society (and the structure of rights implied), national state structures pursue public health as part of their collective national responsibility. This structural tension with private interests in the context of global governance informs zoonoses because (a) the processes of contemporary globalization are changing accepted notions of the role of the state (of which more later), and (b) scholars of public health have increasingly come to recognize that the definition of health itself (which the state must act to protect and promote) is flexible, constantly changing, and subject to the transformations occurring in modern, globalized consumer societies (1) . Indeed, a review of competing definitions of public health by the Institute of Medicine (IOM) in the United States in 1988 led the IOM to define public health as "fulfilling society's interest in assuring conditions in which people can be healthy" (2) . This broad definition, were it to be taken seriously and public policy aligned with it, would create a politics in which state power potentially could invade every private activity that challenges health. Pursued with vigor the political power of the state could develop into a national "health police". The proposition that the state should regulate directly the "upstream" causes of health/disease is advocated by 'the social production of disease' school 1 and is opposed by those concerned that pursuing public health with such means will lead to excessive state regulation of private interests. Ticklish regulatory issues are also raised by the so-called new public health in which individuals are held to be primarily responsible for their own health. In the latter regulatory regime, responsibility shifts from the state (e.g. to regulate noxious industries) to the individual (e.g. to take responsibility in avoiding the products of noxious industry.) The regulation of smoking falls very much within this paradigm. Current controversy over eating behaviors, resulting obesity, and the impact on the "public's" health are of the same character, as are efforts to hold mothers criminally responsible for ingesting "dangerous" substances during pregnancy (ranging from methamphetamines to alcohol and tobacco). As political scientists look at these issues, health as a value always stands in dynamic tension to other rights and values (e.g. wealth, liberty, rectitude, etc.) What societies negotiate at any given moment is the relative authority to be given one value weighed against another. The ultimate policy questions become how much authority will health values and those who espouse and affect them have in society, and at the opportunity cost of what else (3)? Two oft-conflicting notions of authority contend within the contemporary public health paradigm. Optimizing the public's health in a world of expanding threats requires increased amounts of state intervention. However, in part because of the prospect of this increasing state intervention, and to side-step the challenge to the power of private interests associated with the creation of dangers to the public's health, some would shift the ground for responsibility for health to individuals, who would be required to take on new responsibilities to calibrate their social behaviors (where to go, where not to go, what to eat, what not to eat) to promote their own health. Public health always involves some construction of "the public" and its presumptive interests as defined by state authority. Public authority over the monitoring and control of historic zoonoses has involved a full range of extensions authority of state authority to those areas from which such disease arises. And, while some of these extensions are well established, others reappear in response to changing social demands and technologies. As new diseases emerge, each simultaneously is "fitted" into the prevailing model of regulation, even as they challenge that application. National and transnational regulation of beef herds under threat of a BSE infection is a case in point. Equally important, however, has been the level of state fiscal support for public health. Even as the dynamics of globalization increase health threats to the public, the prevailing ideology of neo-liberalism through which state policy is filtered leads to a diminution of state authority and diminished resources for public health disease control (5). The goal of global governance of zoonoses lies in creating an effective regime of regulation. Below I will examine the notion of global regulatory regimes. Here, I seek to determine where in our understandings of public health as a regime of national regulation we find the necessary properties to be effective. The core argument is that a global governance regime of zoonoses regulation would need to be modeled on these characteristics. This seemingly self-evident proposition conceals a greater subtlety, for the current regime of globalization has relied on the state and its capacity to effect purposeful public health policy. Indeed, political scientists will categorize states as weak or strong in part by their ability to create and maintain effective policy processes. Weak states can create, but not necessarily implement public policy (6) . Much of the world continues to live in "weak" states in which governments (central as well as regional and local) have difficulties ranging from grave to impossible to carry out policy intentions, particularly those regarding public health 2 . In these circumstances, maintaining good public health becomes impossible. Proposition Two: Effective public health is dependent on sufficient social investment Overall, public health is losing comparative budget parity with the rising costs of curative medicine in the developed nations. In the United States, with overall medical costs once gain breaking out into double digit annual increases, the relative share of national budgets devoted to public health declines. Especially in times of economic downturn, and the absence of compelling crisis conditions, social investment in public health is uncertain in most national budgets. Proposition Three: Sufficient social investment is dependent on prevailing political and economic ideology Neo-liberal economic and political ideology call for reduced taxes, a dismantling of welfare state structures, increased individual responsibility for social consumption, deregulation of the private sector, and reduced governmental spending. Those who see global public health structures as incapable of meeting the public health challenges spawned by globalization, find the major culprit in neo-liberal policies weakening the public sector (7) . Adoption of necessary regulatory regimes may be viewed as a kind of social investment. The unwillingness of the U.S. and China to participate in Kyoto for the reason that to do so would threaten continued levels of economic development is a testament to the power of neo-liberalism. The endorsement of the protocols by Russia this week (in addition to the relative advantage this gives Russia within the accord group) is a measure of its relative irrelevance to the current politics and economics of Russia. Proposition Four: Inequality is detrimental to good public health A growing consensus holds that contemporary globalization is increasing inequality. A long-range macro analysis of inequality and health status holds that as inequality declines, overall levels of health improve, and the reverse (8). Proposition Five: Regime corruption is detrimental to good public health Regime corruption manifests itself when particularized interests subvert public policy for their own benefit. Persistent regime corruption diverts public policy from its intended purposes. It follows that public health's successes or levels of regime corruption significantly determine failures. Proposition Six: Good public health is directly linked to positive social, economic, and political development Uneven or ineffective development results in poverty and weak regime states. The 'social determinants of health' school holds, importantly, that good developmental policies contribute more overall to the health of the public than medically oriented individual intervention, no matter how sophisticated and successful the latter. Good development policies lead to improved population health; uneven and unsuccessful development leads to inequality, poverty, and deficient provision of clean water, effective sanitation, adequate shelter and diet, as well as the political problems that follow from these conditions 3 . Proposition Seven: To achieve policy success, public health needs to be able to value its own successes The overall goal of public health is to reduce or eliminate the incidence of specific diseases. When public health practices result in lessened disease threats, the relative value of public health in the policy process wanes. In an odd way, public health is successful when things don't happen, when people do not become ill; it is about negative instances, which are notoriously difficult to "count". Tying this observation to proposition two above, public health's budgetary fates rise during times of crisis and suffer during times of normality (9) . Proposition Eight: Public health suffers from the politics of focused expertise and technology In a related manner, public health funding tends to lose out when public policy is oriented toward producing focused expertise and technology. Public sector investment in "health" has reached very high levels: the National Institutes of Health in the United States received $27,066,782,000 in FY2003. These massive levels of investment have produced spectacular successes in knowledge creation, the invention of non-invasive and minimally invasive surveillance of the body, and a vast array of medical interventions. At the system level, however, the multiplication and diffusion of highly technologized interventions results in rising expectations for medical care, and increased overall medical care costs, which crowd public health spending in national and sub-national budgets. By contrast much important public health work is low tech. In a corollary to Gresham's law, high tech drives out low tech in budget contests (just as specialized medicine trumps primary care, and cutting edge proprietary pharmaceuticals trump generics). Some exceptions to this proposition may exist with emerging tools for micro surveillance devices. Proposition Nine: Achieving public health is a moving target: notions of acceptable levels of health change over time; new diseases are constantly developing Health is a relative value. Achieving it is an uncertain objective. Potentially the demand for health-especially as defined by medical interventions-may be infinite in a social climate in which individuals seek and receive new interventions to extend life or improve some aspect of bodily well being (10) . These observations clash with the languages and perspectives of the policy process in which notions of attacking problems, defeating social ills, or achieving victory in another war on something are commonplace. These rhetorical tropes serve well-recognized strategic and tactical means within the policy processes for mobilizing support, achieving agenda positions, and gaining budgetary allocations. When applied to public health, however, they create unrealistic notions of what can and cannot be accomplished within the frame of health by those practices we term public health. The result is that rhetorically, we are always in some ways losing the public health battle (11, 12) . My hypothesis is that any regime of zoonotic control will be subject as well to all of these propositions. Whatever governance regime is developed for zoonotic control at whatever level (local, regional, national, global), the effectiveness of such controls will depend on the degree to which the values and institutional practices suggested by these propositions is found to apply. The contemporary era of globalization dates from the mid-1960's. A set of similar global transformations took place during the period 1870-1914. In both cases heightened economic integration resulted from increases in international trade, finance, and investment. While this earlier globalization took place within the framework of the national state, contemporary globalization has expanded beyond the nation state through transnational economic actors, multinational corporations. The result has been to create a new political space within which the immense transactions of the global economy take place. While new economic institutions have developed in contemporary globalization, corresponding political governance has not: economic space has expanded beyond political space (13, 14) . The dynamics of contemporary globalization compress time and space, creating a more immediately available world of goods, services, communications and interactions of all kinds (including military) (15) . Massive amounts of capital are aggregated at the global level, circulating in currency flows that dwarf anything previously known. These capital concentrations provide immense power to private capital to transform societies, often at the expense of a relative reduction in the power and authority of national governments, who in comparative terms lose the capacity to control their own policy agendas (14) . Economic interdependence brings unparalleled efficiencies in goods and information exchanges, but at the price of a hyper-sensitivity to negative economic effects within the system, as the 1997 Asian currency crisis revealed. Sudden disease outbreaks such as SARS can also have large and immediate negative economic effects that rapidly ripple throughout the system. The actual collapse of national economies from global financial instability such as that of Indonesia in 1997-98 can drastically reduce national income, with society-wide negative health consequences 4 . Contemporary globalization, operating through its primary vehicle the multinational corporation, shifted the manufacturing centers of the world into the former developing world. Essential to the success of this relocation has been the creation of constantly innovating transportation and communication systems that move goods with steadily increasing rapidity and lowering costs, and a global information system based on constantly innovating communications technology and computer networks (16) . As is well known, these advances enhance the spread of disease throughout the world and vastly complicate efforts to establish effective controls. We are just beginning to glimpse how changes in the ownership and extension of global media are coupling with other elements of our information societies and their networking capacities. Already it is obvious that basic consumption patterns related to health are changing, as societies become more consumer oriented. Changes in diet, work and living patterns are associated with goods through which we fashion our identities and make choices. Castells, like Harvey, sees the rate and nature of change the central element of how globalization impacts society, a process that he calls the creation of network societies. This contemporary process of change has the apparent property of being high recursive; elements of change in one dimension affect another through chains of reciprocal causation, working themselves back to the initial causal elements and producing new, and often unexpected, effects in the process. Social theorists make strong analogies to how ecological systems function recursively in making these operations. When these new and complex recursive processes occur within the social stew of hyper-urbanization, the outcomes are highly unpredictable (17) . Globalization is also marked by new means of wealth creation and distribution, inducing labor forces to move toward aggregations of capital most of which are urban. This process has touched off the largest migration in human history (18) . The 21st century has become the urban century as for the first time in human history more people live in cities than in rural areas. One component of this migration is cross-border, legal and illegal migrants seeking work. This pattern, however, is dwarfed by within-country migrations which have brought hundreds of millions of people into cities, now perceived as the critical nodes of global production, and therefore the locus of jobs. In many of these cities, especially the "newly large" cities of the developing world, or the mega-cities of Asia, living and working conditions resemble those of Dickensian England. At the current edge of the 21st century's world, industrialism coupled with urbanization has reproduced the social conditions of the 19th century. The unchecked growth of mega cities as survival harbors make them the new reservoirs for the lethal combination of poverty, crowding, insufficient sanitation, impure water and disease. The non-urban world has also been fundamentally transformed, as logging, mineral extraction, and oil exploration and production have brought virtually every hectare of global space under the economic gaze of capital exploitation. In search of ever greater food supplies, fisheries throughout the world have been driven to collapse and forests pushed back for agricultural development, especially when monoculture drives out smaller scale diversified agriculture. The agricultural equivalent to global financial interdependence lies in the possibility of irreversible environmental shifts such as drought in China, which under current financial arrangements could cause market forces to eliminate from global demand those who cannot afford grain. Structural inequalities coupled with civil unrest (another kind of inequality) already produce widespread hunger, malnourishment and famine. An environmental disaster in a large population, most obviously China or India, could have devastating effects within the world system. This, of course, is the great fear environmentalists have for dependence on mono-cropping, species loss, and water management 5 . In sum: contemporary globalization has dramatically increased global wealth, through innovation and joining new capital to massive labor forces; it has also produced a distribution system that promotes inequality on a scale previously unseen. Stunning innovations in productive capacity, communication, and transportation have imposed new technologies throughout the world. World populations have increased in number and concentration as the world has rapidly urbanized; ecological imbalances have been intensified and with them the conditions for the spawning and transmission of new diseases increased. Contemporary patterns of global industrialization create profound ecological challenges from resource depletion. One further note on globalization as a market phenomenon. Contrary to traditional theories of economics, the market is not self-governing. Classical and neo-classical economics can assure us that markets will create efficient price levels and send appropriate signals to organize supply and demand, but markets also routinely produce externalities, sometimes of massive, negative proportions, and are inseparable from the business cycle. As global wealth increases, the business cycle produces spectacular wealth during periods of boom and spectacular poverty and despair during periods of bust: the greater the extent and interdependence of the world system, the greater the attenuation of these extremes. In the industrial progression of national economies, the early periods of raw industrial growth were followed by the imposition of regulatory regimes designed to mitigate the human costs of industrial development. Markets may be efficient for exchanging goods, but they are not effective in representing the controls on excess that modern industrial human populations historically have come to demand from the economic schemes within their societies. The replication of these economic excesses at the global level is producing similar movements toward global governance (regulation). In a world still governed by the sovereignty of nation states, however, achieving the required and effective regulation is a daunting challenge. Part Four: Governance regimes: Strong and weak programs of global governance The compelling questions for global governance are what should be governed and how? The fundamental environmental/ bio-health questions for global governance include: whether to seek remedies eliminating the causes of negative health and environmental outcomes; or to pursue limited programs that seek to mitigate effects at the margin. A similar tension exists within the discourse of global governance: should global governance seek to regulate the processes of globalization themselves, or should efforts concentrate on regulating effects? While the two can be seen as opposite sides of the same coin, these can fairly be termed the strong and weak programs of global governance. The strong program envisions institutions that have a direct capacity for regulating (with sanctions) individual actors, whether they are nations acting on behalf of private interests (or their own state corporations), or private interests, and individual actors themselves. The weak case envisions governance mechanisms that operate largely through existing institutions (including states) and require their compliance to effect action. These positions define the antipodes of a governance continuum. The current politics of global governance distributes advocacy for various structures or mechanisms along this continuum. Obviously, the strong governance program most directly addresses the operation of globalization actors at the transnational level; conversely, this position acknowledges the degree to which state sovereignty has already been compromised by cross border globalization (14) . A propensity for the strong program in large part depends on one's view of the global condition. If it is seen as a "ticking time bomb" e.g. unrelenting global warming with catastrophic consequences such as the possible melting of the Ross ice shelf and rising sea levels, one views the world in crisis and moves toward a discourse that promotes more radical changes necessary to sustain society under threat. The scale of the threat, it would be argued, justifies actions that impinge drastically on the traditional institutions of the state and its sovereignty. A paradigmatic example may be the Montreal treaty on CFC regulation. The biological version of this scenario might be a conviction that current global practices previous mirror conditions in human history that led to the emergence of new diseases or their epidemic spread: a warming climate, the rapid and large scale movement of populations, novel mechanisms of transportation that permit rapid communication of peoples, and destruction of existing barriers between native forests, agricultural land, urban concentrations, and compromised water and sanitation systems (19) . Convinced of this scenario, one might seek to impel regulatory settlements intended to reduce population flows into cities, reduce the destruction of forests by agricultural incur-sion, curtail global warming, etc., actions that could only take place by directly interdicting the economic forces creating them. But such measures are unlikely. Policy processes do not work that way. Crises discourse may serve to modify the content of policy talk, but for large scale regulatory efforts to take place, the catastrophe in whose name such actions are taken must already have occurred. People are only convinced of the severity of truly profound crises until their very occurrence has validated that severity. The purpose of crisis dynamics and its role in the strong governance program is to legitimize lesser efforts. It is in this terrain that global governance is emerging. Within weak programs for global governance two basic types prevail; one seeks to reform existing institutions of global governance, the other to fashion novel regulatory regimes. Nayyar and Court indicate the assumptional base of the reform model: The endeavor should be to make the market-driven process of globalization conducive to a more egalitarian and broadbased development pattern. The object of such a design should be to provide more countries with opportunities to improve their development prospects and more people within these countries to improve their living conditions. It would have to be supported by a new institutional set-up. This would mean providing global public goods, such as world peace and a sustainable environment, as well as regulating global public bads, such as international crime whether trade in drugs, arms, people, or [human] organs. It will be necessary to reform existing institutions and to introduce new rules or create new institutions. Some of these would require a system to correct for the failures of unregulated or liberalized international markets, while other initiatives will be needed to build up missing markets (14) . From this proposal spring reforms of existing governance institutions including the United Nations 6 , the International Monetary Fund, the World Bank, and the World Trade Organization. Nayyar and Court are sensitive to the needs for differential application of regulatory rules for countries at different levels of development, for the need to make global governance more inclusive (and not just a rich nation's club), to provide a voice for those currently excluded, and to specify the conditions under which countries can opt out or exit from multi-lateral rules. Their reforms include the desirability for articulating the obligations of transnational corporations as well as their rights, including some version of an international regime of antitrust 7 . Like similar global governance proposals, acknowledgement is made of existing institutions that already constitute core arrangements of governance on which this larger program could be built, e.g. UNCTC, UNCTAD, the Organization for Economic Cooperation and Development (OECD), the International Finance Corporation (IFC), and the International Centre for Settlement of Investment Disputes (ICSID). The regime approach to global governance seeks to identify the existence of these institutions and practices in relation to specific problems they would regulate. Oran Young defines regimes as "sets of rules, decision-making procedures, and programs that define social practices, assign roles to the participants in these practices, and govern their interactions" (21) . Regimes differ in terms of their functional scope, geographical domain, and membership. Regimes are empirical, and may be unstable; they form in response to some situations and not others. Regimes have proliferated within environmental governance largely in response to discrete issues. Examples would include the Great Lakes water quality regime, the Antarctic Treaty System, and the European transboundary air pollution regime. Regimes may be nested in larger institutional structures, e.g. that for high seas fishing, which is subject to the more encompassing law of the sea. Often regimes are "lightly administered", generating compliance with minimal organizational resources ("governance without government"). While regimes often do not seek to provide comprehensive systems of public order for large geographic regions, they may occasion participation by states as well as inter-governmental actors. The result, Young suggests, is that regimes tend to form horizontally rather than vertically; they represent a "complex pattern of decentralized order." (21) Since its inception in 1946 and consistent with its predecessor organizational versions, the WHO has ascribed to the weak regulatory program, utilizing the terms employed in this section. Its constitutional function is the direction and coordination of international health work, including setting international norms and standards for health, and technical cooperation among members. Evidence of its weak program nature can be seen in those instances in which it has abjured political involvement (especially concerning Eastern Bloc membership issues in the 1960's) (22) . An issue beyond the immediate scope of this paper is the examination of current efforts within WHO to ascertain those that are part of regime formation and implementation, a la Young. In an effort to apply these diverse observations about public health, global governance and globalization, I turn to Frederick Murphy's 1998 presentation of emerging zoonoses in the special issue of the journal by the same name. I choose Murphy as a text because of his effort to capture in this brief review not only the range of factors implicated in emerging zoonoses, and the specifics of the most prevalent instances of that date, but also because he focuses on specific proposals to address what he sees as the need for an effective response. I seek to put this conclusion in the context developed in this essay. "…an emerging zoonosis is 'a zoonosis that is newly recognized or newly evolved, or that has occurred previously but shows an increase in incidence or expansion in geographical, host or vector range'. Emerging zoonotic diseases have potentially serious human health and economic impacts and their current upward trends are likely to continue." Examples include: avian influenza, Bovine Spongiform Encephalitis (BSE) and the Nipah virus (23) . Writing before the emergence of SARS, but in the context of BSE in the UK, HIV/AIDS, Sin Nombre and West Nile Virus, Murphy's specific concern is the adoption of "unique strategies" that will build more on fundamental research than "traditional" approaches. Included will be the rebuilding of "a cadre of career-committed professionals with a holistic appreciation of several medical and biologic sciences" (5) . He sees a persistent and to some extent alarming increase in emerging disease episodes (nearly all of which involve zoonotic or species-jumping infectious agents) including at the microbial/virologic determinant level mutation, natural selection, and evolutionary progression. Among individual host determinants he identifies acquired immunity and physiologic factors. Host population determinants include host behavioral characteristics and numbers as well as societal, transport, commercial, and iatrogenic factors. Environmental determinants include ecological and climatologic influences. The remainder of his review focuses on ecologic factors, especially those exemplified by arbovirus diseases. He seeks to identify lessons learned from Venezuelan equine encephalitis epidemics, the equine morbillivirus outbreak in Australia, from Ebola hemorrhagic fever, rabies epidemics, from the hantavirus pulmonary syndrome epidemic and from bovine spongiform encephalopathy in cattle and new-variant Creutzfeldt-Jakob disease in humans. Murphy's policy argument is to extend the discovery-tocontrol continuum to the full range of zoonotic diseases. The discovery-to-control structure draws on elements from fundamental scientific research, to the creation of an effective practitioner community, to creating and sharing data bases drawn from increased national and sub-national surveillance, to the full range of actions required in the final control stages of zoonotic diseases. And at each step, he is sensitive to their costs (without necessarily employing the various languages of public policy that apply). He recognizes the extensive nature of the final phases of the project. More expensive and specialized expertise and resources come into play in the final phases of the discovery-to-control continuum: public health systems, including rapid casereporting systems, surveillance systems, vital records and disease registers, staffing and staff support, logistic support, legislation and regulation, and expanded administration; special clinical systems, including isolation of cases, quarantine, and patient care; and public infrastructure systems, including sanitation and sewerage, safe food and water supplies, and reservoir host and vector control. My interest in Murphy's review and analysis stems precisely from its initial assumption that the beginning point for eventual control begins with discovery. It certainly seems a reasonable place to begin with. However, most of the above argument about globalization suggests, paradoxically, otherwise. That argument says, in effect, that if we continue along the road that globalization is taking us, we may be "producing" diseases in ways that overwhelm the capacity of systems to deal with them if they are actuated at the point of discovery. Is this a silly way to reason? I don't think so. Defining zoonoses in terms of a set of problems within the discovery-tocontrol continuum is analogous to the weak program for global governance. It accepts the premise that important public health situations constitute problems to be solved by intense, expert, scientific driven problem solving techniques. But these issues may more rightly be dilemmas than problems. A dilemma differs from a problem in that the situation it "contains" will not produce a "solution" at the level of analysis at which it stated. Simply: within a dilemma one cannot continue to do what one is doing-the thing that produces the "problem"-and gain an effective solution. In the above analysis, the world cannot continue to embrace and support the activities of contemporary globalization and meet the public health problems it will create. To resolve the dilemmas that constitute global public health requires changing (perhaps in radical ways) the ways that globalization works, so that it will produce different results. Problems say: do more; dilemmas say: do something different. This is what proposals for global governance purport to do. The all-important question is, even were they to succeed, through the reforms proposed, would these be sufficient to create transformations in globalization behavior yielding different outcomes? One should never say never. I am tempted to argue, nevertheless, that we cannot solve the problem of creating an effective regime of zoonotic control because such a regime, as posited, would not change the conditions of the global economy that are producing the very diseases one wants to control. This does not mean however, that one might not embrace projects or obtain results that are important steps in this direction. For example, I have made much of the tension between the current world of infectious disease in which there is a large, and perhaps ever increasing, need for more research, more science, more data bases, more trained personnel and international cooperation on policy implementation, all of which cost money (and involve additional contribution of resources on the part of the state), and the trend under current globalized regimes of neo-liberalism to actually weaken the state, to shrink public spending and to weaken its regulatory hold. Following this logic suggests to me two conclusions. One, we will fall behind in the kind of program Murphy advocates because the political and economic forces that promote global neo-liberalism are stronger than those that promote public health (including zoonoses control). In the contest of values and public policy discussed in section one of this paper, particularized interest will triumph over more generalized interests. Two, some disease will present itself with such threat and virulence that its consequences to existing society cannot be ignored. In the face of this manifest crisis (perhaps a breakout of the current Thai-based avian flu virus), public health intervention will go to the top of the policy list. Not only public monies will be made available, but also private sector funds. William Gates or Warren Buffet, or George Soros or some other combination of the world's largest holders of private wealth may seek to intervene to transform the direction of current public policy. And that might turn the tide on the existing threat. Or not. Within the last century only the flu pandemic of 1918-19 has constituted as great and widespread a threat to human population as the HIV/AIDS pandemic. Yet threats of this magnitude have failed to mobilize an effective interventionist regime. While it is useful to speculate on some of the reasons that such a regime has not emerged (e.g. the disease emerged outside the developed world (non "us"), and then proliferated in stigmatized populations, wide spread denial of the size and extent of the disease by many governments, etc.) the combination of urgency of impact (disease incidence continues to rise) and failures in the establishment of an effective governance regime (typified by the failure of the U.S. government to contribute its full share to the global fund) shows that the HIV/AIDS pandemic has new lessons for us about how crisis dynamics do and do not work to effect appropriate responses. Before proceeding to a conclusion, two final words. On risk. Like health, risk is a socially constructed category that changes within the complex norms and values of a given society or culture. What is an acceptable level of risk at one period of time, may not be in another; what is an acceptable level of risk for one group (e.g. others), may not be for another (e.g. ourselves) 8 . Notions of risk are negotiated within the political process; notions of crises, their relative severity and the amounts of resources to be devoted to them emerge from the complex processes of bargaining within the political process. When the site of the political process is global, rather than national, and when the mechanisms for decision making lack the force of national sovereignty, the complex calibrations of acceptable risk compound exponentially. On technology. We are wise enough to realize that technology can bring astonishing immediate benefits, but also through unexpected recursive feedback loops, create yet new and unanticipated problems, perhaps as severe as those they were designed to fix. Nevertheless, at the very least microtechnology and nano-technology hold enormous promise for assisting with surveillance tasks to control disease spread. Biobased sensors that can be made cheaply and sensitive down to the individual molecule level promise new ways of scanning large numbers of people (e.g. airport arrivals) cheaply and with minimal intrusiveness. Nano-tech sensors operating within the bodies of suspected infectious persons are also being studied. To see the individual as the broadcasting site for disease surveillance strikes many as a further draconian incursion into the oppression of individual liberty…but to return to our opening theme, this tension between private interest (including the self) and the public's health is endemic. This is not a pessimistic paper, although it may read like one. It makes a simple argument. Globalization is about the deployment of capital throughout the world through marketprivileging mechanisms. Globalization has a powerful up side in the vast wealth that has been created through its mechanisms, and the benefits that it brings to millions of people every day. (The World Bank for example estimates that 400 million people in China have been lifted from poverty by the mechanisms of the current capitalist based society.) But globalization also has its dark side, some of which is eluded to above, and which can be easily amplified in reference to the millions throughout the world who work in conditions that are unregulated, dangerous and underpaid 9 ; to the emergence of a world-wide sex industry tied to human trafficking that acts as a particularized HIV/AIDS reservoir; to the wanton disregard of private enterprises in many parts of the world to environmental destruction; to corruption and theft in the higher levels of some of the largest corporations in the world, global trafficking in arms and drugs, etc. 10 My argument is in the end simple: the dynamics of globalization have become the major "factor" in the social production of disease, including those elements that are intimately associated with the emergence of zoonoses. Directly addressing how these globalization dynamics impact human populations to produce disease must be a program for global governance 11 . It must look beyond a discovery-to-control paradigm for emergent diseases, to a research-to-prevention-to-discovery-to-control paradigm. Within these mechanisms of governance lay the difficult tasks of regulating capital in its excesses. This-to repeat my earlier point-was the basic issue in the development of national economies, and it is of necessity the basic issue in the creation of global governance. To succeed, these issues must enter the discourses of global governance in a direct way (and not be viewed as some exotic side issue). I conclude by repeating a portion of the previous quote from Nayyar and Court. Global governance means: "This would mean providing global public goods, such as world peace and a sustainable environment, as well as regulating global public bads, such as international crime whether trade in drugs, arms, people, or [human] organs. It will be necessary to reform existing institutions and to introduce new rules or create new institutions. Some of these would require a system to correct for the failures of unregulated or liberalized international markets…" 9 An especially egregious instance exists in China, in which hundreds of thousands, perhaps millions of peasants moving into cities have over the past decade been employed in myriad construction projects, including many proceeding under the approval of government authority. A current suit is being conducted on behalf of these workers who are owed $42 BILLION in unpaid wages. This is a country in which annual individual income is still a pittance. It is no secret who benefits and who loses from such a situation (18) . 10 The next great crisis in globalization has already begun to emerge. As capacity spread through out the developing world for labor-cheap manufacturing, countries such as Bangladesh became early players in the inexpensive manufacture of world goods. The rapid integration of China and India with their enormous labor forces into world markets is threatening to overwhelm other countries in this role, creating capital movement and resulting unemployment and deepening poverty. One adds to this the opening wedge of service industry outsourcing from the developed word, from low level work, e.g. call centers, to the highest levels of professional training. Radiologists in the United States (average income north of $300,000) may be replaced by those in India (average income of around $25,000); similar issues exist in engineering, computer programming and medical technology. Singapore, acutely aware of the implications for this threat is investing heavily to upgrade the country's professional labor force to better gain niches in scientific and technological research. As the deputy minister for development told me some months ago, "we can make all the capacity we need to stay competitive; we rent brains. It is cheaper and more efficient." (The opportunity in question arose from the U.S. government's decision in 2002 not to pursue stem cell research.) In this view of research, only a portion of it needs to be site specific. 11 This argument accords with the general thrust of Burci and Vignes in their recent review of WHO, in which they argue for a strong step beyond what they see as the legal timidity of the organization in the direction of a stronger "normative stance" (22) .

Generation and Characterization of Novel Human IRAS Monoclonal Antibodies

Imidazoline receptors were first proposed by Bousquet et al., when they studied antihypertensive effect of clonidine. A strong candidate for I1R, known as imidazoline receptor antisera-selected protein (IRAS), has been cloned from human hippocampus. We reported that IRAS mediated agmatine-induced inhibition of opioid dependence in morphine-dependent cells. To elucidate the functional and structure properties of I1R, we developed the newly monoclonal antibody against the N-terminal hIRAS region including the PX domain (10–120aa) through immunization of BALB/c mice with the NusA-IRAS fusion protein containing an IRAS N-terminal (10–120aa). Stable hybridoma cell lines were established and monoclonal antibodies specifically recognized full-length IRAS proteins in their native state by immunoblotting and immunoprecipitation. Monoclonal antibodies stained in a predominantly punctate cytoplasmic pattern when applied to IRAS-transfected HEK293 cells by indirect immunofluorescence assays and demonstrated excellent reactivity in flow immunocytometry. These monoclonal antibodies will provide powerful reagents for the further investigation of hIRAS protein functions.

Imidazoline receptors were first proposed by Bousquet et al., when they studied antihypertensive effect of clonidine [1] . Based on their physiologic and pharmacological properties, imidazoline receptors are classified into three main types: I1R, I2R, and I3R [2] [3] [4] . I1R and I2R have been implicated in hypertension and psychiatric disorder regulation, respectively, while I3R may be involved in insulin secretion [5] [6] [7] [8] [9] . Compared with mitochondrial I2R, which resides within the monoamine oxidase protein [10] , the clonidine-preferring imidazoline binding sites (known as I1R) are localized to plasma membrane fractions [11, 12] and specifically to synaptic plasma membranes [13] . A strong candidate for I1R, known as imidazoline receptor antisera-selected protein (IRAS), has been cloned from human hippocampus [14] . hIRAS is a larger protein of 1504 amino acids consisting of an NH 2 -terminal phox (PX) domain, 5 putative leucine-rich repeats, a predicted coiledcoil domain, and a long COOH-terminal region. Several evidence supported the identity of native I1R and IRAS protein in tissue distributions, ligand binding properties, some cellular functions and downstream signal pathways [14] [15] [16] [17] [18] . The murine form of IRAS, Nischarin, truncated at the N-terminal 244 amino acids including the PX domain compared with the hIRAS, was a soluble cytosolic protein involved in cytoskeletal organization [19] . It has been shown that decreasing the expression of rat IRAS or Nischarin in PC12 rat pheochromocytoma cells could attenuate the activation of extracellular signal-regulated kinase (ERK) or reduce the radioligand binding to I1R, which further supported that hIRAS or Nischarin might serve as I1R itself, or at least a functional subunit of I1R [20] . Recently, Molderings et al. have reported that the "I1-imidazoline receptors" mediating effects of clonidine and moxonidine in PC12 and the transfected HEK293 cells belonged to the S1P-receptor family, in particular, representing a mixture of sphingosine-1-phosphate (S1P)1-and S1P3-receptors and/or heterodimers of both. It was then proposed that an increased expression of IRAS or Nischarin may improve the receptor-trafficking from cytosolic S1P-receptors to the cell membrane and thereby increase the number of binding sites in the plasma membrane for radioligand binding [21] . In our previous study, we also reported that IRAS mediated agmatine-induced inhibition of opioid dependence in morphine-dependent cells [22] . Despite intensive efforts, the molecular base of the I1R had not yet been elucidated. To elucidate the functional and structure properties of I1R, several different epitope-specific antisera against IRAS have been raised in rabbits [23] . Because of IRAS splice variants or nonspecificity of these antisera, more sizes of IRAS (≈33, ≈85, ≈170 KDa) have been visualized in various tissue and cells, which limited their uses on western blot. IRAS was reported to target to the endosomes by a combined action of a PX domain and a coiled-coil region. The PX domain, consisting of 10-130 amino acids, was first identified from the sequence analysis of two SH3 domain-containing cytosolic components of NADPH oxidase, p47phox and p40phox [24] . Therefore, in the present study, we developed the newly monoclonal antibody against the N-terminal hIRAS region including the PX domain (10-120aa) . This development has great utility for immunoblotting, indirect immunofluorescent staining, immunoprecipitation, and flow cytometry. These monoclonal antibodies will provide powerful reagents for the further investigation of hIRAS protein functions. Protein. E. coli BL21(DE3) (F-ompT gal dcm hsdSB (rB-mB-) λ (DE3) (Novagen) cells were transformed with recombinant plasmid, pET43.1-IRAS(10-120aa). Transformants were selected from ampicillin-containing Luria Bertani (LB) lates. Selected colonies were cultured in ampicillincontaining LB media. Isopropyl-β-D-thio galactopyranoside (IPTG) was added to a final concentration of 1 mM. The incubation was continued for 3 hours at 30 • C. Cells were harvested, mixed with lysis buffer (phosphate buffered saline [pH 7.3], 1.0 mM EDTA, 1.0 mM phenylmethylsulfonyl fluoride, 0.5 mg/mL lysozyme (Roche)), and sonicated. The high-speed supernatant which contained the pET43.1-IRAS(10-120aa) soluble protein fraction was loaded onto 10 mL Ni 2+ -NTA agarose columns. The fusion protein was eluted with lysis buffer with 400 mM imidazole. Eluted proteins were electrophoresed on SDS-PAGE gels. Expression yields were analyzed with the Quantity One quantitative software (Bio-Rad) based on relative band intensities of comassie blue stains. Purified proteins were detected with western blot using anti-His mAbs (Santa Cruz). Immunosorbent Analysis (ELISA). The purified NusA-IRAS (10-120aa) , diluted to 10 μg/mL in 50 mM carbonate salt buffer (pH 9.6), was coated on plates at 100 μL aliquot per well, 4 • C overnight. The wells were washed three times with PBS-Tween buffer (0.05% Tween 20 in PBS). The coated wells were blocked with 200 μL of 3% BSA for 1 hour at 37 • C and then incubated with 150 μL monoclonal antibodies against NusA-IRAS (10-120aa) fusion protein with different dilutions (from 1 : 1000 to 1 : 25600). After incubation for 1 hour at 37 • C, the wells were incubated with 150 μL horseradish peroxidase-conjugated goat antimouse IgG (dilution 1 : 5000, Santa Cruz) for 1 hour at 37 • C after thorough washing. Peroxidase activity was detected using o-phenylenediamine and H 2 O 2 as enzyme substrates. Color development was stopped with 2 M of H 2 SO 4 and the absorbance was measured at 490 nm using Microplate Reader (Bio-Rad). The purified NusA-IRAS (10-120aa) protein (50 μg in a volume of 80 μL) was mixed with an equal volume of Freund's complete adjuvant. The antigen-adjuvant mixture was injected into 6 female BALB/c mice and was followed by three booster injections at 2-week interval in incomplete Freund's adjuvant. The mouse with the highest antibody titer tested by ELISA was boosted intraperitoneally with 100 μg NusA-IRAS (10-120aa) protein without adjuvant 3 days before the cell fusion. Feeder layer cells were prepared 1 day prior to fusion. Splenocytes from mice with the highest ELISA antibody titers were fused with murine myeloma cells SP2/0 following standard procedures [25] . Culture supernatants were collected after fusion and initially screened by ELISA with purified NusA-IRAS (10-120aa) fusion proteins as antigens. Positive hybridoma clones were selected with the limiting dilution method, and stable hybridoma clones were obtained after 3 cloning cycles. Isotypes of antibodies were identified with a mouse subisotype panel (Bio-Rad). The pristine-primed BALB/c mice were injected intraperitoneally with 1 × 10 6 hybridoma cells per mouse in order to acquire abundant mAbs. The ascitic fluid was collected and mAbs were purified with a protein A/G affinity column (Amersham Pharmacia Biotech). Transfected cells were harvested 48 hours after transfection. Total cell lysate preparation and western blot analysis were performed according to previously described procedures [26] . In brief, cell lysates were prepared, electrophoresed on SDS-PAGE gels, and transferred to polyvinylidene difluoride (PVDF) membranes. Membranes were blocked with 5% nonfat milk in TBST (20 mM Tris-HCl [pH 7.5], 150 mM NaCl, and 0.05% [v/v] Tween 20) and incubated with IRAS or GFP mAbs (Cell Signaling Technology Inc). Blots were incubated with horseradish peroxidase (HRP) conjugated goat antimouse antibodies (Santa Cruz) after primary antibody incubation, and blots were developed with enhanced electrochemiluminescence (ECL, Cell Signaling Technology Inc). The cellular localization of the IRAS protein was identified according to previously described procedures [27] . Cells on glass cover slips were rinsed with PBS, fixed with 3% paraformaldehyde for 30 minutes, and permeabilized with 0.2% [v/v] Triton X-100/PBS. Permeabilized cells were incubated with IRAS mAbs and TRITC-conjugated affinipure goat antimouse by PCR amplification of the human IRAS cDNA (AF082516) using the following oligomers: sense, 5 -CGGGATCCA-TGGAGCGGGAAGCCGAGCCG-3 , and antisense, 5 -CGGAATTCATAGAAGTGAAAATGCAAGAAGTG-3 . The full-length human IRAS was obtained by PCR amplification of the entire coding region, and the resulting 4512-bp PCR product was ligated into the pEGPC1 and PCMV-myc vectors in a similar fashion using the following oligomers, respectively: sense, 5 -CGCGAATTCTATGGCGACCGC-GCGCACCTTCG-3 , and antisense, 5 -CGGGATCCTAGC-CGGTGAGCTCGACAGGC-3 , sense, 5 -CGCGAATTC-GGATGGCGACCGCGCGCACCTTC-3 , and antisense, 5 -CCGCTCGAGCTAGCCGGTGAGCTCGACAGGC-3 . All plasmid sequences were confirmed by sequencing analysis. IRAS is a large protein comprising of 1504 amino acids. Its NH 2 -terminal phox domain is important for membrane association and cellular localization. The N-terminal of IRAS (10-120aa) covering the PX domain was cloned into the pET-43.1a(+) (Figure 1(a) ) [29] . The supernatant fusion protein was purified by His-tag affinity purification and was subsequently used to generate the monoclonal antibody. The dissolved protein yielded one major band of 78 kDa expected molecular weight (Figure 1(b) ) with high purity and integrity. The NusA protein used as a control generated the 66 kDa expected molecular weight. The recombinant protein was also confirmed with western blot using anti-His mAbs (right column, Figure 1(b) ). BALB/c mice (n = 3) were immunized with the NusA-IRAS (10-120aa) fusion protein, and blood was collected from the mice after multiple injections. Antibody titers were tested by ELISA on plates coated with the NusA-IRAS (10-120aa) protein (data not shown). The 2 mice with the highest titers were sacrificed and spleens from both mice were fused to myeloma cells following standard procedures. Individual hybridomas was grown and 125 hybridomas were further characterized. Supernatant from the growing hybridoma clones was screened with ELISA. Screening was performed on plates coated with NusA-IRAS (10-120aa) , NusA protein, and GST-IRAS (10-120aa) fusion protein to determine antibody specificity. A total of 5 hybridomas (DA041, DD015, BE073, BA022, and AH021) reacted selectively with the NusA-IRAS (10-120aa) protein in all 3 assays and were further evaluated. Isotype analysis revealed that all mAbs were of the IgG1 subtype. The immunoreactivities of the 5 representative mAbs with NusA-IRAS (10-120aa) were shown in Figure 2 , all of which specifically recognized a 78 kDa protein band which corresponded to the purified NusA-IRAS (10-120aa) protein, but not to the 66 kDa NusA protein. Anti-His mAbs reacted with both proteins in the same experiment (Figure 2(a) ) as controls. We evaluated the specificity of mAbs in mammalian cells by inserting IRAS cDNA into expression vectors to allow the production of GFP fusion proteins under the control of a CMV promoter. The pEGFPC1 and pEGFPC1-IRAS plasmids were separately transfected into HEK293 cells. Western blot analysis with the anti-GFP antibody demonstrated that chimeric proteins were expressed and migrated separately at expected molecular masses of approximately 19 or 27 KDa (Figure 2(b) ). However, the expected 190 kDa band whose molecular weight corresponded to the full-length IRAS protein was only detected with the mAbs DA041, DD015, and BE073 in GFP-IRAS expressed cells. The remaining BA022 and AH021 mAbs were negative (Figure 2(b) ). The same samples were also incubated with preimmune serum with no reactivity (data not shown). Results revealed that all 5 selected mAbs specifically recognized bacterially expressed NusA-IRAS (10-120aa) proteins, but only 3 mAbs recognized recombinant IRAS in mammalian cells. IRAS mAbs were used to detect the cellular localization of IRAS protein based on specificity characterized by western blot analysis. The pEGFPC1-IRAS plasmid was transfected into HEK293 cells. IRAS protein localization was tested by fluorescence confocal microscopy after 48 hours transfection. GFP-IRAS was primarily located in the cytoplasm in a punctate pattern (Figure 3(a) ) with concentration onto discrete loci and spot fluorescence (panel A, E, I, Journal of Biomedicine and Biotechnology M, and Q), as confirmed in previous studies [30] . The mAbs DA041, DD015, and BE073 were used as probes to detect a predominantly punctate cytosolic distribution as expected (panel B, F, G). However, the mAb BA022 homogenously detected the fluorescence distribution over the cytoplasm (panel N) and the mAb AH021 scored negatively (panel R). Merged images revealed that the GFP-IRAS fluorescence significantly coincided with the immunostaining with mAbs DA041, DD015, and BE073 (panel D, H, and L), but was not reflected by the mAbs BA022 and AH021 (panel P and T). Further, the punctate fluorescence patterns appeared to be unique to the mAbs DA041, DD015, and BE073. In addition, we also detected the endogenous IRAS protein with monoclonal antibody DA041 in PC12 cells, showing a strong signal associated with the membrane and faint signal in the cytoplasm (data not shown). These results suggest that the 3 mAbs specifically detect IRAS proteins by immunofluorescence, similar to immunoblotting results. Flow cytometry is a rapid, quantitative, multiparameter cellular analysis based on the measurement of visible and fluorescent light emission. We chose the representative mAb DA041 for further characterization based on the strong signals detected by immunoblotting and immunofluorescence. The IRAS cDNA was cloned into the mammalian expression vector PCMV-myc and was expressed in HEK293 cells. The mAb DA041 and the c-myc mAb (used as a positive control) resulted in significant increases in fluorescence intensity when compared to cells incubated with preimmune serum (Figure 3(b) ). The binding efficiencies were 30.77% by the mAb DA041 and 48.26% by the c-myc mAb. No reactivities were detectable in preimmune sera (0.01%). These results indicate mAb DA041 is effective for detecting IRAS protein by flow cytometry. We next characterized IRAS mAb ability to recognize IRAS proteins in their native states by immunoprecipitation. Cell extracts were immunoprecipitated with the mAb DA041 and analyzed with mAb DA041 immunoblotting after 48-hour transfection with PCMV-myc-IRAS. IRAS was selectively immunoprecipitated from cell lysates that expressed the myc-IRAS protein (approximately 175 kDa), and those which corresponded to the predicted size of the full-length IRAS ( Figure 4) . Immunoprecipitation of lysates with mouse normal IgG did not result in detection of protein species. Lysates were immunoprecipitated with the c-myc mAb to confirm mAb specificity, and similarly sized bands were detected. No corresponding band was visualized when the negative control c-myc mAb was used to immunoprecipitate cell lysate of the transfectant expressing the empty PCMV-myc vector. Similar results were obtained with the mAbs DD015 and BE073 under the same conditions, and the remaining mAbs BA022 and AH021 scored negatively. Therefore, the 3 isolated antibodies specifically recognized native full-length IRAS protein products. Our results suggested that the mAbs DA041, DD015, and BE073 were reactive for both immunofluorescence and immunoblotting. Therefore, these mAbs likely recognized linear epitopes in the IRAS protein. In addition, the 3 mAbs were capable of recognition of native full-length IRAS proteins by immunoprecipitation. However, the mAbs BA022 and AH021 specifically recognized bacterially expressed IRAS immunogens and did not detect recombinant IRAS in mammalian cells. This could be related to the different IRAS folding patterns, since misfolded IRAS proteins could result in the exposure of unique immunogenic epitopes different from native proteins. The mAbs BA022 and AH021 could be used as backup reagents to safeguard against antibodyspecific artifacts. In addition, analyzing the overall homology of the amino acid sequence of the PX domain, Sorting Nexins 13 (SNX13) is the most relative protein to IRAS [31] . We found monoclonal antibody DA041 against hIRAS could not react with SNX13 by immunofluorescence assay (data not shown). It was proposed that these antibodies developed here were specific to the PX domain of IRAS. In summary, we generated specific mAbs directed against the human IRAS N-terminal. The mAb DA041 exhibited the best performance in a variety of assays including immunoblotting, immunoprecipitation, indirect immunofluorescence stainning, and flow cytometry. Specific mAbs may provide an ideal reagent for further investigation of the function of IRAS proteins.

Molecular and physiologic basis of quinoline drug resistance in Plasmodium falciparum malaria

30 years before the discovery of the pfcrt gene, altered cellular drug accumulation in drug-resistant malarial parasites had been well documented. Heme released from catabolized hemoglobin was thought to be a key target for quinoline drugs, and additional modifications to quinoline drug structure in order to improve activity against chloroquine-resistant malaria were performed in a few laboratories. However, parasite cell culture methods were still in their infancy, assays for drug susceptibility were not well standardized, and the power of malarial genetics was decades away. The last 10 years have witnessed explosive progress in elucidation of the biochemistry of chloroquine resistance. This review briefly summarizes that progress, and discusses where additional work is needed.

Human malarias are caused by infection with one of five malarial (Plasmodium) parasites (Plasmodium ovale, P. malariae, P. knowlesi, P. vivax and P. falciparum), which occurs during a female Anopheles mosquito blood meal (P. knowlesi is zoonotic, the others pass human to human). P. falciparum infections account for 95% of malaria mortality worldwide, so historically most malaria research has focused on this species. However, there is also a critical need for basic research on the others, particularly P. vivax. In P. falciparum malaria, sporozoites injected during the blood meal migrate through the skin via multiple routes, localize to the liver within hours, invade hepatocytes and are then released into the blood after approximately 2 weeks as large merosomes containing 5-10,000 new haploid merozoites. These then rapidly invade red blood cells (RBCs) [1] and proceed through an amazing differentiation and feeding cycle that is 45-50 h long. An initial 'ring' differentiates into a single trophozoite, which then differentiates to 8-36 new mero zoites. These are then released from lysed infected RBC (iRBC) and re-invade RBCs once again. The resulting increased parasitemia generates the well known symptoms of fever, anemia, and so on, which without treatment can ultimately progress to severe or cerebral malaria, coma and death. Sporadically, iRBCs commit to gametocyte differentiation via poorly understood pathways. Haploid gametocytes ingested by another mosquito feeding on an infected human, differentiate to male and female and combine in the mosquito to produce diploid stages that ultimately result in the injection of new sporozoites via mosquito saliva during another blood meal. Most clinical manifestations of malaria are associated with the iRBC cycle. Not coincidentally then, nearly all known antimalarial drugs act against the iRBC. Understanding unique parasite biochemistry within the iRBC is essential since it is intricately linked to the molecular mechanisms of drug action and drug resistance for virtually all known antimalarial drugs. The past 10 years has shown that as these mechanisms are increasingly defined in molecular detail, this guides the rapid development of additional inexpensive antimalarials active against chloroquine (CQ) resistant (CQR) P. falciparum. Promising advances with partially protective vaccines notwithstanding, new antimalarial drugs will be direly needed for at least the next two decades, and probably longer. The above is just a glimpse at a very complex host-parasite biology that varies for different species and that protects multiple parasite stages from effective immune-system intervention in multiple ways. Childhood versus adult versus pregnancy-associated malarias differ, and common co-infections involving malarial parasites and other microbes pose additional therapeutic challenges. As a consequence, multiple vaccines will most likely be needed to eradicate malaria. Thus, it is imperative that equal emphasis be placed on inexpensive drug development and other approaches for controlling malaria. This review briefly summarizes progress related to inexpensive antimalarial drug development on three fronts: Molecular and physiologic basis of quinoline drug resistance in Plasmodium falciparum malaria It is a comprehensive picture of all three that teaches us how to circumvent quinoline antimalarial drug resistance. To date, three of the best five antimalarials of all time have been quinolines, and recent work suggests that, with a comprehensive understanding, additional inexpensive, effective quinolines can be developed quickly. Rapid RBC growth and subsequent multiplex asexual division presents severe metabolic challenges. The parasite has adapted to these challenges by acquiring the ability to rapidly digest millimolar levels of hemoglobin (Hb) acquired from the RBC cytosol. Hb digestion via elegant internalization and proteolysis within a unique lysosomal-like organelle termed the digestive vacuole (DV) is essential for parasite survival. The toxic by product, heme, released upon Hb proteolysis, cannot be degraded by the parasite because it lacks the heme oxygenase pathway. Instead, the parasite is one of only a handful of organisms on the planet that detoxifies copious amounts of heme by crystallization to hemozoin (Hz), as described later. Thus, this pathway is a specific, biochemically unique drug target. Key molecular details of this pathway have been elucidated in recent years. In parallel, drug-resistance researchers are defining how drug-pathway interactions are 'hijacked' by the drug-resistant parasite via subtle alterations in DV bio chemistry and physiology. In a desperate time, some optimism is on the horizon. The most successful antimalarial drug to date is the 7-chloro-4-amino-quinoline, chloroquine, which targets the Hb digestion pathway. The mechanism is not completely understood [1-6] but a critical facet is the inhibition of ferriprotoporphyrin IX (FPIX) heme crystallization to Hz [5-10]. As early as 1964 [11] the likelihood of important CQ-FPIX interactions was realized, and the idea gained momentum via the work of Fitch [12] . The Hz crystal unit is a curious 'head to tail' FPIX dimer, wherein oxygens, coordinating adjacent irons in the two tetrapyrrole rings, are donated by FPIX propionic acid side chains [13] . Formation of this dimer within the DV competes with the formation of the better studied µ -oxo dimer. In simple aqueous solutions the µ-oxo dimer is favored over the head to tail dimer. Overall, equilibria between multiple chemical forms of FPIX within the DV is poorly understood. However, it is known that both FPIX dimerization and Hz crystallization are highly pH dependent, and that Hz formation is accelerated by lipids and is obligatory for parasite survival. A number of drugs, including CQ, potently inhibit the formation of Hz in vitro. Provocatively, in vitro IC 50 for Hz crystallization sometimes (but not always) correlates with antimalarial IC 50 in vivo [14] . However, this does not mean that every quinoline drug acts on the same chemical form of FPIX or at the same pH optimum. The kinetics of Hz inhibition for different drugs against various CQ sensitive (CQS) or CQR strains of malaria are just beginning to be defined in vivo, because crystals form in an irregular shape and are exceedingly difficult to image quantitatively within the live parasite [15] . Different molecular models for how CQ inhibits Hz formation have been proposed, including the lysosomotrophic model wherein changes in DV pH caused by CQ diffusion titrate either Hb digestion, FPIX derivatization, Hz crystallization, or some combination. Alternatively, CQ might inhibit a heme/enzyme complex that forms as an intermediate. However, most evidence suggests that CQ and other anti malarials bind directly to one or more chemical forms (monomers, dimers) of FPIX [5, 6, 14, [16] [17] [18] [19] [20] . These data suggest that quinolines directly inhibit Hz formation by sequestering one or more precursors of the crystal unit, or perhaps by interacting with growing faces of the Hz crystal [17] . It then follows that the CQ resistance mechanism involves the disruption of one or more of these CQ-FPIX interactions. Importantly, since FPIX is made by the host (not the parasite), the target cannot be mutated, and since its release is an inevitable consequence of essential Hb digestion its availability cannot be reduced. Thus two of the most common routes to drug resistance are eliminated (e.g., mutation or decreased expression of target). To become CQR the parasite must therefore either compartmentalize the drug in a different way, alter signal transduction related to death or modify the target in some biochemically unique way. Indeed, Bray et al. extrapolated CQ-binding data and determined that intracellular target affinity differs for CQS versus CQR parasites [9] , so some biochemical transformation of FPIX for CQR versus CQS parasites seems probable. Another study demonstrated this directly in vivo [17] . Recently, we have been the first to determine the atomic-level structures of noncovalent CQ, quinine (QN), future science group amodiaquine (AQ) and quinidine (QD)-µ-oxo dimer FPIX complexes [18, 19] and these structures reveal a number of routes for altering FPIX-drug interactions. Remarkably, we have also discovered that when CQ and FPIX are added together under conditions that mimic that of the DV, a covalent (dative Fe-quinolinal N) FPIX-CQ complex is formed [20] . These discoveries enhance our understanding of how drugs inhibit Hz crystallization, and define important molecular features of CQR malaria. That is, noting the presence of at least two different CQ-FPIX complexes, two different FPIX dimers and the different pH dependencies for complex and dimer formation, we learn that there are several quite simple routes that the parasite can take to titrate multiple drugheme target interactions. Indeed, slightly altered DV pH, volume, ionic or lipid composition for CQR versus CQS parasites [21, 22] can easily alter the ratio of monomeric versus dimeric FPIX, the ratio of µ -oxo versus head to tail dimer, or ratios of multiple drug-heme complexes. Since even closely related quinoline antimalarials have different preferences and affinities for various chemical forms of FPIX, these data suggest that heme target-affinity differences for CQR parasites seen in vivo are easily explained via alterations in physiologic parameters (DV pH, volume, ionic and lipid composition and so on) known to exert enormous influence on Hz and drug-heme chemistry [23] . FPIX monomers and dimers have different affinities for lipids. Lipids and/or monoacyl glycerols found within the DV provide convenient catalysts or scaffolds for the formation of Hz [24, 25] , and different quinoline drugs have different affinities for both monomer and dimer in aqueous solution. Thus, understanding heme monomer-dimer equilibria is essential [14] . De Dios and colleagues have recently explored pH and drug perturbations of the monomer-µ-oxo dimer equilibrium in some detail, and have also measured how this is influenced by drugs and micelles of different charge that, to a first approximation, model various possible lipid phases [26, 27] . In brief, remarkably, CQ is now known to favor binding to FPIX dimers whereas QN clearly favors the FPIX monomer [14] . In free solution, CQ actively promotes the formation of the µ-oxo dimer whereas QN stabilizes the monomer, both in a highly pH-dependent fashion [26] . Addition of a model lipid phase reveals pH-dependent lipid head group dependencies for monomer-dimer equilibria in the presence of quinoline drugs [27] . These are due to different aqueous versus lipid solubilities of monomeric versus dimeric FPIX, as well as different solubilities for these species complexed with quinoline drugs. This partitioning behavior is remarkably drug specific, even for related quinolines such as CQ compared with QN. Thus, even subtle changes in lipids or pH has profound effects on the efficiency of Hz production and has vastly different influences on the ability of even closely related drugs (e.g., CQ vs QN) to inhibit Hz formation. This makes sense since a nondrug-associated form of FPIX successfully partitioned into (or on) a lipid phase is likely to be essential for Hz production at the rate that has recently been directly measured in vivo [15] , and because drugs, pH and lipid composition individually and synergistically perturb FPIX-lipid partitioning. Evidenced by reciprocal patterns of CQ resistance versus QN resistance in some strains (see later), these data further highlight that CQ and QN are really quite different drugs, and that the design of inexpensive second generation quinolines active against CQR malaria should independently target CQ and QN pharmacophore scaffolds. While this progress in elucidating molecular details of quinoline pharmacology is critical, importantly, a recent report suggests that drug-heme interactions are likely not the only way in which the drug exerts a toxic effect -it is only one layer of a more complex collection of effects [28] . In this paper, cytotoxicity of CQ against different stages of iRBC parasites was quantified for both CQS and CQR parasites. Surprisingly, CQ was found to be nearly as toxic to rings and schizonts as to trophozoites. Although recent work by Elliot and colleagues shows that ring parasites do indeed begin degrading some small amounts of Hb, as proposed earlier, schizonts do not metabolize Hb nor produce Hz as trophozoites do [29, 30] . The simple conclusion then is that different forms of free FPIX are not likely to be the only targets of quinoline drugs. Even more surprisingly, this study also demonstrated that P. falciparum CQ-resistance transporter (PfCRT) mutations confer approximately similar levels of CQ resistance to all three iRBC stages [28] . Since a fully formed DV membrane does not exist for ring stage parasites, this suggests that the PfCRT protein (see later section on PfCRT) is likely to have additional locations within the parasite, and more than one physiological function. The next few years of research will yield a more complex, but more complete molecular picture of quinoline antimalarial pharmacology. Although some new pharmacophores are in development, the three major classes of antimalarial drugs are the quinolines (e.g., CQ, mefloquine [MQ] and QN), the antifolates (pyrimethamine and cycloguanil), which both poison nucleotide biosynthesis by limiting essential folate cofactors, and the reactive endoperoxides (artemisin and derivatives) whose action is not well understood. Resistance to the first two classes is profound and widespread, and resistance to the third may be beginning to appear [31] . CQR parasites were first noticed 10-20 years after the introduction of the compound and CQ resistance was well documented in 1959 in Asia. Attempts to eliminate malaria (via insecticides and prophylactic use of CQ) possibly further promoted CQ resistance. CQ resistance is both spreading and continuing to evolve via additional selective pressure. Continued evolution of CQ resistance, artemisin resistance, MQ resistance, QN resistance and so on, continuously produces new genotypes and pharmacologic phenotypes. When resistance to multiple drugs is seen, the strain is often called multidrug resistant (MDR). Molecularly speaking there are many MDR genotypes. Thus, a sometimes overlooked aspect of antimalarial drug discovery is that we are only in the midst of an ongoing phenomenon; what is relevant for one subspecies of CQR P. falciparum is not necessarily relevant for another. It is crucial to understand 'subtypes' of CQ resistance and to develop inexpensive therapy that is effective against multiple subtypes. Several important subtypes can now be described pharmacologically, genetically and (to some extent) in terms of their physiology and biochemistry. As elucidated in several other reviews there are many parallels between drug resistance in tumor cells versus P. falciparum [6, 8, 23] . A key role for ATP-binding cassette (ABC) proteins was emphasized early on, however, this cannot explain all phenomena [8, 23, 32, 33] . Many reports underscore an important role for ion transport in drug resistance phenomena [34] [35] [36] [37] and it is now generally well accepted that dysregulation of ion transport is likely to be central to multidrug resistance in tumor cells, malarial parasites and some bacteria. This suggests that altered drug accumulation in MDR cells may be due, at least in part, to consequential alterations in pH gradients, volume and/or other biophysical parameters that control diffusion or accumulation of drugs; that signal transduction regulated by ion transport is a key player in resistance (this is certainly now clearly recognized for apoptotic signal transduction vs tumor multidrug resistance); and that the chemistry of drug-target interactions is manipulated via ion transport such that drug activity is then altered (e.g., the special case of highly pH-dependent FPIX-quinoline interactions described earlier). Biochemical and physiological features shared among MDR tumor cells, malarial parasites and certain Gram-negative bacteria include altered intracellular accumulation of the drugs to which cells or micro-organisms are resistant, and that certain 'chemo-modulators' (e.g., the calcium channel blocker, verapamil [VPL]) can partially reverse the drug resistance. Although it is still not precisely understood how VPL functions as a chemoreversal agent, it is reasonable to expect that concepts from the study of MDR tumor cells will apply to MDR malarial parasites, and vice versa. Such was the hope when pfmdr genes were first cloned [38, 39] and shown to be homologous to the hsmdr1 gene, which is overexpressed in MDR tumor cells. It was initially thought that an ABCprotein drug pump for CQ must therefore exist in drug-resistant P. falciparum, similar to the drug pump proposed for tumor cells (human multidrug resistance protein 1 [HsMDR1], also known as Pgp) believed by many investigators to directly translocate vinblastine and many other anti-tumor drugs. However, subsequent work showed that some drug-resistant P. falciparum strains do not encode mutated or overexpressed pfmdr [40, 41] , and that other genetic events must therefore be important. In parallel once again to tumor drug resistance, at the same time an increasing importance of the overexpression of other genes (e.g., ion transporter and mutant pro-and antiapoptotic genes) was recognized for MDR tumor cells [42, 43] . Similar to P. falciparum multidrug resistance (PfMDR) protein versus malarial multidrug resistance (see later), the precise role of HsMDR1 protein in tumor multidrug resistance has been questioned for some time as pathology data have not correlated HsMDR1 overexpression with clinically relevant tumor multidrug resistance as strongly as was initially suspected [36, 43, 44] . It is now clear that there are multiple molecular layers to multidrug resistance pathways and that these continue to evolve. Elucidating the molecular function of proteins that appear to be membrane transporters and that are mutated in CQR malarial parasites (which are often also MDR) remains a critical area of study and is briefly summarized in the last section of this review. future science group Initial genetic studies: cg2 versus Na + /H + exchange; the wrong gene but the correct physiology A genetic definition of CQ resistance began with the cloning of pfcg2 [45] , a technical tour de force essentially started years earlier via the successful creation of CQS × CQR cross progeny [46] . Correlation between specific pfcg2 sequence and CQS or CQR status was found for all but one strain of P. falciparum examined. The one exception, strain Sudan 106/1, turned out to have a special utility in characterizing the true CQ resistance determinant (see later section on PfCRT). Prior to that work, Lanzer et al. concluded that mutated PfCG2 protein was a dysregulated Na + /H + exchanger (NHE) that also pumped CQ out of the cell [47, 48] . Wellems et al. questioned this because subcellular localization revealed PfCG2 in vesicle-like structures near the parasitophorous vacuolar space and the DV [49, 50] , not within the plasma membrane as envisioned by Lanzer et al [47] . Follow-up studies by Bray and Ward did not measure any Na + dependency for CQ resistance, arguing against a strong role for NHE in influencing CQ transport or any other feature of CQ resistance [51] . Most recently, quantitative trait loci (QTL) analysis, availability of the P. falciparum genome and novel single-cell imaging in an extensive series of drug-resistant progeny suggested that altered NHE and cytosolic pH (pH cyt ) are indeed related to resistance, but that the relevant ion exchanger is not PfCG2 (see later section on P. falciparum Na + /H + exchanger [PfNHE]), and that resistance to QN is more influenced by these changes than CQ resistance [52] . Wellems and colleagues subsequently proved, via direct transfection experiments, that mutant PfCG2 did not confer CQ resistance [53] . Simultaneously, the Roepe group measured alkaline pH cyt for some but not all CQR parasites, and early NHE data by that group [54] led to conclusions similar to those of Bray and Ward [51] . Thus, attention then focused on another gene found within the same 36 kbp fragment that harbored pfcg2, namely, pfcrt [55] . Results described in this and additional papers show that PfCRT protein is the ultimate determinant of CQ resistance in P. falciparum malaria, that it resides within the DV membrane, and is not directly involved in pH cyt regulation or plasma membrane NHE (see later section on PfCRT) [56, 57] . Along with the now recognized dominant role of PfCRT, P. falciparum MDR protein 1 (PfMDR1) is now thought to only modulate cross-resistance patterns in CQR parasites (see later). Meaning, this transporter modifies the rank order of resistance to other drugs such as MQ [58, 59] . Other proteins, including PfNHE1, which is a true NHE, are now suspected to complement drug resistance even further [52, 60] . This degree of genetic complexity involving at least three membrane transporter genes is daunting, but is to be expected. Early explanations for MDR phenomena in either tumor cells, bacteria or parasites were (in hindsight) a bit too optimistically reductionist. More genetics; the elusive role of PfMDR1 As mentioned, early studies of CQ resistance in P. falciparum showed resistance was associated with decreased drug accumulation that was reversed by the ion-channel blocker VPL [33, 61] . Similar phenomena had been seen in MDR tumor cells. Thus, Wirth and colleagues screened P. falciparum for hsmdr1 homologs and identified pfmdr1 and pfmdr2 [38] . While both encode proteins expressed in drug-sensitive P. falciparum, a MQR strain showed elevated pfmdr1 [54] . MQ is chemically similar to CQ, so at the time resistance pathways were expected to overlap. In support of this, another group found pfmdr1 to be upregulated in some CQR P. falciparum [39] . However, subsequent experiments showed that pfmdr1 overexpression did not correlate with CQ resistance [62] . This was not surprising since Wellems et al. had earlier shown that CQ resistance did not segregate with the pfmdr1 chromosome 5 locus in progeny from a CQS × CQR genetic cross [46] . Proposals arguing against a dominant role for HsMDR1 protein in tumor resistance were also being made during this period, as ion transporters were found to be alternatively expressed in some MDR tumor cells well before overexpression of HsMDR1 [42] . On the other hand, polymorphisms in pfmdr1 were also associated with CQ resistance early on [63] . While CQS isolates had identical PfMDR1 sequences, there were five changes in CQR isolates. In strains K1 and ITG2, N86Y was the only change. CQR strain 7G8 had four changes; Y184F, S1034C, N1042D and D1246Y [63] . The Y184F mutation was postulated as not likely to be involved in CQ resistance since it was also found in CQS strains. Thus, the pfmdr1 overexpression hypothesis was revised to suggest that CQR strains expressed mutant PfMDR1 but did not necessarily overexpress wild-type [63] . (Which pfmdr alleles are considered 'mutant' versus 'wild-type' is open to interpretation, wild-type is used here to refer to the allele found in the CQS strain HB3.) Interestingly, CQR K1 versus 7G8 polymorphisms appear to be geographically biased [64] [65] [66] . future science group Subsequently, when MQR P. falciparum were selected to higher levels of MQ resistance, pfmdr1 was found to be amplified [67] . Halofantrine resistance and QN resistance increased with increasing pfmdr1 whereas AQ resistance did not [67] . However, when CQR strain K1 was selected against halofantrine it did not result in MQ resistance or amplification of pfmdr1 [68] . In a more recent study, which used allelic exchange of pfmdr to probe these questions, incorporation of pfmdr1 7G8 polymorphisms into a CQS strain not previously exposed to a drug had no effect on CQ resistance, but incorporating wild-type pfmdr1 into a CQR strain expressing mutant pfmdr1 did decrease the level of resistance by half [58] . CQS strains expressing mutant pfmdr1 alleles showed some mild QN resistance and altered sensitivity to MQ [58] . Variations on this theme have also been described by Fidock and colleagues [59] . Collectively, these data suggest that PfMDR1 effects are subtle and strain-specific. This brings us to our current understanding: it seems unlikely that mutations in pfmdr1 confer CQ resistance in and of themselves [69] , but they (or overexpression of certain isoforms) provide an important modulatory effect [70] . As previously mentioned, early on, Wellems et al. showed that pfmdr1 was unlikely to directly cause CQ resistance since the relevant region of chromosome 5 did not segregate with the CQ resistance phenotype in genetic cross progeny [46] . A follow-up paper suggested that CQ resistance segregated with the pfcg2 gene on chromosome 7 [45] , but this paper also showed that one CQS strain (Sudan 106/1) carried CQ resistance-associated pfcg2 yet was nonetheless CQS. The 36 kbp chromosome 7 locus harboring pfcg2 that segregated with CQ resistance was thus re-examined and a previously unrecognized gene, now known as pfcrt was found [55] . Mutations in pfcrt are the central determinant of P. falciparum CQ resistance. The 13 exons span 3.1 kbp and encode a 424 amino acid, 48.6 kDa protein. Mutant pfcrt alleles found in CQR parasites contain a number of point mutations that confer amino-acid substitutions, with the pattern depending on the region of the globe from which the CQR parasite originates. In general CQR parasites from southeast Asia and Africa carry seven to eight point mutations, whereas South American CQR strains carry five [55] . Novel patterns continue to be discovered [65] and now at least 12 distinct CQ resistance-associated PfCRT isoforms have been described. The pattern of mutations provides identification of the probable geographic origin of a CQR isolate. The number of mutations apparently required for CQ resistance explains two riddles, namely, why CQ resistance took so long to appear on a large scale and why it had been impossible to create CQR strains from CQS strains in the laboratory via drug selection. However, if one begins with Sudan 106/1, a strain that harbors all but one of the mutations required to complete a CQR-pfcrt allele, CQR strains can be rapidly created in the laboratory by CQ selection [57] , and the final mutation required to complete the CQR-pfcrt allele [55] is found in these selected strains. In performing this experiment, Cooper et al. also found PfCRT substitutions that are not known to exist in the wild but that confer unusual and scientifically informative drug-resistance profiles [57] . Similar to PfMDR1, PfCRT is localized to the DV membrane and is a polytopic, integral membrane protein that is likely to perform some type of transport function [21, 22, 55, [71] [72] [73] [74] [75] [76] . Similar to the case for HsMDR1 most initial hypotheses for its function suggest either ion or drug transport or both, since, again, CQR parasites accumulate less antimalarial drug in a given time relative to CQS and quinolinal antimalarial drugs (like anti-tumor drugs), which are hydrophobic weak bases. In fact, CQ and related drugs are dibasic and the DV is now known to be quite acidic. So, fold concentration of CQ within the DV (where the heme-drug target is found) is dependent upon the square of the net pH gradient and would be as much as 10 5 -10 6 -fold by the predictions of weak base partitioning theory, depending on relative permeabilities of neutral, +1 and +2 CQ species [6]. Thus, even very subtle changes in DV pH would have very significant consequences. Progress on these and related questions is summarized later. As mentioned previously, one additional genetic event that tailors CQ resistance caused by PfCRT, may be the mutation and/or over expression of PfMDR1, but this cannot fully explain all MDR phenotypes. Another contributing event, recently identified by QTL analysis [60] , involves one or more genes encoded by a segment of chromosome 13. This fragment is hypothesized to contain genes encoding for proteins with unclear homology to known proteins, however, it also encodes homologs to a V -type ATPase subunit and to the NHE protein family. By combining future science group QTL analysis with improved pH cyt measurements, we have recently shown that Pf NHE dysregulation is likely to be linked to one route to increased QN resistance [52] . In addition, additive QTLs on chromosomes 5 and 7 were found as expected (the identified fragments contain pfcrt and pfmdr1) [60] . Pairwise effects were also detected between chromosome 13 and a chromosome 9 locus. The NHE homolog encoded within chromosome 13 was named PfNHE1. This protein is the second largest eukaryotic NHE yet identified (surpassed only by a NHE for the related apicomplexan Toxoplasma gondii), and has several unusual features. Bioinformatic ana lysis and cross reactivity with anti-TgNHE antibody [Chen D, Pleeter P, Roepe PD, Unpublished Data] show that PfNHE is localized to the plasma membrane. Polymorphisms that encode variable DNNND repeats in the predicted PfNHE protein sequence have been found in progeny of a CQS × CQR cross, as well as a range of field isolates and additional laboratory strains showing variable QN resistance [60] . Prior to the availability of the P. falciparum genome, Ginsburg et al. quite logically suggested a plasma-membrane NHE must exist to decrease cytosolic acid caused by anaerobic glycolysis [77] . As mentioned earlier, Lanzer et al. suggested that pfcg2 found close to pfcrt on chromosome 7 encoded this NHE [48] , but this was disputed, since PfCG2 is actually peri-vacuolar [50] , not plasma membrane localized, and because the putative PfCG2-NHE homology was based on sequence ana lysis that did not account for very high AT content in malarial genes [49] . Regardless, the field has come full circle and the initial conclusions by Lanzer et al. were indeed partly correct (the physiology was correct, but not the genetic explanation). In our hands single-cell photometry (SCP) analysis of pH cyt for intra-erythrocytic parasites under continuous physiologic perfusion indeed show elevated pH cyt for some CQR parasites, but not all [54, 55, 57] . A corollary we suggested is that the relative size of the net cytosolic-DV pH gradient might be a more important parameter for CQ resistance versus QN resistance, rather than steady-state pH cyt or digestive vacuolar pH values alone [54] . Overall, a range of pH phenomena and genetic changes consistent with changes in pH regulation are associated with QN resistance and CQ resistance. As described [52] , we recently optimized localization of the pH probe BCECF exclusively to the parasite cytosol to avoid complexities in interpretation from earlier studies, and showed that elevated pH cyt is well correlated with QN resistance and increased apparent PfNHE activity [60] . We now believe there are at least two physiological signatures for QN resistance that segregate with the two genetic descriptions, and that one is alkaline pH cyt [52] . Technical details of PfNHE measurements have recently been questioned [78] , but these issues have hopefully been clarified [79] . Genetic definition of CQ resistance and related phenomena is a major breakthrough, but to develop drugs and other therapies the altered biochemistry and physiology linked to that genetics must also be elucidated. As described in the first part of this review a chief drug target is FPIX in the DV, so studies of DV biochemistry are particularly important. A central characteristic of DV biochemistry and physiology is the high pH gradient (acid inside) that the organelle has. Regulation of DV pH is not fully understood, but it includes a V -type H + ATPase that hydrolyzes cytosolic ATP to pump H + into the DV [80] . Interestingly, although conflicting data have been published [81] it is now generally accepted that changes in DV pH and volume are linked to CQ resistance caused by PfCRT [21, 22, 74, 76] . These are further predicted to affect drug, lipid, metabolite and osmolyte traffic in and out of the DV. However, since the DV is a subcellular organelle for an intracellular parasite, precise quantification of these perturbations is quite challenging. It requires the development of novel, overlapping approaches, as described later. Our group attempted the first DV pH measurements for living intra-erythrocytic parasites under physiologic perfusion using the pH probe acridine orange (AO) and novel SCP [71, 82] . Our initial hypothesis was that, owing to the dibasic nature of CQ, even subtle increases in DV pH would significantly lower DV concentrations of CQ and thus cause CQ resistance. Therefore, relative to CQS, CQR parasites might show slightly elevated DV pH or different DV pH behavior upon addition of CQ [83] , or perhaps both. We and others have actually found that mutant PfCRT in CQR parasites causes more acidic (lower) DV pH. The initial AO data in support of this conclusion generated controversy (as most new technologies tend to do), but a number of different, complementary methods from several labs were subsequently developed and strongly supported the initial AO conclusions [21, 22, 74, 76, 79] . Nonetheless, lower DV pH for CQR parasites appeared paradoxical, because simplistically it is predicted to concentrate more drug within the DV by weak base effects. It is difficult to future science group see how concentrating more drug at the site of action works to confer drug resistance, but as pointed out [6, 82] in hindsight the physiology is obvious when examined alongside the chemistry of the principle CQ target, FPIX released from Hb. As mentioned, iRBC parasites detoxify FPIX by crystallization to Hz. In aqueous solution, FPIX dimerizes to a µ-oxo dimer. FPIX has two propionic acid side chains, thus µ-oxo FPIX is a tetraprotic amphipathic weak acid with four equivalent titratable groups, and a pKa near DV pH [6, 82] . As pH is dropped even subtly (0.1-0.3 units), because the titration curve of FPIX dimer with four equivalent pKa is exceedingly steep (shown in [82] ), even low levels of soluble dimer convert to insoluble (aggregated) dimer over a very narrow pH range (<0.5 units [82] ). CQ and other drugs bind well to soluble FPIX but not to aggregates. In addition, acid aggregation of FPIX accelerates conversion to Hz (to which drugs also do not bind well). For these reasons, lowering DV pH by as little as 0.1 units is actually predicted to be a potent pathway to CQ resistance [6] . Ingeniously, CQR parasites titrate one drug target to lower levels, and also bias pHdependent FPIX chemistry. That is, even though at DV pH FPIX monomer is likely to be more abundant than µ-oxo dimer [26] , FPIX equilibria are pulled away from monomer by acid aggregation phenomena. As explained in the first section of this review, it is now known that this simple picture is a bit more complex, and also involves quinoline drug-specific effects (e.g., CQ vs QN) on monomer-dimer equilibria and drug-FPIX aqueous versus lipid partitioning [26, 27] . The overall point is that ion dependent CQ resistance DV biochemistry drives a number of chemical conversions that will act to disrupt quinoline drug-FPIX interactions in very potent ways. However, importantly, lower DV pH caused by CQ resistance-associated mutant PfCRT that correlates with a VPL reversible CQR phenotype can only be one aspect of the CQ resistance mechanism and cannot explain all quinoline drug resistance. For example, evidence suggests lower DV pH may not be directly related to QN resistance, and might even be related to MQ hypersensitivity in some cases [56, 57] . As described earlier, added effects of proteins encoded by other identified loci (PfMDR, PfNHE) may hold clues to this complex spectrum of multidrug resistance phenomena. In addition, better definition of the molecular origins and repercussions of this altered pH is required. How does it occur? Is PfCRT a H + -coupled metabolite transporter that when mutated to a CQR form becomes partially uncoupled? Does PfCRT interact with the DV H + ATPase? Does PfCRT transport a counterion (i.e., Cl -) to shunt membrane potential in the presence of a high change in pH, such that CQ resistance mutations alter anion flux to indirectly cause a greater pH change? Do PfCRT CQ resistance mutations change substrate specificities for a facilitative diffusion transporter? These questions all have very different predicted consequences. Some cell-based experiments have attempted to address these questions. Lanzer and colleagues successfully expressed PfCRT in oocytes and measured pH, membrane potential and certain ion conductances across the oocyte membrane [83] . They noted that oocytes expressing PfCRT exhibited an altered transmembrane pH gradient and membrane potential due to H + leak and somewhat nonspecific cation conductance. They proposed that PfCRT activates endogenous oocyte ion transporters in some way. Multiple molecular models that explain these observations are possible, but overall the data further highlight a role for PfCRT in ion and/or osmolyte traffic. Another study followed ion dependencies for DV pH and volume regulation by imaging these parameters for live parasites under perfusion with a medium of altered salt composition [22] . Importantly, CQR parasites showed increased DV volume relative to CQS parasites, suggesting that pH and volume regulation are linked for the organelle, as is found for other acidic vesicles or lysosomes. However, fast transient changes in Clgradients across the DV membrane did not lead to rapid changes in the DV transmembrane pH gradient, indicating no direct coupling of Cland H + transport. On the other hand, fast transient changes in DV Clgradients were found to strongly influence DV volume. These effects were strongly CQ-and VPL-dependent and differed dramatically for CQS versus CQR parasites. The overall conclusion was that PfCRT mediates transport of an important DV osmolyte (presumably peptides, di-peptides or amino acids released from Hb digestion) and that this transport is altered for CQR parasites. Taken together, the bulk of the evidence suggests that altered DV pH for CQR parasites is an indirect consequence of altered osmolyte traffic promoted by PfCRT mutation. This might be an explanation for why parasites treated in different ways (resulting in various levels of relevant osmolytes produced by a finely tuned metabolism) could perhaps show different DV pH in some studies [81, 84] . If altered osmolyte traffic that then indirectly perturbs normal pH regulation is the explanation for altered DV physiology future science group linked to PfCRT mutations, it begs the obvious question: what is the normal physiological function of PfCRT? An obvious attractive possibility suggested early on is that PfCRT might transport products of Hb digestion (peptides, dipeptides and/or free amino acids) [73, 22] , since these are among the most important DV osmolytes, since products of Hb digestion are unique osmolytes and PfCRT is a unique transporter with a unique sequence, and because when CQ, QN and other drugs are superimposed upon amino acid structures (e.g., CQ vs the cationic amino acid lysine), many interesting similarities can be noted. For example, the superposition (overlay) of a dipeptide N-terminal lysine (a common Hb amino acid) and CQ shows similar charge and a similar carbon chain flanked by nitrogen atoms. As another example, an OH group is bound to one chiral center of QN that is two s bonds away from the positively charged quinuclidine nitrogen, and a near identical spatial arrangement of stereochemically sensitive atoms is also found for one isomer of serine. Interestingly, QN and QD differ in their stereochemistry at this center yet show different PfCRT-isoform dependent pharma cology [57] . Perhaps not coincidentally then, if PfCRT did transport peptides or amino acids, substrate recognition would also be stereo-selective since only l-amino acids are found in Hb. Ultimately, genetics and cell physiology can only take us so far, and cannot fully resolve speculation regarding substrates of PfCRT or define the thermodynamics and kinetics of any proteinmediated transport. Elucidating the remaining questions will require detailed molecular studies with PfCRT, PfMDR1, PfNHE and perhaps other proteins yet to be described. However, the P. falciparum genome is anomalously AT rich, and some genes (including pfcrt) can have significant stretches of AT content that are 80% or more. These do not translate well in convenient heterologous systems such as bacteria or yeast. In fact, in the case of pfcrt and pfmdr1 they do not translate at all. That is unfortunate, since techniques and strategies for defining transporter function have been elegantly laid out in bacteria and yeast model systems for the past four decades. Proteoliposomes, kabackosomes, Goffeau membranes, and Menendez phase separated yeast vesicle preparations would all prove incredibly useful for defining the molecular mechanism of CQ resistance, if only pfcrt cDNA could be expressed in either bacteria or yeast. Hanbang Zhang working the Roepe laboratory succeeded in a brute force resolution to this dilemma. Noting the success of earlier gene design that allowed MSP-1 expression in Escherichia coli [85] , Zhang back-translated the PfCRT amino-acid sequence using preferred yeast codons, then designed a set of overlapping 40-mer oligonucleotides that encoded both strands of the theoretically optimized sequence, and constructed an entirely synthetic gene via nested PCR methods. This enabled the expression of high levels of PfCRT protein in either Saccharomyces cerevisiae or Pichia pastoris, in either a constituitive or inducible fashion, respectively [72] . Analysis of ion transport for plasma membrane inside-out vesicles produced from these strains provided further evidence that PfCRT protein plays a role in ion transport [72] , and purified membranes along with equilibrium binding studies, using 3H-CQ, provided the first direct evidence that PfCRT protein indeed binds CQ [73] . The latter is a central prediction for nearly all molecular hypotheses for PfCRT. Although not often cited, the same paper also provided the first direct molecular evidence that PfCRT likely transports CQ, via flow dialysis experiments using inside-out yeast plasma membrane vesicle with or without PfCRT protein [73] . The simplest model consistent with these results is that CQ transport by PfCRT is facilitative downhill diffusion at a rather low turnover. Different thermodynamic models for CQ transport (e.g., PfCRT as a CQ exchanger or active transporter) have also subsequently been proposed based on different approaches that measure 3H-CQ flux for live parasites [86, 87] . Distinction between these models will require additional direct transport experiments at much higher kinetic resolution. At the time, pfcrt was (to our knowledge) the largest synthetic gene ever constructed by overlapping PCR, nonetheless, success enticed others to attempt the reconstruction of a yeastoptimized pfmdr1 gene, which is over three times the size of pfcrt. After numerous challenges were overcome, this too was achieved and unusual ATPase activity of PfMDR1 was rather quickly quantified via plate-based phosphate release assays [88] . Relative to other ABC transporters involved in drug-resistance phenomena, PfMDR1 has unusually high K m and V max for ATP hydrolysis. Also, interestingly, antimalarial drugs only mildly stimulated PfMDR1 ATPase activity, consistent with the notion that PfMDR1 exerts only mild affects on antimalarial drug resistance. With methods for in vitro future science group analysis of PfMDR1 in hand, a follow-up study quickly defined which amino acid substitutions in the curious 7G8 PfMDR1 isoform confer its unusual and quite specific ATPase activity [89] . The S1034C mutation in 7G8 PfMDR1 is now known to abolish antimalarial drug-stimulated PfMDR1 ATPase activity. Although this molecular-level work is only just beginning, it has already significantly refined proposals for the molecular mechanisms of CQ resistance and antimalarial multidrug resistance. Direct demonstration of CQ binding to PfCRT was a very important result, but equilibrium binding studies with radio-labeled CQ obviously do not allow for definition of the drug binding site. To answer this question, the Roepe and Wolf laboratories designed and synthesized various CQ photoaffinity probes. One particularly novel probe, haboring both per-fluoro-azido and biotin tags via flexible linkers attached to the aliphatic terminus of CQ, proved to be enormously useful. It has recently led to the identification of the CQ binding site in HB3 isoform PfCRT, and has also quickly quantified relative affinities of other drugs (QN, VPL and so on) versus various PfCRT isoforms [90] . An obvious extension of this work would be the definition of CQ binding-site differences (if any) for other PfCRT isoforms, and perhaps similar studies with PfMDR1. The chemistry developed for construction of these photoaffinity probes leads to a rather straightforward design and synthesis of fluorescent CQ derivatives. Some dansyl-CQ derivatives actually turn out to be more active against CQR strains of P. falciparum than CQ, and are useful for imaging subcellular localization of CQ [91] . A more convenient probe that can be followed with more routinely available lasers and fluorescent microscopes, NBD-CQ, has also recently been developed [92] . Combined with vesicles and proteoliposomes as described for drug binding studies [73, 90] this probe is currently allowing us to define putative CQ transport by PfCRT in very detailed molecular and thermodynamic terms. In principle, genetics, physiology and biochemistry related to altered FPIX chemistry and membrane transport now defines CQ resistance. It is my suspicion that permutations of the above will ultimately be shown to define resistance to all other quinolines (QN, AQ, MQ and so on), to acridines, xanthones and reactive endoperoxides, but probably not to antifolates, for which resistance involves other distinct pathways [93] . That is, at least three genes (pfcrt, pfmdr1 and pfnhe) have been identified that, in various allelic combinations, confer what we now understand to be a spectrum of antimalarial multidrug resistance phenomena. The molecular function of at least two of these is becoming clearer via a combination of new chemical biology, artificial gene construction, and tried and trusted membrane biochemistry techniques. The physiologic and pharmacologic consequences of their function is at least qualitatively defined, and although there is still much more to do at the molecular level, the field has come a very long way in the 9 years since the identification of PfCRT. We can be hopeful that, based on this information, new quinolines effective against CQR malaria will prove increasingly easier to design and synthesize. In fact, the author suggests that this hope is even now being realized [91, 94, 95] . This work complements continued uses and combination therapies involving other quinoline derivatives such as pyronaridine (an aza-acridine) [96] , piperaquine [97] and isoquine [98] as well as other heme-binding pharmacophores such as the very exciting xanthones pioneered by Riscoe and colleagues [99] . History repeatedly teaches us that even with the emergence of CQ resistance, inexpensive quinolines and related compounds can still be effective. In summary, the PfCRT protein binds CQ as is predicted from several models for its function. Azido-biotin-CQ now clearly defines that binding site and distinguishes drug binding preferences for various PfCRT isoforms [90] . CQ resistance-associated mutant PfCRT shows altered CQ binding, promotes altered diffusion of CQ and perturbs osmolyte equilibrium within the DV, which then alters DV pH and volume regulation. The latter influences FPIX chemistry and multiple FPIX-drug interactions in critical ways. Taken together, a more complex, but clearer molecular model for the mechanism of trophozoite CQ resistance is now possible [Paguio M, Cabrera M, Roepe PD, Submitted]. However, importantly, we should not neglect that trophozoite/DV-specific models for CQ resistance are clearly only one part of the story. Other effects of CQ and other layers of CQ resistance need to be elucidated [28] . Other aspects of CQ pharmacology must be present at the ring and schizont stages of P. falciparum development. Additional physiological functions for PfCRT at the ring and schizont stages must also exist and when altered by mutation also cause CQ resistance for these stages. A key question future science group is how are these physiologic functions related to putative ion or osmolyte transport? For ring stages, based on the recent work of Elliot et al. it seems possible, at least in theory, that controlling Hz chemistry is again, at least, part of the explanation [29] . But for schizonts, the answer must be different. Perhaps mutant PfMDR1 and PfNHE further modulate drug resistance in the presence of mutant PfCRT for these stages as well. Their molecular level function can now also be studied using 'yeast-optimized' recombinant proteins and novel drug probes as described for PfCRT, so additional rapid progress is on the horizon. It is true that great strides have been made in understanding antimalarial drug resistance at a molecular level, and it is also true that this information is now turning out to be of great value in developing new, inexpensive, secondline antimalarials. 'Molecular medicine', well developed for western cancer and cardiology clinics, has (in theory) finally come to the malaria clinic -genes that dictate how the disease might respond to therapy have now been identified, and additional molecular tools developed with that knowledge are leading to new clinical successes. The function of proteins encoded by these genes is becoming understood. High-throughput screening methods such as the SybrGreen assay have led to fast, inexpensive ways to screen for novel antimalarials active against CQR malaria, at least at the pre-clinical stage. Multiple plasmodial genomes have become available, giving us unique insight into comparative genomics and many new drug targets to explore. New anti malarial drug leads are indeed rapidly appearing. Yet, the expense and often difficult logistics of Phase II and III drug trials still prevent full implementation of these advances. The entire world yearns for vaccine development, but optimistically that is still many years in the future. In the meantime, inexpensive drugs are still desperately needed. In 5-10 years, we hope that additional emphasis will have been placed by the entire global community on the rapid development and deployment of inexpensive anti malarial drugs. More funding and effort directed towards a multi-tiered, balanced approach to malaria treatment and prevention is very much needed. For over 60 years, inexpensive anti malarial drugs have saved the lives of many millions of malaria victims. In theory, there is no scientific roadblock that cannot be overcome such that this continues to be the case well into the future. It has proven quite difficult to either increase or decrease expression of Pf NHE protein to further test how Pf NHE may be linked to QN resistance, but a new paper reports decreased PfNHE levels for three parasite strains (GC03, 1 BB5, 3 BA6) by stable transfection with a truncated 3´-untranslated region [100] . The transfectants from these experiments do not show statistically significant changes in resting (steady state) pH cyt as would be expected based on previous work [52] , but they do show small increases in QN sensitivity consistent with Pf NHE protein being involved in QN resistance as proposed [60, 52] . Unfortunately, in [100] only strains with low levels of QN resistance were amenable to transfection, whereas decreased Pf NHE expression for strains with Executive summary n Over the past 9 years, a clear picture of the genetic alterations that accompany evolution of chloroquine (CQ) resistance in Plasmodium falciparum malaria has become available. P. falciparum CQ-resistance transporter (PfCRT) mutations must be present and these confer a tenfold decrease in CQ-mediated growth inhibition and toxicity. Mutations and/or elevated expression of P. falciparum multidrug-resistant protein 1 (PfMDR1), perhaps along with mutations in P. falciparum Na + /H + exchanger, further tailor quinoline drug resistance, and perhaps resistance to other classes of compounds that also target the parasite digestive vacuole. n To understand the physiological repercussions of these genetic alterations, additional cell culture and computerized microscopy methods have been developed. These are now rapidly providing new information on a number of key cell biological questions, including organellar biogenesis, CQS versus CQR parasite fitness, and metabolism. n To uncover the molecular details behind how mutant proteins link to resistance phenomena function, heterologous expression, purification and reconstitution of at least two of them (PfCRT and PfMDR1) has recently been accomplished. Combined with new 'chemical biology' approaches, the field is now poised to fully describe the molecular mechanism(s) of CQ resistance. n Taken together, these advances and others are accelerating development of new second-line antimalarial drugs active against CQR malaria. future science group high levels of QN resistance would have been particularly informative. Also, the authors of [100] were unable to consistently clamp parasite pH cyt to acid values in the absence of Na + , formally leaving open the question of whether decreased levels of Pf NHE do indeed (as obviously expected) lead to a lower rate of Na + /H + exchange. More work is needed with these and other transfectants, and better methods for analyzing Na + /H + exchange for the intracellular parasite would clearly be helpful. Regardless, it is essential to recognize that effects on H + transport and steady state pH due to lowering Pf NHE expression [100] are certainly not expected to be analogous to the effects predicted from analysis of multiple pairwise loci influences on PfNHE activity [52] . Clearly, levels of QN resistance span a wide range, and Pf NHE effects on QN resistance are complex, require multiple loci [52] and, as reported earlier, do not necessarily segregate exclusively with elevated pH cyt [52] . Overall, as emphasized throughout this review, we now realize that multiple genetic alterations confer the molecular and physiologic basis of quinoline resistance patterns, and so it is to be expected that resistance phenotypes will involve multiple overlapping pathways.

Interaction Between Humans and Poultry, Rural Cambodia

Because avian influenza H5N1 infection risks are associated with exposure to infected poultry, we conducted a knowledge, attitudes, and practices survey of poultry-handling behavior among villagers in rural Cambodia. Despite widespread knowledge of avian influenza and personal protection measures, most rural Cambodians still have a high level of at-risk poultry handling.

Because avian influenza H5N1 infection risks are associated with exposure to infected poultry, we conducted a knowledge, attitudes, and practices survey of poultryhandling behavior among villagers in rural Cambodia. Despite widespread knowledge of avian influenza and personal protection measures, most rural Cambodians still have a high level of at-risk poultry handling. T he circulation of the highly pathogenic H5N1 avian influenza (AI) strain throughout Asia since late 2003 (1) , and more recently in Europe and Africa, has resulted in considerable concern for the potential of a new pandemic. In Cambodia, outbreaks of HPAI A/H5N1 infection were first reported in poultry in early 2004 (2) . Since 2005, 6 human cases have occurred (100% fatal); the 2 most recent cases occurred in early 2006 (3, 4) . Most Cambodians live in rural areas and raise animals for consumption (2) , typically keeping poultry, swine, or cattle close to the home. Because H5N1 infection has been associated with exposure to infected poultry (5-10) and little is understood of the perceptions of rural farmers regarding AI (11), we conducted a knowledge, attitude, and practices survey of poultry handling in rural Cambodia to estimate the extent of interactions between humans and poultry, to understand practices in poultry handling among villagers, and to develop interventions designed to increase reports of poultry deaths and safe poultry handling. We conducted a 2-stage household based cluster survey (12) with a goal of 500 participants: 20 persons >15 years of age in each of 25 villages from Prey Veng and Kampong Cham Provinces. The sampling frame of eligible villages within these provinces were those located in H5N1 high-risk communes, as defined by the Food and Agriculture Organization of the United Nations training program for village animal health workers. The villages were selected with probability proportional to size. For the second stage, we randomly selected the first household within each village. Subsequently, households were selected by proximity until 20 eligible participants were enrolled in each cluster. Verbal consent was obtained from all participants. All were interviewed by using a structured questionnaire designed to collect information on demographics, basic hygiene practices, quantity of poultry owned, poultry death reporting, practices when deaths occurred, knowledge and attitude of sick and dead poultry, and knowledge of AI. Twenty-three villages were included in Kampong Cham (11) and Prey Veng (12) Provinces ( Figure 1 ). Four hundred sixty respondents from 269 households completed the questionnaire. Most were women (60%), farmers (88%), and persons who had completed less than primary schooling (57%). The median number of household members was 5 (range 1-16), and 77% of all households included children <15 years of age. Many households owned chickens (97%) and ducks (39%) (Figure 2 ), although the size of most poultry flocks was small (Table) . Almost all poultry were free ranging (100% of chicken flocks; 96% of duck flocks), and mixing of the poultry with pigs and other domestic animals was common. Respondents reported that they use poultry feces for manure (77%), touch sick/dead poultry with bare hands (75%), eat poultry that died from illness (45%), eat wild birds (33%), let children touch sick/dead poultry with bare hands (20%), and gather dead wild birds for consumption (8%). During the previous 6 months, of the 260 households that owned poultry, 162 (62%) experienced poultry deaths; however, only 18 (7%) reported these deaths to local authorities. Half of the respondents (n = 231) believed that it was important to report any poultry deaths because the death may be due to AI (61%) or because the poultry owners may receive management advice from the village veterinarians (39%). Among these 231 respondents, many did not report poultry deaths because they did not know how (41%), were in the habit of not reporting poultry deaths (31%), believed they would have a problem selling poultry if they reported deaths (18%), did not know the risks of AI (7%), or feared poultry culling (5%). Among those respondents who did not believe reporting deaths was important, the reasons provided included the following: "the number of poultry deaths were too few" (62%), "poultry are not as important as cattle" (18%), "no help would be provided from veterinary staff or authorities" (13%), or "because mortality was similar to previous years" (7%). Of respondents that experienced poultry deaths, 62% buried or burned dead poultry, 53% prepared them for food, 22% threw away the dead poultry, 3% used them to feed other animals, and 2% prepared them for sale or gave them to their neighbors. Participants had learned about AI from television (81%) and radio (78%). Thirty-one percent of respondents were able to describe AI symptoms in humans, and 72% believed that AI is a fatal disease among poultry that can be transmitted to humans. Most respondents believed it is unsafe to touch sick or dead poultry with bare hands (67%), eat wild birds (70%), let children touch sick or dead birds with bare hands (83%), and eat meat or eggs that are not fully cooked (86%). Sixty-one percent of respondents mentioned at least 1 of the recommended behavioral practices that protect against AI infection. General media reports about AI through radio and television broadcasts appear to have been effective at reaching rural people. However, despite high awareness and widespread knowledge about AI and personal protection measures, most rural Cambodians still often practice atrisk poultry handling. Anecdotally, we also reported that family members of H5N1-infected patients, who knew about AI risks, still prepared dead or sick poultry for household consumption during massive die-offs, because they observed that neighbors with the same behavior did not become sick (Institute Pasteur in Cambodia, unpub. data). These findings provide evidence that high awareness does not necessary lead to behavior change. Behavior change involves comprehensive and multidisciplinary intervention, which combines risk perception communication and feasible and practical recommendations, including economic considerations. We speculate that it is hardly feasible to sustain good poultry-handling practices if access to personal protective equipment is cost prohibitive, particularly when disease occurrence poultry die-offs are common. Further studies are needed to determine appropriate behavior change strategies in Cambodia. We did find that many of the villagers were willing to report poultry deaths but did not know how. However, this finding should be interpreted in light of some limitations. We observed difficulties and frustrations among farmers whose flocks underwent culling after identification of H5N1 viruses in their flocks because compensation has not yet been approved by the government of Cambodia. In contrast, Thailand and Vietnam have introduced compensation along with the introduction of poultry vaccination in Vietnam and the reduction of backyard poultry ownership in Thailand in an effort to protect the commercial poultry industry. Thus, it is difficult to envision effective control strategies in Cambodia based exclusively on culling. Coincidentally, Vietnam has reported far fewer H5N1 outbreaks in poultry and humans since the introduction of the vaccination program, while Cambodia detected 4 outbreak sites in domestic poultry and 2 unrelated human cases in 2006. The real effect of a no-compensation policy on willingness to report poultry deaths needs to be assessed. Not surprisingly, direct contact with poultry and poultry products was common among household members. Transmission of H5N1 from poultry to humans, even in circumstances in which human-poultry interactions are regular and intense has been limited; however, as the virus continue to circulate and evolve among poultry, bird-tohuman transmission may increase. In this context, improvement in risky practices can only be achieved through relentless behavior change efforts. Because lack of knowledge does not appear to be a factor, intervention programs must include feasible options for resource-poor settings that have limited materials for personal protection (water, soap, rubber gloves, masks) and must offer farmers alternative methods to safely work with poultry on a daily basis.

Bird Migration Routes and Risk for Pathogen Dispersion into Western Mediterranean Wetlands

Wild birds share with humans the capacity for moving fast over large distances. During migratory movements, birds carry pathogens that can be transmitted between species at breeding, wintering, and stopover places where numerous birds of various species are concentrated. We consider the area of the Camargue (southern France) as an example to highlight how ad hoc information already available on birds’ movements, abundance, and diversity can help assess the introduction and transmission risk for birdborne diseases in the western Mediterranean wetlands. Avian influenza and West Nile viruses are used as examples because birds are central to the epidemiology of these viruses.

B irds are the only terrestrial vertebrates that share with humans the peculiarity of traveling in a few hours across national and intercontinental borders. The record for distance covered in a single year belongs to the arctic tern (Sterna paradisaea), which travels ≈50,000 km between Antarctica and northern Scandinavia. As a whole, billions of birds travel between continents twice a year in only a few weeks (1) . During these yearly migrations, birds have the potential of dispersing microorganisms that can be dangerous for public as well as animal health (2, 3) . For instance, birds are believed to be responsible for the wide geographic distribution of various pathogens, including viruses (e.g., West Nile, Sindbis, influenza A, Newcastle), bacteria (e.g., borrelia, mycobacteria, salmonellae), and protozoa (e.g., cryptosporidia). Insight into the ecology of bird populations is necessary to understand the epidemiology of bird-associated diseases. Furthermore, data about avian movements might be used to improve dis-ease surveillance schemes or to adapt preventive measures. However, solid bridges between ecology and human medicine are still lacking. We explored the bird sector, in an attempt to provide general ideas on bird abundance, migration, geographic origin, and interspecies mingling. We focused on the Camargue area, an alluvial lowland covering some 140,000 ha in the Rhône Delta. As other Mediterranean wetlands (Figure 1) , the Camargue is a major rallying point for Palearctic birds that are migrating between the great continental masses of Eurasia and Africa. This area is the current focus of intense sampling to study 2 pathogens closely associated with wild birds: avian influenza (AI) virus and West Nile virus (WNV). These 2 viruses have very different transmission cycles and ecology: AI viruses have a waterborne transmission, and ducks are their main natural reservoirs (4) (5) (6) (7) (8) ; WNV has a vectorborne transmission, and passerines are believed to play a major role in the amplification cycle (9) (10) (11) . However, both viruses are known to be carried by reservoir birds during migration and have been associated with emerging disease transmission risk for humans and domestic animals (2, 5, 7, 11, 12) . For both of them, the avifauna abundance, diversity, and departure origin may be of key importance in terms of disease transmission. We use these 2 viruses as examples in our discussion of the risk for dispersion of bird-carried pathogens into Mediterranean wetlands. We address the following questions: 1) What are the main geographic origins of birds observed in western Mediterranean wetlands? 2) How abundant and diverse in species are they during the year cycle? 3) When are interspecies contacts between birds from different origins most likely to occur? To address these issues, we used crude empiric indexes, which are known to have biases yet prove valuable within the scope of our objectives. Readers interested in modern ecologic methods used to study wildlife diseases in natural populations may refer to general publications on host-parasite systems (13) (14) (15) . Migration research is constantly changing, and new methods are always emerging. Historically, information about the movements of individual birds was first acquired through ringing studies. Bird ringing (also known as bird banding) consists of catching birds and attaching a small individually numbered metal or plastic ring to their legs or wings. Ring-recovery data are obtained when ringed birds are resighted, recaptured, hunted, or found dead. In Europe, large-scale ringing projects have been conducted, mostly between the 1950s and 1980s, and they represent a wealth of information that has not yet been fully exploited. Data recovered from birds ringed from 1950 to 1975 at the Station Biologique de la Tour du Valat in the Camargue were collected from annual reports. Seven species of waterbirds were chosen to illustrate various migratory patterns. We selected 4 species of the Anatidae family, known to have different geographic origins, including 3 dabbling ducks, i.e., ducks that search for their food primarily in surface water (mallard, Anas platyrhynchos, n recovered = 434; green-winged teal, A. crecca, n = 3,903; garganey, A. querquedula, n = 181) and 1 diving duck, i.e., a species that mostly searches for its food under water (tufted duck, Aythya fuligula, n = 313). We also took the example of the common coot (Fulica atra, n = 99), a diving bird of the Rallidae family that frequently shares ponds with ducks. The common snipe (Gallinago gallinago, n = 54) is an example among waders, i.e., shorebirds that feed in muddy swamps and coastlines. Finally, the purple heron (Ardea purpurea, n = 39) is an ardeid species that lives in reed beds and marshes. All these species are large or hunted, which explains the high number of rings recovered. We only considered data recovered from birds ringed in the Camargue area and later reported outside France. Since the 1950s, a large amount of data have been collected at the Station Biologique de la Tour du Valat thanks to bird counts, netting records, and field ornithologists' observations (see supplemental, online Technical Appendix Table l, indicating the methods used for each bird genus; available from www.cdc.gov/EID/content/ 13/3/365-appT1.htm). This information was used to create a database with a row for each of the 289 avian species regularly observed in the Camargue (16) . Strictly pelagic birds were not taken into account as they do not have any contact with terrestrial vertebrate species. Quantitative data were completed on the number of birds (abundance) and number of bird species (diversity) observed monthly in the Camargue. Three categories of migrating birds were considered, depending on the area from which they come: incoming birds from sub-Saharan Africa in spring and those arriving in autumn either from continental Europe or from Scandinavian and the Siberian tundra and taiga. Analyses were performed for all species and separately for species of the Anatidae family (ducks, swans, geese) and waders (shorebirds of the families Scolopacidae and Charadriidae), which are essentially associated with wetlands or coastlines. Regular bird counts provide information on bird populations for the studied area and therefore give an idea of potential contacts between species that share similar biotopes. Since September 1964, the Camargue duck and coot populations have been estimated every winter (17) . The count was made monthly by the same observer from a plane flying at an altitude of 200 feet. One hundred brackish lakes and marshes used by waterbirds as resting places were counted. The arrival of the plane made dabbling ducks fly off, which is necessary for detecting them and identifying their species. To count diving ducks, it was necessary to turn the group of birds around by using the plane. Results of the winter 2004-05 counts were used as examples. Ringing recoveries provide a valuable insight into the origins and dispersion areas of bird species. Figure 2 illustrates that western Mediterranean wetlands provide habitat for birds from a wide geographic range: all European countries but also other areas in the Mediterranean Basin, central and northern Asia, and sub-Saharan Africa. Ringed common coots and common snipes were mostly reported from continental Europe and Mediterranean areas, whereas mallards and common teals were also found in more northern places, including the former Soviet Union and Scandinavia. The pattern was slightly different for tufted ducks, for which >40% of recoveries were located in areas of taiga and tundra. Garganeys were recaptured in very distant places far north (Siberia, Finland), far east (Kazakhstan, Altai), and far south (Senegal, Mali) of the Camargue. In contrast to the previously described species, purple heron rings were recovered only from areas located south, including 4 countries in the Guinea Gulf in Africa (Benin, Côte d'Ivoire, Ghana, and Sierra Leone). As a whole, we discerned 3 broad areas from which Mediterranean waterbirds come and potentially disperse pathogens: continental Europe, northern Siberia and Scandinavia, and sub-Saharan Africa. Monthly abundance (number of individual birds) and diversity (number of species) in the Camargue are presented respectively in Figures 3 and 4 for birds originating from the 3 major areas of provenance described above. These figures show how many birds are in the Camargue, just as monthly photographs of bird populations do. A corresponding table indicates monthly abundance of each species (online Technical Appendix Table 2 , available from www.cdc.gov/EID/content/13/3/365-appT2.htm.). As many as 111 bird species might disperse pathogens from sub-Saharan Africa into the Camargue. Broadly speaking, birds coming from sub-Saharan Africa become rapidly and simultaneously abundant and diverse in spring, are still numerous in summer, and decrease in winter. The pattern is different if one considers solely ducks, as only 3 duck species fly south to tropical Africa, namely, the northern pintail (Anas acuta), the garganey, and the northern shoveler (Anas clypeata). Conversely, numbers and species diversity are high for waders, which are mainly passage visitors, especially in spring and late summer. A total of 53 species might introduce pathogens from northern areas into the Camargue. Abundance is highest in April and October-November with a higher peak in autumn, notably because of juvenile birds. Species diversity is high during winter and low from May to July. The opposite pattern was observed for sub-Saharan species. This pattern is even clearer for birds of the Anatidae family: They are abundant from October to January and in very small numbers from March to September. In contrast to ducks, waders are mainly transient visitors, and only a few individual birds spend the winter in the Camargue. Their number is greatest in spring and autumn. Up to 135 species could be involved in pathogen dispersion from continental Europe. Their abundance is highest from February to April and later from September to November. Species diversity remains high year-round with peaks in spring and autumn due to migrating passage visitors. The pattern observed for Anatidae species is the same as the 1 we described for Arctic species: birds are abundant in autumn and winter and in very small numbers in spring and early summer. However, the number of duck species remains stable year-round. Indeed, in species such as the mallard or the red-crested pochard (Netta rufina), some birds are sedentary whereas others are migratory. Waders show a constant level of species diversity because migration staggers over several months, but numbers are highly variable throughout the year. The results of the winter 2004-05 waterfowl counts are presented in Figure 5 for the species mentioned in Methods. Other species are also present, such as the northern pintail or the common shelduck (Tadorna tadorna). Garganeys are present in small numbers in September and February-March, but from an airplane they cannot be distinguished from common teals. These counts show that numerous species, with various migratory patterns, congregate on the same wetlands during the long winter period and therefore easily transmit waterborne pathogens such as AI virus. Most wintering birds are still present in March, when the first African migratory birds have already returned to breed in the Camargue or make a stop for refueling before flying further north. For instance, as many as 11,550 black-tailed godwits (Limosa limosa) were counted in the Camargue in March 2005. Moreover, the movement and abundance of birds vary greatly from 1 year to another because of movements influenced by weather conditions. For example, the duck population in the Camargue was estimated at ≈60,000 ducks in March 2005 compared with only ≈40,000 the previous year, when climatic conditions in Europe were warmer. Maps of ring-recovery data and graphs of monthly variations in bird abundance and diversity show that western Mediterranean wetlands such as the Camargue are a hub for birds from all origins (Central Asia, Siberia, northern and Eastern Europe, western Africa, and the Mediterranean basin) and that numerous birds of various species are seasonally aggregated in similar habitats. Under the hypothesis that risk for dispersion of pathogens into the Camargue is correlated with the number of birds and bird species encountered in a given area, these indices are helpful to determine periods at higher risk for introduction and emergence of birdborne diseases. We recall that these empiric estimates are skewed, which is briefly discussed with the perspectives below. The risk for introduction of African pathogens in Mediterranean wetlands would be highest from March to July, which corresponds with spring migration and breeding for birds. Conversely, in autumn, birds return to Africa and are more likely to introduce pathogens originating from the north than from the south (Table) . Of the 111 species that come every year to the Camargue from different countries of sub-Saharan Africa, most are insectivorous passerines that spend winter in Africa and breed in Europe; among aquatic birds, waders are the most numerous. Up to now, no evidence exists that birds migrating from sub-Saharan regions play a major part in the epidemiology of AI viruses. However, under the assumption that this area became an important epicenter for AI viruses, ducks would likely have the highest probability of introducing AI viruses in Mediterranean wetlands, even if they are less numerous than waders. Indeed, recent studies in Europe showed that overall AI virus prevalence in waders is really low compared with that in dabbling ducks (7) . WNV, which is transmitted by arthropod vectors, could potentially be introduced by any species of bird that comes from disease-endemic areas in Africa, is exposed to mosquito or tick bites (18) , and sustains high viremia levels. Insectivorous passerines are the most numerous and thus may be particularly suspected. WNV dispersion by birds migrating from sub-Saharan Africa might explain why an outbreak occurred in 2000 in the Camargue, even though the virus had not been observed there since the 1960s (19) . Pathogens may be introduced into Mediterranean wetlands by birds coming from northern areas of Scandinavia and Siberia. The risk would be higher from September to December when Arctic bird abundance reaches its peak (Table) . In spring, the northern birds observed in the Camargue have recently spent a long time in southern lands so that their associated probability of introducing pathogens originating from Scandinavia or Siberia is rather low. Waterbirds and granivorous passerines, which do not need to fly further south to find food supplies throughout the cold season, could introduce pathogenic microorganisms that could be transmitted later between wintering birds when densities are high. Waders, which migrate from Siberia and stop in the Mediterranean wetlands in autumn before crossing the Mediterranean Sea, could contaminate other bird species before pursuing their flight. As a whole, 53 species seen in the Camargue come from Arctic areas, which is half the number of species that come from sub-Saharan Africa or continental Europe. As a result, the probability of pathogens being introduced from Arctic areas should be lower than from birds of these 2 other areas. Another scenario can nevertheless be considered: if birds coming from northern areas disseminate a pathogen all along their migration route, then this pathogen would also infect continental European species and the probability of its being introduced into the Mediterranean wetlands would depend on the arrival of both Arctic and continental birds. AI viruses are likely to be introduced in autumn by ducks that breed in northern Europe and Siberia, especially since numbers are high because of the presence of juveniles. Furthermore, surveillance studies of wild ducks showed that the prevalence of AI viruses is primarily high in juveniles (5, 7, 20) . Conversely, WNV activity has never been reported in Scandinavia and Siberia, probably because the transmission cycle cannot be maintained in these northern biotopes. Autumn and winter are the 2 seasons during which the transmission of bird pathogens originating from continental Europe would be most likely (Table) . Indeed, in spring, the introduction of pathogens from continental Europe is less probable because birds have been absent from this area for 5 or 6 months. As previously seen, up to 135 species have the potential to introduce pathogenic agents in the Camargue. Granivorous passerines, birds of prey, and waterfowl are among the species that come in large numbers to take advantage of the Mediterranean wetlands' temperate climate during winter. Aquatic birds, which need unfrozen ponds to feed, show variations in their movements, depending on climatic conditions. For instance, if a cold spell occurs in eastern or northern Europe, the number of green-winged teals in the Camargue increases (17) . These weather-associated movements might at certain times prove essential in pathogen dispersion within European and Mediterranean wetlands. Surveys of wild waterbirds in Europe have shown that AI viruses are frequently found (21) (22) (23) (24) , which means that waterbirds arriving from continental Europe might often be carriers of AI viruses. Similarly, because WNV activity was recently reported in Romania (25) and the Czech Republic (26) , wild birds migrating in autumn from these countries to the Mediterranean basin could introduce WNV, either because of a high viremia level or because they carry infected ectoparasites. If the virus managed to overwinter in a reservoir host or a vector, it could then be responsible for an outbreak the next summer, when mosquito vectors are abundant (27) . Several factors affect the risk for bird-to-bird transmission: bird abundance or density, bird diversity, species receptivity and sensitivity to pathogens, interspecies interactions, and environmental conditions (14) . For watertransmitted pathogens such as AI viruses, risk for transmission may be associated with the number of ducks congregated in the same pond, particularly in autumn and winter ( Figure 5 ). This crowding of wintering species, in addition to the permanent presence of a transient population of birds using wetlands to stop off during migration, could allow AI viruses to circulate and be maintained because of rapid dissemination on shared water. For vector-transmitted pathogens such as WNV, transmission possibilities depend both on the bird reservoir density and on the dispersion abilities and activity periods of the arthropod vectors. The risk for interspecies transmission of disease is particularly problematic when wild and domestic species are involved. Ducks are aquatic birds that are most likely to come in contact with free-range poultry, especially because the presence of congeners can induce migrating wild ducks to make a stopover. Captive-bred mallards, used for hunting purposes and voluntarily put in the wild to attract other ducks, are particularly likely to share pathogens with their migratory congeners and facilitate the transmission of diseases to other domestic species. The risk is different for domestic chickens or turkeys, which are more likely to have contact with granivorous birds. Conversely, waders are rarely in direct contact with human-raised species. Bird-carried pathogens are above all susceptible to being spread worldwide because of human activities such as legal or illegal trade of wild and domestic birds or bird products (28) . The mechanism for the introduction of WNV into America in 1999 is not known with certainty, but a plausible scenario is the importation of an infected bird (29, 30) . Similarly, the highly pathogenic AI strain H5N1 was isolated in Belgium from crested hawk-eagles (Spizaetus nipalensis) smuggled by air travel (31) . In Asia, transmission of H5N1 influenza virus has mainly been the result of human activity such as live-poultry markets and the international trade of birds, bird products, or contaminated equipment (32) (33) (34) (35) . The ornithologic data we have presented are merely crude estimates. Ring-recovery data, for instance, are subject to strong biases related to where and when the ringing was conducted but also to high variability in the probability of reporting marked animals among areas (36) . Similarly, our estimates of bird abundance and diversity are basic indices associated with the number of birds heard, seen, or caught in the Camargue (see online Technical Appendix Table 1 ). These estimates do not take into account 2 important sources of error: detection error, related to the fact that the probability of detecting a bird is <1, and survey error, associated with spatial and temporal variability (37) . Since our motivations were merely to show that information already available on birds may lead to better understanding of animal and human health issues associated with birdborne pathogens, these biases do not invalidate our objectives. The results obtained were helpful to identify key groups of species likely to introduce pathogens from a given area at a given time of year. We voluntarily chose to focus on birds and leave pathogens aside, but studies of diseases in natural bird populations are obviously critically needed. Ecology, the science of interactions between living organisms and their physical environment, has been extended to include microorganisms. Understanding the relationships between organisms (such as hosts, pathogens, predators, competitors) and their environment is the aim of disease ecology. As studying the dynamics of systems with many hosts and pathogenic agents is complex, efforts should primarily focus on a few specific bird-pathogen models. Mathematical modeling may help to predict specific bird-pathogen interactions and to identify key parameters that need to be better estimated through additional research. Long-term records enable establishment of databases, which would illustrate bird-pathogen relationships in natural conditions. These data would focus on hosts, their migration, population age, behavior, and so forth. Host-pathogen interactions should be described by using data such as antibody prevalence in different age classes, frequency of virus isolation, and characterization of the strains involved. Complementary laboratory and field experiments within a controlled environment might also provide relevant information. All these investigations should gradually make it possible to gather valuable baseline data to test specific hypotheses and gain new insights in bird-pathogen relationships in Mediterranean wetlands.

Prolonged extracorporeal membrane oxygenation therapy for severe acute respiratory distress syndrome in a child affected by rituximab-resistant autoimmune hemolytic anemia: a case report

INTRODUCTION: Autoimmune hemolytic anemia in children younger than 2 years of age is usually characterized by a severe course, with a mortality rate of approximately 10%. The prolonged immunosuppression following specific treatment may be associated with a high risk of developing severe infections. Recently, the use of monoclonal antibodies (rituximab) has allowed sustained remissions to be obtained in the majority of pediatric patients with refractory autoimmune hemolytic anemia. CASE PRESENTATION: We describe the case of an 8-month-old Caucasian girl affected by a severe form of autoimmune hemolytic anemia, which required continuous steroid treatment for 16 months. Thereafter, she received 4 weekly doses of rituximab (375 mg/m(2)/dose) associated with steroid therapy, which was then tapered over the subsequent 2 weeks. One month after the last dose of rrituximab, she presented with recurrence of severe hemolysis and received two more doses of rrituximab. The patient remained in clinical remission for 7 months, before presenting with a further relapse. An alternative heavy immunosuppressive therapy was administered combining cyclophosphamide 10 mg/kg/day for 10 days with methylprednisolone 40 mg/kg/day for 5 days, which was then tapered down over 3 weeks. While still on steroid therapy, the patient developed an interstitial pneumonia with Acute Respiratory Distress Syndrome, which required immediate admission to the intensive care unit where extracorporeal membrane oxygenation therapy was administered continuously for 37 days. At 16-month follow-up, the patient is alive and in good clinical condition, with no organ dysfunction, free from any immunosuppressive treatment and with a normal Hb level. CONCLUSIONS: This case shows that aggressive combined immunosuppressive therapy may lead to a sustained complete remission in children with refractory autoimmune hemolytic anemia. However, the severe life-threatening complication presented by our patient indicates that strict clinical monitoring must be vigilantly performed, that antimicrobial prophylaxis should always be considered and that experienced medical and nursing staff must be available, to deliver highly specialized supportive salvage therapies, if necessary, during intensive care monitoring.

Autoimmune hemolytic anemia (AIHA) in children is usually characterized by a severe course with a mortality rate of approximately 10% [1] . The required prolonged immunosuppressive therapy often leads to steroid dependence [2] . The administration of non-steroidal immunosuppressive drugs such as cyclosporine A, cyclophosphamide and azathioprine, has been used in the past [1] [2] [3] [4] . Nowadays, the use of monoclonal antibodies such as rituximab, has given promising results for pediatric refractory AIHA [5] [6] [7] , with sustained remissions in the majority of patients. Nevertheless, potentially life-threatening infections are known to occur with rituximab [7] . In the event of rituximab failure, there is no general consensus or guidelines available indicating precisely how to manage resistant forms of AIHA. Heavy immunosuppression consisting of the combined use of cyclophosphamide and high-dose steroids may be considered [8, 9] . We report the case of an 8-month-old Caucasian girl referred to us for observation due to intense pallor, jaundice, lethargy and fever. Serological evaluations revealed severe anemia (Hb = 2.8g/dL) with a strongly positive direct antiglobulin test and high-titer warm IgG autoantibody. AIHA was diagnosed and steroid therapy with intravenous methylprednisolone at 2mg/kg/day was administered for 5 days (Figure 1 ). An adequate Hb increase was obtained and the child was discharged after 10 days with oral prednisone at 2mg/kg/day. During the subsequent months, several attempts were made to taper off the prednisone, but the patient had developed steroid dependence. Considering this dependence on high steroid doses, a therapeutic course with four doses of rituximab was performed (375mg/m 2 /dose) at weekly intervals ( Figure 1 ). Before rituximab infusion, serum immunoglobulin levels were normal and subpopulation lymphocyte counts were within the normal range. The treatment with rituximab was well tolerated and the patient received intravenous substitutive therapy with commercially available immunoglobulin preparations (400mg/kg, every 3 weeks for 6 months). One month after the end of the first course of rituximab, while still receiving low-dose steroids, the patient presented with a clinical relapse of AIHA, so prednisone was increased to 2mg/kg/day and two further rituximab infusions were performed ( Figure 1 ). After these infusions, B lymphocytes became undetectable and the count returned to normal values 8 months after treatment. The patient remained in clinical remission and free from immunosuppressive drugs for 7 months, before presenting with a further relapse. A more intensive treatment was performed ( Figure 1 ) with cyclophosphamide 10mg/kg/ day for 10 days and methylprednisolone 40mg/kg/day for 5 days, which was tapered over 20 days. Hb level increased and the patient was discharged 10 days later in good clinical condition, without any antifungal or antiviral prophylaxis. Two weeks later, the child was referred to the Emergency Room for respiratory failure, persistent fever and abdominal pain. Laboratory examination showed an Hb level of 12.8g/dL, total leukocyte count (WBC) of 710/µL, absolute neutrophil count (ANC) of 90/µL, a platelet count (PLT) of 339,000/µL, and low levels of immunoglobulin (IgG = 360mg/dL, IgA = 10mg/dL, IgM = 33mg/ dL). Chest X-ray and CT scan revealed an interstitial pneumonia ( Figure 2 ). Therapy with amikacin, ceftazidime, G-CSF and voriconazole was started. Within a few hours, her clinical condition deteriorated and the patient developed Acute Respiratory Distress Syndrome (ARDS), which required immediate admission to the intensive care unit (ICU). Acceptable gas exchange was initially maintained by non-invasive continuous positive airway pressure ( Figure 3 ). Serologic tests showed a level of Aspergillus galactomannan antigen of 0.8. All tested virus and microbial antigens were negative. On day 4, concomitantly with an elevation of WBC from 400 to 10,400/µL (ANC = 4900/µL), respiratory conditions precipitated and endotracheal intubation and mechanical ventilation were started ( Figure 3) . A protective ventilatory strategy with tidal volume of 6mL/kg and positive end expiratory pressure (PEEP) of 10cmH 2 O was instituted. In the following days, gas exchange deteriorated and PEEP levels rose to 17cmH 2 O. Recruitment manoeuvres, prone positioning, and high doses of inhaled nitric oxide (NOi) were necessary to maintain viable gas exchange. Endotracheal instillation of porcine surfactant and a trial with High Frequency Oscillation were ineffective. On day 10, owing to the refractory hypoxia, worsening hypercapnia, and chest X-ray evidence of a pneumomediastinum, the patient was placed on venous-venous extracorporeal membrane oxygenation (ECMO) (Figure 3) . A double lumen 15 french catheter (Maquet, Jostra Medizintechnik AG, Hirrlingen, Germany) was inserted into the right internal jugular vein. The ECMO circuit consisted of a polymethylpentene membrane oxygenator, permanent life support and a centrifugal Rotaflow pump (Maquet, Jostra Medizintechnik AG, Hirrlingen, Germany). ECMO was started with a blood flow of 0.8 to 0.9 L/minute and gas flow of 1 L/minute of 100% oxygen. Following the institution of ECMO, respiratory rate decreased from 45 to 10 breaths/minute, and it was possible to stop NOi. After commencing caspofungin with voriconazole, WBC, ANC and C reactive protein (CRP) slowly decreased, while pulmonary function slightly improved. On day 19, a multidrug-resistant Pseudomonas aeruginosa was isolated from bronchoaspirate ( Figure 3 ). In spite of antibiotic reinforcement with levofloxacin, Pseudomonas aeruginosa antibiogram showed increased resistance to all antibiotics and to colimicine which was started on day 26. On day 28, a sudden increase in resistance on the return part of the circuit caused a massive thrombosis in the oxygenator. The entire circuit and the cannula were immediately changed and ECMO restarted within 2 hours. Following the development of pulmonary embolism, the gas exchange rapidly worsened and NOi had to be restarted. An echocardiographic assessment showed right ventricular dilatation, with paradoxical septal wall motion and pulmonary hypertension (systolic pressure 90mmHg). Prostacyclin and sildenafil improved the right heart function and effectively attenuated pulmonary hypertension. In the following days, WBC, ANC and CRP slowly decreased, while pulmonary function improved. Thirty days from ICU admission, ECMO was stopped, the patient rapidly restored her spontaneous ventilatory functions and she was extubated 10 days later. She was finally discharged from the ICU on day 44. At 16-month follow-up, the patient is alive and free from immunosuppressive drugs. At the time of last follow-up, Hb level was 13.5g/dL, reticulocyte count was 11 × 10 9 /L, with total bilirubin, lactic dehydrogenase and haptoglobin all within the normal ranges. AIHA in children younger than 2 years is in some cases characterized by a resistance to corticosteroids or dependence on high steroid doses and subsequent development of severe side effects [2] . Splenectomy or immunomodulating agents have frequently been used, but there is no consistent demonstration of their efficacy in controlling hemolysis [3] . Immunosuppressive drugs such as azathioprine, cyclosporine A, or cyclophosphamide, alone or in combination reduce steroid dependence and sometimes control hemolysis [1] [2] [3] [4] . Clinical experience with monoclonal antibodies appears encouraging. In particular, rituximab is increasingly being used off-label, for difficult-to-treat auto-immune diseases and presents the advantage of inducing a selective B-cell depletion, sparing cellular immunity mediated by T cells and natural killer cells. Even though prospective controlled studies are not currently available, the efficacy of rituximab has been shown in pediatric studies. Quartier et al. [5] treated five pediatric refractory AIHA patients, who achieved a complete remission within 15 to 22 months after rituximab therapy. These results were confirmed by Zecca et al. [6] in a group of 15 children treated with rituximab. Four other children were treated by Motto et al. [7] , with the achievement of complete remission. Nevertheless, the prolonged impairment of antibody production leads to an increased risk of viral and bacterial infections. For this reason, monthly intravenous immunoglobulin infusions are recommended for a minimum of 6 months following completion of therapy and prophylaxis for P. jirovecii pneumonia is also suggested [7] . The patient described in our report received four rituximab infusions in an off-label setting, followed by two additional doses over 6 months. Clinical remission was achieved for 7 months after which it was possible to interrupt steroid treatment. The pattern of immune reconstitution after rituximab therapy revealed persistently low immunoglobulin levels, partially corrected by the substitutive therapy. Immunoglobulin levels reached their normal range in 18 months and the lymphocyte subpopulations returned to normal range in 16 months. It is, nevertheless, difficult to quantify the real role of rituximab in this heavy immunosuppression, since a combined therapy with high doses of methylprednisolone and cyclophosphamide was subsequently started. Even though our patient did not present any early side effects related to the rituximab infusions, a prolonged follow-up should be carried out to monitor and prevent long-term side effects of rituximab, which are still unknown. When our patient relapsed, an alternative treatment was required, since therapies with steroids, rituximab and intravenous immunoglobulins proved to be ineffective. The role of splenectomy in refractory AIHA is still controversial [1] [2] [3] [4] . Although effective vaccinations are available, this surgical treatment should be avoided in children younger than 6 years of age, due to the risk of developing severe bacterial infections. According to local policy, drug-based immunosuppressive therapy is to be preferred. We therefore decided to adopt a combined therapy approach, with high doses of methylprednisolone (40mg/kg/day for 5 consecutive days), which was then tapered down over 20 days, and cyclophosphamide (10mg/kg/day for 10 consecutive days). This approach appeared to be feasible and encouraging, since we had previously successfully treated two pediatric cases of refractory AIHA with an identical approach [8, 9] . The administration of methylprednisolone and cyclophosphamide increased the already significant immunodepression which had resulted from prior therapies and further contributed to the severity of the infectious complication presented by our patient that required ECMO therapy. While in the ICU, the patient underwent various ventilatory treatments, some of which are not considered conventional. Modern ventilatory strategy in ARDS aims to provide viable gas exchange with high oxygen concentration and PEEP, while minimizing the injurious effects of mechanical ventilation by using low tidal volume ventilation (6mL/kg) [10] . Although other techniques such as the prone position, NOi and recruitment manoeuvres are effective in improving gas exchange, they did not prove effective in terms of survival [10] . Nevertheless, before ECMO, the only means of providing minimal acceptable oxygenation was to use both NOi and the prone position. Despite using low tidal volumes, a respiratory rate of up to 45 breaths/minute was necessary to obtain acceptable CO 2 levels, and the occurrence of pneumomediastinum demonstrated that we were unable to provide an effective protective ventilatory strategy. Thus, ECMO was the only real means of providing such a strategy, while allowing adequate gas exchange. Refractory AIHA in pediatric patients is a challenging disease that forces us to weigh up the risks and benefits of heavy and prolonged immunosuppressive therapies that can reduce or even eradicate the hemolysis, despite the risk of infectious complications. For this reason, we feel that prolonged viral and fungal prophylaxis therapy should always be considered, during and after the immunosuppressive therapy. Furthermore, strict clinical monitoring should be carried out, even when no evident symptoms are present. In our patient, we did not administer any prophylaxis and clinical monitoring was probably delayed for too long after discharge. Resolution of the infectious complication was possible thanks to an advanced intensive care assistance, which consisted of ECMO and the management of its related complications. This case study shows that rituximab-resistant AIHA in young children represents a significant challenge, requiring aggressive immunosuppressive therapy, which may potentially cause severe life-threatening complications. Nowadays, it is not clear which is the best immunosuppressive agent to be administered in the event of rituximab failure. We found that the combination of methylprednisolone and cyclophosphamide could be a valid alternative, based on previous experience. Nevertheless, a universal therapeutic flow-chart is still lacking and should be defined, which considers new therapeutic strategies such as alemtuzumab [11] or hematopoietic stem cell transplantation [12] . What is clear, however, in the case of heavy immunosuppressive therapy, is the importance of strict patient monitoring during and after immunosuppressive therapy and an antimicrobial prophylaxis, particularly for fungal agents and P. jirovecii. AIHA, autoimmune hemolytic anemia; ANC, absolute neutrophil count; CT, computed tomography; WBC, leukocyte count; PLT, platelet count; ARDS, acute respiratory distress syndrome; ICU, intensive care unit; PEEP, positive end expiratory pressure; NOi, inhaled nitric oxide; ECMO, extracorporeal membrane oxygenation; CRP, C reactive protein.

Regulated Ire1-dependent decay of messenger RNAs in mammalian cells

Maintenance of endoplasmic reticulum (ER) function is achieved in part through Ire1 (inositol-requiring enzyme 1), a transmembrane protein activated by protein misfolding in the ER. The cytoplasmic nuclease domain of Ire1 cleaves the messenger RNA (mRNA) encoding XBP-1 (X-box–binding protein 1), enabling splicing and production of this active transcription factor. We recently showed that Ire1 activation independently induces the rapid turnover of mRNAs encoding membrane and secreted proteins in Drosophila melanogaster cells through a pathway we call regulated Ire1-dependent decay (RIDD). In this study, we show that mouse fibroblasts expressing wild-type Ire1 but not an Ire1 variant lacking nuclease activity also degrade mRNAs in response to ER stress. Using a second variant of Ire1 that is activated by a small adenosine triphosphate analogue, we show that although XBP-1 splicing can be artificially induced in the absence of ER stress, RIDD appears to require both Ire1 activity and ER stress. Our data suggest that cells use a multitiered mechanism by which different conditions in the ER lead to distinct outputs from Ire1.

The ER is responsible for folding and processing proteins entering the secretory pathway and uses a variety of mechanisms to adjust its capacity in response to changes in the folding burden. This collection of mechanisms, termed the unfolded protein response (UPR; Ron and Walter, 2007) , is activated by stress caused by environmental stimuli and diseases such as viral infection (Tardif et al., 2004; Bechill et al., 2008) and multiple myeloma (Carrasco et al., 2007) . In metazoans, components of the UPR are essential for developmental processes requiring high levels of secretion such as the differentiation of plasma cells (Gass et al., 2002; Iwakoshi et al., 2003) and pancreatic  cells (Harding et al., 2001; Zhang et al., 2002) . The UPR restores ER function both by increasing its capacity and decreasing the load of new proteins through transcriptional induction of secretory pathway components and general translational attenuation. One of the key players in the UPR is Ire1 (inositolrequiring enzyme 1), a conserved transmembrane protein with a luminal domain that senses protein misfolding in the ER. The resulting oligomerization of Ire1 leads to activation of its cytoplasmic kinase and endoribonuclease domain. The main function of the kinase appears to be activation of the nuclease, which requires binding of ATP or ADP in the active site of the kinase (Papa et al., 2003) . The nuclease in turn cleaves two specific sites in the mRNA encoding XBP-1 (X-box-binding protein 1), a conserved UPR transcription factor, which leads to XBP-1 activation and translation through removal of a regulatory intron. We have recently shown that Ire1 in Drosophila melanogaster cells independently mediates the cleavage and degradation of mRNAs encoding proteins that traverse the secretory pathway (Hollien and Weissman, 2006) . This new branch of the UPR, which we call regulated Ire1-dependent decay (RIDD), has the potential to selectively relieve the burden on the ER while clearing the translation and translocation machinery for the subsequent influx of new proteins induced by the UPR. Studies on the regulation of specific messages suggest that the RIDD pathway also operates in mammalian cells (Tirasophon et al., 2000; Iwawaki et al., 2001; Oikawa et al., 2007; Iqbal M aintenance of endoplasmic reticulum (ER) function is achieved in part through Ire1 (inositolrequiring enzyme 1), a transmembrane protein activated by protein misfolding in the ER. The cytoplasmic nuclease domain of Ire1 cleaves the messenger RNA (mRNA) encoding XBP-1 (X-box-binding protein 1), enabling splicing and production of this active transcription factor. We recently showed that Ire1 activation independently induces the rapid turnover of mRNAs encoding membrane and secreted proteins in Drosophila melanogaster cells through a pathway we call regulated Ire1-dependent decay (RIDD). In this study, we show that mouse fibroblasts expressing wild-type Ire1 but not an Ire1 variant lacking nuclease activity also degrade mRNAs in response to ER stress. Using a second variant of Ire1 that is activated by a small adenosine triphosphate analogue, we show that although XBP-1 splicing can be artificially induced in the absence of ER stress, RIDD appears to require both Ire1 activity and ER stress. Our data suggest that cells use a multitiered mechanism by which different conditions in the ER lead to distinct outputs from Ire1. ments of RIDD. By expressing wild-type and mutant variants of Ire1-, we find that the nuclease activity of Ire1 is required for both splicing and RIDD. However, these two outputs can be differentially triggered, revealing an unexpected complexity in Ire1 activation. To determine whether the RIDD pathway functions in mammals as well as in Drosophila, we used a mouse embryonic fibroblast (MEF) cell line established by Lee et al. (2002) in which the Ire1- gene has been disrupted. We stably introduced human Ire1- (hIre1) into these cells using flippasemediated, site-directed recombination , which allowed us to insert the wild-type and hIre1 variants discussed in the next sections into the same sites within the genome. In response to various forms of chemically induced ER stress, the reconstituted cells (referred to here as hIre1 R ) but not the Ire1 knockout (Ire1 / ) cells induced the splicing of XBP-1 et al., 2008; Lipson et al., 2008) . Mammals express two isoforms of Ire1: Ire1- is expressed ubiquitously, and Ire1- is expressed in intestinal epithelial cells. Overexpression of Ire1- leads to cleavage of its own message in COS-1 cells (Tirasophon et al., 2000) and reduced levels of the message encoding CD59 (complement defense 59) in HeLa cells (Oikawa et al., 2007) . Ire1- also appears to mediate the degradation of insulin transcripts in pancreatic  cells under chronic high glucose conditions, perhaps promoting cell survival during extreme chronic stress (Lipson et al., 2008) . Ire1- appears to have alternative targets as well, mediating cleavage of the 28-S ribosomal RNA (Iwawaki et al., 2001) and the mRNA encoding microsomal triglyceride transfer protein, an ER chaperone important for the assembly of lipid transport vesicles (Iqbal et al., 2008) . Together, these examples of Ire1 function in mRNA decay may explain observations that Ire1 in metazoans has a broader range of physiological outputs than XBP-1 splicing (Zhang et al., 2005) . In this study, we take advantage of mouse fibroblasts lacking Ire1 activity, both to confirm that RIDD is conserved in mammalian cells and to investigate the functional require- and Fig. S1 ). Although 120 mRNAs fell into this category in the array data, the magnitudes of the changes in expression were generally small (many were twofold or less) compared with those seen in Drosophila cells, where many RNAs were downregulated by 5-10-fold. Despite the relatively small changes for individual messages, the effect of down-regulating mRNAs at the ER surface may profoundly impact the folding burden of the ER. For example, the redistribution of certain nascent proteins from the ER to the cytosol during ER stress significantly impacts cell survival, although the effect on translocation is similar in magnitude to RIDD (Kang et al., 2006) . To limit false positives, we applied a series of strict criteria for identifying the most likely RIDD targets (see Materials and methods); the results are shown in Table I . We confirmed the regulation of several of these targets by reverse transcription followed by quantitative real-time PCR (qPCR; Fig. 1 and Fig. S1 ). To determine whether the down-regulation was a general response to ER stress or was restricted to DTT, we also treated cells with tunicamycin (Tm), which inhibits N-linked glycosylation or thapsigargin, which disrupts calcium homeostasis. All candidate RIDD (see Fig. 3 B). Both cell lines strongly induced immunoglobinbinding protein (BiP), an abundant ER chaperone, in response to ER stress ( Fig. 1) , which is in agreement with previous observations that BiP induction is Ire1 independent in mouse fibroblasts (Lee et al., 2002) . As expected, induction of ERdj4, a transcriptional target of the Ire1/XBP-1 branch of the UPR , was stronger in the hIre1 R cells compared with Ire1 / cells, especially in response to DTT, a reducing agent which disrupts disulfide bond formation in the ER (Fig. 1 A) . To investigate the RIDD pathway, we took an unbiased, microarray-based approach. We induced ER stress in the Ire1 / and hIre1 R cells with DTT, purified total RNA from treated and untreated cells, amplified and labeled these samples, and hybridized them to whole-genome MEEBOChip arrays. As expected, treatment of our hIre1 R cells with DTT led to the induction of classical UPR targets to levels comparable with that seen by others in wild-type cells ( Fig. S1 ; Lee et al., 2003) . Consistent with the RIDD pathway functioning in these cells, we observed that several RNAs were down-regulated in response to ER stress in the hIre1 R but not the Ire1 / cells ( Fig. 1 genome encode membrane or ECM proteins, as do 6 of the 7 confirmed targets that displayed increased decay rates. The seventh confirmed target, Blos1, encodes a protein that does not appear to contain a membrane-spanning domain itself but is part of a complex detected in both cytosolic and peripheral membrane fractions (Starcevic and Dell'Angelica, 2004) . We consistently observed partial splicing of XBP-1 in our hIre1 R cells in the absence of ER stress, which is likely caused by overexpression of hIre1 in these cells. To confirm that the mRNA decay we observed was not an artificial product of additional stress or Ire1 activity caused by overexpression, we measured the mRNA decay rates of two confirmed RIDD targets in NIH-3T3 cells (Fig. 2) . As in our reconstituted cells, NIH-3T3 cells displayed increased decay rates and lower abundance for RIDD targets in the presence of DTT. Collectively, our data indicate that RIDD is a general, conserved pathway associated with folding stress in the ER. Because the parent Ire1 / cell line lacks Ire1 activity, we could probe specific functions of Ire1 by inserting variants of hIre1 targets tested were down-regulated in response to these three forms of ER stress in an Ire1-dependent manner (Fig. 1 , A-C). Several observations indicate that the RIDD pathway, if not the specific targets, is conserved in mammalian cells and Drosophila. First, the down-regulation of mouse target mRNAs was independent of XBP-1 (Fig. 1 D) . Using RNAi to deplete Ire1 from cells lacking a functional XBP-1 , we found that Ire1-dependent down-regulation of several target mRNAs occurred in the absence of XBP-1. As a control, in cells lacking XBP-1, ERdj4 was induced only to the level seen in Ire1 / cells, and this induction was insensitive to Ire1 depletion ( Fig. 1 D) . Second, the down-regulation of many mouse targets was achieved through an increase in their decay rates. We inhibited transcription with actinomycin D and monitored mRNA levels over time in the absence and presence of ER stress ( Fig. 2 and Fig. S2 ). For 7 of the 11 targets tested, the decreases in mRNA abundance were dependent not on transcription but on increased rates of mRNA degradation. Third, as in Drosophila cells, the set of targets in mouse cells is highly enriched for mRNAs encoding membrane proteins. 17 of the 22 targets in Table I that were assigned localization annotations in the mouse little effect on splicing or RIDD targets. However, cells treated with both 1NM-PP1 and ER stress degraded RIDD targets to similar levels as those expressing the wild-type Ire1 (Fig. 4 , D and E). As a further control, we confirmed the XBP-1 independence of RIDD in these cells. XBP-1 is transcriptionally induced in our cells in a largely Ire1-independent manner and therefore does not occur to a significant extent in cells treated with 1NM-PP1 alone. Therefore, we depleted XBP-1 from cells expressing Ire1-I642G using RNAi, and although this blocked containing point mutations into these cells in an isogenic manner. Using this approach, we expressed an hIre1 variant containing a single point mutation (D847A) that has been previously shown to compromise the nuclease activity of hIre1 while leaving the kinase activity intact (Tirasophon et al., 2000) . For these experiments, we used C-terminally Flag-tagged versions of both the wild-type and mutant Ire1 and confirmed by Western blotting that the hIre-D847A variant was expressed at similar levels to the wild type (Fig. 3 A) . Consistent with a lack of nuclease activity, cells expressing hIre1-D847A did not support XBP-1 splicing in the absence or presence of ER stress (Fig. 3 B) and induced ERdj4 only to the level seen in Ire1 / cells (Fig. 3, C and D) . All cell lines induced BiP by similar amounts, indicating that they are experiencing a similar level of ER stress. However, unlike the Ire1 R cells, cells expressing hIre1-D847A displayed no change in the abundance of RIDD targets upon induction of ER stress (Fig. 3, C and D) . These results indicate that the nuclease activity of Ire1 is required for degradation of RIDD targets, which is consistent with a mechanism in which Ire1 directly cleaves these RNAs in response to ER stress. Recently, it has been shown that for yeast and human Ire1, the cytoplasmic kinase activity can be bypassed using a small molecule that binds to the enlarged ATP pocket of an engineered Ire1 variant (Papa et al., 2003; Lin et al., 2007) . This variant, I642G in hIre1, has very little activity in the absence of the ATP analogue 1NM-PP1 (4-amino-1-tert-butyl-3-[1'-naphthyl methyl]pyrazolo[3,4-d] pyrimidine). Binding allows activation of the nuclease while simultaneously inhibiting the kinase activity of Ire1-I642G. In HEK293 cells (Lin et al., 2007) and MEF cells (Han et al., 2008) , hIre1-I642G splices XBP-1 and induces the UPR upon binding 1NM-PP1 alone, even in the absence of ER stress. To test whether activation of Ire1 is sufficient for degradation of RIDD substrates, we introduced hIre1-I642G into Ire1 / cells and tested its activity in the absence and presence of 1NM-PP1 and ER stress (Fig. 4) . Treatment of mouse cells expressing hIre1-I642G with 1NM-PP1 was sufficient to induce splicing of XBP-1 (Fig. 4 A) and transcriptional induction of the Ire1/XBP-1 target ERdj4 (Fig. 4 B) . This Ire1 activity was not caused by a general induction of ER stress, as cells lacking Ire1 or expressing the wild-type hIre1 were insensitive to 1NM-PP1 (Fig. 4) , although the drug did suppress the Ire1-independent induction of ERdj4 by DTT in all strains through an unknown mechanism (Fig. 4 B) . The 1NM-PP1induced activation of hIre1-I642G was not caused by increased levels of the protein, as assayed by Western blotting (Fig. 4 C) . Thus, our data are consistent with a 1NM-PP1-induced conformational change in the Ire1 variant itself, as proposed previously (Papa et al., 2003) . In contrast to XBP-1 splicing and ERdj4 induction, no degradation of RIDD substrates was observed with 1NM-PP1 alone, indicating that XBP-1 splicing and RIDD are separable functions of Ire1 and have distinct requirements for activation (Fig. 4, D and E) . Treatment of cells expressing Ire1-I642G with DTT or Tm alone (in the absence of 1NM-PP1) had very BiP, an ER chaperone which binds Ire1 and is titrated off by misfolded proteins during ER stress. A second possibility is that other factors activated by ER stress are necessary for RIDD. Although new transcription of such a factor is not required (Fig. 2) , there may be proteins recruited or activated by ER stress that are important for mRNA decay but not for splicing. Finally, the complement of mRNAs available for targeting to this decay pathway may change upon subjecting cells to ER stress. Certain signal sequences are sensitive to the ER folding environment and influence how associated ribosomes partition between the ER and cytoplasm (Kang et al., 2006) ; thus, a substantial change in the pool of mRNAs associated with the ER membrane under stress conditions could affect the RIDD pathway. The fact that the two outputs of Ire1's nuclease activity, RIDD and XBP-1 splicing, can be differentially activated reveals an unanticipated complexity in the UPR. Growing evidence suggests that various sensors associated with the UPR are activated under different conditions, allowing for specific responses to different forms of ER stress in different cell types (Gass et al., 2002; DuRose et al., 2006) . This customization of responses may extend to activation within Ire1 itself. Activation of XBP-1 leads to a protective remodeling and expansion of the secretory pathway, whereas RIDD reduces the load of incoming proteins. Under certain physiological situations, e.g., during plasma cell development, activating the XBP-1-dependent remodeling pathway without inducing RIDD may be more beneficial; thus, cells may induce an active state of Ire1 that is similar to the 1NM-PP1-mediated activation described in this study. In contrast, conditions such as viral infection may call for a more destructive response that limits the load of incoming proteins, which is analogous to the effects of reduced translation mediated by the PERK (PKR-like ER kinase) branch of the UPR (Harding et al., 2000) . Intriguingly, both hepatitis C virus and human cytomegalovirus induce Ire1 but appear to block downstream effects of XBP-1 (Tardif et al., 2004; Isler et al., 2005) . However, hepatitis C virus protein production is increased in the absence of Ire1 (Tardif et al., 2004) , suggesting that an alternate activity of Ire1 such as RIDD may be attenuating viral protein synthesis. It remains to be determined what role RIDD may play in viral infection and other physiological stress conditions, but the ability of this pathway to function in mammalian cells and the potential to decouple it from XBP-1 splicing could allow for a more specific and effective response to changing conditions within the ER. Establishment of Ire1-expressing cell lines and RNAi Ire1- / MEFs were obtained from R. Kaufman (University of Michigan Medical Center, Ann Arbor, MI), and XBP1 / MEFs were obtained from L. Glimcher (Harvard School of Public Health, Boston, MA). All cells were maintained at 37°C and 5% CO 2 in DME supplemented with 10% fetal calf serum, Gln, and antibiotics. We generated Ire1- / MEFs containing induction of ERdj4 by DTT, RIDD remained intact when cells were treated with both 1NM-PP1 and DTT (Fig. 4 F) . To determine whether the lack of RIDD in our 1NM-PP1induced cells is merely a threshold effect, we treated cells with higher concentrations of 1NM-PP1, such that no further XBP-1 splicing or ERdj4 activation could be achieved with 1NM-PP1 alone (Fig. 4, G and H) . The levels of XBP-1 splicing in cells expressing hIre1-I642G treated with 30 µM 1NM-PP1 were as high or higher than those seen in wild-type cells treated with low concentrations of DTT (Fig. 4 G) or with Tm ( Fig. 4 A) , where RIDD was clearly functioning. Note that although XBP-1 splicing appears insensitive to low concentrations of DTT in Fig. 4 G, we observed partial splicing shortly after the addition of DTT followed by rapid recovery before changes in gene expression could be reliably measured. For cells expressing wildtype hIre1, ERdj4 and RIDD appeared to be activated in parallel by low DTT concentrations, suggesting that RIDD is not limited to extreme ER stress (Fig. 4 H) . In contrast to splicing and activation of ERdj4, no change in mRNA abundance for RIDD targets was detected in cells expressing hIre1-I642G and treated with 1NM-PP1 alone, even at high concentrations, although the cells were still capable of inducing RIDD when treated with 1NM-PP1 and DTT together (Fig. 4 H) . In summary, although XBP-1 splicing and downstream transcriptional induction can be artificially induced with 1NM-PP1, RIDD cannot. Our results indicate that in mammalian cells, RIDD requires an active Ire1 nuclease domain but does not, per se, depend on Ire1's kinase activity. ER stress-dependent degradation of RNAs by hIre1-I642G was seen at concentrations of 1NM-PP1 1.5-fold higher than those that inhibited the kinase activity of the protein in vitro (Papa et al., 2003) and sixfold higher than those that inhibited kinase activity in HEK293 cells (Lin et al., 2007) . These data, together with the observation that RIDD does not require new transcription, are consistent with a direct mechanism of mRNA target cleavage by Ire1, although we cannot rule out an alternative role for Ire1's nuclease activity and/or the involvement of other nucleases. However, our data also indicate that activation of Ire1's nuclease is not sufficient for RIDD. This observation represents a mechanistic divergence between RIDD and Ire's well-established role in XBP-1 splicing, which is induced by 1NM-PP1 activation of hIre1-I642G. There are several potential explanations for this. It may be that Ire1 assumes a distinct conformation or oligomerization state when activated by 1NM-PP1 versus ER stress and that although the former is sufficient for splicing XBP-1, the latter is required for RIDD. For example, it was recently found that yeast Ire1 forms higher order oligomers that lead to higher levels of RNase activity (Aragon et al., 2009; Korennykh et al., 2009) . Such oligomers may form only in the presence of misfolded proteins or, conversely, in the absence of binding to relative mRNA abundance for ERdj4 and Blos1 (H) in various concentrations of 1NM-PP1 and/or DTT. Measurements are as described for A and B. (A-H) Cells were treated with 2 mM DTT, 3 µg/ml Tm, and/or 7 µM 1NM-PP1 for 5 h unless otherwise indicated. The means and SD for three to five independent experiments are shown. proteins on NuPage Bis-Tris 4-12% acrylamide gels (Invitrogen), transferred them to nitrocellulose membranes, and probed blots using primary polyclonal anti-Flag antibodies (Sigma-Aldrich) at a 1:2,000 dilution followed by secondary peroxidase-conjugated anti-rabbit antibodies (Jackson Immuno-Research Laboratories) at a 1:5,000 dilution. As a loading control, we also probed blots with polyclonal anti-GAPDH (glyceraldehyde 3-phosphate dehydrogenase; ProSci, Inc.). We visualized the immunoblots using a chemiluminescent assay (ECL; GE Healthcare). Online supplemental material Fig. S1 shows the array data for hIre1 / and hIre1 R cell lines, a comparison with published data for UPR targets, and confirmation of RIDD target regulation measured by qPCR. Fig. S2 shows decay rate measurements for 11 RIDD candidates in the presence and absence of DTT. ferritin-like protein recombination target (FRT) sites for site-specific transgene expression by transfection with pFRTlacZeo (Invitrogen), and stable clonal integrants were selected with Zeocin . We then transfected the Ire1- / FRT cells with the pOG44 ferritin-like protein recombinase vector (Invitrogen) and FRT vectors containing hIre1 under the control of the human ER-1 promoter (Lin et al., 2007) . Quik-Change mutagenesis (Agilent Technologies) was used to make the point mutations in the hIre1 coding sequence. Multiple independent isogenic clones were analyzed for transgene mRNA and protein expression with identical findings. The ATP analogue 1NM-PP1 was a gift from C. Zhang and K. Shokat (University of California, San Francisco, San Francisco, CA; Bishop et al., 2000) . To deplete cells of Ire1 or XBP-1 by RNAi, we transfected 3 × 10 4 cells with 150 ng total of a mixture of four different siRNAs (QIAGEN). After 48 h, we treated cells with DTT and/or 1NM-PP1, and purified RNA as described in the next section. We typically attained 90% knockdown as measured by qPCR. We passaged and plated Ire1 / and hIre1 R cells into 150-cm 2 flasks for collection of microarray samples. After a 3-d growth (at 70% confluency), we either left cells untreated or induced ER stress by adding 2 mM DTT for 6 h. We then harvested the cells and purified total RNA using TRIZOL (Invitrogen). To generate labeled samples for array hybridization, we amplified 0.5 µg RNA in a single round using the amino allyl Message-Amp II aRNA kit (Applied Biosystems). As a reference for the array hybridizations, we also amplified and labeled a pool of RNA samples from NIH-3T3, Ire1 / , and hIre1 R cells (both untreated and DTT treated). In parallel, we synthesized unlabeled cDNA from 2 µg of total RNA for qPCR measurements to confirm the array data. We repeated the entire experiment for a total of three times. We hybridized 5 µg each of a Cy5-labeled experimental sample and the Cy3-labeled reference pool to whole-genome mouse arrays at 65°C for 48 h. The arrays were produced in-house using the MEEBOChip platform. We extracted and processed image data using GenePix 6 (MDS Analytical Technologies), normalized each array to achieve a mean Cy5/ Cy3 ratio of 1.0, and removed spots containing low signal or poor signal uniformity. To simplify the analysis, we disregarded spots for which one or more samples did not pass these quality control measures and for which the SD of the three measurements was >30% of the mean. For the remaining spots (representing 25% of the total spots on the array), we calculated the log 2 ratio of DTT treated/untreated for each of the three replicates in the two cells lines. We then performed hierarchical clustering of the averaged data (Fig. S1 ). To select robust RIDD candidates, we applied the following criteria to the 122 spots in clusters displaying Ire1-dependent down-regulation. First, we required the candidate RNA to be down-regulated by 1.5-fold (log 2 [DTT/untreated signal] ≤ 0.58) in at least two of the three replicates and in the mean of the three replicates. Second, to select those RNAs whose down-regulation was truly Ire1 dependent, we required that the targets be down-regulated 1.5-fold more in the hIre1 R cells compared with the Ire1 / cells and display no more than a 25% decrease in signal in response to ER stress in the Ire1 / cells. Lastly, to rule out artifacts caused by higher expression of the candidate RNAs in the hIre1 R cells in the absence of ER stress, we also required that the mean signal intensity in the presence of ER stress was lower (by 15%) in the hIre1 R cells compared with the Ire1 / cells. 26 RIDD candidates fit these criteria; these are listed in Table I. qPCR and XBP-1-splicing assays We purified RNA samples for all experiments using TRIZOL and synthesized cDNA from total RNA samples using Superscript II (Invitrogen). We then performed qPCR measurements using the primers shown in Table S1 . We measured each sample in triplicate using the Opticon (Bio-Rad Laboratories) or Realplex (Eppendorf) qPCR machines and normalized them using the signal from Rpl19, which did not change significantly relative to total RNA input concentration in any of the treatments used. We used mock cDNA samples containing no reverse transcription to ensure that the qPCR signals arose from cDNA and not from contaminating genomic DNA or other sources. To quantify the amount of XBP-1 splicing in each experiment, we amplified cDNA using primers surrounding the splice site (Table S1 ) and ran products on 2% agarose gels. We washed cells in PBS and lysed them in radioimmunoprecipitation assay buffer (25 mM Tris, pH 7.6, 150 mM NaCl, 1% NP-40, 1% Nadeoxycholate, and 0.1% SDS) with added protease inhibitors. We resolved

Empirical Relationship between Intra-Purine and Intra-Pyrimidine Differences in Conserved Gene Sequences

DNA sequences seen in the normal character-based representation appear to have a formidable mixing of the four nucleotides without any apparent order. Nucleotide frequencies and distributions in the sequences have been studied extensively, since the simple rule given by Chargaff almost a century ago that equates the total number of purines to the pyrimidines in a duplex DNA sequence. While it is difficult to trace any relationship between the bases from studies in the character representation of a DNA sequence, graphical representations may provide a clue. These novel representations of DNA sequences have been useful in providing an overview of base distribution and composition of the sequences and providing insights into many hidden structures. We report here our observation based on a graphical representation that the intra-purine and intra-pyrimidine differences in sequences of conserved genes generally follow a quadratic distribution relationship and show that this may have arisen from mutations in the sequences over evolutionary time scales. From this hitherto undescribed relationship for the gene sequences considered in this report we hypothesize that such relationships may be characteristic of these sequences and therefore could become a barrier to large scale sequence alterations that override such characteristics, perhaps through some monitoring process inbuilt in the DNA sequences. Such relationship also raises the possibility of intron sequences playing an important role in maintaining the characteristics and could be indicative of possible intron-late phenomena.

The apparent lack of pattern of composition and distribution of bases in DNA sequences have been one of the enduring problems of molecular biology. Chargaff's rule provided a clear relationship between the numbers of guanines and cytosines, and between adenines and thymines in duplex DNA sequences, understood well by the Watson-Crick model. However, no clear relationship has been found as yet between the occurrences of these four bases in an individual strand of a gene sequence, although much work have been done on understanding nucleotide frequencies and base distributions in DNA sequences. To cite a few examples, Sueoka [1] proposed a unitary theory based on genetic and evolutionary considerations which attempted to account for the main characteristics of the distribution of DNA base composition in nature. The author considered the GC content of DNA sequences of organisms and mutations of GC base pairs to AT base pairs and vice versa over many generations from which a distribution function of the GC content of DNA molecules at equilibrium was obtained based on assumptions of rather uniform mutation and selection pressures affecting base pair conversions. Goldman and Yang [2] proposed a codon-based model for the evolution of protein-coding DNA sequences using a Markov process to describe substitutions between codons. They used codon level information to model synonymous and asynonymous nucleotide substitution applicable to homologous sequences with no insertion/deletion gaps or with gaps removed. A 61661 matrix of codon substitution rates (excluding the three stop codons) is used, assuming that mutations occur at the three codon positions independently and only single-nucleotide substitutions are permitted to occur at any instant. Using several constraints and refinements on the nucleotide substitution rates, the transition/ transversion bias and amino acid differences, they show that their codon based model gives better phylogenies than simple nucleotide substitution model. However, the possible patterns arising from this model is very large and computationally very slow requiring Monte Carlo simulations, maximum likelihood methods and other approximations to arrive at quantitative results. While the model is useful for pairwise distance measures and for phylogenies, a relationship defining base composition in a DNA sequence is not clearly realised. In a recent paper, Qi Ding et al [3] formulated an approach to determine a linear regression model for DNA sequences. By regarding a DNA primary sequence as a random process in time and defining the nucleotides' random distribution functions in three ways based on chemical structures they proposed two methods to measure their similarities. Relating the random distribution functions by a linear regression equation enabled them to construct six new models to analyse the DNA sequences and quantify their similarities and dissimilarities. The optimal model can be chosen based on the amount of information contained or lost in the process. Several other studies have also focussed on nucleotide asymmetries in DNA sequences to gain an insight into correlations, if any, in base composition and distribution. Arnold et al [4] used tetranucleotide frequencies in a third order Markov chain to predict the frequencies of longer oligonucleotides in the yeast genome and observed that the oligonucleotide frequencies depended strongly on base composition. Freeman [5] considered several prokaryotic DNA genomic sequences and found, from base composition asymmetries like purine excess over pyrimidines and coding strand excess, that the global minima of the purine excesses correlated with the origin of replication, and the maxima with the likely terminus for prokaryotic genomes and that such a prominent correlation between the purine excess and replication direction probably leads to excess pyrimidine accumulation in the sense strand and accordingly should increase the less mutationally vulnerable purine content of the coding strand. Prusak and Grzybowski [6] observed that there is a strong non-random distribution of nucleotides in the cytochrome b sequence in several species with the highest differences at the third codon position which is also the codon position of the strongest compositional bias; some species like quail, frog, python and elk appeared to prefer C over A in the light DNA strand whereas species belonging to the artiodactyls contained fewer pyrimidines compared to other species investigated. Such studies serve to also highlight the complexity of base arrangements in a DNA sequence and the difficulties in finding any inherent pattern or signal sequences in such arrangements, especially in a character based representation of the hundreds and thousands of nucleotides comprising the sequences. However, a graphical representation can be expected to provide visual clues to any inherent pattern or regularity as distinct from a purely random distribution of the bases which could be expected to generate a corresponding random distribution in the representative plot. Indeed, a purine-pyrimidine sequence map proposed by Peng et al [7] provided a visual rendition of the growth of purine and pyrimidine numbers in a DNA sequence, which was interpreted by the authors as indicative of an inherent fractal nature in the purine-pyrimidine structure of the DNA sequences. The chaos generator representation of Jeffreys [8] with its double-scoop depletion pattern reflected largely the abundance, or sparseness, of various dinucleotides and higher combinations, and also showed characteristic variations for different classes of organisms. Graphical representations of DNA sequences assigning the four individual nucleotides to designated axes on a Cartesian grid provide a more direct visual indication of the progression of nucleotides in a sequence, akin, in some respects, to the Wilson cloud chamber for particle tracks. The first graphical representation of a DNA sequence was proposed by Hamori and Ruskin [9] based on a 3D Cartesian axis system that generated a visual map of the sequence of bases in a selected DNA sequence. A 2D representation was proposed by Gates [10] and rediscovered independently by Nandy [11] and Leong & Morgenthaler [12] , with different axes identifications for the four bases. These were followed subsequently by many different approaches to render a visual representation of the DNA sequences [13, 14] , and their applications have been found to be very useful in elucidating different characteristics of DNA sequences that are not easily accessible in other ways. For example, Gates [10] referred to large scale structures seen in the plots of some sequences; such structures were also reported by Nandy and Nandy [15] , and recently by Larionov et al [16] who reported, inter alia, the presence of gigantic palindromes in mouse and human chromosomal sequences. Liao et al [17] have shown that 2D graphical methods that convert a DNA sequence to a series of co-ordinates and therefrom construct distance matrices can be used for computation of molecular phylogeny without need for multiple alignments; Wang et al [18] constructed a 3D representation in which a DNA sequence could be denoted mathematically and a similarity matrix constructed for multiple sequences to derive a phylogenetic tree by virtue of the fuzzy theory. Lo et al [19] have shown using a 3D trajectory method that global views of DNA sequences can be obtained such that different types of DNA sequences can be easily distinguished and any local differences and similarities between two DNA sequences can also be easily observed. Furthermore, numerical characterisation techniques based on graphical representations have enabled quantitative estimations of sequence similarities and dissimilarities [see review 14] . Basically there have been two approaches for numerical characterization, both of which use the graphical representation to map a DNA sequence into a set of numbers. One approach using geometrical mapping proposed by Raychaudhury and Nandy [20] have been found to be useful for several calculations based on the 2D graphical representation [14] , and extended recently to an abstract 20D modelling for protein sequences [21] , where individual sequences are indexed by numerical descriptors. The other approach is to use matrix methods by forming ratios of graph theoretic and Euclidean distances between nodes of the graphical plots, first formulated for DNA sequences in Randic et al [22] . Since invariants associated with matrix formulation are well-known, individual sequences can be indexed by one or more such invariants of various orders; it is expected that these would be sufficiently characteristic of the underlying sequences to enable unique characterization. This technique has been the most widely used method of choice for the researchers in this field who have defined different types of matrices to construct various invariants to describe the DNA sequences. However, the difficulties associated with computing various parameters for very large matrices that are natural for large sequences have restricted the numerical characterizations to leading eigenvalues and the like [14] . In principle, however, many of the indices used to characterize numerically DNA representations are graph invariants that describe the distribution of nodes and/or node-node connections in these graphs. In the parlance of graph theory, many authors have referred to some of these indices as Topological Indices (TIs) and applications have been made not only to DNA sequences but also to proteins, viral surfaces, RNA secondary structures and small molecules [23, 24] . Consequently, the method is of more general application taking into consideration that the type of graph representations referred to above have been extended from DNA/ RNA to the study of other types of relevant biological sequences. In particular, González-Díaz et al. extended these representations to the study of protein sequences [25] and Mass Spectra outcomes of proteins and/or protein serum profiles in parasites [26] , toxicoproteomics and diagnosis of cancer patients [27, 28] . In any case, the various numerical parameters of DNA/RNA graph representations (TIs or otherwise) may be used not only to study sequence-sequence similarity but also to fit Quantitative Structure-Activity Relationship (QSAR) models. These QSAR connect structural information with the biological function of a molecular entity under study and may be used to predict unknown entries. Structure here refers not only to drug structure but also to DNA sequence, RNA sequence or secondary structure, and protein sequences or 3D structure [see review 28] . Thus, the utility of graphical methods in revealing different types of hidden structures and similarities/dissimilarities in and between DNA and other biological sequences can be considered to be well demonstrated. In this light, a perusal of the representative patterns of conserved gene sequences appears to indicate a possible relationship between the numbers of the various nucleotides in conserved gene sequences. Here we use the 2D graphical representation method to show that plots of the conserved gene sequences trace out apparently curved paths that are also visually similar across species for the same gene. The nature of these curves is seen to generally imply a so far undescribed quadratic relationship between intra-purine and intra-pyrimidine differences, whereas the null hypothesis would have indicated random directionless walks. With such empirical relationship between the two basic nucleotide differences, we propose a probable mutation path to explaining the relationships. We then hypothesize that the parameters of such relationships could be a property of the underlying gene sequences and speculate that extensive alterations of such genes by accretion or deletion of DNA fragments would be s only if the modified sequences subscribe to the same basic parameterised relation. Here we use the random walk system envisaged in the Nandy plot [11] based on a 2D Cartesian grid where the four bases are assigned to the four cardinal directions: guanine (g) to the positive x-direction, thymine (t) to the negative y-direction, adenine (a) to the negative x-direction and cytosine (c) to the positive y-direction. The method to plot a DNA sequence is to start at the origin and take a step in the positive x-direction for a guanine base, in the negative x-direction for an adenine, positive y-direction for a cytosine and the negative y-direction for a thymine, and proceed likewise for each succeeding base in the sequence, starting each step from the end of the last one taken. This way a succession of bases in the original DNA sequences is represented by a succession of points in the 2D plot, the overall trace being a representation of the distribution of bases in the DNA sequence (see e.g., Fig. 1 human beta globin). The axes essentially represent the excess of guanine over adenine along the x-axis and the excess of cytosine over thymine along the y-axis; thus the plots are basically of instantaneous values of intra-pyrimidine, intra-purine differences as we proceed along the sequence. The end point of such a curve will be given by (N G -N A , N C -N T ), where by N A , N C , N G, N T we mean the total number of adenines, cytosines, guanines and thymines in the sequence being plotted. GC-rich sequences therefore plot mostly in the first quadrant, AT-rich sequences in the third quadrant on this axes system. Once the co-ordinates are available, we use an Excel spreadsheet to plot the graph and apply the Add Trendline feature of the Excel software to fit the best polynomials for our analysis, with axes transformation where required to conform to the software's curve fitting engine. Fig. 1 shows a plot of the human beta globin gene complete cds generated using the above algorithm. We note that a DNA sequence that consists of a succession of short segments each having a complete mix of a,g,c,t with equal contributions of each of the bases within each of the segments would be expected to generate a dense cluster of points around the origin; a random distribution of the a,g,c,t along the sequence could be expected to generate a random walk. The human beta globin gene sequence complete cds inclusive of all introns and exons (Fig. 1) shows a distinct pattern where the bases appear to follow one another with some regularity, with the total extent of the representative plot arising from the non-equal composition of the bases in the sequence; other beta globin sequences produce similar plots implying that the human beta globin gene cds is not an arbitrary random sequence. Fig. 2 shows the close similarity of the shapes of the plots of three sequences of histone H4 genes of wheat, maize and chicken, demonstrating that these sequences are not random but have a close kinship in base distribution. A randomisation of the bases in the human beta globin gene [29] produces, on the other hand, a simple linear plot (Fig. 3) in the third quadrant of the axes system as can be intuitively expected for an unorganised mixture of the four bases along the sequence. To further demonstrate that the base distribution in these gene sequences are non-random, and generally true for conserved gene sequences, we have generated graphical representations of several conserved genes, a selection of which are shown in Fig. 4 . Plots of several alpha globin genes are included here to show that there is shape similarity between the same genes from different species indicating that entire gene sequences inclusive of introns and exons have close similarities. This can also be seen in the plots of several histones, tubulins and heat shock proteins, of which some representative samples are given in Figs. 4 and 5. With our observation that the sequences of the conserved genes, both intronless and with-introns, have close similarities, we can start to enquire whether these sequences have any discernible patterns. We notice that the general nature of the DNA walks on the 2D representation as per the Nandy prescription [11] shown in the above figures is directional and curvilinear. A simple 2 nd degree polynomial produces reasonable fits. A selection of the plots with the trendlines is shown in Fig. 6 and 7 corresponding to the sequences shown in the previous two figures. A list of the details of the fits and statistics are given in Table 1 . While this nature of the base distributions is found across many different sequences of eukaryotic genes, it is also very much evident in the case of viral sequences like the H5N1 neuraminidase which are known to mutate very rapidly; other plots from a sample of over 600 sequences of the H5N1 neuraminidase show similar quadratic forms. Interestingly, plots of the wheat, maize and chicken histone H4 genes, which are also intronless genes, can also be fitted by polynomials of degree 2, similar to the case of the viral genes. Chicken beta globin gene with intron sequences that are quite different compared to the mammalian genes plot in the first quadrant, but it also can be fit by a quadratic polynomial. This is however not true of all gene sequences. The mouse beta globin gene sequence representation in the 2D framework gives a very poor fit for the quadratic function (R 2 = 0.06) but much better statistics with a cubic polynomial (R 2 = 0.53) ( Table 1) ; the human delta globin gene also shows very good statistics when fitted with a cubic polynomial (R 2 = 0.98). Some sequences where significantly large segments are at variance with the overall pattern too cannot be put into such slots. The rat myosin heavy chain gene sequence where the intron sequences have been hypothesized to have grown through accretion of large fragments [30] , for example, presents a highly compact form on the 2D plot [15] and cannot be fit by the simple polynomials we have used so far. However, sufficiently large numbers of sequences are seen to follow the apparent quadratic relationship with good statistics that it is of interest to try to understand the underlying pattern. The equations that fit the curvilinear patterns of the base distributions with reasonable statistics are of the form and where a,b,c,d are parameters, and n A , n C , n G , n T are the instantaneous values of the numbers of a, c, g, t present up to the particular position (x,y) on the sequence, starting the count from the beginning, i.e. the 59-end, of the sequence. These are our empirical equations connecting the intrapurine (x) and intra-pyrimidine (y) numbers obtained from the observations of the patterns on the 2D graphical plots. While the majority of the plots shown are well-represented by such polynomials of the second degree, the fits could be improved in some instances by fitting higher degree polynomials as mentioned earlier; e.g., in the case of the human beta-globin gene a polynomial of the third degree yields better statistics (R 2 = 0.94) than the second degree (R 2 = 0.81), for the human delta globin gene the statistics for the quadratic and cubic fits are R 2 = 0.87 and R 2 = 0.98, respectively. We, however, consider the second degree form for now for conformity without excessive loss of statistical significance. The origin of such a relationship as in equations 1 and 2 could be traced to mutational changes in a sequence, where we restrict our analysis for the moment to transitional types of mutations since this is the dominant mode. Consider a mutation of a cytosine to thymine in one strand of a DNA in a GC-rich sequence. The opposite strand, calling it strand 2 for convenience, will initially have a bulge for the original paired guanine, and the event leads to following possibilities [31] : (a) the DNA repair mechanism reverse mutates the thymine to cytosine in strand 1, thus negating the effect of the original mutation; (b) the guanine in strand 2 is replaced by an adenine; and (c) a third possibility in case the damage repair coincides with replication, that the DNA is elongated by pairing the mutated thymine in strand 1 with a new adenine in strand 2 and addition of a cytosine in strand 1 to match the guanine in strand 2 left over after the original mutation, i.e. creation of a T-A pair and addition of a new C-G pair. Such an event would be quite rare, especially in coding regions since it will alter the reading frames unless the total change leads to addition of three base pairs; several intronless gene sequences indeed show very small contribution from the quadratic term compared to the interrupted genes, e.g. petunia hsp70G ( Table 1 ). The example of the mutation event of cytosine to thymine in strand 1 can be considered to change the intra-pyrimidine difference, n C-T , in strand 1 and trigger a change in the intrapurine difference, n G-A , in strand 2. The two changes can be related by where the first term on the right relates to the probability of the G to A change in strand 2 and the second term to the probability of the third type of response, i.e. DNA elongation by addition of a C in strand 1; the (1) and (2) in equation 4 are strand identifiers and we have defined We expect q,p since the probability of DNA elongation will be significantly lower than effecting only a replacement of the Using recursion and keeping up to first order terms in q/p, equation (7) reduces to where by O3 we mean terms of higher orders (dropping the conventional form O(3) so as not to confuse with the strand number indicators). For many mutations over a long sequence, this takes the form where a,b are redefined constants with b,a. This equation relates changes in the intra-pyrimidine numbers in strand 1 to intrapurine numbers in strand 2 where the n C-T and n G-A are as defined in equations 5 and 6 above. Now, Chargaff's rule states that and since from Watson-Crick rule we know n C 1 ð Þ~n G 2 ð Þ and n C 2 ð Þ~n G 1 ð Þ, and similarly for A,T, then from the above we have as a consequence of which, Excluding some pathological cases or very short segments where one or the other type of nucleotide is conspicuous by its absence, e.g., as in poly-adenylation segments where only A's dominate and the others are absent, in a real gene sequence we can expect the second term within each of the square brackets above to be%1, implying that which transforms Eq. (9) to or, dropping the strand number indicator since all quantities now refer to the same strand, From similar considerations for the AT-rich sequences considering the mutations of adenine to guanine for example, we can obtain the equation where n C-T and n G-A are as defined in Eqs. 5 and 6. Equations 16 and 17 are similar to equations 1 and 2 and conform to the general shapes of the sequence plots. This shows that the path traced out by the nucleotide sequence of a gene follows some pattern that can be ascribed to the accumulated effects of spot mutations over evolutionary time scales. The actual plots of the gene sequences in the 2D grid representation show reasonable conformity with the predicted curves with variations that could be ascribed in part to the higher order terms and other factors described below. It is to be noted that equations 16, 17 have been derived independent of any graphical representations or reference frames, and they are functions of the instantaneous values of the base counts only. The 2D Nandy representation here used happens to be a natural and convenient reference frame to plot the outcome of these equations. Equations 16-17 have been developed on the basis of transition type mutations only. Transverse mutations are much less frequent than the transition mutations and will affect these equations to a smaller extent. Consider a mutation of a cytosine to an adenine. This will result again in a bulge in the guanine in the opposite strand and the repair mechanism will either re-establish the cytosine or replace the guanine with a thymine, or replace the guanine in strand 2 with a thymine and add a CG base pair to the DNA thus elongating it. This will reduce the n C-T (1) by 1 unit and also reduce the n G-A (2) by 1 unit as well as have, as before, a small probability of adding an extra guanine in strand 2 by virtue of the additional base pair elongating the DNA. Thus we will have an equation of the type of Eq.4 again except that the coefficients in this instance will be much lower in value than for these parameters in the transition type mutation case. Again, we end up with equations of the form 16 to 17, this time for transverse type mutations, but with much smaller parameter values, so those equations can be considered to be quite general. It is also possible that as a result of mutations more than one base pair are added during the repair and replication phases. Such actions can be expected to lead to further higher order terms being added to equations 16, 17 . That the third order terms seem to be required for good fits to some of the gene sequences has already been noted earlier. Such modifications to the intra-purine, intrapyrimidine relationship can result in deviations from the smooth distribution that first order approximation equations like Eq.16, 17 would imply. This could be the underlying reason for the deviations between the actual representative plots of the various sequences and their fitted curves observed in several graphs. There is a further assumption that underlies the applicability of this analysis. We assume that in the case of the conserved gene sequences inclusive of introns and the coding segments, extensive restructuring through recombinations have not taken place. Indeed the first order formalism developed here implicitly assumes that all changes to a DNA sequence that have resulted in the form that we observe now have taken place through mutational changes only, and recombinations and transpositions have not had any major impact on the sequences. A couple of points of interest arising from these equations may be mentioned here. First, homologous sequences of conserved genes of the intronless or with-introns varieties that have the same general shape on the 2D plots therefore have similar describing equations, and the coefficients of the variables have similar order of magnitude values. This would seem to indicate that these gene sequences have inherent characteristics that are expressed by the values of the parameters of the describing equations, whereby major deviations in base distributions that necessitate large departures from the characteristic values could be inimical to the functioning of the gene and thus would either be rejected, or would render the gene ineffective. A case could be made from the human alpha globin 1 pseudogene: although it shares a reasonable degree of homology with the functional alpha 1 globin gene and has the 3 exon-2 intron architecture of the globin family, its 2D plot shows a wide variation in base distribution from the alpha globin gene (Fig. 8) , and the coefficients of fitted polynomial also are quite different from the other mammalian alpha globin family fits (Table 1) . It may also be mentioned that transposons are found in many instances to insert segments into genes which are then excised out in successive replication cycles. If DNA sequences have inherent characteristics, which are encapsulated by the polynomial expression, and the inserted segments lead to incompatibility with such characteristics, then such excisions can be understood. We note that reversion of mutated genes to ancestral forms is not totally unknown. A case in point is the recent study of reversion of mutant hothead gene in Arabidopsis thaliana to genes that existed in plants of two or more generations ago [32] . The authors have hypothesized that a template-directed restoration of ancestral DNA passed on in an RNA cache could underlie the mechanism of such reversion; the existence of such a mechanism that lies outside the DNA genome could lead to the high-frequency modification of DNA sequences in a template-directed manner, perhaps by the postulated RNA cache that could allow it to persist for several generations. While Lolle et al's RNA cache [32] would need to carry an exact duplicate of the ancestral sequence, in the case of gene sequences of the types considered here that may or may not contain intron segments and could be quite large, we could postulate existence of some as yet unknown mechanism for monitoring conformity with the overall intra-purine intra-pyrimidine base distribution pattern, as perhaps for long range correlations [7, 33] . In this connection it is also interesting to consider the possibility of the existence of some error correcting code in DNA sequences as speculated upon by Liebovitch et al [34] . They considered the DNA sequence as a digital code of four symbols and speculated that since the integrity of modern information encoding is secured by having error correcting codes built in, DNA sequences might also have such codes to allow repair enzymes to protect the fidelity of nonreplicating DNA and increase the accuracy of replication. In such a case if a linear block error correcting code is present in DNA then some bases would be a linear function of the other bases in each set of bases. Although the authors were unable to find any such simple code in the lac operon and cytochrome c gene they investigated, the suggestion remains an intriguing possibility nevertheless. Given that we are considering highly conserved genes whose functions are important to the survival of the organism, mechanisms such as these would provide a survival advantage and could be used under conditions that compromised the continued functioning of the organism or the requirements of the monitoring process. Second, the basic patterns, and therefore the describing equations, have the same form irrespective of the intron content of the genes considered here. This could be indicative of the monitoring process hypothesized above also functioning irrespective of whether the sequences are intronless or intron-rich. Since mutations and other evolutionary changes have led to modifications in the coding sequences within the requirements of maintaining protein functionality, the introns may have a role to play such as maintaining the structure of base arrangements so that the restrictions implied by the equations 16, 17 can continue to apply. We could perhaps consider for support for this contention the globin genes where the beta globins separated from the alphas quite late, but the introns of the beta globin cluster are generally longer than the alpha globins while the coding sequences, though with differences, remain at almost the same length. If sustained, such a hypothesis could lend support to the intron late theory. We note in passing that for the majority of the gene sequences of the vertebrates, the 2D plots display an overall shape that is concave going in the clockwise direction; correspondingly, the coefficient of the second degree variable is negative. This is seen even in the case of the chicken beta globin gene where the large intron component makes it GC-rich, the human beta globin gene which is AT-rich, as also the alpha globin genes of the horse, rhesus monkey and human which are all GC-rich. Plots of some viral sequences such as the H5N1 neuraminidase, on the other hand, have concavity in the opposite direction and the sign of the coefficient of the second degree variable is positive. Interestingly this is also seen in the case of the wheat histone H4 gene, which is claimed by some authors to have viral features [35] ; their origins from before the eukaryote split could also be a factor. Whether this is just chance coincidence or whether it is symptomatic of some deeper characteristic arising from base composition and distribution differences in prokaryotic and eukaryotic sequences remains an intriguing question. In summary, our observations have shown that the 2D plots of intra-purine versus intra-pyrimidine differences in conserved gene sequences exhibit an apparent pattern in base distribution of the sequences that mimic the behaviour essentially of a polynomial of degree 2, and in some cases of degree 3. This is found over a wide cross section of sequences, from e.g., the members of the globin family, the histones, tubulins and heat shock proteins. Viral sequences such as the H5N1 neuraminidase, although known to mutate rapidly, also exhibit similar structure. We have seen that this may arise from the non-symmetrical mutation repair mechanism where e.g., a cytosine mutating to thymine in a GCrich sequence could lead to negating the mutation, to replacing the original paired guanine with adenine, or elongating the DNA by addition of a CG pair along with coupling the thymine with an adenine. Equivalent considerations apply to AT-rich sequences as well. Since these observations appear true for intron-rich sequences also, the intron sequences may play a regulatory role in preserving sequence integrity as indicated by the intra-purine intra-pyrimidine relationships permitting greater flexibility in changes in coding sequences. Not unexpectedly, we have seen that homologous genes have characteristic equations where the coefficients of the describing polynomials are quite close. Assuming that the DNA fidelity processes fit to this scheme of preferential arrangement of bases in conserved segments, our observations raise the possibility that DNA fragments, introduced into such segments by processes such as transpositions, that do not conform to the overall fit may be preferentially excised by the replication machinery to retain the integrity of the host sequence. If our observations here in gene sequences are extendible further to genomic sequences then it would imply that not all genetic modifications would be sustainable.

Willingness of Hong Kong healthcare workers to accept pre-pandemic influenza vaccination at different WHO alert levels: two questionnaire surveys

Objective To assess the acceptability of pre-pandemic influenza vaccination among healthcare workers in public hospitals in Hong Kong and the effect of escalation in the World Health Organization’s alert level for an influenza pandemic. Design Repeated cross sectional studies using self administered, anonymous questionnaires Setting Surveys at 31 hospital departments of internal medicine, paediatrics, and emergency medicine under the Hong Kong Hospital Authority from January to March 2009 and in May 2009 Participants 2255 healthcare workers completed the questionnaires in the two studies. They were doctors, nurses, or allied health professionals working in the public hospital system. Main outcome measures Stated willingness to accept pre-pandemic influenza vaccination (influenza A subtypes H5N1 or H1N1) and its associating factors. Results The overall willingness to accept pre-pandemic H5N1 vaccine was only 28.4% in the first survey, conducted at WHO influenza pandemic alert phase 3. No significant changes in the level of willingness to accept pre-pandemic H5N1 vaccine were observed despite the escalation to alert phase 5. The willingness to accept pre-pandemic H1N1 vaccine was 47.9% among healthcare workers when the WHO alert level was at phase 5. The most common reasons for an intention to accept were “wish to be protected” and “following health authority’s advice.” The major barriers identified were fear of side effects and doubts about efficacy. More than half of the respondents thought nurses should be the first priority group to receive the vaccines. The strongest positive associating factors were history of seasonal influenza vaccination and perceived risk of contracting the infection. Conclusions The willingness to accept pre-pandemic influenza vaccination was low, and no significant effect was observed with the change in WHO alert level. Further studies are required to elucidate the root cause of the low intention to accept pre-pandemic vaccination.

In 2005 the World Health Organization recommended its member states to revise or construct a preparedness plan for pandemic influenza. The WHO also set up a system of influenza pandemic alert levels. Phases 1-3 include capacity development and response planning, while phases 4-6 signify the need for response and mitigation efforts. 1 By August 2008, 47 countries had prepared such a plan. 2 The recent spread of infection with a novel influenza A virus (H1N1 subtype) of swine origin ("swine flu") has prompted governments to review and carry out their pandemic responses, including vaccination strategies. Modelling studies have shown that vaccination is an effective measure to reduce infection, hospitalisation, mortality and morbidity. 3 Since the supply of vaccines will be limited at the beginning of the influenza pandemic, prioritisation in the administration of the vaccines has been one of the major components in pandemic preparedness. Governments in different countries have issued consultation documents outlining their proposed vaccination policies. 4 5 In nearly all countries with a preparedness plan, healthcare workers are listed as the priority group for mass vaccination. 6 7 The American College of Physicians position statement supports measures to increase the supply of influenza vaccine and antiviral drugs in the strategic national stockpile. 8 Pre-pandemic H5N1 vaccine is now ready and stockpiled by many countries. The first batch of H1N1 vaccine is expected to be available by the end of 2009 or early 2010. 9 More than 30 governments have placed their orders. 9 For example, Hong Kong has decided to buy five million doses for its population of seven million, and the UK Department of Health has ordered 130 million doses of vaccine for its population of 61 million. However, healthcare workers' acceptance of the H5N1 and H1N1 vaccines is unknown. A survey conducted in May, when the pandemic level was already at phase 5, revealed that the general public in Hong Kong did not perceive swine flu (H1N1 influenza) as a threatening disease and did not think an outbreak to be highly likely. 10 The uptake of pre-pandemic vaccination among healthcare workers is a concern as the uptake of seasonal influenza vaccine is often low. In most studies, fewer than 60% of healthcare workers were vaccinated against seasonal influenza in various clinical settings. The most common barriers were fear of side effects, uncertainty about the vaccine's efficacy, and misconceptions about the vaccination and the infection. [11] [12] [13] [14] [15] One study in the UK of 520 participants found that 57.7% of healthcare workers expressed willingness to accept vaccination with the stockpiled H5N1 vaccine. Willingness was associated with perceived risk and benefits, and previous seasonal vaccination. 16 The aim of our study was to investigate the stated acceptability of pre-pandemic vaccination (H5N1 or H1N1 vaccine) and its associated factors among public hospital based healthcare workers in the departments of accident and emergency medicine, internal medicine, and paediatrics in Hong Kong. We also investigated the effect of escalation of the WHO pandemic influenza alert level on the acceptability of pre-pandemic vaccination. The first survey was conducted from January 2009 to March 2009. The WHO influenza pandemic alert level assigned to H5N1 during that period was phase 3. Phase 3 signifies an animal or human-animal influenza reassortant virus that has caused sporadic cases or small clusters of disease in people but has not resulted in human to human transmission sufficient to sustain community level outbreaks. Limited human to human transmission may occur under some circumstances, such as when there is close contact between an infected person and an unprotected carer. However, limited transmission under such restricted circumstances does not indicate that the virus has gained the level of transmissibility among humans necessary to cause a pandemic. We recruited healthcare workers only in public hospitals in this study because 94% of secondary and tertiary healthcare services in Hong Kong are provided by these hospitals. Pandemic influenza patients will be primarily treated in these hospitals-as occurred in the outbreak of SARS (severe acute respiratory syndrome), when public healthcare workers were at the highest risk for contracting infection. All department heads (42 departments with 8508 doctors and nurses) of emergency medicine, internal medicine, and paediatrics units under the Hong Kong Hospital Authority were contacted through emails, invitation letters, or telephone calls to obtain approval to send the questionnaires to their staff. These departments were chosen as their staff are at the highest risk of exposure to patients with influenza. We also sent invitation letters to the physiotherapy and occupational therapy departments under the Hospital Authority. Self administered, anonymous questionnaires were used in both surveys. The questionnaire consisted of five sections with 17 questions: (1) demographics, patient contact, and history of seasonal influenza vaccination in 2008-9; (2) willingness to accept pre-pandemic vaccination with H5N1 vaccine; (3) willingness to accept pandemic vaccination with H1N1 vaccine (the second survey only); (4) perception of risk and seriousness of the pandemic influenza; and (5) suggestions on deployment of duties during pandemic flu, opinion on mandatory vaccination, and the possible ways of disposal for nearly expired vaccines. The participants were requested to send the completed questionnaires via internal mail. The second survey was conducted in May 2009 when the WHO pandemic influenza alert level assigned to H1N1 influenza (swine flu) was phase 5. Phase 5 signifies human to human spread of the virus into at least two countries within one WHO region. Although most countries are not affected at this stage, the declaration of phase 5 is a strong signal that a pandemic is imminent and that the time to finalise the organisation, communication, and implementation of the planned mitigation measures is short. 1 During this phase 5 period, we repeated our questionnaires in the three specialties in one hospital. All questionnaires were collected within two weeks, before the announcement of phase 6 by the WHO. Descriptive statistics were performed. Using cross tabulations, we analysed univariate associations between intention to accept vaccine and the following variables: sex, age (≤30 v >30 years), specialty, job title, years working in health services, work site, weekly number of contacts with patients, whether the respondents had seasonal flu vaccination in 2008-9, how likely they thought they were to get flu if there was a pandemic, and how seriously they thought a pandemic would affect their lives. We tested the statistical significance of the associations using Pearson's χ 2 test, Fisher's exact test, or trend tests where appropriate. Trend tests were used to test associations between one binary and one ordinal variable, and for two ordinal variables if a trend was apparent in the data. Multiple logistic regression was used to evaluate independent predictors of intention to accept vaccine. Demographic variables and variables on vaccination history and attitudes were tried in the models. A flexible modelling approach was adopted, and variables were retained in the model if they had P<0.1. We compared the results in the second survey with the results from the same hospital in the first survey to assess the effect of escalation in pandemic alert level on the willingness to accept vaccination. Differences in characteristics between the two surveys were evaluated using Fisher's exact test. Of the 4006 questionnaires distributed for the first survey, 1866 were completed and returned, giving a response rate of 46.6%. Of the 42 targeted units, 31 (73.8%) participated (including emergency medicine, internal medicine, and paediatrics units), representing 20% of all doctors and nurses working in these units in Of the 810 questionnaires distributed in the second survey, 389 were completed and returned. The response rate was 48.0%. Details of the second survey are shown in figure 2. All three invited departments had also participated in the first survey. The age and sex distribution of respondents were similar to those of the respondents in the same hospital in the first survey. Demographics of all respondents in both surveys are shown in table 1. The overall intention to accept pre-pandemic vaccination (H5N1 vaccine) was only 28.4% for the first survey, which was conducted at WHO influenza pandemic alert phase 3. The level of intention to accept increased to 34.8% in the second survey, when the WHO alert phase was level 5. Responses from three departments in the hospital where both surveys were conducted are shown in table 2. No significant changes in the level of intention to accept pre-pandemic vaccination (H5N1 vaccine) was observed, despite the escalation to phase 5 because of the wide spread of H1N1 virus (swine flu). The proportion of healthcare workers intending to accept pre-pandemic vaccination (H1N1 vaccine) was 47.9% when the WHO alert level was at phase 5. The respondents who were willing to accept H5N1 vaccines were more likely to accept H1N1 vaccines as well (91%); in contrast only 23.6% of those who declined H5N1 vaccination expressed an intention to accept H1N1 vaccination (P<0.0001). The most common reasons for intending to accept vaccination were "wish to be protected" and "following Health Authority's advice" (fig 3) . The most common reason for refusal was "worry about side effects," and other reasons included "query on the efficacy of the vaccine," "not yet the right time to be vaccinated," and "simply did not want the vaccine" (fig 4) . More than half of the respondents thought that nurses should be the first priority group to receive the vaccines, followed by doctors and allied health professions, and then similar ratings for non-clinical and administrative staff. About half of the respondents (52.2% in the first survey and 56.0% in the second) wanted their family members to receive the vaccines as well. All the above responses remained constant in the different WHO alert phase levels. Univariate associations between intention to accept H5N1 vaccination and other variables at WHO alert phase 3 are shown in table 3. Male sex, working in a specialty other than internal medicine, being a doctor, having fewer years of work in the health services, having received seasonal influenza vaccination in 2008-9, and perceptions that they were likely to contract the influenza and that a pandemic would seriously affect their lives were all significantly associated with greater intention to accept vaccination. In multiple logistic regression models for intention to accept vaccination (table 4), all of these variables remained significant except for specialty, which became marginally significant. At WHO alert phase 5, only having received seasonal influenza vaccination in 2008-9 and younger age were found as significant associated factors for intention to accept H5N1 vaccination in multiple logistic regression (table 4) . For H1N1 vaccination, the factors showing a significant association with intention to accept at WHO alert phase 5 after adjustment by multiple logistic regression included younger age; having received seasonal influenza vaccination in 2008-9, and the perception that they were more likely to contract the pandemic influenza. The results are shown in table 5. The surveys conducted at WHO pandemic alert phases 3 and 5 found a consistently low level of willingness to accept pre-pandemic influenza vaccination among hospital based healthcare workers, especially in those working in the allied health professions. This is particularly surprising in a city where the SARS outbreak had such a huge impact. The intention to accept vaccination against H1N1 influenza (swine flu) in our study was less than 50% even at WHO alert phase 5. On 21 May 2009, the WHO stated that 41 countries had officially reported 11034 cases of swine flu, including 85 deaths. 17 This was the time when our survey was conducted. It was surprising that neither this information nor the escalated WHO alert phase affected the intention to accept pre-pandemic vaccination. Vaccination is one of the potentially effective measures that can reduce mortality and morbidity from pandemic influenza. On 13 July 2009, the WHO also recommended that all countries should immunise their healthcare workers as a first priority to protect the essential health infrastructure. 18 However, the effectiveness of this measure depends heavily on the uptake rate in those groups assigned high priority. Therefore, the low potential acceptance of this vaccine warrants our attention, with a view to improving acceptance. The factors with the strongest association with intention to accept pre-pandemic or pandemic vaccine were history of seasonal influenza vaccination and perceived risk of contracting the infection. The strong association between acceptance of seasonal and pre-pandemic vaccination also suggests similar barriers exist for both vaccines. Efforts to improve the uptake of seasonal influenza vaccination by healthcare workers should therefore be a part of the pandemic preparedness plan, as disseminating correct information may be more difficult at time of crisis, and the health belief model could be applied to improve the acceptance of pre-pandemic vaccine as in seasonal influenza vaccination. 19 The major barriers to vaccination we identified were fear of side effects and questions about its efficacy. This suggests that public and hospital health agencies need to provide more information to the staff, especially to those with higher levels of anxiety and doubt. In our study younger staff and staff whose working experience was less than 5 years were more willing to accept vaccination. This again implies that the experience of the SARS outbreak did not enhance the willingness to accept the vaccination. Comparison with another study The willingness of Hong Kong healthcare workers to be vaccinated against H5N1 virus is very low when compared with the study done in a UK NHS trust, where more than half of the staff were willing to accept pre-pandemic vaccination when surveyed at a similar WHO alert level (phase 3). 16 The uptake rates for seasonal influenza vaccine were higher among our participants than in the UK study (32.9% for Hong Kong and 15.6% for UK healthcare workers), whereas the proportion willing to be vaccinated against H5N1 influenza in the UK study (57.7%) was double that in our survey (28.4%). Whether this higher willingness to accept was only temporary as a result of the well publicised H5N1 outbreak at a poultry farm in the UK when the survey was started remains to be explored. The uptake rate for seasonal influenza vaccine in Hong Kong varies among the target groups. A previous study on patients attending a general outpatient clinic, of whom half had chronic illnesses, found a vaccination rate of 27%, but without correlation with sex, occupation, or household income. 20 The uptake rate among elderly people aged >65 years living in the community was also low (26.9%-36.4%). 21 In contrast, more than 90% of elderly people living in institutions received influenza vaccine, which was delivered on site. 22 The overall vaccination rate for elderly people in Hong Kong could be within the range reported from surveys in the UK, Italy, France, Germany, and Spain conducted from 2003 to 2007 (41.3%-78.1%). 23 As for healthcare workers, both Hong Kong and European countries face a low uptake rate. The 32.9% vaccination recorded in the current study is close to the ranges reported for the UK (15.9%-25.2%), Italy (13.3%-23.2%), Spain (20.5%-28.9%), Germany (22.0%-27.5%), and France (15.8%-22.2%). 23 The Department of Health in Hong Kong provides a comprehensive immunisation programme from birth to 12 years old. There seems to be no major generic barrier to vaccination in Hong Kong, as the uptake for childhood immunisation programmes is high (84.7%-99.6%), 24 and a recent survey indicates a high level of willingness (88%) to accept human papillomavirus vaccine, 25 similar to that recorded in the UK (89%). 26 While cultural differences could affect the acceptance of vaccines in general, we believe there are common barriers to influenza vaccination that exist across geographical regions and racial groups. 27 The findings of this study can therefore serve as a reference for other countries that are planning to offer the H1N1 vaccine to their healthcare workers. Strengths and weaknesses of this study To our knowledge, this is the largest study conducted to assess the willingness of healthcare workers to accept pre-pandemic influenza vaccination, and it provides important information on barriers to vaccination. Campaigns to promote vaccination should consider addressing the knowledge gap of staff and the specific target groups for intervention. Our study also captured the effects of change in WHO alert level on people's perception and willingness to accept vaccination. The main limitation of this study is the response rate of below 50%. The low response rate may have resulted in a biased sample. Another caveat is the lack of details for the non-responders. Nevertheless, the characteristics of the participants matched the overall staff profile, and the participating specialty departments represented over 70% of our target population. An additional limitation is that the study only documented what people said they would do and thus may not reflect the actual vaccine uptake rates. A follow-up study will be needed to capture the true uptake rates and factors associated with vaccine uptake when it is available. Further qualitative studies such as focus group discussion or semi-structured interviews could help to consolidate and supplement the findings. We believe this information can assist governments to design their pandemic vaccination plan for healthcare workers, taking into account their opinions on these contentious issues. A successful vaccination strategy does not just protect the health of healthcare workers but also can limit the transmission between the health sector and the community, a lesson from the SARS outbreak. Qualitative studies are being conducted by our group to explore barriers faced by healthcare workers in the uptake of vaccination. With the reported low level of willingness to accept pre-pandemic vaccination in this study, future work on intervention to increase vaccination uptake is warranted. We thank Dr Ian Stepheson for his invaluable advice and experience in constructing the questionnaires and generous support from the chiefs of service in the participating departments. We also thank Ms Kate TC Ng for organising the questionnaire administration. Contributors: JSY Chor and PKS Chan designed the study, interpreted the findings, and wrote the manuscript; KLK Ngai collated the survey data; WB Goggins and SYS Wong performed statistical analyses; MCS Wong, N Lee, TF Leung, TH Rainer administered the survey; S Griffiths supervised the study. All authors, external and internal, had full access to all of the data (including statistical reports and tables) in the study and can take responsibility for the integrity of the data and the accuracy of the data analysis. Funding: This study did not receive any external funding. Competing interests: None declared. Ethical approval: The study was approved by the Survey and Behavioural Research Ethics Committee of the Chinese University of Hong Kong. Data sharing: no additional data available. The WHO recommended that healthcare workers in all countries should get top priority for vaccination against the influenza A H1N1 virus of swine origin The acceptance of seasonal influenza vaccination in healthcare workers worldwide is low. One study conducted at WHO alert phase 3 showed 57.7% of UK healthcare workers would accept the pre-pandemic H5N1 vaccine The willingness to accept vaccination against both influenza A subtypes H5N1 and H1N1 among hospital based healthcare workers in Hong Kong was low Neither the change in WHO pandemic alert levels nor the experience of the SARS outbreak affected the potential acceptance of the vaccines Barriers identified include fear of side effects and doubts about the efficacy of the vaccines The strongest associations with the intention to accept vaccination were a history of seasonal influenza vaccination and perceived likelihood of being infected

Evolutionarily Conserved Herpesviral Protein Interaction Networks

Herpesviruses constitute a family of large DNA viruses widely spread in vertebrates and causing a variety of different diseases. They possess dsDNA genomes ranging from 120 to 240 kbp encoding between 70 to 170 open reading frames. We previously reported the protein interaction networks of two herpesviruses, varicella-zoster virus (VZV) and Kaposi's sarcoma-associated herpesvirus (KSHV). In this study, we systematically tested three additional herpesvirus species, herpes simplex virus 1 (HSV-1), murine cytomegalovirus and Epstein-Barr virus, for protein interactions in order to be able to perform a comparative analysis of all three herpesvirus subfamilies. We identified 735 interactions by genome-wide yeast-two-hybrid screens (Y2H), and, together with the interactomes of VZV and KSHV, included a total of 1,007 intraviral protein interactions in the analysis. Whereas a large number of interactions have not been reported previously, we were able to identify a core set of highly conserved protein interactions, like the interaction between HSV-1 UL33 with the nuclear egress proteins UL31/UL34. Interactions were conserved between orthologous proteins despite generally low sequence similarity, suggesting that function may be more conserved than sequence. By combining interactomes of different species we were able to systematically address the low coverage of the Y2H system and to extract biologically relevant interactions which were not evident from single species.

Herpesviruses are subdivided into three taxonomic subfamilies (a, b and c) based on both genomic composition and biology according to a well-known phylogeny [1, 2, 3] (Figure 1A ). While all herpesviruses are structurally similar, the different subfamilies are highly divergent in genome size, content and organization. The genome size ranges from 120 kbp for varicella-zoster virus (VZV), which belongs to the a-herpesviruses, to 240 kbp for human cytomegalovirus (hCMV), a member of the b-herpesviruses [4, 5] . Gene-coding potential is reflected in the size of the genomes with VZV containing 70 open reading frames (ORFs) and hCMV containing ,170 ORFs. The overlap between the protein sets of the five viruses clearly supports the known phylogeny, but there are also some proteins shared among viruses not consistent with the phylogeny (Figure 2A ). Although the three subfamilies are thought to have diverged from a common ancestor around 400 million years ago (McGeoch 2006) , they still contain a set of 41 core orthologs present in all herpesviruses [6, 7] . Herpesviral core proteins are generally involved in fundamental aspects of viral morphogenesis (e.g. DNA replication, DNA packaging, structure and egress), and are consequently often essential for replication in cell culture [8, 9, 10] . Several genome-wide yeast-two-hybrid (Y2H) studies of proteinprotein interactions in eukaryotes have been published over the last years, including Saccharomyces cerevisiae [11] , Caenorhabditis elegans [12] , Drosophila melanogaster [13] , Plasmodium falciparum [14] and Homo sapiens [15, 16] . The first complete genome-wide interaction study, however, was published for the E.coli phage T7 [17] . With their relatively small genomes and few genes, viruses seem the ideal candidates for studying protein-protein interactions on a genome-wide level and to address the generally low coverage of Y2H measurements in a systematic way. It is therefore surprising that not more genome-wide studies of intraviral interactions have been performed to date. With the exception of bacteriophage T7 [17] and Vaccinia virus [18] , most of the studies of viral interactions have been performed on small RNA viruses [19, 20, 21, 22] . Recently more studies have been focusing on larger DNA viruses. In addition to our previous studies of VZV and Kaposi's sarcoma-associated herpesvirus (KSHV) [23] , a study by Calderwood and colleagues identified 43 interactions between viral proteins in Epstein-Barr virus (EBV) [24] . Two Y2H studies on herpes simplex virus 1 (HSV-1) and KSHV have also focused on interactions between structural components of these viruses [25, 26] . To add to our understanding of intraviral interactions in herpesviruses we present in this article the first interactomes for herpes simplex virus I and murine cytomegalovirus (mCMV), in addition to a second and independent interactome for Epstein-Barr virus. Based on these data we are able to compare five related interactomes, obtained using a standardized experimental setup for all five species. From the comparison and extensive experimental testing by CoIP we conclude that i) genome-wide interaction studies are sufficiently sensitive for between-species comparisons to identify the basic sunflower structure of the interaction networks and their common core, ii) interactions are to a large degree conserved between orthologs in herpesviruses, iii) comparing interactomes from several species can improve the low coverage of individual Y2H measurements and iv) biologically relevant interactions which may not be apparent from the interactome of a single species, often become obvious when multiple interactomes are aligned and compared. To study intraviral protein-protein interactions of herpesviruses we recombinatorially cloned the individual open reading frames of HSV-1, mCMV and EBV into the yeast-two-hybrid (Y2H) vectors pGBKT7-DEST and pGADT7-DEST and tested all pairwise intraviral protein interactions using an array-based Y2H strategy [27] . To address the issue of false negative interactions, viral proteins containing transmembrane domains were cloned both as full-length and as intracellular and/or extracellular domains. From the mCMV Y2H analysis we observed that 33% of the tested preys, and 40% of the baits, gave positive interactions. Similar results were observed with HSV-1 and EBV with ,1/3 of the clones yielding positive interactions (Table S1). In total, the Y2H analysis revealed 111 interactions for HSV-1, 406 for mCMV and 218 for EBV (Figures 1B, S1 and S2, Tables S2 and S13) . Combined with our previously published interactomes for VZV (173 interactions) and KSHV (123 interactions), we obtained altogether 1,031 intraviral interactions in five herpesviral species (Tables S2 and S13 ). To evaluate the coverage of our five interactomes we performed an extensive literature search which identified 257 previously published interactions for these herpesviruses (including human cytomegalovirus (hCMV) homologues). Of these 257 interactions we were able to detect 24 (9.3%) in at least one virus ( Figures S3 and S4 and Tables S3 and S14). When comparing our EBV interactome with the recently published EBV network by Calderwood et al., 6 out of 43 (13.9%) interactions could be confirmed [24] . Such low confirmation rates are common to Y2H studies, even for studies within the same species, which in general suffer from low coverage [28, 29, 30] . For instance, in a previous study of human interactions only 2.3-8.4% of known interactions were identified [15] . On the other hand, this implies that ,3% of the interactions found in the present study have been published so far in the literature or identified in previous genomewide screens in the case of EBV. In the case of HSV-1 our study added 102 new interactions to the network of already known interactions (coloured grey in Figure 1B ). As is typical for such interaction networks, no apparent structure can easily be recognized. A comparison of the five herpesviral networks revealed that the degree distribution differed from cellular networks, local clustering was not as high as expected in small-world networks of this size (Figures 1C and S5 and Table S4 ), and attack tolerance and robustness were increased compared to cellular networks ( Figure 1D and S6), probably reflecting that the viral interactome in itself only represents a minor part of the complete interactome of the infected cell. In a previous study we observed that the topology of the KSHV and VZV networks approached that of cellular networks as the viral interactomes were connected into a human interactome [23] . The observations presented here confirm our previous findings, and indicate that herpesviral PPI networks share an evolutionarily conserved topology. Apart from general topological features, herpesviral interactomes were also compared on the level of individual interactions. For this purpose, we used the orthology assignments based on sequence similarity and gene order (Table S5 ) [31] . Species within the same subfamily are generally characterized by higher sequence similarity between orthologous proteins. They also share more orthologous proteins with each other than species from different subfamilies ( Figure 2A ). In previous inter-species comparisons [32] , very few interactions were found to be shared between different species (yeast, worm, fly). Unlike the previous comparative studies, the five different interactomes analysed in this study were obtained using exactly the same experimental protocols. Nevertheless, we still observed little overlap between the networks of the five herpesviruses. Of 488 (409 non-redundant, i.e. conserved interactions are only counted once) interactions between proteins conserved in more than one species, 140 (61 non-redundant) (28.7% or 14.9% non-redundant, respectively) interactions were conserved between at least two species. For any two herpesvirus species, we compared the number of interactions between proteins conserved in both species against the number of interactions found in both species ( Figure 2B ). Although the pair Herpesvirus proteins interact with each other in a complex manner throughout the infectious cycle. This is probably best exemplified in the process where a large number of viral proteins come together to form new viral particles which are subsequently released from the infected cell. A more detailed understanding of how viral proteins interact with each other might assist the development of drugs which may inhibit these interactions and consequently block viral replication. Here we present three genomewide studies of protein-protein interactions in the herpesviruses herpes simplex virus I, murine cytomegalovirus and Epstein-Barr virus. Altogether we identified 735 interactions in the three viruses, most of which have not previously been reported. By combining these studies with our previously published studies for Kaposi's sarcomaassociated herpesvirus and varicella-zoster virus we were able to perform a comparative analysis of interactions in five related viral species. We observed that a high proportion of interactions were conserved between the different species, despite a low degree of sequence conservation. This implies that by comparing interaction data, we were able to increase the coverage of our viral networks and thus obtain a better and more complete picture of interactions between herpesviral proteins. [6, 7] . (B) Intraviral protein-protein interaction network for HSV-1. The proteins are coloured according to their conservation in the herpesvirus phylogeny: the blue nodes are core proteins conserved in all five viruses, two nodes (pink) are conserved in a and c herpesviruses, several red ones in a herpesviruses and the grey ones are specific to HSV-1. Edges indicate observed interactions in HSV-1, and red edges indicate previously reported interactions. The protein interaction network was generated using the Cytoscape software (www.cytoscape.org) [58] .. (C) node degree distribution on a linear or logarithmic (inset) scale. The herpesviral networks can be approximated by power law distributions (Table S3 ). (D) Simulations of deliberate attack on HSV-1 in comparison to two human networks by removing their most highly connected nodes in decreasing order. After each node is removed, the new network characteristic path length (average distance between any two nodes) of the remaining network is plotted as a multiple or fraction of the original parameters. The herpesviral networks consistently exhibited a higher attack tolerance, as the increase in path length is considerably smaller. doi:10.1371/journal.ppat.1000570.g001 wise overlaps observed were small, they were nevertheless significantly higher than observed with randomized orthology assignments ( Figures 2C and S7 ). Randomized orthology assignments for each pair of herpesviruses were obtained by first selecting the sub-network of conserved proteins between the two species, and then randomizing the orthology assignments for these sub-networks. A similar analysis was performed for all five networks taken together. First, networks were divided into interactions conserved within a subfamily or between different subfamilies, and the number of interactions conserved in 2, 3 or 4 species in each category was evaluated. We furthermore compared the number of interactions conserved in 2, 3, 4 and 5 species against the results for randomized orthology assignments and found in each case a significant enrichment ( Figure 2D ). This shows that despite the low coverage of the Y2H system significant conservation can still be observed. Herpesviruses share a set of 41 core orthologous proteins which are conserved throughout the three subfamilies (Table S5 ) [31] . These core orthologs comprise approximately half of the genome of HSV-1, VZV, EBV and KSHV but less than 25% of mCMV. They can be further subdivided into a group of 31 orthologs with relatively high sequence similarity (approximately 30-60% sequence similarity), and a group of 10 orthologs with little similarity (approximately 16-30% similarity) (Table S6 ). Based on this orthology assignment, we generated an overlay of all protein interactions between the core orthologs detected in any of the five herpesviruses (Core network, Figure 3A ). Of a total of 283 (218 In each rectangle, the value above the lines indicates the observed number of homologous interactions detected in both herpesviruses (in green). The value below the line (in black) gives the total number of interactions detected in the first species (indicated in columns) between proteins which have orthologs in the second species (indicated in rows). On the diagonal, the total number of interactions is shown for each virus. (C) Distribution of the number of conserved interactions between HSV-1 and mCMV for 1000 random orthology assignments (blue line) in comparison to the true number of conserved interactions (red vertical line). For each pairwise comparison, subnetworks were selected between proteins conserved in both viruses and then the orthology assignments between the proteins were randomized. Accordingly, the size and degree distribution of the subnetworks does not change. (D) Comparison of the number of interactions conserved in 2, 3, 4 and 5 species for 1000 random orthology assignments (yellow boxes) to the true number of interactions conserved in that many viruses (red line). Random orthology assignments were created in a similar way as for Figure 2C . non-redundant) core protein interactions detected, 113 (48 nonredundant, 39.9%) were found in more than one species (Table S7) . For the core network, we did not observe a correlation between sequence similarity and the number of conserved interactions detected ( Figures 3B and 3C ). For example, the interaction between the two tegument proteins UL11 and UL16 in HSV-1 was also detected in mCMV and EBV, although sequence similarity of UL11 and its orthologs across subfamilies is quite low (28%). This interaction was interestingly also observed for HSV-1 in a recent report by Vittone and colleagues [33] . In addition, interactions were not preferentially conserved between closely related species ( Figures 3D and S8) . Accordingly, overlaps between the interaction sets in the core network were not correlated to the true phylogeny of herpesviruses ( Figure 3E ). Indeed, the highest overlap was observed between HSV-1 (a-subfamily) and mCMV (b) which belong to lineages separated early in herpesvirus evolution [7] . However, since our phylogenetic trees are based on relative overlaps between the different species, we cannot exclude that a more complete set of core interactions might have allowed for better separation of the subfamilies. In contrast, when also including subfamily-and species specific interactions (i.e. the complete interaction network of the five herpesviruses with the characteristic sunflower structure, see Figure 4A ), the analysis yielded a phylogeny that was consistent with the known evolutionary relationships ( Figure 4B ). This indicates that the presence of conserved subfamily specific interactions provides sufficient conserved and non-conserved interactions to accurately separate the subfamilies from each other. In the overlay of all five herpesviral networks ( Figure 4A , sunflower structure), the core network is indicated as a central node common to all herpesviruses. Subfamily-and species-specific networks are attached (as leafs) to this core. Only few connections exist between the subfamily-specific networks due to few shared proteins outside of the core. Our data provides evidence that the viral core network is extremely dense while the non-core network appears relatively sparse. However, since non-core interactions were tested in at most two species, and not in five as the core interactions, the non-core network may be equally dense. Indeed, no consistent difference was observed between the number of intraviral core and non-core interactions when considering each network separately (Table S8) . To further evaluate whether interactions between orthologous proteins are conserved we used co-immunoprecipitation to test 92 interactions predicted from 55 interactions detected in KSHV for the corresponding orthologs in HSV-1, mCMV and EBV. 11/19 (58%) of the predicted interactions could be confirmed by CoIP in HSV-1, 12/18 (67%) in mCMV and 36/55 (65%) in EBV, in comparison to 29/55 (53%) in KSHV itself ( Figures 5A and S9A and Table S9 ). The percentage of core-derived orthologs that were confirmed by CoIP significantly correlated with the number of species in which the interactions were detected in Y2H screens, suggesting that the accuracy increases with the number of positive assays ( Figure S9B , Table S10 ). As negative controls, ten interactions which were not detected in any of the Y2H screens were tested in four viruses (39 interactions in total, Table S10 ). Although the confirmation rate of these negative controls seems relatively high (6/39 (15%)), it is still significantly smaller than for the predicted interactions and correlates well with the confirmation rates of interactions observed in 2, 3 and 4 species ( Figure S9B ). Due to the low coverage of the Y2H system many (particular weak) interactions were most likely missed, and the positively tested controls may be examples of such interactions. It also suggests that, although our core interactome is very dense, it has not yet reached full coverage. Since the confirmation rate by CoIP for the Y2H interactions in KSHV is not higher than for the predicted interactions in HSV-1, mCMV and EBV, we conclude that a high percentage of interactions between core orthologs are conserved despite low sequence similarity of some of the orthologs across subfamilies. To further assess the level of completeness of our core network, we evaluated the average number of new interactions added to the core network with each new Y2H screen ( Figure 5B ). If core interactions indeed are conserved, as indicated by our predicted interactions, we would expect the coverage to increase with each new herpesviral interactome. Although the number of newly discovered interactions steadily decreased with each new screen, saturation does not seem to be reached yet. Thus, although coverage for the core network could be increased, a significant fraction of interactions might still be missing. Finally, to determine if conserved intra-viral interactions allow viral proteins to interact across different species, we tested four interactions which were detected in at least two herpesviruses in the original screens by Y2H and LUMIER (luminescence-based mammalian interactome mapping) pull-down assays ( Figure S10A and B) [34] . While Y2H in general yielded few cross-species interactions, we detected a larger number of interactions by LUMIER ( Figure S10B ). The cross-species interactions between the HSV-1 UL11 and UL16 tegument and between the HSV-1 UL19 and UL35 capsid orthologs were mainly observed within a specific subfamily, in accordance with previous observations by Schnee et al. [35] . For the two other interactions, involving orthologs with both a high and low degree of sequence similarity based on Table S6 , we saw a more promiscuous interaction pattern. HSV1 UL14 for example was able to interact with HSV-1 UL33 and its orthologs in all five species, suggesting that sequence similarity might be a poor predictor of interspecies interactions in herpesviruses. Additionally, we tested 4 core and 4 noncore VZV baits against prey libraries of all five viruses. As expected, the intraspecies analysis (VZV baits against VZV preys) yielded the highest fraction of positive interactions (2.8%), compared to 0.5% positive interactions in the cross-species screens. Of the positive crossspecies interactions we observed 4 core-core, 15 core-noncore and 2 noncore-noncore interactions (Table S11 ). When the number of positive interactions was correlated to the number of interactions tested for each class, we observed a significant enrichment of positive interactions for the core-core and core-noncore classes compared to the noncore-noncore class ( Figure S11 ). Most core proteins are essential, and a majority can be found in herpesvirus virions composed of an icosahedral capsid of 162 capsomers, an amorphous tegument layer and a lipid bilayer membrane with embedded glycoproteins. Using the high-coverage core network, a map of conserved protein interactions in herpesviral particles was generated ( Figure 6A and S12). One outstanding example for a highly connected protein in this virion map is the mCMV M51 ortholog (HSV-1 UL33, VZV Orf25, EBV BFRF4 and KSHV Orf67.5), which interacted with 14 tegument proteins. Since (i) 11 of the 14 interactions (79%) of this protein were found in more than 1 species, (ii) most Y2H interactions were confirmed even under high concentrations of the competitive HIS3 inhibitor 3amino-1,2,4-triazole which can be used to suppress non-specific Y2H interactions ( Figure S13A and S13B), and (iii) a majority of interactions were confirmed by CoIP (Table S12) , we considered them as high-confidence interactions. Furthermore, 4 of the 5 interactions conserved in 4 herpesviral species are M51 interactions. One example is the interaction of mCMV M51 orthologs (HSV-1_UL33/VZV_25/EBV_BFRF4/KSHV_67.5) with M53 orthologs (HSV-1_UL31/VZV_27/EBV_BFLF2/KSHV_69), which also interact in 4 species with M50 orthologs (HSV-1_UL34/ VZV_24/EBV_BFRF1/KSHV_67). M50 and M53 and their orthologs are involved in the nuclear egress of viral capsids and are well-characterised in mCMV, HSV-1, EBV and pseudorabiesvirus [36, 37, 38, 39] . Both M50 in mCMV and its ortholog UL34 in HSV-1 recruit protein kinase C to the nuclear membrane, which subsequently phosphorylates lamins to dissolve the nuclear lamina allowing the capsids to reach the inner nuclear envelope [38, 40] . In mCMV, we confirmed 17/22 (.75%) (Table S12) of M51 interactions by CoIP, and showed that M51 is targeted to the nuclear membrane by M50 and co-localizes with both M50 and M53 ( Figure 6B ). Our results suggest that M51 and its orthologs are part of a larger protein complex and may be involved in nuclear egress. Since most of its interaction partners are present in the virion tegument we hypothesize that it plays a role in tegument formation, and represents a possible link between DNA packaging, nuclear egress and tegumentation. Here we present an extensive study of intraviral protein-protein interactions for the three herpesviruses HSV-1, mCMV and EBV, using Y2H as the main experimental method. By combining the results with our previous studies of interactions in VZV and KSHV we were able to compare the interactomes of five related herpesviral species. Although there was little overlap between the five viral networks according to the Y2H maps, we were able to show that interactions between core orthologous proteins are to a large degree conserved between species of different subfamilies. By generating a separate network of interactions between core proteins of five herpesviruses, we were also able to overcome the coverage problem of Y2H and to identify interactions of interest from the common network which were not apparent in each single network. While the overlaps between the different interactomes were generally quite low, there was still a significant enrichment of conserved interactions between orthologous proteins for any pair of the five species ( Figures 2C and S7A to I). The same holds true for the conservation between all five viruses. Interactions observed in two, three or four species were all enriched significantly as compared to background expectations ( Figure 2D ). This argues that, although troubled with false negative and false positive interactions, Y2H as a technique is still sufficiently sensitive and specific to obtain data for a comparative analysis of related interactomes. Similar observations were also made for the recently published Campylobacter jejuni interaction network, where highly significant overlaps were found with both the Escherichia coli and Helicobacter pylori interactomes [41] . It is very likely that true overlaps between the herpesvirus interactomes are higher, but that due to false negative interactions we only observe modest overlaps. There are numerous reasons why interactions may be missed in the Y2H system, including improper folding of fusion proteins and post-translational modifications. In an attempt to address some of these issues we cloned all the viral proteins containing transmembrane domains as both full length and extra/intracellular fragments, which has been reported to increase sensitivity [42] . Our observations indicate that intraviral interactions between core proteins are conserved, and as a result we are not able to separate the Y2H interactomes into their phylogenetic subfamilies solely based on their core interactions. However, when interactions involving subfamily-specific proteins present in at least two of the virus species were included, we were able to generate a correct phylogenetic tree. This implies that interactions involving subfamily-specific proteins are at least partly conserved. Indeed, several of the interactions predicted from KSHV and confirmed in EBV by CoIP involved subfamily specific proteins. From published literature there are several examples of core interactions being conserved between species of different herpesviral subfamilies, e.g. the interactions between HSV-1 UL31 and UL34 [36, 37, 38, 39] , HSV-1 UL15 and UL28 [43, 44] and the HSV-1 UL54 self-interaction [45, 46] . Indeed, much of what is currently known about herpesvirus biology is derived from studies of Herpes Simplex Virus and extrapolated to other species. Our study indicates that it is effectively possible to transfer intraviral interactions between orthologous proteins from one species to another. Thus, by generating an overlay network from several genome-wide Y2H screens in related species, the large number of false negative interactions within each individual analysis can be overcome and a more complete picture of the core interaction network obtained. In general, interactions are transferred between different species based on the sequence similarity between the corresponding proteins. In addition, one might expect interactions among orthologous proteins with high sequence similarity to have a higher likelihood of being conserved. Yu and colleagues found that interactions could be confidently transferred from one species to another if the joint sequence identity of the interacting orthologs was .80% [47] . However, since none of the herpesviral core proteins shares such a high degree of sequence similarity across subfamilies, these criteria cannot be applied to herpesviruses. Furthermore, no correlation was observed between sequence similarity and the number of species in which an interaction was observed in the Y2H experiments. Thus, our results show that sequence similarity alone seems to be insufficient for predicting herpesviral interactions from one species onto another. Our analysis of cross-species interactions indicates an enrichment of interactions involving core proteins (either core-core or core-noncore). The detailed cross-species analysis of the interaction between the major capsid protein (MCP) and the smallest capsid protein (SCP) (HSV-1 UL19-UL35, Figure S10B ) only yielded 1 intraspecies interaction by Y2H, however 4 by LUMIER, indicating that this interaction is conserved despite being observed in only one species by Y2H. While capsid proteins and interactions are thought to be highly conserved, most of them were indeed only observed in one species in our genome-wide Y2H screens. However, three of the four observed capsid interactions (HSV-1 UL19-UL35, UL18-UL38, UL18-18 and UL35-UL35) have been published previously (Tables S3 and S14) , and, in addition, the LUMIER analysis resulted in an increased number of cross-species interactions. The cross-species interactions between the two tegument proteins HSV-1 UL11 and UL16 (and their orthologs), as well as between the two capsid proteins HSV-1 UL19 and UL35 (and their orthologs), were mainly observed between species within the same herpesviral subfamily. A similar observation has recently been reported by Schnee and colleagues [35] , and may indicate that some binding sites are more conserved within herpesvirus subfamilies. The other two interactions could be detected in a larger number of cross-species interactions by both Y2H and LUMIER. HSV-1 UL14, for example, was observed to interact with all orthologs of HSV-1 UL33 by LUMIER. Previous reports suggested that HSV-2 UL14 shares certain similarities with cellular chaperones which may account for its promiscuous binding pattern [48] . In the core network derived from the overlap of all five herpesviruses, mCMV M51, and its orthologs HSV-1 UL33, VZV ORF 25, EBV BFRF4 and KSHV ORF 67.5, show up as intraviral hubs with a number of conserved interactions. For instance, the interaction between M51 and M53 was observed in all species apart from HSV-1. Interestingly, when retesting UL33 interactions under more stringent conditions (Figure S12) , the corresponding interaction between UL33 and the M53 ortholog in HSV-1, UL31, is clearly one of the positive interactions on both 2.5 and 5 mM 3AT. These interactions were not included in the HSV interactome to prevent an overrepresentation of interactions tested more than once. While not much is known about M51, M53 has been extensively documented to be involved in nuclear egress through its binding to M50 [35, 38, 49, 50] . The interaction between M53 and M50 was confirmed in this study in four viruses. In addition, from our study of interactions in VZV we observed an association between the ortholog of M51 (ORF25) with the M50 ortholog (ORF24) [23] , and retesting of HSV-1 UL33 interactions also revealed its binding to HSV-1 UL34 lacking the transmembrane region ( Figure S12 ). Finally, immunofluorescense studies indicated that M51 co-localizes with both M53 and M50 when using fluorescent fusion proteins. These results suggest a possible role for M51 in nuclear egress through its interactions with M53 and/or M50. As interactions between orthologs of M51 and M53 were observed in members of all three subfamilies, it is likely that this represents a conserved function of M51. Previous studies have indicated that the M51 ortholog in HSV-1 (UL33) is involved in packaging of DNA [51] , and that it interacts with at least one of the subunits of the terminase complex (UL28) [52] . In the data presented here, UL33 was observed to interact with UL15 and UL28 in three different species. These results suggest that UL33 represents an association between packaging and egress. Studies done in HSV-1 have indicated that UL33 is associated with the external surface of capsids [53] , which would make such a dual role reasonable. While it is not known exactly how UL33 associates with the capsid, the interaction observed between M51 and the smallest capsid protein (m48.2) in mCMV and EBV suggests a possible manner of association. In summary, this study suggests that a distinctive network topology is still present in all vertebrate herpesvirus species although herpesviruses co-evolved with their hosts for millions of years. Moreover, it provides evidence (i) that interactions and hence functions of proteins may be more conserved than their sequence and (ii) that a common core of protein interactions is conserved in all herpesviruses. We hope that the data presented will inspire future herpesvirus research and facilitate the selection of potential targets for antiviral therapy [54] . The nucleotide sequences for all ORFs were obtained from the ncbi (http://www.ncbi.nlm.nih.gov/), and cloned into the Y2H vectors pGBKT7-DEST and pGADT7-DEST by recombinatorial cloning [55] (Protocol S1). All clones were sequence verified. Yeast strains AH109 and Y187 were transformed using 1 mg of prey (pGADT7-DEST) or bait (pGBKT7-DEST) plasmid DNA, respectively, and grown on SD medium lacking either leucine (-leu) or tryptophane (-trp). Prey-and bait-expressing yeast were arrayed in a 384-pin format using a Biomek 2000 workstation (Beckman-Coulter) (4 replicas for each interaction tested), and mated in an all-against-all matrix approach [27] . Diploid colonies were grown for 2 days at 30uC on SD -leu-trp plates, and subsequently transferred to selective SD -leu-trp-his plates. Interactions were considered positive if at least 3 out of 4 colonies grew (Protocol S1). pGBKT7-DEST and pGADT7-DEST were co-transfected into HEK-293 cells by means of calcium phosphate, and superinfected with recombinant vaccinia virus (vTF-7) expressing T7 RNA polymerase (NIH AIDS repository) at a MOI of 10. After 24 h cells were lysed, and precipitation of proteins was done using 1 mg of either anti-myc (Santa Cruz) or anti-HA (Roche) antibodies in addition to protein G Sepharose beads. Precipitates were separated by SDS-PAGE, and western blots initially reacted with the anti-myc and anti-HA antibodies, and secondary, peroxidaseconjugated anti-mouse IgG or anti-rat IgG antibodies (Jackson). The CoIP was scored positive if a co-precipitate was detected in at least one direction (Protocol S1). HeLa cells were grown on a cover slip until ,50% confluence, and subsequently transfected with 1 mg of DNA for each of the fluorescent vectors analyzed, either alone or in combinations, by means of Effectene (Qiagen). Cells were incubated for 24 h, and fixed by incubating with 4% paraformaldehyde for 30 min at RT. Coverslips with fixed cells were mounted in Vectashield Mounting Medium (Vector Labs), and imaged on an OLYMPUS BX61 microscope. Literature interactions were identified by combining automatic text mining and manual curation. A set of ,87000 MEDLINE abstracts on herpesviruses was screened using ProMiner [56] for occurrences of proteins of any of the five viruses considered. Subsequently, 565 abstracts were selected containing a reference to interactions and at least two different proteins of the same virus. Physical interactions were then extracted manually from the corresponding articles. From the five individual networks an overlay network was created by merging orthologous proteins and interactions between orthologous proteins. Orthology relationships were assigned based on Davison [31] . The overlay network was then used to predict interactions between core proteins and to analyze network characteristics (Protocol S1). For all core orthologous proteins the average pairwise global sequence similarity across all five viruses was calculated. Global similarity was used to avoid a distortion of the results by short but high local similarities between orthologous proteins. For an interacting pair of core proteins, the similarity was calculated as the geometric mean of the average similarities for the corresponding proteins. The distance metric used to construct the phylogenetic tree ( Figure 3E ) for the complete and core network, respectively was based on the relative interaction overlaps. Accordingly, where S ij is the number of shared interactions between species i and j and C i and C j are the total number of interactions for species i and j. In this case, only interactions between proteins conserved in at least two species or all species for the core network were considered for each species-specific network. The phylogenetic tree was generated using the neighbour-joining algorithm of the PHYLIP package [57] with 10,000 bootstrap samples. Protocol S1 Found at: doi:10.1371/journal.ppat.1000570.s001 (23 KB PDF) Figure S1 Protein network in mCMV. The protein interaction network was generated using Cytoscape software (www.cytoscape. org) [58] . Interactions previously reported in the literature are indicated with red edges. The colours of the nodes indicate in which herpesviral species a specific protein is conserved. [16, 17] . After each node is removed, the new network characteristic path length (average distance between any two nodes) of the remaining network is plotted as a multiple or fraction of the original parameters. The herpesviral networks consistently exhibited a higher attack tolerance, as the increase in path length is considerably smaller. Figure S13 Protein interaction partners of HSV-1 UL33 tested by Y2H. HSV-1 UL33 cloned in pGADT7 (prey) was tested against a variety of interaction partners cloned in pGBKT7 (bait). As negative controls, each bait was tested against the empty pGADT7 prey vector (left plates) while the UL33 prey was tested against the empty pGBKT7 bait vector (right plates). (A) Evaluation of mated yeast clones on double and triple selective plates with empty pGADT7 vector used as a control. (B) Evaluation of mated yeast clones on increasing amounts of 3-AT (0, 2.5, 5, 10 mM) with empty pGADT7 vector as a control. Self-activation of UL15, UL16 and UL21 at 0 mM 3-AT was suppressed at 5 mM 3-AT, while the interactions with UL33 were still found to be positive. Found at: doi:10.1371/journal.ppat.1000570.s014 (0.08 MB PDF) Table S1 Summary of prey and bait hit-rates for HSV-1, mCMV and EBV. Overview of the total number of preys and baits included in the Y2H screens, including the number of preys and baits which yielded interactions. The total number of preys and baits exceed the total number of proteins tested due to many of the proteins being cloned as both fragments and full-length proteins. Table S8 Average degree values for core and non-core proteins in all viruses. Average degree of core vs non-core proteins for all five interactomes. P-values were calculated with a Wilcoxon rank test between the degree values of core and non-core proteins. Found at: doi:10.1371/journal.ppat.1000570.s022 (0.06 MB PDF) Table S9 Ortholog protein interactions (predicted from KSHV) tested by Y2H and CoIP. List of orthologous interactions predicted from the KSHV interactome [23] which were tested by co-immunoprecipitation in HSV-1, mCMV and EBV. Results from the Y2H analysis of the predicted interactions are also indicated. Found at: doi:10.1371/journal.ppat.1000570.s023 (0.00 MB PDF) Table S10 Negatively predicted orthologous protein interactions (predicted from core interaction network) tested by Y2H and CoIP. Ten interactions were predicted to be negative, based on the fact that they were not observed in any of the five viral interactomes, and analysed by co-immunoprecipitation. Found at: doi:10.1371/journal.ppat.1000570.s024 (0.02 MB PDF) Table S11 Analysis of interspecies interactions. VZV core and noncore baits were analysed for Y2H interactions against prey libraries of VZV, HSV-1, mCMV, EBV and KSHV. The species of the interacting prey is included, in addition to whether it was a core or a noncore protein. Table S12 M51 interactions tested by CoIP. Interactions observed with mCMV M51, or with other orthologs of M51, with a subset of interaction partners from the Y2H analysis. Tegument proteins, and other virion components, were determined based on whether they were reported to be present in the CMV virion [59, 60] .

Insertion/Deletion Polymorphism of Angiotensin Converting Enzyme Gene in Kawasaki Disease

Polymorphism of angiotensin converting enzyme (ACE) gene is reported to be associated with ischemic heart disease, hypertrophic cardiomyopathy, and idiopathic dilated cardiomyopathy. In this study, we investigated the relationship between Kawasaki disease and insertion/deletion polymorphism of ACE gene. Fifty five Kawasaki disease patients and 43 healthy children were enrolled. ACE genotype was evaluated from each of the subjects' DNA fragments through polymerase chain reaction (PCR). Frequencies of ACE genotypes (DD, ID, II) were 12.7%, 60.0%, 27.3% in Kawasaki group, and 41.9%, 30.2%, 27.9% in control group respectively, indicating low rate of DD and high rate of ID genotype among Kawasaki patients (p<0.01). Comparing allelic (I, D) frequencies, I allele was more prevalent in Kawasaki group than in control group (57.3% vs. 43.0%, p<0.05). In Kawasaki group, both genotype and allelic frequencies were not statistically different between those with coronary dilatations and those without. ACE gene I/D polymorphism is thought to be associated with Kawasaki disease but not with the development of coronary dilatations.

Kawasaki disease is an acute systemic vasculitis and its diagnosis is made on clinical features. Main complication of the disease is coronary artery lesion that may result in myocardial infarction or sudden death. Coronary artery aneurysms or ectasia develop in approximately 15-25% of untreated children (1) . Since the intravenous immune globulin (IVIG) and aspirin therapy have been introduced, its mortality rate has decreased to 0.1%, but cardiac sequelae continues to occur in about 13% of Kawasaki disease patients (2) . The etiology of Kawasaki disease is largely unknown despite the various suggested hypotheses. Based on epidemiologic and clinical manifestations, it is thought that Kawasaki disease is caused by some infectious agents (3) (4) (5) (6) (7) (8) (9) (10) . Hypercytokinemia and hyperchemokinemia have also been observed and are thought to cause vascular injuries by inflammatory reaction and immunologic activation (11) . Genetic factors are also thought to have influences on the development and progress of Kawasaki disease (12, 13) . Up to present days, Kawasaki disease is thought to be an infectious disease manifested by immunologic reaction in genetically susceptable person. Angiotensin converting enzyme (ACE) breaks down the potent vasodilator, bradykinin to its inactivate metabolite and catalyzes angiotensin I to angiotensin II. Angiotensin II promotes hyperplasia and hypertrophy of vascular smooth muscle cells, induces the production of proinflammatory cytokines and causes endothelial dysfunction by free radical generation. By playing an important role in cardiovascular regulatory system, ACE gene has been proposed to be associated with various cardiovascular diseases, such as ischemic heart disease, hypertrophic cardiomyopathy, idiopathic dilated cardiomyopathy, and vascular hypertrophy (14) . Among the different ACE genetic loci, the insertion/deletion polymorphism coded within intron 16 has been studied in many literature to find the association with such diseases. However, there has been few studies regarding the association between the polymorphism of ACE gene and Kawasaki disease. The present study investigates whether the I/D polymorphism of ACE gene (DD, ID, II) is associated with the prevalence and severity of Kawasaki disease among Korean pediatric populations. Fifty five Kawasaki patients (mean age 28.2±25.2 months) diagnosed at Ewha Womans University Mokdong Hospital from January 2001 to June 2003, and 43 healthy children (mean age 28.5±17.2 months) were enrolled. Kawasaki disease was diagnosed by its clinical features, that is fever lasting for at least 5 days, accompanied by 4 of the 5 classical signs: 1) bilateral bulbar conjunctival injection; 2) pharynx, injected and/or dry fissured lips, strawberry tongue; 3) changes of the peripheral extremities in the acute phase or periungal desquamation in the subacute phase; 4) nonvesicular rash; 5) cervical adenopathy, ≥1.5 cm. The diagnosis was made if he/she had typical manifestations even before less than 5 febrile days of illness. All patients were treated with immunoglobulin (2 g/kg) on the day of diagnosis. High dose (50 mg/kg/day) aspirin was given from the day of diagnosis and its dosage was changed to 5 mg/kg/day after 2 nonfebrile consecutive days. Two dimensional echocardiography was done to evaluate cardiac complications. Coronary arterial lesion was defined as following: 1) inner diameter that is >3 mm in children <5 yrs old and >4 mm in children ≥5 yrs old; 2) internal diameter of a segment ≥1.5 times that of an adjacent segment; or 3) lumen with irregular surface. Among these patients, 17 showed coronary dilatations and 38 did not. Informed consent was obtained from his/her parents prior to the participation in the study. The study was approved by the hospital's ethics commitee. Each of the subjects' DNA was extracted from whole blood at the time of diagnosis using QIAamp DNA Blood Mini Kit (Gene Company LTD., Chai Wan, Hong Kong). DNA fragments were amplified through polymerase chain reaction (PCR), which was carried out in a total volume of 10 L containing 50 ng of genomic DNA, 200 mM dNTPs, 0.3 mM/mL of each primers (5′ -CTGGAGACCACTCCCATCCTTTCT); (5′ -GATGTGGCCATCACATTCGTCAGAT) in PCR buffer with 0.5 units Taq DNA polymerase (Takara, Shiga, Japan). After the initial denaturation step (10 min at 95℃), 35 cycles were repeated for 30 sec at 94℃, 30 sec at 52℃, 90 sec at 72℃, and 5 min at 72℃. DNA fragments were then separated by electrophoresis on 2.5% agarose gel. Student's t-test was used to compare the demographic characteristics between the two groups. Chi square test and Fisher exact test were performed to compare the genotype and allelic frequencies between the two groups. Its frequencies were compared by odds ratio. p value less than 0.05 was considered significant. The mean age of Kawasaki group (28 boys and 27 girls) was 28.2±25.2 months, and that of the control group (30 boys and 13 girls) was 28.5±17.2 months. Among 55 Kawasaki disease patients, coronary dilatation was observed in 17 patients (8 boys and 9 girls, Table 1) Frequencies of ACE genotypes (DD, ID, II) were 12.7%, 60.0%, 27.3% in Kawasaki group, and 41.9%, 30.2%, 27.9% in control group respectively, indicating low rate of DD genotype (p<0.01, odds ratio=0.2) and high rate of ID genotype (p<0.01, odds ratio=3.3) among Kawasaki patients (Table 2) . Comparing allelic (I, D) frequencies, I allele was more prevalent in Kawasaki group than in control group (57.3% vs. 43.0 %, p<0.05, odds ratio=1.78, Table 3 ). In Kawasaki group, both genotype (DD, ID, II) and allelic frequencies were not statistically different between those with coronary dilatations and those without ( The ACE gene is localized on chromosome 17q23 and is characterized by a major insertion/deletion polymorphism consisting of the presence or absence of a 287-base pair Alu repeat sequence within intron 16 (15) . Angiotensin converting enzyme is an ectoenzyme found on the external surface of the endothelial and epithelial cell membranes. It enhances the synthesis of angiotensin-II, that promotes proliferation, migration, and hypertrophy of vascular smooth muscle cells. Angiotensin II also induces the production of proinflammatory cytokines and matrix metalloproteinases (16, 17) . Moreover, the increased free radical generation by angiotensin-II contributes to endothelial dysfunction (18, 19) . In humans, the ACE activity is partly under genetic control. It is suggested that about half of the interindividual difference in ACE levels may be accounted for its polymorphism (19, 20) . It is reported that mean ACE levels were lowest for II homozygotes, highest for DD homozygotes, and intermediate for ID heterozygotes (18, 21) . Danser et al. explained the higher ACE levels observed in subjects with D allele than those with the II genotype by the sequence harbored in the insert of ACE gene. This sequence was said to be very similar to a silencer element (22) . But it is not clear whether increased levels of ACE actually affects the levels of angiotenisn II because the renin-angiotensin system is regulated by feedback mechanism (22) . Angiotensin converting enzyme is mainly produced by vascular endothelial cells (23) . In Kawasaki disease, the associated endothelial cell damage subsequently lowers the ACE level. It is reported in some literature (24, 25) that serum ACE levels are significantly attenuated during the acute phase, and recovered during the convalescent phase of Kawasaki disease. Slowik et al. (26) reported that the II genotype of ACE gene contributes to vascular dilatation at the site of aneurysm by 1) increased bradykinin activity, 2) another polymorphism responsible for vascular dilatation that is in linkage disequilibrium with ACE I/D polymorphism, 3) degeneration of endothelial cells, or 4) lack of vascular remodeling. There is pathological difference between adult coronary artery disease (CAD) and that caused by Kawasaki disease. CAD caused by Kawasaki disease is characterized by vascular intimal thickening, whereas the adult CAD is characteri-zed by atherosclerotic lesions initiated by atheroma and plaque formation. Thus, the pathophysiologic mechanism of myocardial ischemic development is also different; in adult CAD, plaque rupture and thrombus formation plays the important role, whereas in Kawasaki disease, coronary arterial narrowing by intimal hyperplasia is responsible (24) . In this study, the ID genotype was more prevalent (59.3 vs. 30.2%, p<0.01), and the DD genotype (12.9 vs. 41.9%, p< 0.01) was less prevalent in Kawasaki group than in control group. And I allele was more prevalent in Kawasaki group than in control group. However, both genotype (DD, ID, II) and allelic frequencies were not statistically different between Kawasaki disease patients with coronary dilatations and those without. Therefore, ACE gene I/D polymorphism was thought to be associated with the prevalence of Kawasaki disease but not with the development of coronary lesions. These results are similar to the report by Wu et al. (13) that the DD genotype was present in lower frequency among Kawasaki patients and ACE polymorphism was not associated with coronary aneurysmal formation. These results are different from the study by Takeuchi et al. (27) that the II genotype of ACE gene is more prevalent in Kawasaki disease and those with coronary aneurysm. These inconsistent results may be accounted for different ethnic traits of the individual population. In conclusion, the ID genotype is present in a significantly higher frequency and DD in lower frequency among Kawasaki disease patients than in control subjects. In addition, there are no significant association between ACE I/D polymorphism and the coronary artery aneurysm formation in Kawasaki disease patients. Because the study groups are relatively small size in number, it is difficult to generalize these results. In this study, the coronary arterial dilatation was observed in 17 out of 55 Kawasaki patients. This is higher than that in previously reported literature. It can be accounted for either 1) longer duration from the onset of illness to the infusion of immunoglobulin, 2) other factors that may affect the development of coronary arterial lesions in Kawasaki disease, such as different ethinicity or 3) selection bias caused by limited number of enrolled institution and small number of patients. IVIG is recommended to be given within the first 10 days of illness and, if possible, within 7 days of illness. Since the duration of febrile days before the diagnosis and the IVIG infusion were 4.6±1.6 days (range, 2-8 days) in our study groups, the first explanation seems less likely. Further study is required to clarity the association. Table 5 . Deletion/insertion allelic prevalence of angiotensin converting enzyme gene of Kawasaki patients with and without coronary dilatation

Differential Regulation of Type I Interferon and Epidermal Growth Factor Pathways by a Human Respirovirus Virulence Factor

A number of paramyxoviruses are responsible for acute respiratory infections in children, elderly and immuno-compromised individuals, resulting in airway inflammation and exacerbation of chronic diseases like asthma. To understand the molecular pathogenesis of these infections, we searched for cellular targets of the virulence protein C of human parainfluenza virus type 3 (hPIV3-C). We found that hPIV3-C interacts directly through its C-terminal domain with STAT1 and GRB2, whereas C proteins from measles or Nipah viruses failed to do so. Binding to STAT1 explains the previously reported capacity of hPIV3-C to block type I interferon signaling, but the interaction with GRB2 was unexpected. This adaptor protein bridges Epidermal Growth Factor (EGF) receptor to MAPK/ERK pathway, a signaling cascade recently found to be involved in airway inflammatory response. We report that either hPIV3 infection or transient expression of hPIV3-C both increase cellular response to EGF, as assessed by Elk1 transactivation and phosphorylation levels of ERK1/2, 40S ribosomal subunit protein S6 and translation initiation factor 4E (eIF4E). Furthermore, inhibition of MAPK/ERK pathway with U0126 prevented viral protein expression in infected cells. Altogether, our data provide molecular basis to explain the role of hPIV3-C as a virulence factor and determinant of pathogenesis and demonstrate that Paramyxoviridae have evolved a single virulence factor to block type I interferon signaling and to boost simultaneous cellular response to growth factors.

Viruses need to interact with host macromolecules to hijack the cellular machinery and replicate. These interactions are essential for viruses to target endocytic pathways and penetrate host cells, to recruit cellular transcription and/or translation machinery, and to achieve intracellular migration and viral particles assembly. But viruses also encode virulence factors that induce a substantial alteration of host cell functions and genetic programs to increase virus replication and spreading. For example, specific viral factors stimulate survival pathways to prevent apoptosis of infected cells or inhibit cell signaling involved in immune response. Among these pathways, IFN-a/b signaling represents a prime target for viruses because of its critical role in the induction of both innate and adaptive antiviral immune responses [1] . IFN-a/b transduce signals through direct binding to a cell surface receptor composed of two transmembrane subunits, IFNAR1 and IFNAR2c [2] . This interaction activates IFNAR1/IFNAR2c associated kinases Tyk2 and Jak1 that subsequently phosphorylate STAT2 and STAT1 transcription factors. Activated STAT1 and STAT2, altogether with IRF9, form the Interferon-Stimulated Gene Factor 3 that binds IFN-stimulated response element (ISRE) promoter sequences to induce a large antiviral gene cluster. As a consequence, most viruses that are pathogenic in vertebrates have evolved virulence factors both to block IFN-a/b expression and signal transduction downstream of IFN-a/b receptor. Human parainfluenza virus type 1 (hPIV1) and human parainfluenza virus type 3 (hPIV3) are important human pathogens that belong to Respirovirus genus (Paramyxoviridae family; [3] ). These viruses are responsible for upper respiratory tract infections and colds, but often spread to the lower respiratory tract causing bronchitis, bronchiolitis and pneumonia in young children and immuno-compromised patients. hPIV3 infection is also suspected to exacerbate chronic airway inflammatory diseases like asthma [4] . Sendai virus and bovine parainfluenza virus type 3 (bPIV3) are animal counterparts of hPIV1 and hPIV3 that infect mouse and cattle, respectively. Respirovirus genome is a single-strand, negativesense RNA molecule that encodes six structural proteins (Mononegavirales order). While hemagglutinin-neuraminidase (HN) and fusion (F) are membrane glycoproteins associated with the envelop of hPIV3 particles, the nucleoprotein (N), the phosphoprotein (P) and the viral polymerase (L) form the ribonucleocapsid complex. The matrix protein (M) is at the interface between glycoprotein tails and ribonucleocapsids. The P gene of Respirovirus encodes for P but also for a panel of accessory proteins by site-specific editing of P mRNA and usage of overlapping open reading frames (ORFs). In all Respirovirus except hPIV1, the co-transcriptional insertion of one G residue at an editing motif midway of P mRNA leads to the expression of a chimeric protein called V. The V proteins of bPIV3 and Sendai virus bind MDA5 and suppress double-stranded RNA-stimulated IFN-b production, thereby contributing to the virus evasion of host immune response [5] . Surprisingly in hPIV3, multiple stop codons localized downstream of the editing site prevent the normal expression of a full-length V protein. As a result, P mRNA molecules edited by the addition of one G residue encode for the 242 amino acid (AA)-long N-terminal residues of P followed by only six additional AA (see Materials and Methods and [6] ). But P mRNA molecules edited by the addition of five G residues encode for D, a protein exhibiting a large and specific C-terminal domain of unknown function ( Figure 1A ). Besides co-transcriptional edition, an overlapping ORF embedded in the first half of the P mRNA allows the expression of a single C protein (hPIV3 and bPIV3) or a nested set of four proteins called C9, C, Y1 and Y2 (Sendai virus and hPIV1). The C proteins of Sendai virus and hPIV1 have a high degree of sequence homology and have been studied in details. They are involved in the regulation of viral RNA synthesis [7, 8] , the inhibition of innate immune response [9] and potentially contribute to the budding of viral particles [10] [11] [12] . In particular, the C protein of Sendai virus both inhibits IFN-b production [13, 14] and blocks interferon signaling downstream of IFN-a/b and IFN-c receptors [15] [16] [17] . The C proteins of hPIV3 and bPIV3 only share ,35% of sequence homology with the C proteins of Sendai virus and hPIV1, but they have also been shown to target interferon expression and signaling [5, 18] . Although expression of the C protein of hPIV3 (hPIV3-C) is essential to virulence in vitro and in vivo [19] and explains hPIV3 ability to block IFN-a/b signaling [20] , host proteins that bind hPIV3-C remain unknown. (A) Organization of the gene P of hPIV3 that encodes for three proteins: P, D and C. Whereas conventional transcription and translation lead to the expression of the phosphoprotein P, co-transcriptional insertion of five G residues at the editing site by the virus RNA polymerase leads to the expression of a chimeric protein called D. Insertion of one G residue can also occur during transcription but two stop codons immediately downstream of the editing site prevent the expression of the protein V that is specific of Paramyxoviridae. The protein C is encoded by an overlapping opened reading frame (ORF) embedded in P mRNA. (B and C) HEK-293T cells were transfected with expression vectors encoding GST alone or fused to the C proteins of measles virus (MV-C), hPIV3 (hPIV3-C), or Nipah virus (Nipah-C), and tested for the interaction with endogenous STAT1 (B) or GRB2 (C). Total cell lysates were prepared 48 h post transfection (cell lysate; middle and lower panels), and copurifications of endogenous cellular proteins were assayed by pulldown using glutathione-sepharose beads (GST pull-down; upper panel). GST-tagged C proteins were detected by immunoblotting using anti-GST antibody, while endogenous STAT1 and GRB2 were detected with specific monoclonal antibodies. doi:10.1371/journal.ppat.1000587.g001 Respiroviruses are important pathogens responsible for acute respiratory tract infections associated with severe airway inflammation in children, elderly and immunocompromised individuals. Their RNA genome encodes for structural proteins that compose viral particles, but also for virulence factors that alter cell biology to enhance virus replication and spreading. Among them, the C protein plays a critical role by blocking cellular response to type I interferons, the main antiviral cytokines secreted during virus infections. To provide molecular basis to this activity, we found that the C protein of human parainfluenza virus type 3 (hPIV3-C), the most frequent human Respirovirus, interacts with STAT1, a key component of type I interferon receptor complex. But hPIV3-C was also found to interact with GRB2, an adaptor molecule involved in cellular response to Epidermal Growth Factor (EGF), and to enhance cell response to this cytokine. This pathway increases protein translation, promotes cell survival and contributes to airway inflammation and mucus secretion. Thus, our findings show that hPIV3-C not only inhibits the antiviral response but also stimulates cellular response to EGF, which benefits virus replication and induces an excessive inflammation of airways during infection. In an attempt to answer this question, we performed a yeast two-hybrid screen and we report here the identification of STAT1 and GRB2 as direct interactors of hPIV3-C. Although binding to STAT1 accounts for hPIV3-C ability to block IFN-a/b signaling, the interaction with GRB2 was unexpected. This adaptor protein bridges growth factor receptor tyrosine kinases (RTKs), like Epidermal Growth Factor (EGF) receptor, to the mitogenactivated protein kinase/extracellular signal-regulated kinase (MAPK/ERK) pathway. Upon engagement by their ligands, RTKs autophosphorylate on tyrosine residues to recruit adaptor proteins containing phosphotyrosine binding (PTB) or Src homology 2 (SH2) domains like GRB2 [21] . Once associated to RTKs by its SH2 domain, GRB2 recruits the guanine nucleotidereleasing factor son-of-sevenless (SOS) to activate Ras. Downstream events include MAPK/ERK kinase (MEK1/2) activation, which in turn phosphorylates ERK1/2. Finally, phosphorylated ERK1/2 directly or indirectly activates numerous cellular targets including transcription factors (e.g. Elk1, SAP1, SAP2, c-Fos, CREB, SRF) but also cellular factors that control mRNA translation like eukaryotic initiation factor 4E (eIF4E) or small ribosomal subunit S6 protein [22, 23] . Growth factor binding to RTKs regulates a multiplicity of cellular processes including proliferation, differentiation and survival. In the respiratory tract, this signaling cascade has been shown to trigger inflammation and mucus secretion by epithelial cells [24] [25] [26] [27] , suggesting a critical role in innate immunity [28] . However, excessive activation of this pathway could benefit to virus replication by inhibiting IFN-a/b signaling [29] and promoting infected cell survival [25] . Altogether, these data provided a rational to investigate the functional impact of hPIV3-C expression on IFN-a/b vs EGF receptor and MAPK/ERK signaling pathways. The C protein of hPIV3 interacts directly with STAT1 and GRB2 To identify cellular targets of hPIV3-C, this viral protein was used as bait in the yeast two-hybrid system to screen a human cDNA library. The screen was performed at saturation with a 10-fold coverage of the library (50.10 +6 diploids), and positive yeast colonies growing on selective medium were analyzed by PCR and sequencing to identify binding partners of hPIV3-C. STAT1 and GRB2 were the main interactors of hPIV3-C identified in the screen with 5 and 150 yeast colonies corresponding to these cellular proteins, respectively. In both cases, cDNA clones retrieved from the screen corresponded to full-length STAT1 and GRB2 in frame with the Gal4-AD transactivation domain. To validate these interactions in human cells, GST-tagged hPIV3-C was expressed in HEK-293T cells and purified with glutathion-sepharose beads. As shown in Figure 1B and 1C, endogenous STAT1 and GRB2 co-purified with hPIV3-C. Highly divergent C proteins from measles virus (MV-C) and Nipah virus (Nipah-C) failed to do so, thereby demonstrating the specificity of identified interactions. Binding to STAT1 provides molecular basis to the inhibition of IFN-a/b signaling by hPIV3-C [18] , and parallels the interaction previously identified between Sendai virus C protein and mouse STAT1 [15] . Altogether, this suggests that STAT1 is a specific cellular interactor of Respirovirus C proteins. In contrast, binding to GRB2 is unexpected and suggests a new function for hPIV3-C that we decided to investigate. hPIV3-C has opposite effects on IFN-a/b and EGF signaling pathways The adaptor protein GRB2 plays a critical role in coupling signal from growth factor receptors to MAPK/ERK signaling pathway. To address the question of hPIV3-C interference with this pathway, we used a trans-reporter gene assay that measures Elk1 activation by ERK1/2. In this system, Elk1 transcription factor is fused to the DNA binding domain of Gal4 (Gal4-DB) and binds the promoter sequence of a luciferase reporter gene. Upon stimulation with a growth factor like EGF, Elk1 is activated as assessed by a significant increase in luciferase expression. Surprisingly, we observed a 6-fold enhancement in this cellular response to EGF when 36FLAG-tagged hPIV3-C was expressed in HEK-293T cells (Figure 2A ). Same results were obtained when using hPIV3-C without a tag (14-fold enhancement) or tagged with the red fluorescent protein Cherry (7-fold enhancement). In contrast to hPIV3-C, neither MV-C nor Nipah-C enhanced Elk1 activity upon EGF stimulation (Figure 2A ) whereas expression levels of hPIV3-C, MV-C and Nipah-C were similar in this system ( Figure 2F , left panel). Elk1 activity was also enhanced by hPIV3-C expression in Vero and Hela cells as well as BEAS-2B and A549, two epithelial cell lines that originate from the respiratory tract, which is the tissue targeted by hPIV3 in vivo ( Table 1 ). The effect of hPIV3-C in these different cell lines was highly significant (see p-values in Table 1 ) although relatively modest when compared to HEK-293T cells. This is probably because our reporter system requires the co-transfection of four plasmids and Vero, Hela, BEAS-2B and A549 cells are more difficult to transfect than HEK-293T. In parallel experiments, cellular response to IFN-a/b was monitored using a cis-reporter gene, of which expression is controlled by five ISREs. As previously reported [18] , we found that hPIV3-C efficiently blocked IFN-a/b signaling ( Figure 2B ) as opposed to what we observed for the EGF pathway. Again, MV-C or Nipah-C was unable to do so. Altogether, these results show that hPIV3-C enhances the cellular response to EGF in addition to its ability to block IFN-a/b signaling. We also determined if similar effects on the EGF pathway were observed in infected cells. HEK-293T cells were infected with hPIV3 (MOI = 3) and then transfected with Elk1 activity reporter plasmids. Infection of HEK-293T cells was confirmed by anti-hPIV3 hemagglutinin-neuraminidase (hPIV3-HN) immunostaining and flow cytometry analysis ( Figure 2E ). Like hPIV3-C alone, hPIV3 infection enhanced Elk1 activity upon EGF stimulation ( Figure 2C ). Interestingly, hPIV3 infection induced a significant level of Elk1 activity in the absence of EGF stimulation. This suggests that in addition to hPIV3-C interaction with GRB2, other mechanisms modulate MAPK/ERK pathway during hPIV3 infection. Finally, to demonstrate that enhancement of Elk1 activation by hPIV3-C is completely dependent on ERK1/2 activation, HEK-293T cells were pre-treated with MEK1/2 inhibitor U0126 before stimulation with EGF. This molecule targets MEK1/2 and totally abrogates downstream phosphorylation and activation of ERK1/2 [30] . As shown in Figure 2D , Elk1 activation was blocked by U0126, whereas hPIV3-C expression was maintained ( Figure 2F , right panel). This demonstrates that hPIV3-C is acting through ERK1/2 stimulation. Altogether, these results support a model where hPIV3-C interaction with GRB2 enhances cellular response to growth factors as assessed by an increased activation of MAPK/ ERK pathway. Phosphorylation of ERK1/2, eIF4E and small ribosomal subunit S6 protein are stimulated by hPIV3-C expression or hPIV3 infection To further document hPIV3-C impact on MAPK/ERK signaling pathway, we compared the kinetic of ERK1/2 phosphorylation in HEK-293T cells expressing hPIV3-C or not. Cells were transfected hPIV3-C Impact on IFN-a/b and EGF Pathways In addition to these three plasmids, cells were co-transfected with an expression vector encoding 36FLAG-tagged MV-C, hPIV3-C or Nipah-C or the corresponding empty vector pCI-neo-36FLAG. 12 h after transfection, cells were starved and 6 h later EGF was added at a final concentration of 100 ng/ml. After 24 h, relative luciferase activity was determined. (B) HEK-293T cells were transfected with pISRE-Luc, a plasmid containing a luciferase gene of which expression is controlled by five ISREs, and pRL-CMV. In addition to these two plasmids, cells were co-transfected with an expression plasmid encoding 36FLAG-tagged MV-C, hPIV3-C or Nipah-C or the corresponding empty vector pCI-neo-36FLAG. 24 h after transfection, 1000 IU/ml of recombinant IFN-b were added. After 24 h, relative luciferase activity was determined. (C) HEK-293T cells were infected with hPIV3 (MOI = 3) and then transfected with pFA2-Elk1, pGal4-UAS-Luc, pRL-CMV vectors. 12 h later, cells were starved during 6 h and stimulated with EGF at a final concentration of 100 ng/ml. After 24 h, relative luciferase activity was determined. (D) Same experiment as (A) but 20 mM of MEK1/2 specific inhibitor U0126 was added as indicated. (A-D) All experiments were achieved in triplicate, and data represent means6SD. (E) HEK-293T cells were infected as in (C), and hPIV3-HN expression determined by immunostaining and flow cytometry analysis. (F) HEK-293T cells were transfected to express 36FLAG-tagged MV-C, hPIV3-C or Nipah-C as described in (A) and (B), and relative expression levels were determined 36 h later by western blot analysis (left panel). In a parallel experiment, HEK-293T cells were transfected to express hPIV3-C and were cultured with or without EGF in the presence or absence of U0126 as described in (D). hPIV3-C expression level was determined by western blot analysis (right panel). doi:10.1371/journal.ppat.1000587.g002 with 36FLAG-tagged hPIV3-C or a control plasmid and 24 h post transfection, they were starved before stimulation with EGF. ERK1/ 2 phosphorylation was determined at 10, 30 and 120 min after stimulation. As illustrated by one representative experiment in Figure 3A , EGF stimulation induced ERK1/2 phosphorylation in control cells but signal was markedly and reproducibly increased by hPIV3-C expression at maximum phosphorylation time point (1.4 to 2.8 fold increase; p = 0.005; n = 4). We then determined the phosphorylation level of two downstream targets of this pathway that are involved in the control of mRNA translation, the translation initiation factor eIF4E and the ribosomal protein S6 ( Figure 3B ). Before EGF stimulation, low levels of phosphorylated eIF4E and S6 were detectable in mock-treated cells ( Figure 3B and 3D). hPIV3-C expression had virtually no effects on this background. Thus, eIF4E and S6 phosphorylation levels were determined at different time-points after EGF stimulation. Because ERK1/2 activation precedes eIF4E and S6 phosphorylation, maximal phosphorylation occurs at later time points and was determined at 30 min, 2 h, 6 h and 24 h after stimulation. As observed for ERK1/2, phosphorylation levels of eIF4E and S6 were enhanced by hPIV3-C expression when stimulating the cells with EGF. To validate these observations in infected cells, HEK-293T cells were infected with hPIV3 (MOI = 3) and 24 h later, cells were starved for 12 h before stimulation with EGF. Like hPIV3-C expression alone, hPIV3 infection enhanced ERK1/2 phosphorylation at the peak of induction, i.e. 10 min after adding EGF to the cells ( Figure 3C ). Interestingly, hPIV3 infection of A549 cells also enhanced ERK1/2 phosphorylation but the induction profile was different. Indeed, ERK1/2 phosphorylation was not significantly increased at the peak of induction, but the signal was boosted by hPIV3 infection at late time points ( Figure S1 ). The same profile was observed when eIF4E and S6 phosphorylation levels were analyzed in infected HEK-293T cells. hPIV3 infection sustained the phosphorylation of these two translation factors at late time points, but showed no increase at the peak of stimulation, i.e. 30 min after adding EGF to the cells ( Figure 3D ). This could relate to the fact that hPIV3 infection also induces low levels of eIF4E and S6 phosphorylation in the absence of EGF stimulation ( Figure 3D ). This is reminiscent to what was observed for Elk1 ( Figure 2C ), and suggests that hPIV3 infection induces a basal activation of MAPK/ERK pathway leading to the constitutive phosphorylation of downstream targets. Altogether, these data demonstrate that hPIV3 infection or hPIV3-C expression alone both enhance MAPK/ERK pathway activation in EGF-stimulated cells. Several RNA viruses require an activated MAPK/ERK pathway to produce viral components and replicate properly (for review see [31] ). To test if the same was true for hPIV3, cells were treated for 2 h with MAPK/ERK pathway inhibitor U0126 and infected with hPIV3 (MOI = 1). Two days after infection, cell surface expression of hPIV3-HN was detected by immunostaining and flow cytometry. U0126 completely blocked the expression of hPIV3-HN in hPIV3-infected cells (Figure 4 ), whereas the same inhibitor had no effect when cells were infected with MV ( Figure S2 ). Altogether, this demonstrates that MAPK/ERK signaling is essential for the expression of hPIV3 proteins and suggests that hPIV3 manipulates this pathway to increase replication efficiency. The C-terminal region of hPIV3-C binds STAT1 and GRB2 To better understand how hPIV3-C targets both the IFN-a/b and EGF signaling pathways, we characterized the functional domains of hPIV3-C that bind STAT1 and GRB2. To do so, we generated by PCR a full matrix of hPIV3-C overlapping fragments and tested their ability to interact with either STAT1 or GRB2 in the yeast two-hybrid system ( Figure 5 and 6 ). Both forward and reverse primers were designed every 75 nucleotides along hPIV3-C sequence and fused to appropriate tails to allow gap-repair recombination with linearized Gal4-DB yeast two-hybrid vector ( Figure 5A ). All possible combinations of forward and reverse primers were used to amplify hPIV3-C fragments. Finally, corresponding PCR products were transformed in a yeast strain expressing Gal4-AD fused to either STAT1 or GRB2, and growth on selective medium was used to detect potential interactions. A 124 (AA)-long peptide encompassing position 76 to 199 located in the C-terminal half of hPIV3-C was sufficient to bind STAT1 ( Figure 5B ) or GRB2 ( Figure 6A ). In an iterative process, we then generated a second, a third and a fourth set of hPIV3-C fragments corresponding to one-by-one AA deletions ( Figure 5C -E and Figure 6B -D), allowing to further reduce the STAT1 and GRB2 binding motifs to minimal peptides. A 106 AA peptide encompassing residues 90 to 195 of hPIV3-C was sufficient to observe the interaction with STAT1 ( Figure 5E ). The binding region to GRB2 was virtually the same, encompassing AA 97 to 195 ( Figure 6D ). The C-terminal region of hPIV3-C required to bind STAT1 and GRB2 in the yeast two-hybrid system is highly conserved among Respiroviruses ( Figure 7A ) and suspected to fold into a structured coiled-coil domain [18] . Furthermore, virtually the same C-terminal region of Sendai virus C protein (AA 85-204) was previously reported to mediate the interaction with mouse STAT1 [32] . To further validate our observations performed in the yeast two-hybrid system, we retested by co-affinity purification the ability of hPIV3-C fragment encompassing AA 90-195 (hPIV3-C 90-195 ) to interact with STAT1 and GRB2 in HEK-293T cells. GST-tagged hPIV3-C 90-195 was expressed together with 36-FLAG-tagged STAT1 or GRB2, and purified with glutathionsepharose beads. Full-length hPIV3-C and the N-terminal region encompassing AA 1-89 (hPIV3-C 1-89 ) were used as positive and negative controls, respectively. As shown in Figure 7B and 7C, hPIV3-C 90-195 interacted with STAT1 and GRB2, whereas hPIV3-C 1-89 did not. Although hPIV3-C 90-195 interacted with GRB2 as efficiently as full-length hPIV3-C ( Figure 7C ), interaction with STAT1 was weaker suggesting that more residues contribute to the stabilization of this interaction ( Figure 7B ). Altogether, these results confirm that AA 90-195 of hPIV-3 include both STAT1 and GRB2 binding sites. As described in Figure 2 , HEK-293T, Hela, Vero, A549 or BEAS-2B cells were transfected with pFA2-Elk1, pGal4-UAS-Luc, pRL-CMV to measure the activation level of MAPK/ERK signaling pathway. Cells were co-transfected with plasmids encoding 36FLAG-tagged MV-C, hPIV3-C or Nipah-C or the corresponding empty vector pCI-neo-36FLAG. 12 h after transfection, cells were starved and stimulated 6 h later with 100 ng/ml of EGF. After 24 h, expression of luciferase was quantified. Results were normalized so that reporter activity in cells transfected with a control vector equals 1. Experiments were performed in triplicates and data represent means6SD. doi:10.1371/journal.ppat.1000587.t001 hPIV3-C Impact on IFN-a/b and EGF Pathways Although STAT1 and GRB2 essentially bind to the same region of hPIV3-C as demonstrated above, it remained unclear whether these interactions are mutually exclusive. To answer this question, a competition experiment was designed where GST-tagged hPIV3-C was co-expressed with STAT1 in the presence or absence of GRB2 ( Figure 7D ). In this setting, GRB2 expression prevents STAT1 copurification together with GST-tagged hPIV3-C. This validates our finding that STAT1 and GRB2 interact with the same region of hPIV3-C, and demonstrates that STAT1 and GRB2 compete for hPIV3-C binding. Interestingly, GRB2 interaction with hPIV3-C was not affected by STAT1 expression (Figure 7D and data not shown), suggesting that GRB2 has a higher affinity for hPIV3-C than STAT1. Both the N-and C-terminal domains of hPIV3-C are required for its activity We finally tested if hPIV3-C 90-195 was able, like full-length hPIV3-C, to block IFN-a/b signaling and enhance cellular response to EGF stimulation. First, cells were transfected with 36FLAG-tagged hPIV3-C, hPIV3-C 90-195 or hPIV3-C 1-89 together with the IFN-a/b reporter plasmid, and stimulated 24 h later with recombinant IFN-b. Reporter gene expression was determined 24 h post transfection and found to be inhibited exclusively by full-length hPIV3-C ( Figure 8A ). Although this may reflect the weakness of hPIV3-C interaction with STAT1 ( Figure 7B ), this also indicates that both the N-terminal and C-terminal regions of hPIV3-C are required to block IFN-a/b signaling, even if only the C-terminal region is required for the binding to STAT1. The same constructs were tested using the Elk1 activity reporter plasmids ( Figure 8B ). Again, only full-length hPIV3-C was able to enhance Elk1 activation upon EGF stimulation whereas full-length hPIV3-C and hPIV3-C 90-195 were expressed at similar levels in transfected cells ( Figure 8B , upper right panel). Because GRB2 binding to hPIV3-C and hPIV3-C 90-195 were essentially equivalent in co-affinity purification experiments, we hypothesized that the N-terminal region of hPIV3-C was required for its proper subcellular localization. Thus, hPIV3-C, hPIV3-C 90-195 and hPIV3-C 1-89 were expressed in fusion downstream of the red fluorescent protein Cherry. As shown in Figure 8C , Cherry alone or fused to hPIV3-C 90-195 localized both in the nucleus and the cytoplasm of transfected cells. In contrast, full-length hPIV3-C essentially accumulated at the cellular membrane whereas hPIV3-C 1-89 was in the nucleus. Although we have no explanation for this unexpected localization of hPIV3-C 1-89 , these observations show that only full-length hPIV3-C is able to target the cellular membrane where both IFN-a/b and EGF signaling are triggered. Paramyxoviridae have evolved various mechanisms to block IFNa/b response, in particular signaling downstream IFNAR1/ IFNAR2c receptor [20, 33] . Although members of Pneumovirinae subfamily have specific genes to encode inhibitors of IFN-a/b signaling pathway, those expressed by other Paramyxoviridae (i.e. Paramyxovirinae subfamily) are encoded by overlapping reading frames embedded within the gene P. Rubulaviruses express V proteins that target STAT1 and/or STAT2 for ubiquitination and degradation, while Morbilliviruses and Henipaviruses V proteins essentially impair STAT1/2 phosphorylation, activation and nuclear translocation. In addition, Morbilliviruses and Henipaviruses also encode for C proteins of which role in the inhibition of IFNa/b response has been a matter of debates [34] [35] [36] [37] . Recent reports showed that Morbillivirus C proteins only have a minor role in the inhibition of IFN-a/b signaling [37] , but are essential to block IFN-a/b induction [38] . Whether Henipavirus C proteins can directly block IFN-a/b or promote viral replication through alternative mechanisms is unclear [35] . In contrast, it has been clearly established that Respirovirus C proteins are potent inhibitors of IFN-a/b signaling [5, 18, [39] [40] [41] . In this report, we show that hPIV3-C, but not MV-C or Nipah-C, directly interacts with STAT1 and efficiently inhibits IFN-a/b signaling. In addition, we identified a minimal STAT1 binding domain that encompasses AA 90-195 of hPIV3-C, a region suspected to fold into a coiled coil. Interestingly, this conserved domain is localized within the STAT1 binding region shared by all four isoforms of Sendai virus C protein [32] . Together, these results confirm the capacity of hPIV3-C to block IFN-a/b signaling pathway [18] , provide molecular basis to this inhibition and clarify the fact that Respirovirus C proteins are functionally distinct from Morbillivirus and Henipavirus C proteins. In addition to STAT1, we show that hPIV3-C interacts directly with GRB2 and enhances MAPK/ERK signaling downstream of EGF receptor (EGFR). Our data give the first example of a Paramyxoviridae protein that contributes to the stimulation of EGFR and MAPK/ERK pathway and provides molecular basis to this activity. This pathway has been known for decades as a prime target of DNA tumor viruses and oncogenic retroviruses, and its activation represents an essential step toward carcinogenesis [3, 42] . But recent data demonstrate that non-oncogenic RNA viruses also activate this signaling cascade to support viral replication and spreading [31] . Whether it is activated upon EGFR engagement or other means, MAPK/ERK pathway regulates a multiplicity of cellular processes including proliferation, differentiation, development, cell survival and inflammation. As a consequence, how the activation of MAPK/ERK pathway promotes viral replication is a complex question. Interestingly, two non-oncogenic RNA viruses associated with acute respiratory tract infections have been recently reported to modulate the EGFR pathway. Both human respiratory syncytial virus (hRSV), a member of Paramyxoviridae like hPIV3, and a rhinovirus that belongs to Picornaviridae family activate EGFR and MAPK/ERK pathway [25, 27] . Infection of epithelial cells by these viruses stimulates the processing and activation of EGFR ligands by membrane matrix metalloproteinase and subsequent engagement of EGFR through autocrine/paracrine mechanisms. Experiments performed on rhinovirus show that viral replication and TLR3 engagement by viral RNA are both required to activate the EGFR and MAPK/ERK pathway [27] . In this report, we show that hPIV3 infection also activates MAPK/ERK pathway in the absence of external stimuli, a phenomenon that possibly relies on the engagement of pathogen recognition receptors. Although hPIV3-C alone is unable to activate this pathway, our data suggest that expression of this virulence factor enhances MAPK/ERK activation above normal level in infected cells, thereby contributing to viral replication and pathogenesis. Induction of MAPK/ERK pathway by RNA viruses has numerous consequences on cell biology. First, it results in increased expression of inflammatory factors, in particular cytokines and chemokines that recruit cellular effectors of immunity [27, [43] [44] [45] [46] [47] . MAPK/ERK pathway was also reported to block the antiviral response induced by IFN-a/b, making its activation beneficial to virus replication [29] . Another consequence of MAPK/ERK pathway activation is the induction of mucin production by infected epithelial cells [24, 27] . Although mucin expression is a critical innate defense system, excessive production of mucus results in the obstruction of airways and delays the elimination of pathogens. Finally, it has been demonstrated in vitro that upon hRSV infection, activation of EGFR and MAPK/ERK pathway sustains viral replication by retarding the death of infected cells [25] . Altogether, these data suggest that although a moderate activation of MAPK/ERK pathway contributes to the innate response against viruses, an excessive activation leads to deleterious inflammation, inhibition of IFN-a/b response, airway obstruction and infected cell survival Figure 6 . Identification of a minimal hPIV3-C region interacting with GRB2. Fragments of hPIV3-C were generated and tested for their ability to interact with GRB2 following the procedure described in Figure 5A . (A) The first iteration identified a GRB2 binding region of 124 AA. After two (B), three (C) and four additional rounds (D), this domain was finally reduced to a minimal GRB2 binding motif of 99 AA. This binding domain, encompassing position 97 to 195, is contained in the STAT1 binding region of hPIV3-C previously identified in Figure 5 . doi:10.1371/journal.ppat.1000587.g006 hPIV3-C Impact on IFN-a/b and EGF Pathways [28] . Therefore, it is tempting to speculate that hPIV3-C interaction with GRB2 and EGFR pathway participates in such deregulation of airway epithelium homeostasis to promote hPIV3 replication and spreading. A consequence of such perturbations could be an aggravation of chronic inflammatory airway diseases like asthma or chronic obstructive pulmonary disease as already suggested by epidemiological links with Paramyxoviridae infections and in vivo models [48, 49] . Besides its effects on immune response, activation of MAPK/ ERK pathway has direct consequences on viral replication as assessed by in vitro experiments. It is now well documented that MAPK/ERK pathway inhibition with U0126 or PD098059 deeply impairs the replication of numerous RNA viruses including hRSV ( [50] ; and for review see [31] ). Similarly, we show in this report that MAPK/ERK pathway inhibition prevents hPIV3 protein expression in infected cells as assessed by hPIV3-HN detection. In influenza virus infected cells, membrane accumulation of influenza virus hemagglutinin (HA) induces lipid-rafts clustering that leads to MAPK/ERK pathway activation and nuclear export of viral ribonucleoprotein complexes to achieve A and B) hPIV3-C 90-195 and hPIV3-C 1-89 were tested for their ability to modulate IFN-a/b or EGF signaling. (A) As described in Figure 2 , HEK-293T cells were transfected with pISRE-Luc and pRL-CMV to determine the activation level of IFN-a/b signaling pathway. Cells were co-transfected with expression plasmids encoding 36FLAG-tagged full-length hPIV3-C (C FL ) or fragments. 24 h after transfection, 1000 IU/ml of recombinant IFN-b were added. After 24 h, relative luciferase activity was determined. Experiments were performed in triplicates, and data represent means6SD. (B) As described in Figure 2 , cells were transfected with pFA2-Elk1, pGal4-UAS-Luc and pRL-CMV to determine the activation level of MAPK/ERK signaling pathway. Cells were co-transfected with expression plasmids encoding full-length hPIV3-C (C FL ) or fragments. 12 h later, cells were starved during 6 h and stimulated with EGF at a final concentration of 100 ng/ml. After 24 h, relative luciferase activity was determined. Experiments were performed in triplicate, and data represent means6SD. As a control, relative expression levels of C FL , C 90-195 and C 1-89 were determined by western blot analysis (upper right panel). (C) Full-length hPIV3-C, hPIV3-C 90-195 or hPIV3-C 1-89 was expressed in fusion downstream of the red fluorescent protein Cherry to determine subcellular localization in HEK-293T cells. 24 h after transfection, cells were fixed with PFA, permeabilized and labeled with DAPI to stain nuclei. Single confocal sections show Cherrytagged protein fluorescence in red and DAPI staining in blue. doi:10.1371/journal.ppat.1000587.g008 viral particles assembly [51] [52] [53] . Because hPIV3 replication cycle is only cytoplasmic, mechanisms involved are necessarily distinct. A possible link between MAPK/ERK pathway and hPIV3 protein expression lies in the fact that among downstream targets of this pathway are essential factors of cellular translational machinery. We show that hPIV3-C expression enhances the phosphorylation of S6 and eIF4E. The small ribosomal subunit protein S6 is the major phosphoprotein of eukaryotic ribosomes with five phosphorylation sites (Ser235, Ser236, Ser240, Ser244, and Ser247). Two families of serine/threonine kinases phosphorylate S6 in vitro: S6K1/2 and p90 ribosomal S6 kinase (RSK). Recently it has been shown that MAPK/ERK signaling pathway activates RSK family members that contribute to S6 phosphorylation on Ser235/236 thereby stimulating cap-dependent translation [23] . In addition, eIF4E that interacts with the cap structure and brings translation initiation factors together with the small ribosomal subunit via the scaffold protein eIF4G, undergoes regulated phosphorylation on Ser209 upon MAPK/ERK pathway activation. This phosphorylation event is dependent on eIF4Gassociated MAPK signal-integrating kinases, Mnk1 and Mnk2 [22] . eIF4E is believed to be the least abundant of all initiation factors and therefore considered as a perfect target to regulate protein synthesis. Even though there is no direct link between eIF4E phosphorylation and the enhanced translation observed, the fraction of phosphorylated eIF4E dramatically increases following treatment of the cells with growth factors, hormones and mitogens. Therefore, eIF4E phosphorylation has been associated with increased translation rates. hPIV3 mRNAs are capped and polyadenylated like their host counterparts. Thus, S6 and eIF4E phosphorylation together with a high level of viral gene transcription may contribute to a rapid switch toward viral protein synthesis within infected cells. Specific biochemical investigations are still required to decipher how hPIV3-C can both inhibit IFN-a/b signaling and enhance EGFR and MAPK/ERK pathway. When searching the literature for viral proteins that target GRB2, we found specific reports on NS5A from hepatitis C virus and ORF3 from hepatitis E virus [54, 55] . Although NS5A inhibits MAPK/ERK activation induced by exogenous EGF, ORF3 was described as an activator of MAPK/ERK pathway like hPIV3-C. Both NS5A and ORF3 exhibit a proline-rich motif (PXXP) to bind the Src homology 3 (SH3) domains of GRB2, but there is no such motif in hPIV3-C suggesting that other mechanisms mediate STAT1 and GRB2 binding. Interestingly, these two cellular proteins exhibit SH2 domains. Such domains typically bind a phosphorylated tyrosine residue in the context of a longer peptide motif within a target protein. Although there is no evidence that hPIV3-C becomes phosphorylated, we have tested hPIV3-C interaction with mutant STAT1 and GRB2 exhibiting SH2 domains disabled for the interaction with phosphotyrosine residues. These mutants were not affected for the interaction with hPIV3-C (data not shown). This suggests that hPIV3-C either binds distinct regions of STAT1 and GRB2, or interacts with a region of the SH2 domain that does not involve the phosphotyrosine binding site. Finally, our results also show that full-length hPIV3-C is required to modulate IFNa/b and MAPK/ERK pathways since AA 90-195 that bind STAT1 and GRB2 are unable to do so when expressed alone. Full-length hPIV3-C was also required to observe a localization at the cell membrane, suggesting a link with its activity. Interestingly, the N-terminal 23 residues of Sendai virus C protein act as a membrane targeting signal [56] . But the N-terminal residues of hPIV3-C (AA 1-89) were unable to do so, and sequence analysis did not show any conservation with the C protein of Sendai virus. Thus, hPIV3-C tertiary structure is apparently required to target this protein at the cell membrane. This specific localization could both sequester STAT1 to prevent the stimulation of IFN-target genes and contribute to the aggregation of GRB2-SOS complexes to enhance MAPK/ERK signaling [57] . Altogether, this suggests that hPIV3-C interaction with STAT1 and GRB2 represents a potential target for the development of antiviral molecules against hPIV3 and possibly other members of Respirovirus genus. Plasmid DNA constructs P-encoding sequence from hPIV3 wild-type strain (DF042505) was amplified by RT-PCR (Titan One tube; Roche Applied Science) from total RNA purified from infected cells (RNeasy kit; Qiagen). Amplification was performed using the following hPIV3-P specific primers flanked with Gateway cloning sites: 59-gggga-caactttgtacaaaaaagttggcatgGAAAGCGATGCTAAAAACTATC-AAA and 59-ggggacaactttgtacaagaaagttggttaTTGGCAATTATT-GACATCTTCATTGAAC. PCR products were cloned using TOPO TA Cloning kit (Invitrogen) into TOPO vector. A total of 21 clones were analyzed to establish the sequence of hPIV3-P (GenBank ID: EU719627). Interestingly, 8 clones were not edited, 11 clones were edited by the addition of one G residue, and 2 clones were edited by the addition of 5 G residues. One of the plasmids containing the unedited sequence of hPIV3-P was selected and subsequently used as a template to clone hPIV3-C. DNA sequences encoding full-length hPIV3-C or fragments corresponding to AA 1-89 or 90-195 were amplified by PCR from p(hPIV3-P)-TOPO and cloned by in vitro recombination into pDONR207 (Gateway system; Invitrogen) as previously described [58] . Similarly, MV-C was amplified from p(+)MV323 that contains the full genome of measles virus wild-type strain Ichinose (kindly provided by Dr. K. Takeuchi, [59] ). Nipah-C was amplified from NiV-P plasmid (kindly provided by Dr. TF. Wild; [60] ). GRB2 coding sequence was amplified from the human spleen cDNA library used to perform the yeast two-hybrid screen (Invitrogen). The pDONR207 plasmid containing STAT1 was previously described [58] . Viral or cellular coding sequences were subsequently transferred by in vitro recombination from pDONR207 into different Gateway-compatible destination vectors (see below) following manufacturer's recommendation (LR cloning reaction, Invitrogen). To perform yeast two-hybrid experiments, coding sequences were recombined into pPC86 (Invitrogen) to be expressed in fusion downstream of the activation domain of Gal4 (Gal4-AD) or into pDEST32 to be expressed in fusion downstream of the DNA binding domain of Gal4 (Gal4-DB). In mammalian cells, GST-tag and 36FLAG-tag fusions were achieved using pDEST27 (Invitrogen) or pCI-neo-36FLAG vector, respectively [61] . We also used pCI-neo (Promega) and pmCherry-C1 (Clontech) to express proteins without a tag or in fusion downstream of Cherry, respectively. These two plasmids were made Gateway-compatible using the Gateway vector conversion system (Invitrogen). HEK-293T, Hela and Vero cells were maintained in Dulbecco's modified Eagle's medium (DMEM; Gibco-Invitrogen) containing 10% fetal bovine serum, penicillin, and streptomycin at 37uC and 5% CO 2 . A549 and BEAS-2B cells were maintained in F-12K medium (Gibco-Invitrogen) containing 10% fetal bovine serum, penicillin, and streptomycin at 37uC and 5% CO 2 . hPIV3 (strain C243) was amplified and titrated on Vero cells following recommendations of ATCC (American Type Culture Collection). Recombinant MV-EGFP virus used in Figure S2 has been hPIV3-C Impact on IFN-a/b and EGF Pathways previously described [62] . Infections were performed for 2 h at 37uC in Optimem (Gibco-Invitrogen). Later on, cells were washed and incubated in fresh culture medium for 24 or 48 h. To detect viral replication, cells were recovered and incubated in PBSparaformaldehyde 3.2% for 20 min. After extensive washing with PBS, cells were permeabilized with PBS-Triton 0.05% for 15 min, and then incubated with a monoclonal antibody specific to hPIV3-HN (M02122321, Abcam). Cells were washed and incubated in the presence of an anti-mouse Cy3-conjugated antibody (Jackson Immunoresearch). After extensive washing, cellular immunostaining was analyzed using a FACSCalibur flow cytometer (BD). When specified, cells were pre-treated with MEK1/2 specific inhibitor U0126 (20 mM final; Promega) for 2 h before, during and after infection to study the impact on hPIV3 infection. To perform co-affinity purification experiments, cloned ORFs were transferred from pDONR207 to pDEST27 expression vector (Invitrogen) to achieve GST fusion, and to pCI-neo-36FLAG vector [61] for 36FLAG-fusion. Cell transfections were performed using Lipofectamine 2000 (Invitrogen). Unless specified otherwise, 5610 5 HEK-293T cells were dispensed in each well of a 6-well plate, and transfected 24 h later with 600 ng of each plasmid DNA per well. Two days post transfection, HEK-293T cells were washed in PBS, then resuspended in lysis buffer (0.5% Nonidet P-40, 20 mM Tris-HCl at pH 8, 120 mM NaCl and 1 mM EDTA) supplemented with Complete Protease Inhibitor Cocktail (Roche). Cell lysates were incubated on ice for 20 min, and then clarified by centrifugation at 14,0006g for 10 min. For pull-down analysis, 400 mg of protein extracts were incubated for 1 h at 4uC with 25 ml of glutathione-sepharose beads (Amersham Biosciences) to purify GST-tagged proteins. Beads were then washed 3 times in ice-cold lysis buffer and proteins were recovered by boiling in denaturing loading buffer (Invitrogen). Purified complexes and protein extracts were resolved by SDSpolyacrylamide gel electrophoresis (SDS-PAGE) on 4-12% NuPAGE Bis-Tris gels with MOPS running buffer (Invitrogen), and transferred to a nitrocellulose membrane. Proteins were detected using standard immunoblotting techniques. 36FLAGand GST-tagged proteins were detected with a mouse monoclonal HRP-conjugated anti-36FLAG antibody (M2; Sigma-Aldrich) and a rabbit polyclonal anti-GST antibody (Sigma-Aldrich), respectively. Specific antibodies were used to detect endogenous STAT1 (clone-1; BD Biosciences), GRB2 (clone-81; BD Biosciences), phospho-ERK1/2 (clone-12D4; Upstate), ERK1/2 (CT; Upstate), phospho-eIF4E (Ser209; Cell Signaling), eIF4E (Cell Signaling), phospho-S6 235-236 (Ser235/236; Cell Signaling), S6 (clone-54D2; Cell Signaling) and hPIV3-HN (M02122321; Abcam). Secondary anti-mouse and anti-rabbit HRP-conjugated antibodies were from GE-Healthcare. Densitometric analysis of the gels was performed using a specific module of Photoshop CS3 Extended (Adobe Systems Inc.). Yeast two-hybrid screening and gap-repair procedure Our yeast two-hybrid protocols have been described in details elsewhere [58] . Briefly, pDEST32 plasmid encoding Gal4-DB fused to hPIV3-C was transformed in AH109 yeast strain (Clontech), and used to screen by mating a human spleen cDNA library cloned in the Gal4-AD pPC86 vector (Invitrogen) and previously established in Y187 yeast strain (Clontech). Yeast cells were plated on a selective medium lacking histidine and supplemented with 10 mM 3-amino-triazole (3-AT; Sigma-Aldrich) to select for interaction-dependent transactivation of HIS3 reporter gene. AD-cDNAs from [His+] colonies were amplified by PCR and sequenced to identify the host proteins interacting with hPIV3-C. The gap-repair procedure was used to map the minimal portion of hPIV3-C interacting with STAT1 and GRB2. As previously described [63] , both forward and reverse PCR primers were designed along the sequence of hPIV3-C and fused to specific tails allowing yeast-based recombination in Gal4-DB two-hybrid vector. Matrix combinations of forward and reverse primers were used to amplify fragments of hPIV3-C by PCR. AH109 yeast cells expressing AD-fused STAT1 or GRB2 were co-transformed with 5 mL of each PCR product in the presence 50 ng of linearized pDEST32 vector to achieve recombinatorial cloning by gaprepair. Fragments of hPIV3-C fused to Gal4-DB were then tested for interaction with AD-STAT1 or AD-GRB2 by plating yeast cells on selective medium lacking histidine and supplemented with 10 mM of 3-AT. Luciferase reporter gene assay HEK-293T, Hela or Vero cells were plated in 24-well plates (2610 5 cells per well). One day later, cells were transfected with either pISRE-Luc (0.3 mg/well; Stratagene) or pFA2-Elk1 (0.3 mg/well; Stratagene) and pGal4-UAS-Luc plasmids (0.3 mg/ well; provided by Dr. Y. Jacob) together with pRL-CMV reference plasmid (0.03 mg/well; Promega). Cells were simultaneously cotransfected with 0.3 mg/well of pCI-neo-36FLAG, pCI-neo or pmCherryC1 expression vectors encoding viral proteins as specified. 24 h after transfection, cells were stimulated with IFNb (Biosource) at 1000 IU/ml or starved for 6 h then stimulated with EGF (Upstate) at 100 ng/ml. 24 h post transfection, cells were lysed, and both firefly and Renilla luciferase activities in the lysate were determined using the Dual-luciferase Reporter Assay System (Promega). Reporter activity was calculated as the ratio of firefly luciferase activity to reference Renilla luciferase activity, and normalized so that positive control activity equals 100. When indicated, cells were treated with U0126 (Promega) at 20 mM final concentration upon EGF stimulation. Subcellular localization of Cherry-tagged hPIV3-C FL , C 1-89 and C 24-well plates containing coverslips were seeded with HEK-293T cells (2610 5 cells per well). One day later, cells were transfected with pmCherryC1 expression vector alone or encoding hPIV3-C FL , hPIV3-C 1-89 or hPIV3-C 90-195 . 36 h after transfection, cells were incubated with PBS-PFA 4% for 20 min at RT, then treated with PBS-Triton 0.05% for 15 min at RT to permeabilize the cells. Finally, cells were incubated for 10 min at RT in a PBS-PFA 4% solution containing DAPI (49-6-Diamidino-2-phenylindole) at 10 mg/ml. Preparations were mounted using Fluoromount-G (Southernbiotech), and imaging performed using a SP5 confocal miscroscope (Leica). Figure S1 ERK1/2 phosphorylation is enhanced by hPIV3 infection in A549 cells. A549 cells were infected with hPIV3 (MOI = 3) and after 24 h, cells were starved for 12 h before stimulation with 100 ng/ml of EGF. Phosphorylation of ERK1/2 was determined by western blot analysis at 10 min, 30 min and 2 h post stimulation. hPIV3 infection was confirmed by anti-hPIV3 hemagglutinin-neuraminidase (hPIV3-HN) immunoblotting. Found at: doi:10.1371/journal.ppat.1000587.s001 (7.34 MB TIF) hPIV3-C Impact on IFN-a/b and EGF Pathways Figure S2 MEK1/2 inhibitor U0126 has no effect on MV protein synthesis. HEK-293T cells were left untreated (A) or treated with 20 mM of U0126 for 2 h (B). Then, cells were mocktreated or infected with a recombinant MV strain expressing EGFP (MOI = 1) and cultured with or without U0126 (A and B, respectively). 48 h after infection, EGFP expression was quantified by flow cytometry analysis. Found at: doi:10.1371/journal.ppat.1000587.s002 (7.34 MB TIF)

A host type I interferon response is induced by cytosolic sensing of the bacterial second messenger cyclic-di-GMP

The innate immune system responds to unique molecular signatures that are widely conserved among microbes but that are not normally present in host cells. Compounds that stimulate innate immune pathways may be valuable in the design of novel adjuvants, vaccines, and other immunotherapeutics. The cyclic dinucleotide cyclic-di–guanosine monophosphate (c-di-GMP) is a recently appreciated second messenger that plays critical regulatory roles in many species of bacteria but is not produced by eukaryotic cells. In vivo and in vitro studies have previously suggested that c-di-GMP is a potent immunostimulatory compound recognized by mouse and human cells. We provide evidence that c-di-GMP is sensed in the cytosol of mammalian cells via a novel immunosurveillance pathway. The potency of cytosolic signaling induced by c-di-GMP is comparable to that induced by cytosolic delivery of DNA, and both nucleic acids induce a similar transcriptional profile, including triggering of type I interferons and coregulated genes via induction of TBK1, IRF3, nuclear factor κB, and MAP kinases. However, the cytosolic pathway that senses c-di-GMP appears to be distinct from all known nucleic acid–sensing pathways. Our results suggest a novel mechanism by which host cells can induce an inflammatory response to a widely produced bacterial ligand.

To sense infection, the innate immune system preferentially responds to conserved molecular signatures of microbes that are absent from normal host cells. There is currently great interest in determining what molecular features of pathogens are detected, and how these features are sensed by the innate immune system. A better understanding of the fundamental principles of how immune responses are stimulated is critical for the design of more effective vaccines, adjuvants, and immune therapeutics. One important class of ligands sensed by the innate immune system is nucleic acids. Several distinct nucleic acid sensors have been described and have been found to be distributed to distinct subcellular locations and exhibit distinct specificities for different nucleic acid ligands. A common theme is that nucleic acid sensors tend to induce expression of type I IFN genes and, thus, appear to be particularly important for initiating immune responses to viruses. including Listeria monocytogenes, Mycobacterium tuberculosis, Legionella pneumophila, Francisella tularensis, group B Streptococcus, and Brucella abortus, it appears that induction of type I IFN is via a novel cytosolic pathway and is independent of TLRs. It has been suggested that these pathogens trigger the cytosolic DNA-sensing pathway, but neither the host sensors nor the bacterial ligands that trigger cytosolic induction of type I IFNs by these pathogens have been identified. Intriguingly, evidence from L. monocytogenes suggests that bacterial multidrug efflux pumps are required for induction of type I IFN , and raises the possibility that a small molecule substrate of these pumps, rather than DNA, is the bacterial trigger of type I IFN gene expression. Nucleotide second messengers are critical transmitters of signaling in all living things. cAMP is used by bacteria and eukaryotes alike, whereas certain nucleotide second messengers are unique to bacteria. For example, guanosine tetraphosphate (ppGpp) is a key regulator of the stringent response in bacteria (Srivatsan and Wang, 2008) . Another nucleotide second messenger, cyclic-di-GMP (c-di-GMP), is a relatively recently appreciated cyclic ribonucleotide (Ross et al., 1987) synthesized by bacteria from two GTP precursors that are hydrolyzed and ultimately circularized via 5-to-3 monophosphate linkages. The bacterial diguanylate cyclases that synthesize c-di-GMP all contain a characteristic GGDEF domain, whereas the phosphodiesterases that specifically degrade c-di-GMP to pGpG contain an EAL domain. The GGDEF domain, and presumably c-di-GMP, appears to be limited to Bacteria (Galperin et al., 2001) and is not found in Archaea or Eukarya. c-di-GMP appears to play complex roles as a second messenger in most bacterial species, and regulates diverse processes, such as motility, biofilm formation, and virulence gene expression (Tamayo et al., 2007) . The molecular mechanisms by which c-di-GMP regulates these biological processes in bacteria are only beginning to be understood. It appears that c-di-GMP is capable of specific binding to regulatory proteins containing the PilZ protein domain (Cotter and Stibitz, 2007) . The PelD protein of P. aeruginosa is an additional non-PilZ-containing c-di-GMP sensor protein (Lee et al., 2007) . Several recent reports have suggested that c-di-GMP can stimulate a variety of signaling pathways in mammalian cells in vivo. One early report demonstrated that 50 µM of exogenous c-di-GMP inhibited growth of human colon cancer cells (Karaolis et al., 2005a) without toxicity against normal kidney cells. Two additional reports have indicated that c-di-GMP can function as an adjuvant. One group demonstrated a highly significant (>200-fold; P < 0.001) induction of anti-ClfA IgG2a titers when mice were vaccinated with ClfA and c-di-GMP as compared with vaccination with ClfA alone (Karaolis et al., 2007a) . Another group found a similar induction of anti--Gal titers when c-di-GMP was used as an adjuvant (Ebensen et al., 2007) . Both groups also found increased T cell responses in mice injected with c-di-GMP. Karoalis et al. (2007a) also showed that c-di-GMP has immunostimulatory effects in vivo and in vitro on innate cell populations, Toll-like receptor (TLR) 3, 7, 8, and 9 are transmembrane nucleic acid receptors that reside in intracellular compartments, where they detect endocytosed or autophagocytosed nucleic acids Medzhitov, 2007) . TLR7, 8, and 9 all require the signaling adaptor MyD88 to initiate downstream signaling, whereas TLR3 requires the signaling adaptor TRIF. Thus, MyD88 / Trif / double-knockout cells are deficient in signaling through all known TLRs that sense nucleic acids (Hoebe et al., 2003; Yamamoto et al., 2003) . Interestingly, recent work has established that MyD88 / Trif / cells are capable of responding to nucleic acid ligands via cytosolic immunosurveillance pathways. Cytosolic RNA appears to be sensed by two RNA helicase-containing proteins, RIG-I and Mda5 (Yoneyama et al., 2004) . Interestingly, RIG-I and Mda5 do not perform redundant functions and appear to respond to distinct classes of viruses (Gitlin et al., 2006; Hornung et al., 2006; Kato et al., 2006; Pichlmair et al., 2006; . Signaling by RIG-I and Mda5 requires a common adaptor molecule called MAVS (also known as IPS-1, Cardif, or VISA; Kawai et al., 2005; Meylan et al., 2005; Seth et al., 2005; Xu et al., 2005; Sun et al., 2006) . MAVS recruits the TBK1 kinase that phosphorylates and activates the IRF3 transcription factor. Other transcription factors such as ATF2/c-Jun and NK-B form a coordinated DNA-bound complex with IRF3, and are together required for robust activation of the IFN- promoter (Maniatis et al., 1998; Panne, 2008) . Many cell types also appear to respond to the cytosolic presence of DNA Stetson and Medzhitov, 2006a) . The cytosolic response to DNA does not require MAVS, but does require TBK1 and IRF3 in most cell types Stetson and Medzhitov, 2006a; Sun et al., 2006) . Recently, a putative cytosolic sensor of DNA was identified, and was named DNA-dependent activator of IFN regulatory factors (DAI; previously known as DLM-1 or ZBP1; Takaoka et al., 2007) . Knockdown experiments indicated that DAI was involved in sensing cytosolic DNA in the L929 fibroblast-like and RAW macrophage-like cell lines (Takaoka et al., 2007; Wang et al., 2008) . DAI was expressed in other cell types, such as mouse embryonic fibroblasts (MEFs), but was dispensable for the response to cytosolic DNA in these cell types, suggesting that additional uncharacterized nucleic acid sensor proteins also exist . Indeed, cells from DAI knockouts appear to respond normally to cytosolic DNA (Ishii et al., 2008) , possibly because of redundancy with other cytosolic DNA sensors. Although type I IFNs are primarily considered to be antiviral cytokines (Stetson and Medzhitov, 2006b) , there is growing appreciation for their complex role in bacterial infections as well. Indeed, it is clear that type I IFNs are induced by many, if not all, bacterial pathogens, and can contribute to diverse outcomes in vivo (Coers et al., 2000; O'Riordan et al., 2002; Opitz et al., 2006; Stetson and Medzhitov, 2006a; Henry et al., 2007; Roux et al., 2007; Stanley et al., 2007; Charrel-Dennis et al., 2008) . For several bacterial pathogens, (a) the c-di-GMP compound was chemically synthesized and >98% pure by HPLC (not depicted), (b) phosphodiesterase treatment of c-di-GMP resulted in a product that is no longer stimulatory ( Fig. 1 E) , and (c) macrophages deficient in all TLR signaling and unable to respond to LPS still responded to c-di-GMP (see below). Thus, these observations suggest that c-di-GMP is sensed in the cytosol, resulting in a potent induction of type I IFN in bone marrow macrophages. We performed microarray experiments to compare the global transcriptional response induced by cytosolic c-di-GMP with that induced by cytosolic DNA (Fig. 1 F and Table S1 ). Whole transcriptome spotted MEEBO microarrays (Verdugo and Medrano, 2006) were used in these experiments. The results indicated that the transcriptional profile of cells stimulated by cytosolic c-di-GMP was very similar to the transcriptional profile of cells stimulated with cytosolic DNA. Induction of type I IFNs and known type I IFN-inducible genes dominated the transcriptional response, but other genes (e.g., Il6, Cd86, and Cxcl9) were also strongly induced ( Fig. 1 F and Table S1 ). Although the transcriptional profiles of cells stimulated with c-di-GMP and DNA were highly similar, a small number of genes were reproducibly preferentially induced by c-di-GMP as compared with DNA (e.g., HtrA serine peptidase 4 and claudin 23; Table S1 ). Collectively, these results suggested that similar but not identical downstream signaling pathways are triggered by DNA and c-di-GMP. It was previously reported that HEK293 cells stably transfected with various TLRs failed to respond to c-di-GMP (Karaolis et al., 2007a) , but the reason for this failure was unclear. For example, HEK293 cells might lack an essential accessory protein. To rule out a role for TLRs in sensing of c-di-GMP, we tested responses in Myd88 / Trif / double-knockout macrophages, which are deficient in all TLR signaling (Yamamoto et al., 2003) . Both wild-type and Myd88 / Trif / macrophages showed a robust induction of IFN- after c-di-GMP stimulation (Fig. 2 A) . As expected, the Myd88 / Trif / macrophages did not respond to LPS (Fig. 2 A) . These results indicated that TLR signaling is not required for responsiveness to c-di-GMP, and are consistent with c-di-GMP signaling via a cytosolic surveillance pathway. Robust transcription of the IFN- gene requires the coordinate activation of several transcription factors, including IRF3 and NF-B (Panne, 2008) . We therefore expected that the c-di-GMP-induced signaling pathway would also require these factors to induce type I IFN. IRF3 activation requires phosphorylation by the TBK1 kinase, which leads to IRF3 dimerization and nuclear translocation (Fitzgerald et al., 2003; Sharma et al., 2003; Hemmi et al., 2004; McWhirter et al., 2004; Perry et al., 2004) . To address whether TBK1 is required for activation of type I IFN by c-di-GMP, we tested including monocytes, macrophages, and granulocytes. c-di-GMP induced dendritic cell maturation and led to the production of various cytokines and chemokines, including TNF, IL-1, IP-10, Rantes, and CXCR4 (Karaolis et al., 2007a) . However, c-di-GMP did not stimulate induction of type I IFNs by plasmacytoid dendritic cells. Impressively, preinjection of 2.5 mg/kg c-di-GMP protected against subsequent lethal challenge with Klebsiella pneumoniae or Staphlyococcus aureus (Karaolis et al., 2005b; Karaolis et al., 2007b) . Despite the impressive in vivo biological effects of c-di-GMP, the mechanism by which c-di-GMP stimulates host immunity remains unknown. In this study, we provide evidence that mammalian cells survey their cytosol for the presence of c-di-GMP. We find that c-di-GMP triggers a transcriptional response virtually indistinguishable from the response triggered by cytosolic DNA. Like the response to DNA, the response to c-di-GMP requires the TBK1 kinase and IRF3 transcription factor. However, the pathway for sensing c-di-GMP can be distinguished from that for sensing DNA in certain cell types and, moreover, appears to be distinct from all other known cytosolic sensing pathways. Our results suggest a novel mechanism by which host cells can induce an inflammatory response to a widely produced bacterial ligand. We hypothesized that c-di-GMP might be sensed by a cytosolic sensor leading to the induction of type I IFNs. To test this hypothesis, we compared the ability of overlayed versus transfected c-di-GMP to induce type I IFN in bone marrow macrophages. At the concentrations used (up to 25 µg/ml), overlay of c-di-GMP did not elicit a significant response ( Fig. 1 A) , although IFN- production could be detected at higher concentrations (50 or 100 µg/ml; not depicted). In contrast, delivery of c-di-GMP to the cytosol by transfection elicited robust activation of IFN- in a dose-dependent manner ( Fig. 1 A) . A significant response was detected with a concentration of c-di-GMP as low as 2.5 µg/ml (equivalent to 3.6 µM). These results suggested that c-di-GMP signals via a cytosolic immunosurveillance pathway. We tested whether compounds related to c-di-GMP were capable of stimulating IFN- expression when delivered to the cytosol. We measured induction of endogenous type I IFN protein by bioassay ( Fig. 1 B) , and induction of IFN- and IFN-5 mRNA levels by quantitative RT-PCR ( Fig. 1 , C and D). No detectable IFN- production was observed upon transfection with GTP, cGMP, ppGpp (a different bacterial nucleotide second messenger), or pGpG (hydrolyzed c-di-GMP), whereas the cytosolic response to 3.3 µg/ml of transfected c-di-GMP was of a similar magnitude to that of 3.3 µg/ml of transfected DNA (pdA:dT) and RNA (pI:C), which are known cytosolic inducers of type I IFN. We do not believe that the cytosolic induction of type I IFNs in response to c-di-GMP is caused by a contaminant (such as LPS) because a role in IFN--mediated antiviral immunity (Tenoever et al., 2007) , and found that Ikbke / macrophages responded normally to c-di-GMP (Fig. 2 C) . To examine the role of IRF3 in c-di-GMP induction of IFN-, we tested the ability of c-di-GMP to induce signaling in macrophages lacking Irf3. We observed that Irf3 / macrophages failed to induce IFN- in response to c-di-GMP ( Fig. 2 D) . We also tested Irf7 / macrophages and observed a partial requirement for IRF7 by the c-di-GMP pathway (Fig. 2 D) . Loss of Irf7 is likely compensated for by Irf3 in macrophages. The partial requirement for Irf7 may reflect a c-di-GMP responses in Tbk1 / macrophages (Fig. 2 B) . We observed very little IFN- induction by c-di-GMP in cells lacking Tbk1, suggesting that TBK1 is required for c-di-GMP signaling. We also observed that Tbk1 is required for the induction of IFN- by pdA:dT ( Fig. 2 B) , in agreement with a recently published report (Miyahira et al., 2009) . Although the data indicate a key role for Tbk1 in the response to c-di-GMP, it is formally possible, albeit unlikely, that Tbk1 protein rather than Tbk1 kinase activity is required for the response. We also tested the requirement for the TBK1related kinase Ikbke (also called IKK or IKK-i), which plays (Jiang et al., 2005) using recombinant IFN- as a standard. *, P < 0.01 as compared with transfected c-di-GMP (Student's t test). (B) Bone marrow macrophages were transfected with 3.3 µg/ml of the indicated molecules (or overlaid with 100 ng/ml LPS), and type I IFN was assessed after 6 h by bioassay as in A. *, P < 0.001 as compared with unstimulated cells (Student's t test). (C) Bone marrow macrophages were transfected with 3.3 µg/ml of the indicated molecules (or overlaid with 100 ng/ml LPS), and transcription of the IFN- gene (Ifnb) was assessed by quantitative RT-PCR, with normalization to ribosomal protein rps17 message. n.d., not detectably induced above background (lipofectamine alone). *, P < 0.001 as compared with unstimulated cells (Student's t test). (D) As in C, but transcription of the IFN-5 gene was assessed. n.d., not detectably induced above background (lipofectamine alone). *, P < 0.01 as compared with unstimulated cells (Student's t test). (E) c-di-GMP, treated or untreated with snake venom phosphodiesterase, was transfected at 3.3 µg/ml into B6 bone marrow macrophages and analyzed after 6 h for IFN- production by bioassay, as in A. *, P < 0.02 as compared with c-di-GMP treatment alone (Student's t test). Results in A-E are representative of at least three independent experiments. Data are means ± SD (n = 3). (F) Bone marrow macrophages were transfected with poly dA:dT (DNA) or with c-di-GMP. Total RNA was isolated after 6 h of stimulation. Probes were amplified and hybridized to whole-transcriptome spotted MEEBO microarrays. Each dot represents a single gene, and its position is determined by its induction in response to DNA (x axis) or c-di-GMP (y axis). Most genes lie on the diagonal, indicative of similar induction ratios by both treatments. Selected highly induced genes are labeled. Two independent microarray experiments gave similar results. transcription factors, as well as phosphorylation of the p38, JNK, and ERK MAP kinases, in response to c-di-GMP. Activation of IRF3 involves phosphorylation, dimerization, and translocation into the nucleus. We tested whether treatment with c-di-GMP resulted in significant nuclear accumulation of IRF3. As expected, we found that c-di-GMP resulted in significant translocation of IRF3 to the nucleus (Fig. 3 A) . We also tested whether transfected c-di-GMP is capable of activating NF-B. This was particularly important to establish in light of a previous report that exogenous (untransfected) c-di-GMP failed to activate NF-B (Karaolis et al., 2007a) . There have also been inconsistent reports in the literature as to whether cytosolic DNA induces NF-B Stetson and Medzhitov, 2006a; Sun et al., 2006) . We found that c-di-GMP treatment led to the activation of NF-B, as assessed by using an electromobility gel shift assay, with similar kinetics to that of transfected pI:C and pdA:dT ( Fig. 3 B) . The probe-bound complex was confirmed to contain NF-B p65 by supershifting with an anti-p65 antibody ( Fig. 3 B) . We confirmed these results by using a RAW role for Irf 7 signaling as part of a positive feedback loop involving signaling through the type I IFN receptor (Ifnar). Indeed, Ifnar / macrophages exhibited a diminished response to c-di-GMP (Fig. 2 E) . Irf5 and Irf1 did not seem to be required for responsiveness to c-di-GMP (Fig. 2 F) . In addition, macrophages deficient in Rip2, Nalp3, or doubly deficient in Nod1/2 responded normally to c-di-GMP (unpublished data). We validated our results by examining production of a second secreted molecule known to be a target of the Tbk1-Irf3 signaling axis, the chemokine Rantes (Ccl5). Rantes production was assessed by ELISA (Fig. S1 ). As expected, we found that Rantes was induced by c-di-GMP in a manner dependent on Tbk1, Irf3, and Ifnar, but largely or entirely independent of MyD88/Trif, Irf1, Irf5, Irf7, or Mavs (Fig. S1 ). Our genetic data suggested that the cytosolic responses to DNA, RNA, and c-di-GMP share a common downstream signaling pathway. If this were true, we would expect to be able to detect biochemical activation of the NF-B and IRF3 Myd88 / Trif / mice were transfected with 3.3 µg/ml c-di-GMP or pdA:dT (DNA), or overlayed with 100 ng/ml LPS as indicated. After 6 h of stimulation, induction of IFN- mRNA was analyzed by quantitative RT-PCR, with normalization to ribosomal protein rps17 message. n.d., not detectable. (B) As in A, except Tnfr1 / or Tnfr1 / Tbk1 / macrophages were analyzed. Tbk1 deficiency is embryonic lethal, but viability can be rescued on the TNF receptor 1 (Tnfr1)-deficient background. (C) As in A except Ikbke (IKK/IKK-i)-deficient mice were analyzed. (D) as in A, except Irf3 / and Irf7 / macrophages were analyzed. (E) as in A, except macrophages deficient in the type I IFN receptor (Ifnar / ) were analyzed. (F) as in A, except Irf5 / and Irf1 / macrophages were analyzed. Results are representative of at least three independent experiments. Data are means ± SD (n = 3). *, P < 0.05; and **, P < 0.001 as compared with wild-type bone marrow macrophages (Student's t test) . discrepancy between our results and those of Karaolis et al. (2007a) , it is more likely that we were able to detect MAP kinase activation because transfection of c-di-GMP triggers more robust responses than overlayed c-di-GMP, and because we examined later time points. Cell type-specific responses to cytosolic DNA and c-di-GMP Collectively, these results suggest that there are strong similarities in the signaling pathways triggered by the cytosolic presence of c-di-GMP and other nucleic acids, such as DNA and RNA. However, c-di-GMP is not structurally similar to these nucleic acid ligands. The DNA molecules capable of provoking cytosolic responses are double stranded and >25 deoxyribonucleotides in length Stetson and Medzhitov, 2006a) . In contrast, c-di-GMP is composed of two guanosine ribonucleotides (not deoxyribonucleotides) in an unusual cyclic conformation. c-di-GMP also lacks features consistent with recognition by RIG-I or Mda5, such as 5triphosphate, polyuridine, or double strandedness . Therefore, we considered whether distinct sensors might be responsible for triggering responses to DNA/RNA and c-di-GMP, and if so, whether the responses could be distinguished in certain cell types. macrophage cell line stably transfected with an NF-B luciferase reporter. We observed activation of the NF-B luciferase reporter by transfected c-di-GMP (Fig. 3 C) . Both the gel shift and reporter assays showed that the activation of NF-B by transfected c-di-GMP was not as strong as observed with transfected pdA:dT and pI:C. The subtle difference in NF-B induction between cytosolic DNA and c-di-GMP may suggest that there are distinct signaling components upstream of NF-B in the two pathways. Nonetheless, we concluded that transfected c-di-GMP is capable of activating NF-B. The relatively low levels and late time course of NF-B induction by c-di-GMP (as compared with induction by LPS), as well as the fact that we transfected c-di-GMP into cells, may explain why NF-B induction was not previously observed in response to c-di-GMP (Karaolis et al., 2007a) . We also tested whether MAP kinase pathways are activated by c-di-GMP. A previous report found that overlay of c-di-GMP induced ERK but not p38 in human macrophages (Karaolis et al., 2007a) . In contrast, we found that cytosolic c-di-GMP stimulated p38, ERK1/2, and JNK phosphorylation in mouse bone marrow macrophages transfected with c-di-GMP (Fig. 3 D) . Although a difference between human and mouse macrophages may explain the Data are means ± SD (n = 3). *, P < 0.05 compared with unstimulated cells (Student's t test) . tion, and is downstream of two cytosolic viral RNA sensors, RIG-I and Mda5 (Sun et al., 2006) . To test if c-di-GMP signals through the cytosolic RNA-sensing pathway, we tested c-di-GMP responses in Mavs / and Mda5 / macrophages. Loss of Mavs or Mda5 did not affect the induction of IFN- by c-di-GMP (Fig. 5, A and B; and Fig. S1 C) , suggesting that c-di-GMP does not signal through the cytosolic RNA signaling apparatus (Gitlin et al., 2006; Kato et al., 2008; . DAI (encoded by the Zpb1 gene) is a recently identified molecule that has been proposed to function as a sensor of cytosolic DNA (Takaoka et al., 2007; Wang et al., 2008) . We found that bone marrow-derived macrophages from Zbp1deficient mice (Ishii et al., 2008) still responded to c-di-GMP (Fig. 5 C) . However, as previously reported (Ishii et al., 2008) , Zbp1-deficient macrophages also responded normally to DNA (Fig. 5 C) . These results imply that macrophages express at least one cytosolic DNA sensor that is distinct from In addition to bone marrow macrophages, we observed induction of IFN- by cytosolic c-di-GMP in a variety of other cell types, including peritoneal macrophages, conventional dendritic cells, L929 cells, and RAW 264.7 macrophages (Fig. 4, A-D) . These cells also responded to transfected DNA and RNA. Interestingly, however, we were unable to detect significant induction of type I IFNs by c-di-GMP in MEFs or in 293T cells, even though the cytosolic pathways for detecting DNA and RNA are intact in these cells (Fig. 4 , E and F). These results suggest that at least one component of the host signaling pathway responding to c-di-GMP is distinct from that used for responses to cytosolic RNA or DNA, and is differentially expressed in different cell types. Known DNA and RNA sensing pathways are not required for responses to c-di-GMP MAVS (also known as IPS-1) is an essential adapter protein in the cytosolic RNA pathway leading to type I IFN induc- Data are means ± SD (n = 3). n.d., not detectable. *, P < 0.01; and **, P < 0.001 as compared with unstimulated cells (Student's t test) . and measured their responsiveness to YAC-1 target cells ex vivo. Because NK cells respond to in vivo injection of poly I:C, a prototypical IFN inducer, we expected that c-di-GMP might also activate NK cells. Indeed, NK cells obtained from B6 mice injected with c-di-GMP responded to YAC-1 cells, as measured by intracellular cytokine staining for IFN- (Fig. 6 B) . However, NK cells obtained from Irf3/7 / mice stimulated with c-di-GMP were nonresponsive to YAC-1 target cells. As a negative control, NK cells stimulated ex vivo with PBS did not produce IFN-. These observations indicate that c-di-GMP can stimulate cytokine and NK cell responses in vivo, and these responses require the IRF3/7 transcription factors. To obtain a preliminary indication as to whether adaptive immune responses could be stimulated by c-di-GMP, we immunized mice with human serum albumin (HSA) and tested whether coinjected c-di-GMP could function as an adjuvant, as previously reported (Ebensen et al., 2007; DAI, and at least one of these uncharacterized DNA sensors, or another independent sensor, may respond to c-di-GMP. To validate our findings in vivo, we injected mice intraperitoneally with c-di-GMP (200 nmol per mouse) and measured the production of type I IFNs in the blood 18 h later. B6 mice injected with c-di-GMP produced an IFN response to c-di-GMP, and this response was abolished in Irf3/7 double-deficient mice (Fig. 6 A) . Interestingly, single-deficient Irf3 / mice responded to c-di-GMP in vivo (unpublished data), implying a role for IRF7 in responses to c-di-GMP in vivo. To determine whether cytokine induction by c-di-GMP affected cellular responses in vivo, we collected splenic NK cells from mice injected with c-di-GMP Figure 5 . c-di-GMP does not stimulate known cytosolic pathways for sensing nucleic acids. (A) Mavs / and littermate wild-type macrophages were transfected with 3.3 µg/ml c-di-GMP, poly dA:dT (pdA:dT, DNA), or poly I:C (pI:C, RNA), and transcription of the IFN- gene (Ifnb) was assessed by quantitative RT-PCR, with normalization to ribosomal protein rps17 message. n.d., not detectable. (B) Mda5 / and wild-type control macrophages were stimulated and analyzed as in A. (C) Zbp1 / (DAI knockout) and wild-type control macrophages were stimulated and analyzed as in A. SeV, Sendai virus. Results are representative of at least three independent experiments. Data are means ± SD (n = 3). *, P < 0.01; and **, P < 0.001 as compared with wild-type bone marrow macrophages (Student's t test) . (A and B) B6 and Irf3/7 / mice were injected intraperitoneally with 200 nmol c-di-GMP, and (A) serum IFN was measured by bioassay 18h later, or (B) 36 h after injection, splenic NK responsiveness to YAC-1 target cells (or PBS control) was measured ex vivo by intracellular staining for IFN-. (C) B6 and Irf3/7 / mice were injected intraperitoneally with HSA ± 200 nmol c-di-GMP, and 2 wk later serum IgG1 specific for HSA was determined by ELISA. The experiments in A and B were repeated twice, and the experiment in C was performed once. Horizontal bars represent means. **, P < 0.01; and *, P < 0.05 (Student's t test) . respond well to c-di-GMP. Given the marked chemical dissimilarities among DNA, RNA, and c-di-GMP, it is perhaps not surprising that these nucleic acids appear to be recognized by different sensors. Indeed, we favor the idea that there are multiple cytosolic sensors for nucleic acids leading to induction of type I IFNs. The specificities of these various sensors will require further dissection. For example, although DAI knockout cells still responded to c-di-GMP, it remains formally possible that DAI is just one of several redundant sensors for c-di-GMP. These possibilities can be addressed once additional cytosolic nucleic acid sensors are identified. Our results suggest that cytosolic sensing of c-di-GMP is relatively specific. Other related small nucleic acid compounds such as GTP, cGMP, ppGpp, and pGpG did not trigger transcriptional induction of type I IFNs (Fig. 1) . In addition, c-di-GMP that was hydrolyzed by snake venom phosphodiesterase did not induce type I IFN (Fig. 1) . However, our results cannot eliminate the possibility that a metabolite of c-di-GMP, rather than c-di-GMP itself, is the true molecular moiety that is sensed in the cytosol. Elimination of this possibility awaits identification of the c-di-GMP sensor and biophysical or crystallographic characterization of its binding to c-di-GMP. Our results also cannot eliminate the possibility that c-di-GMP triggers type I IFN expression indirectly, for example, by stimulating the synthesis of a host ligand that functions as the "true" proximal trigger of type I IFN gene expression. It is also possible that c-di-GMP acts "pharmacologically" by disrupting host physiology in a way that results in type I IFN expression. Moreover, because the signaling triggered by c-di-GMP and cytosolic DNA are very similar, it is possible that the c-di-GMP pathway is a branch off of the DNA-sensing pathway rather than a fully independent pathway. However, even if these alternative models are correct, our data indicate that the pathway downstream of c-di-GMP contains at least some novel components and is at least partially independent of known cytosolic or TLR signaling pathways. Several bacterial pathogens, including L. monocytogenes, L. pneumophila, F. tularensis, M. tuberculosis, and group B Streptococcus, have been reported to induce type I IFNs (Coers et al., 2000; O'Riordan et al., 2002; Opitz et al., 2006; Stetson and Medzhitov, 2006a; Henry et al., 2007; Roux et al., 2007; Stanley et al., 2007; Charrel-Dennis et al., 2008) . The mechanism by which these pathogens induce type I IFN resembles that of c-di-GMP: in all cases, type I IFN induction is independent of MyD88/Trif and TLRs, requires TBK1 and IRF3, and (with one exception; unpublished data; Opitz et al., 2006) is independent of the cytosolic RNA-sensing pathway. It is widely assumed that bacterial DNA, perhaps released after bacterial cell lysis, is the relevant ligand that triggers type I IFNs in response to pathogens. There are no data that eliminate or confirm this possibility, as sensors required for IFN induction in response to cytosolic bacteria have not yet been reported. However, it is noteworthy that in all cases, induction of type I IFNs by cytosolic bacteria requires expression of an auxiliary secretion system, namely the ESX-1 system of M. tuberculosis, the Francisella pathogenicity island-encoded 2007a). Serum was collected from immunized mice 2 wk after a single intraperitoneal injection. Mice immunized with HSA alone did not mount a significant antibody response to HSA. In contrast, mice injected with HSA plus c-di-GMP produced variable but significant titers of HSA-specific IgG1 antibodies, confirming that c-di-GMP can function as an adjuvant in vivo. Importantly, our preliminary data indicated that the adjuvant effect of c-di-GMP depended on IRF3/7, as antibody responses in Irf3/7 / mice immunized with HSA and c-di-GMP were significantly (P < 0.05) reduced as compared with immunized B6 mice (Fig. 6 C) . Further studies will be required to establish the mechanism by which c-di-GMP functions as an adjuvant, but our results are consistent with a recent report indicating that the IRF3/7 kinase, TBK1, is required for adaptive immune responses in a DNA-vaccine model (Ishii et al., 2008) . It has been well established that innate immune responses are initiated in response to certain microbial ligands that are evolutionarily conserved and that can be distinguished from selfligands. Nucleic acids appear to be a favored target of immune recognition, and are sensed by a variety of endosomal and cytosolic sensor proteins. The cyclic dinucleotide c-di-GMP is a bacterial second messenger that exhibits several characteristics that are desirable in an immunostimulatory ligand: it is produced by numerous species of bacteria, it is a critical regulator of bacterial physiology, and it is not similar to host molecules. Indeed, several previous reports have demonstrated that c-di-GMP can provoke potent immune responses when injected in vivo into mice (Karaolis et al., 2005a; Karaolis et al., 2005b; Karaolis et al., 2007a; Karaolis et al., 2007b) . However, the mechanism by which c-di-GMP stimulates immune responses remained unclear. Our data provide evidence that c-di-GMP is sensed by a novel cytosolic immunosurveillance pathway. Responsiveness to c-di-GMP is independent of TLRs that monitor the extracellular/endosomal compartments and is strongly potentiated by transfection of c-di-GMP into the host cell cytosol. Moreover, c-di-GMP provokes a transcriptional response highly reminiscent of that triggered by the cytosolic presence of DNA or RNA, and involves activation of TBK-1, MAP kinases, and the IRF-3 and NF-B transcription factors. Thus, we propose that c-di-GMP is sensed in the cytosol. How many different cytosolic sensors exist for nucleic acids? Our data suggest that sensing of c-di-GMP occurs via a novel cytosolic immunosurveillance pathway. Known nucleic acid sensors include Mda5 and RIG-I, which sense viral RNA and signal via MAVS, as well as DAI, a putative sensor of DNA that signals independently of MAVS. Based on the finding that most cell types, including MEFs, can still respond to DNA in the absence of DAI (Ishii et al., 2008) , it has been proposed that an additional DNA sensor must exist and that MEFs express both of these sensors . Our results now suggest that there is at least one additional nucleic acid sensor in the cytosol, because we find that MEFs do not B. Beutler (The Scripps Research Institute, La Jolla, CA), and RAW-B cells were obtained from G. Barton. Animal protocols were approved by the University of California, Berkeley Animal Care and Use Committee. Reagents. c-di-GMP was synthesized as previously described (Kawai et al., 2003) , poly I:C was obtained from GE Biosciences, pdA:dT (poly(dA-dT): poly(dA-dT)) was purchased from Sigma-Aldrich, ppGpp (guanosine-3,5bisdiphosphate) was obtained from TriLink Biotechnologies, pGpG (diguanosine) was purchased from IBA GmbH, GTP and cGMP were obtained from Sigma-Aldrich, purified LPS was purchased from InvivoGen, and Sendai virus was obtained from Charles River Laboratories. Theiler's virus was the gift of M. Brahic (Stanford University, Stanford, CA). Cell culture. L929, RAW 264.7, and MEF cell lines were cultured in DMEM containing 10% FBS, glutamine, and penicillin-streptomycin. For bone marrow-derived macrophages, bone marrow cells from femurs and tibias were cultured for 7 d in RPMI 1640 media containing 10% FBS, glutamine, penicillin-streptomycin, and10% CSF from 3T3 cells, with feeding on the fourth day of growth. For conventional dendritic cells (GM-CSFdendritic cells), bone marrow cells were cultured for 5 d with RPMI 1640 containing 10% FBS, glutamine, penicillin-streptomycin, -mercaptoethanol, and GM-CSF, with fresh media added on the second and fourth days of growth. Peritoneal macrophages were elicited by injection of 2 ml 4% thioglycollate (Fluid Thioglycollate Medium; BD), and were obtained 4 d later by lavage of the peritoneal cavity with RPMI 1640. Cell stimulations (transfections). Cells were transfected using Lipofectamine 2000 (LF2000; Invitrogen) according to the manufacturer's protocol. All nucleic acid stimulants were mixed with LF2000 at a ratio of 1 µl LF2000/1 µg nucleic acid, incubated at room temperature for 20-30 min, and added to cells at a final concentration of 3.3 µg/ml (96-well plates) or 4 µg/ml (6-well plates). For pI:C, 2 mg/ml of the stock solution was heated at 50°C for 10 min and cooled to room temperature before mixing with LF2000. Transfection experiments were done for 6 h, unless otherwise stated in the figures. For phosphodiesterase treatment of c-di-GMP, 0.5 µg/µl c-di-GMP was incubated for 2 h at room temperature in Optimem buffer (Invitrogen), with 15 mM MgCl 2 and 1 U Phosphodiesterase I (GE Healthcare). Quantitative PCR. Stimulated cells were overlayed with RNAlater (Applied Biosystems) and stored. RNA was isolated using the RNeasy kit (QIAGEN) according to the manufacturer's protocol, and was treated with RQ1 RNase-free DNase (Promega). 0.5 µg RNA was reverse transcribed with Superscript III (Invitrogen). Platinum Taq DNA polymerase (Invitrogen) and EvaGreen dye (Biotium) were used for quantitative PCR assays and analyzed with a real-time PCR system (StepOnePlus; Applied Biosystems). All gene expression values were normalized to Rps17 (mouse) or SP9 (human) levels for each sample. The following primer sequences were used: mouse Ifnb, (forward) 5-ATAAGCAGCTCCAGCTCCAA-3 and (reverse) 5-CTGTCTGCTGGTGGAGTTCA-3; mouse Ifn5, (forward) 5-TGACCTCAAAGCCTGTGTGATG-3 and (reverse) 5-AAG-TATTTCCTCACAGCCAGCAG-3; mouse Rps17, (forward) 5-CGCC-ATTATCCCCAGCAAG-3 and (reverse) 5-TGTCGGGATCCACC-T CAATG-3; human IFN, (forward) 5-AAACTCATGAGCAGTCT-GCA-3 and (reverse) 5-AGGAGATCTTCAGTTTCGGAGG-3; and human SP9, (forward) 5-ATCCGCCAGCGCCATA-3 and (reverse) 5-TCAATGTGCTTCTGGGAATCC-3. Type I IFN bioassay and luciferase reporter assay. Cell-culture supernatants from stimulated cells were overlayed on top of ISRE-L929 IFN reporter cells (Jiang et al., 2005) and incubated for 4-6 h (96-well plate). The reporter cells were lysed in Passive Lysis Buffer (Promega) for 5 min at room temperature, mixed with firefly luciferin substrate (Biosynth), and measured on a luminometer (LmaxII 384 ; MDS Analytical Technologies). Levels of type I IFN were calculated from a standard curve using recombinant mouse IFN- (R&D Systems). secretion system of F. tularensis, the Dot/Icm system of L. pneumophila, or the MdrM multidrug efflux pump of L. monocytogenes. It is tempting to speculate that a small, bacterially derived molecule such as c-di-GMP could be transported or leak through these secretion systems. It has not yet been possible to test this idea directly because all of these bacterial species encode numerous c-di-GMP synthases, and a strain lacking all c-di-GMP synthesis has not been reported. In any case, interpretation of these experiments would be complicated by the fact that c-di-GMP plays important regulatory roles in bacterial physiology and pathogenesis. Nevertheless, our results with synthetic purified c-di-GMP suggest that in addition to DNA, c-di-GMP is a candidate for a conserved molecule unique to bacteria that is responsible for triggering transcription of type I IFN genes. In light of a recent report that bacteria appear to be able to synthesize c-di-AMP (Witte et al., 2008) , it is interesting to consider whether additional nucleic acids produced specifically by bacteria might also trigger host immunosurveillance pathways. Non-nucleic acid small molecules, such as the drug DMXAA, also appear to be able to stimulate the TBK1-IRF3 axis (Roberts et al., 2007) . Indeed, there may be multiple redundant pathways for cytosolic sensing of bacteria. Collectively, the available evidence suggests that host cells may sense a wider array of bacterial ligands than was previously appreciated. Our results may have important implications for the design of new adjuvants and vaccines. DNA vaccines have attracted considerable enthusiasm as an approach for protecting against a variety of infectious diseases (Wang et al., 2001; Yang et al., 2004) , but there are safety concerns about the insertional mutagenic potential of DNA vaccines and/or their potential to trigger pathogenic anti-DNA autoimmune antibody responses (Schalk et al., 2006) . The potency of DNA vaccines appears to derive from their ability to stimulate the TBK1 and innate cytosolic DNA-sensing pathways . Thus, our demonstration that a synthetic nonself non-DNA molecule such as c-di-GMP can stimulate an in vitro and in vivo innate and adaptive immune response (Fig. 6 ) similar to that induced by DNA, without similar autoimmune or mutagenic risks, suggests that c-di-GMP might have valuable application as a small-molecule adjuvant. Understanding the molecular basis of c-di-GMP signaling in mammalian cells will be a crucial step toward achieving this aim. Mice and cell lines. Bone marrow macrophages were derived from various mouse strains, including Mavs / (Sun et al., 2006) , Mda5 / (Gitlin et al., 2006) , and Zbp1 / (Ishii et al., 2008) . Wild-type C57BL/6 and Tnfr1 / mice were from the Jackson Laboratory. Myd88 / /Trif / mice were obtained from G. Barton (University of California, Berkeley, Berkeley, CA). Irf1 / , Irf3 / , Irf5 / , and Irf 7 / mice were obtained from T. Taniguchi (University of Tokyo, Tokyo, Japan). Tbk1 / /Tnfr1 / mice were generated by crossing Tbk1 / mice (provided by W.-C. Yeh, University of Toronto, Toronto, Canada) with Tnfr1 / mice. Rip2 / mice were obtained from R. Medzhitov (Yale University, New Haven, CT). Nod1 / /Nod2 / mice were obtained from D. Portnoy (University of California, Berkeley, Berkeley, CA). Nalp3 / mice were obtained from V. Dixit (Genentech, South San Francisco, CA). ISRE-L929 IFN reporter cells were obtained from

Gene Expression Profiling in Cells with Enhanced γ-Secretase Activity

BACKGROUND: Processing by γ-secretase of many type-I membrane protein substrates triggers signaling cascades by releasing intracellular domains (ICDs) that, following nuclear translocation, modulate the transcription of different genes regulating a diverse array of cellular and biological processes. Because the list of γ-secretase substrates is growing quickly and this enzyme is a cancer and Alzheimer's disease therapeutic target, the mapping of γ-secretase activity susceptible gene transcription is important for sharpening our view of specific affected genes, molecular functions and biological pathways. METHODOLOGY/PRINCIPAL FINDINGS: To identify genes and molecular functions transcriptionally affected by γ-secretase activity, the cellular transcriptomes of Chinese hamster ovary (CHO) cells with enhanced and inhibited γ-secretase activity were analyzed and compared by cDNA microarray. The functional clustering by FatiGO of the 1,981 identified genes revealed over- and under-represented groups with multiple activities and functions. Single genes with the most pronounced transcriptional susceptibility to γ-secretase activity were evaluated by real-time PCR. Among the 21 validated genes, the strikingly decreased transcription of PTPRG and AMN1 and increased transcription of UPP1 potentially support data on cell cycle disturbances relevant to cancer, stem cell and neurodegenerative diseases' research. The mapping of interactions of proteins encoded by the validated genes exclusively relied on evidence-based data and revealed broad effects on Wnt pathway members, including WNT3A and DVL3. Intriguingly, the transcription of TERA, a gene of unknown function, is affected by γ-secretase activity and was significantly altered in the analyzed human Alzheimer's disease brain cortices. CONCLUSIONS/SIGNIFICANCE: Investigating the effects of γ-secretase activity on gene transcription has revealed several affected clusters of molecular functions and, more specifically, 21 genes that hold significant potential for a better understanding of the biology of γ-secretase and its roles in cancer and Alzheimer's disease pathology.

Introduction c-Secretase is an unconventional aspartyl protease (composed of PS1, NCT, Aph-1 and Pen2) with an intramembranous catalytic site that is typical of the class of intramembrane-cleaving proteases (I-CliPs) (for review, see [1, 2] ). Via the processing of its substrates and freeing of their intracellular domains (ICDs), c-secretase regulates a multitude of signaling pathways and biological processes by influencing gene transcription. This is exemplified by the processing of the Notch receptor and the Notch signaling pathway (for a review, see [3] ). After specific ectodomain shedding via tumor necrosis factor a converting enzyme (TACE) (Fig. 1, step 1) , Notch is further cleaved intramembraneously by c-secretase (Fig. 1, step 2) . The intracellular domain of Notch (NICD) is freed to enter the nucleus, where it interacts with the transcription factor CSL (Fig. 1, step 3 ). With help from the coactivator Mastermind, CSL is converted from a transcriptional repressor to a transcriptional activator. CSL as an activator leads to the expression of Notch target genes (Fig. 1, step 4) , like the Hes or Hey family. Hes1, a transcriptional repressor, inhibits the transcription of NC3C1 (Fig. 1, step 5 ). Enhanced c-secretase activity, through its cleavage of Notch, leads to increased transcription of specific genes (Fig. 1, step 4) that repress the expression of other genes (Fig. 1, step 6) to influence a multitude of biological processes. For example, the processing of Notch by c-secretase is crucial for hepatoblast differentiation [4] , epidermis and hair follicle differentiation [5] , alveolar differentiation in mammary glands [6] , maintenance of skin appendages [7] , intestinal stem cell specification [8] , induction of satellite cells after injury and maintenance [9] and neural specification of embryonic stem cells [10] . The directions in which c-secretase activity can up-and downregulate gene transcription following its cleavage of a variety of substrates is further exemplified by the processing of Amyloid-b (Ab) precursor protein (APP), one of the better-known c-secretase substrates. The successive processing of APP by BACE1 and csecretase indeed leads to the production of Ab peptides (a causative agent in the pathogenesis of Alzheimer's disease (AD)), and APP-intracellular domains (AICDs) which, following associ-ation with the adaptor protein Fe65 and nuclear translocation, are able to suppress the expression of the major Apolipoprotein e (ApoE)/lipoprotein receptor LRP1 by binding directly to its promoter [11] . Thus, APP processing is also involved in the regulation of brain ApoE and cholesterol metabolism through LRP1 [11] . As ApoE4 is the major known genetic risk factor for late onset Alzheimer's disease (LOAD) and since AICD production depends on c-secretase, the latter is implicated in the sporadic form as well. In contrast to LOAD, which correlates directly with age, early onset familial Alzheimer's disease (FAD) is genetic and is mainly caused by mutations in presenilin1 or presenilin2 (PSEN1 or PSEN2), leading to loss of physiological or gain of toxic functions. Murine specific loss of Psen1 in the forebrain has been shown to affect certain aspects of memory [12, 13] . However, it remains difficult to correlate the loss of four murine PSEN alleles with the mild single PSEN allele mutations in FAD [14, 15] . c-Secretase is thus directly or indirectly implicated in the pathogenesis of both FAD and LOAD, making this protease an attractive therapeutic target for the prevention and/or treatment of AD. c-Secretase inhibitors/modulators have indeed reached clinical phase III trials [16] . With an increasing number of reports about new c-secretase substrates and the transcriptional effects of their ICDs being potentially implicated in the pathogenesis of AD or several types of cancer, we see a need for a basic overview of genes and molecular functions that are transcriptionally affected by c-secretase activity. In an effort to identify specific alterations of gene transcription as a result of c-secretase activity, the transcriptomes of two CHO cell lines (biological triplicates were used in each case) with To identify genes whose transcription is affected by c-secretase activity, two starkly contrasting conditions were analyzed by cDNA microarray: genetically engineered enhanced c-secretase (left panel) and pharmacologically inhibited c-secretase (right panel) in CHO cell lines. For a schematic depiction of the strategy, the Notch-1 receptor signaling pathway is used as an example. After processing by the Furin protease and when activated by binding to its ligands Notch-1 is cleaved at the S2 position by the TACE protease, generating a substrate for c-secretase (1, 7) . Under enhanced (left panel) or inhibited (right panel) c-secretase activity, the cleavage of the substrate controls the release of the Notch intracellular domain (NICD) (2, 8) . With enhanced c-secretase, increased numbers of NICDs enter the nucleus and interact with CSL (3), leading to the transcription of target genes like Hes1 and Hey (4). The Hes1 transcription repressor inhibits transcription of target genes like NC3C1 (5), with the final consequence being reduced production of NC3C1 mRNA (6) . Thus, enhancing c-secretase leads simultaneously to gene-dependent increase (in the case of Hes/Hey) or decrease (in the case of NC3C1) of mRNA copy numbers. With inhibited c-secretase, reduced numbers of NICDs (9) lead to the transcription of less Hes1/Hey (10), to reduced inhibition of target genes like NC3C1 (11) and consequently to increased production of NC3C1 mRNA (12) . Inhibiting c-secretase thus leads to gene-dependent decrease (in the case of Hes/Hey genes) or increase (in the case of NC3C1) of mRNA copy numbers. Following mouse cDNA microarray analysis of both transcriptomes, top scoring candidates were evaluated and validated by real time PCR and further analyzed for changes of transcript levels between healthy and AD human brain cortices. doi:10.1371/journal.pone.0006952.g001 enhanced and inhibited c-secretase activity were analyzed and compared (strategy depicted in Fig. 1 -exemplified by Notch processing). The S-1 cell line overexpresses the four components of c-secretase (NCT, Aph1a, PS1 and Pen2) and was characterized by a marked increase in the level of PS1 heterodimers and an associated 8-fold increase in c-secretase activity compared to untransfected controls [17] . The other cell line consisted of the original parental wild type CHO cells incubated with DAPT, a well-known c-secretase inhibitor. We strategically chose those two conditions, overexpression of c-secretase and inhibition of its activity, to amplify the activity-dependent effects on gene transcription levels (i.e., amplification of the signal from the cDNA microarray). To reduce potential effects due to changes in the protein levels of the c-secretase subunits as opposed to changes in its activity that we are interested in, we used chemical inhibition (DAPT) of c-secretase activity instead of gene silencing, which ultimately leads to changes in protein levels [18] . Biological functions have indeed been reported mainly for the c-secretase subunit PS1, independently to the c-secretase activity. However, because treatment of CHO cells with DAPT has been recently reported to exacerbate the secretion of exosomes [19] , we cannot exclude at this stage that some detected genes may be exosomerelated in response to the DAPT treatment. The gene transcription levels of the two cell lines were analyzed using a mouse cDNA microarray ( Fig. 1 and Material and Methods) [20] because of the absence of a readily available DNA microarray based upon hamster gene sequences or cDNA clones. This lack has been noticed, and cross-species reactivity of a mouse microarray hybridized with CHO-derived samples has been investigated recently by De Leon Gatti et al. [21] . This group generated an EST-based CHO microarray and compared it with results from a mouse microarray and vice versa. They state that cross-species hybridization yielded 89.6% overlap in their arrays, noncontradicting results and led only to a decrease in sensitivity resulting in detection of fewer differentially expressed genes. Accordingly, we probably have not detected all differentially expressed genes, but we have detected a significant amount, including clusters of functional relevance. For example, Neprilysin, an Ab-degrading enzyme of functional relevance to AD that has been previously shown to be transcriptionally downregulated in PSEN1/PSEN2 double knock out fibroblasts and to exhibit reduced activity under chemical (DAPT) inhibition of c-secretase in mouse neurons [22] , was not detected in the current study. Collectively, this supports the use of CHO cells with a mouse microarray. The microarray data set discussed in this publication has been deposited in the NCBI Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) and is accessible through GEO Series accession number GSE16379. The mouse microarray consistently detected the four human csecretase subunits overexpressed in the S-1 cell line (Table 1) . By applying a cut-off based on the false discovery rate (FDR, i.e., the probability to wrongly accept a difference between the two conditions) with a p value of 0.005, we found 2658 EST clones (1981 genes) to be differentially expressed, with 1241 EST clones of increased and 1417 EST clones of decreased transcription upon enhanced c-secretase activity (Supplemental Material, Dataset S1 and Dataset S2). Mapping clusters of genes of GO functions transcriptionally susceptible to c-secretase activity levels resulted in a GO hierarchydependent tree that will provide further orientation for c-secretase research. Functional clustering of 2658 differentially expressed sequences (1981 genes, Supplemental Material, Dataset S3) was performed using the FatiGo tool [23] . Comparing the representation of functional groups of genes throughout the entire mouse genome with their representation within the group of differentially transcribed genes allowed us to see whether clusters of genes of a specific functional group were enriched in the differentially expressed set. Clusters of over-and underrepresented genes were detected (Fig. 2) . The gene functions ''transcription regulator activity'', ''kinase regulator activity'', ''catalytic activity'' and ''binding '' were found to be overrepresented among the 2658 sequences (1981 genes) that were differentially transcribed. The cluster of ''molecular transducer activity'', through its subclusters in the GO hierarchy: ''receptor activity'' GO0004872, ''transmembrane receptor activity'' GO 0004888 and ''neurotransmitter receptor activity'' GO0030594, as well as the cluster of ''transporter activity'', via its subcluster of ''ion transporter activity'' GO0015075, were underrepresented (Fig. 2 , blue boxes). This is significant since neurotransmitter activity and transmembrane receptors are well within the focus of current AD research [24] . Supporting our hypothesis that c-secretase has a role in multiple transcriptional regulatory activities, the GO cluster of ''transcription regulator activity'' is overrepresented through both its subclusters ''transcriptional activator activity'' GO0016563 and ''transcriptional repressor activity'' GO 0016564 (Fig. 2 , red boxes. Single member genes of each cluster are annotated in Supplemental Material Dataset S3). A well-described gene within the activator cluster is b-catenin (CTNNB1, FC = 3, p = 0.001), whereas an example of a gene in the cluster of ''transcriptional repressor activity'' is HES1. Hes1 (FC = 5.4, p = 7.69E-04) is a transcription factor that has previously been reported as a downstream target of the Notch signaling pathway [25] (Fig. 1) . Like the examples above, 56 other transcription-related genes were found to be differentially transcribed with enhanced csecretase activity (Supplemental Material, Dataset S4). Consistent with these findings, several known substrates of the enzyme were detected on the microarray as well ( Table 2 ). This suggests a possible feedback mechanism by which the augmented processing of these substrates by c-secretase might lead to their altered transcription. The overrepresentation of genes in the clusters of enzymatic activity, such as ''kinase regulator activity'' GO0019207 and ''catalytic activity'', through four distinct GO subclusters (''isomerase activity'' GO 0016853, ''ligase activity'' GO0016874, ''hydrolase activity'' GO0016787 and ''transferase activity'' GO 00167740- Fig. 2 , red boxes), is broad in terms of the type of enzymatic activity and further shows the diversification of the downstream effects of enhanced c-secretase activity. The most complex cluster of molecular function that is overrepresented among the differentially transcribed genes identified in our microarray analysis is the GO function termed ''Binding''. This cluster is overrepresented through six subclusters and several subclusters of these ( Fig. 2 , lower part). Consistent with transcription regulation, the binding subclusters of ''nucleic acid binding'' GO0003676 and ''nucleotide binding'' GO0000166 are overrepresented. The cluster of ''ion binding'' GO0043167 is overrepresented as well as the cluster of ''protein binding'' GO0005515. A consistently overrepresented subcluster of the latter is ''cytoskeletal protein binding'' GO0008092 ( Fig. 2 ). Cytoskeletal proteins have long been known to play a role in AD and Tauopathies. They are targets of the cell polarity Wnt pathway, and their dynamics have recently been shown to be affected by AICD [26] . ''Receptor binding'' GO0005102 also includes the Notch ligand and known c-secretase substrate Jagged 2 [27, 28] , as well as the asecretase ADAM 10 [29] , four members of the Wnt family (Wnt6, 7a, 9b and 10a) and, the aforementioned b-catenin. Indeed, the translocation of b-catenin is mediated by ADAM 10, which is of the same functional cluster [30] . By clustering transcriptionally affected genes, we demonstrate that neurotransmitter, transcription regulator and enzymatic activities, transmembrane receptor and cytoskeletal proteins functional groups are affected by c-secretase activity in their mRNA copy numbers. For specific analysis of single genes, the fifty most prominently transcriptionally altered genes were evaluated by real time PCR. Mouse code based primers worked reproducibly and specifically for 35 genes. Among them, 21 genes were found to be differentially transcribed with enhanced c-secretase activity ( Fig. 3 upper panel, annotations lower panel). The highest increase in transcription Categories within the Molecular Function GO hierarchy that were over-and under-represented among the genes that were differentially transcribed in cells with enhanced c-secretase activity. Red boxes display GO terms that were overrepresented; blue boxes indicate GO terms that were under-represented. Black boxes represent main molecular functional clusters and arrows point toward according subclusters. The clustering of 1981 differentially transcribed genes was performed with DAVID and the FatiGo tool [23] . doi:10.1371/journal.pone.0006952.g002 levels was detected for UPP1, a gene encoding an enzyme (Uridine phosphorylase, UPase) directly implicated in the processing of uridine. UPP1 was confirmed by real time PCR to have a 39.2-fold increase in transcription levels ( Fig. 3 upper panel). Uridine is a strong sleep-promoter and is crucial for RNA, DNA and membrane biosynthesis [31] . Because of the latter, a lack of uridine (caused by increased UPase) would thus first damage cells with a large membrane to cytoplasm ratio, one of the most extreme ratios being found in neurons due to their axon and dendrite structure [32] . Interestingly oral administration of uridine has improved AD phenotypes [33, 34] . The protein Upp1 also interacts with Vimentin [35] , the distribution of which is characteristically altered in FAD fibroblasts [36] . The Notch-dependent transcriptional repressor Hes1 was also confirmed by real time PCR with a 7-fold increase in mRNA levels under enhanced c-secretase activity (Fig. 3) . Importantly, we found several key players of the three Wnt pathways to be transcriptionally altered in response to enhanced csecretase. We confirmed one of these, Wnt3a, to be increased by 2.8-fold in S-1 cells (Fig. 3 ). Aoyama et al. have reported that Wnt3a can influence Notch protein levels and increase Notch1 activation [37] , which increases the effect of enhanced c-secretase even further through substrate enhancement. Thus, enhanced csecretase activity may lead to increased WNT3A transcription, which in turn can increase the protein levels of the NICD-carrying c-secretase substrate. This proposed enhancement of the canonical Wnt pathway is further supported by the recent observation that b-catenin (the central protein that also ties PS1 to the pathway) modulates the level and transcriptional activity of Notch1/NICD through their direct interaction [38] . Several proteins, including Frizzled and Disheveled (Dvl), relay the Wnt signal along the canonical pathway between Wnt and b-catenin. We found that they both show increased gene transcription in our microarray analysis. DVL3 was confirmed by qPCR to increase in mRNA copy numbers by 3-fold (Fig. 3) . Taken together with the microarray data showing differential transcription of PROC, DKK, LRP5/6, GBP, AXIN, b-catenin, C-JUN and CYC D (all part of the canonical Wnt pathway-for further details see 'Discussion'), our data suggest a strong transcriptional effect on this pathway by c-secretase activity and resulting alterations in gene expression. CYC D has also been reported to be downregulated by Protein tyrosine phosphatase receptor type G (PTPRG) [39] . We confirmed by real time PCR that PTPRG transcript level is reduced by 515-fold (Fig. 3) . AMN1 (levels down 978-fold) has also been connected with cell cycle regulation in yeast, but its role in mammals is not well known. TERA, a gene of unknown function (levels down 24-fold), has also been associated with Wnt antagonism (see Fig. 3 and results of human cortex analysis). b-actin served as housekeeping gene. Protein interaction data suggest Wnt pathways as a major target of c-secretase susceptible gene transcription In order to see whether c-secretase affects the transcription of genes encoding interacting proteins, an interaction map of encoded proteins was generated with the string 8.0 data bank exclusively relying on evidence-based data. Clusters of protein interactions suggest the Wnt signaling pathways as a major focus of c-secretase-affected candidates (Fig. 4 , highlighted in grey). Indeed, we found several members of the canonical Wnt pathway, but also some interactors of the planar cell polarity (PCP) pathway and the Wnt/Ca2+ pathway, to have c-secretase activity susceptible gene transcription (Fig. 4) . Some of these genes have been confirmed by real time PCR as well as DIGE experiments (Egger et al., unpublished). The largest decrease in gene transcription occurred for the gene encoding the protein Ptprg. This single-pass type I membrane protein dephosphorylates protein tyrosine phosphate and was recently suggested as a candidate tumor suppressor gene in nasopharyngeal carcinoma [39] . The same group reported functional evidence for a critical interaction of Ptprg with the extracellular matrix, which induces cell arrest, changes in cell cycle status and downregulation of cyclin D1 [39] . The latter is strongly affected by the canonical Wnt pathway. Ptprzeta and beta, structurally similar to Ptprg, interact with Psd95 [40] , which directly interacts with Wnt3a [41] . We could confirm that WNT3A transcripts show an increase of 2.8-fold (Fig. 3) . Further, Wnt3a has also been reported to interact directly with LRP1 ( Fig. 4, lower right) , a stimulator of the Wnt5a signaling pathway [42] and a known c-secretase substrate tying c-secretase to a major AD risk factor, ApoE [43] . Porcn, another protein that interacts with Wnt3a [44] , shows a three-fold increase in transcript level by the microarray experiment. Porcn also interacts with Wnt 6 (4-fold increase in microarray) as reported by the same group and is the first player of the canonical Wnt pathway as displayed by the Kegg database (mmu04310, http://www. genome.jp/dbget-bin/show_pathway?mmu04310). Wnt3a interacts with Frizzled 1 [45] , which showed a 5-fold increase in mRNA levels by our microarray. Following the canonical Wnt pathway, the first intracellular protein of the Wnt signaling cascade is ''Disheveled''. As confirmed by real time PCR, DVL3 mRNA is increased by 3-fold with enhanced c-secretase activity. Next, with the help of Gbp (microarray reports a 4-fold increase of Gbp2), Gsk-3b is inhibited, which in turn inhibits b-catenin. As made apparent by the graphical overview of interacting proteins encoded by genes we found to be transcriptionally susceptible to csecretase activity, b-catenin plays a central role, linking different proteins involved in different Wnt pathways (Fig. 4) . Furthermore, b-catenin has been found to function as a major node connecting PS1 and several proteins that are encoded by genes that we found to be differentially transcribed (Fig. 4) . b-catenin transcription was shown by the microarray to be 3-fold decreased. It interacts directly with cdh15 (which showed a 2.4-fold increase in transcript levels as confirmed by real time PCR, Fig. 3) , with Cdh1 (a known c-secretase substrate [46] ), with PS1 (the c-secretase catalytic subunit) and other proteins encoded by candidate genes reported by the microarray. In the context of the canonical Wnt pathway, b-catenin affects c-myc (CMYBP 3-fold increased on the microarray), c-jun (3-fold decreased on the microarray) and cyclin D (4-fold decreased on the microarray) -the latter, as mentioned, is also downregulated by Ptprg. Dvl3 however also interacts with other proteins encoded by candidate genes, among which Nkd1 is of special interest since it links the canonical with the planar cell polarity pathway where it has a different effect on Dvl. The Planar cell polarity pathway through several players, among them Rac (3fold decreased on the microarray) affects gene transcription, as we hypothesize for c-secretase activity changes. Through a chain of different mediators, the planar cell polarity Wnt pathway affects the cytoskeleton. Our microarray has reported some of these mediators to be differentially transcribed as c-secretase activity is enhanced; RhoA transcript levels for example are 3-fold decreased. Rock, which is known to directly interact with the csecretase substrate CD44 (CD44/Rho Family GTPase/ROCK2) [47, 48] , is transcriptionally affected too. Our top candidate UPP1 has only one interaction partner that was also reported to be differentially expressed by the microarray, the cytoskeleton protein vimentin (Vim) [35] . Vimentin itself is not new to AD research, as altered Vim distribution patterns were observed in FAD fibroblasts [36] . Also, UPP1 transcription is regulated by the transcription factor Oct3/4, as is the transcription of another candidate, called SPP1 [49] . Spp1 is a direct interaction partner of the aforementioned c-secretase substrate CD44 and strongly affects Ca 2+ levels [50] . It directly interacts with several proteins encoded by candidate genes, including PKCA, which itself directly interacts with Aplp2, a well-known c-secretase substrate, and Csnk2b, which directly interacts with b-catenin, thus closing the circle. Csnk2b also directly interacts with Shmt1, which has enhanced transcription of 3.5-fold (Fig. 3) , and has been further confirmed in DIGE experiments (Egger et al., unpublished) . The third Wnt pathway mentioned is the Wnt/Ca 2+ pathway which includes, among others, Plc (Plcb1 5-fold increase on the microarray), CaMKII (4-fold decrease on the microarray) and Calpain (3-fold decrease on the microarray). Our mapping of genes differentially transcribed with c-secretase activity shows that they encode proteins that directly interact with each other, with many of them being members of Wnt pathways. Our modeling of extreme levels of c-secretase activity in CHO cells has revealed c-secretase-dependent differences in transcript levels of specific genes. One of the major known risk factors for developing Alzheimer's disease is carrying the ApoE4 allele. Recently it was shown that ApoE through LRP1 regulation is connected with c-secretase [43] , which supports the hypothesis of a potential role of c-secretase in sporadic AD. c-Secretase is also directly implicated in the inheritable familial early onset forms of AD (FAD), as most cases are caused by mutations in PSEN1, the gene encoding for PS1, the catalytic center of this enzyme. To investigate whether changes in gene transcription that coincide with alterations of c-secretase activity levels also differ between sporadic Alzheimer's and healthy human brain tissue, we evaluated our top scoring c-secretase affected genes in human AD and healthy cortices. Based on b-actin as housekeeping gene, we found one c-secretase affected gene, TERA, to be significantly differentially transcribed in the AD brain relative to the normal brain. Real-time PCR results showed an average two-fold increased TERA transcript levels (P2 = 0.04) in human AD cortices compared to healthy controls (Fig. 5) . Altogether, the Wnt antagonism gene TERA represents a new candidate for differential expression with c-secretase activity as well as in AD brain cortex tissue. Whether it is implicated in the pathogenesis of AD requires further investigation. Since the discovery of the roles for NICDs and AICDs in gene transcription, the notion of c-secretase as a major player in pathologically altered gene transcription patterns has been steadily gaining ground with new substrates and their transcriptionally active ICDs being identified regularly. To investigate the impact of c-secretase activity on gene transcription, we compared two starkly contrasting situations: genetically engineered enhanced human csecretase activity and pharmacologically inhibited c-secretase activity in CHO cell lines. By investigating the effects of enhanced c-secretase activity on gene transcription using cDNA microarray analysis, we could show that the canonical, the planar cell polarity (PCP) and the Ca 2+ /Wnt pathways are transcriptionally affected through more than a dozen of Wnt signaling players (summarized in Fig. 6 ). From Proc and Wnt outside the membrane, through Frizzled and Dvl, to b-catenin and down to cell cycle regulating genes, the canonical Wnt pathway is the most affected of Wnt pathways. Several genes of the PCP Wnt pathway as well as Ca 2+ / Wnt pathways were found to be differentially expressed too (Fig. 6) . One of the cell cycle regulating genes is CYC-D, which itself is Figure 5 . Selected relative gene transcript levels in AD cortices. Real time PCR validated genes differentially transcribed in cells with enhanced c-secretase activity were selected and their gene transcript levels analyzed in ten to twelve AD and healthy human cortical brain tissue samples. Only the transcript levels of TERA, a gene of unknown function, is significantly altered with a two-fold increase in AD cortices. Note that TERA transcript levels were significantly reduced in cells with enhanced c-secretase activity (Fig. 3) . Relative quantification of gene transcription in CHO cells as well as in brain tissue used b-actin as housekeeping gene. Healthy control levels are displayed on the left part of each diagram, AD transcript levels on the right. Dashed lines indicate mean values for healthy controls (green) and AD cases (red). Double-headed arrows indicate tendencies of differences between groups. P2 values obtained from t-test are indicated in black boxes of the upper part of each diagram. doi:10.1371/journal.pone.0006952.g005 regulated by one of the most c-secretase-dependently altered genes reported by us, PTPRG. Functional clustering of the microarray data revealed the overrepresentation of the ''receptor binding'' cluster, which includes four different Wnt signaling molecules and b-catenin. b-Catenin also finds itself in the center of interactions of proteins encoded by strongly differentially expressed genes. Components downstream of the canonical Wnt pathway, like c-myc, c-jun and cycD, influence the cell cycle, the latter as mentioned is downregulated by protein tyrosine phosphatase receptor type c (Ptprg). Interestingly, we found PTPRG transcription to be strongly decreased in cells with enhanced c-secretase. Barnea et al. [51] identified a subfamily of PTPRs, defined by the carbonic anhydrase-like domain in the extracellular region of PTPRG, and described its expression during hippocampal formation, and in septal and midline thalamic nuclei in the cortex of newborn rats (in contrast to the expression pattern in adult rats, which is reduced to the hippocampal formation). Several groups have shown a connection between alterations in receptor tyrosine phosphatases' expression levels and c-secretase [52, 53] . However, we report here for the first time, to our knowledge, the transcriptional connection between the receptor tyrosine phosphatase type gamma and c-secretase. TERA, a gene that we found to be decreased in transcription (down by 23.5-fold), has been connected to brain development and Wnt antagonism as well. TERA is decreased to minimal transcript levels with enhanced c-secretase activity (Fig. 3) . This gene, encoding a phosphoprotein of unknown function, is upregulated in squamous cell carcinoma (SCC), adenocarcinoma (AC), and colon, ovary, rectum and stomach tumors [54] (suggesting associations with Notch?). It has also been reported that TERA gene expression is increased in day 13 embryonic (E13) and decreased in E17 cortex and maintains low, but consistent expression levels in the subventricular zone (SVZ) [55] . The expression pattern in earlier rather than later stages of brain development and in the location of neuronal stem cell niches, like the SVZ, suggest possible roles for Tera in regenerative processes and raise questions about its function if the gene is being shut down in degenerative disorders like AD [55] . Tera expression has further been found to be maintained in neural progenitors and downregulated during non-neural differentiation, and was shown to have appreciable expression in embryonic stem cells in a screen Figure 6 . Involvement of c-secretase-dependently transcribed genes in Wnt pathways. Several key players of the canonical Wnt pathway (green panel) were reported by our microarray to increase (red quadrangles) or decrease (blue quadrangles) in transcript levels under conditions of enhanced c-secretase activity compared to inhibited activity. b-Catenin is a central node connecting Wnt-Frizzled-Dishevelled to a downstream effect influencing the cell cycle (see also Fig. 4 and interactions of encoded proteins). For better understanding, selected genes that were not detected by the microarray are displayed as well (dashed lines black quadrangles). CycD was reported to be regulated by PTPRG, one of the top scoring candidates for c-secretase affected gene transcription. Nkd, which we found to be increased in transcript levels, connects the canonical Wnt pathway with the planar cell polarity pathway (blue panel). CD44, a well-known c-secretase substrate, interacts with SPP1. SPP1 and UPP1, two strong candidates are both under the control of the same transcription factor Oct3/4, as has been suggested for TERA [92] . UPP1 directly interacts with Vimentin (see also Fig. 4) , a known player in AD and a cytoskeletal protein. Crucial genes of the Wnt/Ca 2+ pathway (grey panel) were also found to be differentially expressed in our array. All together, c-secretase activity influences the transcript levels of genes of the canonical, the planar cell polarity and Ca 2+ Wnt pathways. doi:10.1371/journal.pone.0006952.g006 for functional genes in ES cells that implicated Wnt antagonism in neural differentiation [56] . TERA and the anti-mitotic exit network antagonist 1 (AMN1) map to chromosome 12p11, which is interesting when considering the fact that chromosome 12 has been discussed to contain an unknown LOAD locus for over a decade, and in a recent study including 492 LOAD cases [57] [58] [59] . In our study, AMN1 transcription is decreased by 978-fold with enhanced c-secretase activity. The function of AMN1 is not known. However, several expression pattern based studies suggest it functions as a cilia gene in sensory neurons [60] . Another typical cilia gene is intraflagellar transport protein 81 (IFT81) which, among a dozen of known cilia genes, was also shown by the microarray to be differentially expressed with altered c-secretase activity (see also Fig. 3 ). More and more evidence has been emerging over the last years that primary cilia, in parallel to their well-established functions in sight, smell and mechanosensation, are key participants in intercellular signaling [61] . The importance of monocilia for the regeneration of olfactory neurons has only been better understood recently [62] . Subventricular zone (SVZ) astrocytes, providing glia as well as neurons for the mammalian olfactory bulb, have primary cilia [63] . They give rise to type C cells, which in turn generate neuroblasts [64] that migrate in the adult brain from the SVZ to the olfactory bulb along the cerebrospinal fluid (CSF) flow. The CSF flow depends on the beating of the ependymal cilia [65] . Cilia genes are not only relevant to the maintenance of adult regeneration in the brain since they uphold the constant flow of the CSF, but also because they are directly implicated in cell cycle control. Polycystins, for example, control the cell cycle through three major pathways with one depending directly on b-catenin [66] . A study of inversin has further shown that flow shear stress as sensed through cilia may regulate the Wnt signaling pathway through b-catenin [67, 68] . Given that fluid flow is crucial for the transport of neuroblasts in the SVZ, one could expect that bcatenin and the Wnt signaling pathway that connects our candidates are also functionally relevant to the cilia genes found in this study. We found both genes of unknown function TERA and AMN1 to be decreased in transcription with enhanced csecretase activity. TERA and AMN1 can be connected to neural stem cells through several types of cancer, neural differentiation (in the case of TERA) and through the role of monocilia for neurogenesis (in the case of AMN1). All in all, we have demonstrated that AMN1 and TERA are genes of basically unknown function that are worthy of further investigation to understand their roles in neurogenesis, cancer and c-secretase biology. We further report here that UPP1 transcript levels are increased with enhanced c-secretase activity (by 39.2-fold). UPP1 encodes for uridine phosphorylase (UPase), an enzyme that catalyzes the reversible phosphorylytic cleavage of uridine and deoxyuridine to uracil and ribose-or deoxyribose-1-phosphate [69] . UPP1 expression has been extensively connected to cancer, stem cells and inflammation such as multiple sclerosis [70] [71] [72] [73] [74] [75] [76] [77] . UPase is induced by vitamin D3 and a mixture of inflammatory cytokines, Interferon gamma, TNF-alpha and IL-1, with the latter two being upregulators of Ptprg [78] . Increased UPP1 transcript levels, associated with enhanced UPase activity cleaving uridine, would potentially have inhibitory effects on several pathways downstream of uridine, like RNA/DNA and membrane synthesis, as well as protein glycosylation, which would in turn trigger long-term neurodegeneration. Particularly, decreased membrane synthesis, in the case of synaptic membranes, would also reduce synaptic activity and plasticity. In support of that, TNF-a and IL-1, inducers of UPP1, alter lipid metabolism and stimulate production of eicosanoids, ceramide and reactive oxygen species that potentiate CNS injuries and certain neurological disorders [33] . Interestingly, this hypothesis offers an explanation for the multitude of beneficial effects of orally administered DHA and uridine on memory, neuronal health, regeneration and membrane synthesis in traumatic and chronic neuropathological conditions [33, 34] . The presented work demonstrates that c-secretase is capable of influencing single gene transcription. However strong the impact will prove to be on the protein level of each single gene, we have further observed transcriptional effects spanning several genes throughout clearly defined pathways. This puts forth the possibility of much stronger effects on the target functions of these pathways than the small impact on the individual genes transcriptional or translational levels might indicate. In support of this hypothesis, we have observed that the proteins encoded by those genes interact with each other and are part of the Wnt pathways. Evaluation of the impact of these pathway-specific accumulative effects needs further investigation. This should include physiological and pathological in vivo experiments on both the transcript as well as protein levels. For c-secretase to serve as a therapeutic target, it is indeed crucial to sharpen our view of its role and influence over gene transcription and biological functions. The S-1 cell line overexpressing Flag-Pen2, Aph1-a2-HA, PS1 and NCT-GST [17, 79] was derived from the Chinese Hamster Ovary (CHO) c-30 cell line [80] CHO parental cell line triplicates were exposed for 24 hrs to the c-secretase inhibitor DAPT (10 mM) in DMSO (0.05%), and S-1 cells were treated for the same time with DMSO (0.05%). Cells were next washed twice with PBS and total RNA was extracted, amplified, reversely-transcribed, labeled and hybridized to a 17 k mouse cDNA microarray chip produced by the DNA array facility of Lausanne (DAFL, see below). Total RNA extraction: was performed using the RNeasy Mini Kit (Qiagen, Basel, Switzerland), in the absence of DNAse treatment. RNA quality was assessed using the RNA 6000 Nanochip assay (Agilent Technologies, Meno Park, USA) and RNA concentration was determined using the ND-1000 spectrophotometer (Nanodrop Technologies, Wilmington, USA). Three independent experiments were performed. RNA amplification: a single round of amplification was performed with 3 mg of total RNA using the MessageAmp RNA Amplification Kit (Ambion, Austin, USA) and following the protocol provided with the kit. Next, 5 mg of amplified RNA was mixed with 9 mg random primers (Cat. No. 4819001; Invitrogen, Carlsbad, USA) in 19 ml of water, heated for 5 minutes at 70uC and then immediately transferred to ice. Reversed transcription and labeling: was performed for 2 hrs at 42uC in a final reaction volume of 40 ml containing 1X SuperScript II buffer (Invitrogen), 40 units RNasin (Promega, Madison, USA), 10 mM DTT, 0.5 mM dATP, dGTP, dTTP, 0.2 mM dCTP, 0.1 mM of either Cy3-dCTP or Cy5-dCTP (GE Healthcare, Uppsala, Sweden) and 400 units of SuperScript II reverse transcriptase (Invitrogen). The RNA strand was hydrolyzed by adding 2 ml 500 mM EDTA and 4.5 ml 1 M NaOH and heating at 65uC for 15 minutes; the solution was then neutralized by adding 2.5 ml 1 M Tris (pH 6.8) and 4.5 ml 1 M HCl. The labeled cDNA was purified using the Qiagen MiniElute PCR Purification Kit (Cologne, Germany), eluting in 50 ml of EB buffer according to the manufacturer's instructions. The Cy3 and Cy5 labeled targets were combined and mixed with 400 ml of TE, 20 mg Cot 1 DNA (Invitrogen), 10 mg polyadenylic acid (Sigma, St. Louis, USA) and 10 mg yeast tRNA (Sigma). This mixture was concentrated to a final volume of 19.4 ml using a Microcon YM-30 filter (Millipore, Billerica, USA) according to the manufacturer's instructions. 20X SSC and 10% SDS were added to final concentrations of 3X and 0.4%, respectively, in a final volume of 24 ml. This mixture was heated for 2 minutes at 98uC, pipetted immediately onto the cDNA microarray and, after covering with a glass cover slip (Erie Scientific, Portsmouth, USA), placed in a humidified chamber (Telechem, Sunnyvale, USA) and allowed to hybridize at 64uC for 20 hrs. Slides were then washed at room temperature twice for 5 minutes in 2X SSC, 0.1% SDS, twice for 1 minute in 0.2X SSC, once for 1 minute in 0.1X SSC and once for 5 minutes in 0.1X SSC, 0.1% Triton X-100. After drying, slides were scanned on a microarray scanner (Agilent Technologies) and the resulting TIFF images were analyzed using the GenePix Pro 6.0 software (Molecular Devices, Sunnyvale, USA). The mouse cDNA microarrays used in this study consisted of approximately 17,000 PCR products generated from cDNA clones and control DNAs spotted onto Nexterion AL slides (Schott, Mainz, Germany). A complete description of the slides and their content can be obtained from the Lausanne DNA Array Facility (http://www.unil.ch/dafl). The microarray data set discussed in this publication has been deposited in the NCBI Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) and is accessible through GEO Series accession number GSE16379. Note that Hamster genomic sequence information is not yet sufficiently available to the research community. Consequently no commercial hamsterspecific microarrays were available at the time of the experiment. However, the strategy to use a microarray from a closely related species is not new and has proven successful before [81] . The analysis was performed with open source R software packages (http://www.r-project.org/ and http://www.BioConductor.org/). Gene expression was quantified with the array package using print tip group lowess normalization without background subtraction. The resulting measures of expression for each array are the log2 ratios (M values) and the average log2 intensities (A value) of Cy3 and Cy5 signals. Statistics of differential expression between the different groups of samples were calculated with a linear model fitted by the limma package. Total RNA was isolated with the RNeasy mini kit following the manufacturer's protocol for adherent cells in the case of CHO cell cultures. For the isolation of total RNA from brain tissue, the TRIzol reagent was used as described in the human samples section. RNA was dissolved in water, which was followed by ND-1000 spectrophotometer (Nanodrop Technologies, Wilmington, USA) quantification and pico chip quality control analysis (6000 Nanochip assay Agilent Technologies, Meno Park, USA). Total RNA was reverse transcribed with our standard laboratory protocol. 1 mg of total RNA was dissolved in 4 ml of RNase-free water (Ultrapure DNase/RNase free water, Invitrogen Carlsbad, USA)) and premixed with 0.5 mg of oligo dT primer (synthesized by Eurogentec Seraing, Belgium) dissolved in 1 ml RNase-free water. The RNA/oligo dT premix was heated to 70uC for 5 minutes in a standard PCR machine (TProfessional Basic Gradient, Whatman Biometra Goettingen, Germany). The machine was paused to add 4 ml of 5X Buffer (ImProm-II M28A, Promega Madison USA), 4 ml of MgCl 2 (25 mM) (Promega Madison USA), 1 ml dNTP Mix (10 mM U151B, Promega Madison USA), 1 ml RNase inhibitor (RNasin Plus N261A 40 u/ul Promega Madison USA), 1 ml of ImProm-II Reverse Transcriptase (Promega Madison, USA) and 4 ml RNasefree water. The PCR machine program was continued after pausing at 25uC for completion of reaction mixes with 60 min at 42uC and 15 min at 70uC. cDNA was kept at 4uC on wet ice for short-term or at 280uC for long-term storage. Reverse transcription products were used without purification for real time PCR at equivalent of 0.5 ng/ml RNA in 384 well plates. Samples were used as biological triplicates and each one was additionally pipetted as a triplicate. Reaction volumes were 10 ml consisting of 5.02 ml SYBR Green (Power SYBR Green Master Mix #4367660 Applied Biosystems, Cheshire UK), 1.49 ml RT-PCR product at 0.5 ng/ml input RNA equivalent (0.75 ng/rxn) and 3.49 ml of 3 mM Forward and Reverse primer mix. 384 well plates were prepared with a liquid handling robot (Freedom EVOware Tecan Trading AG, Switzerland) and read for relative quantification with Applied Biosystems 7900HT Real-Time PCR System (Applied Biosystems, Cheshire UK). Primers (synthesized by Eurogentec Seraing, Belgium) for CHO cDNA were based on mouse code, which was aligned with rat and human code, preference was given to aligning sequences (Table 3) . Sequence specificity was determined via nBlast. b-actin was used as housekeeping gene [82] [83] [84] [85] [86] [87] [88] [89] for CHO as well as human cortex templates with the forward sequence: CCTTCAACACCCCAGCCATGTACG and the reverse sequence: CCTTCAACACCCCAGCCATGTACG. Results were analyzed by the DDCt method [90] and significance was calculated via students t-test. b-actin was used a normalizer to determine DCts. DDCts were calculated against the mean of DAPT treated WT-CHO DCts or the mean of healthy human brain cortex DCts. Results were expressed as relative quantification by 2ˆ-(DDCt) [90] . Human brain tissue was kindly provided by the Joseph and Kathleen Bryan Alzheimer's Disease Research Center, Duke University Medical Center. The Autopsy and Brain Donation procedures have been approved by the Duke University Institutional Review Board (IRB) and cortical brain tissue was obtained as described by [91] . 12 AD post-mortem confirmed cortical samples as well as 12 healthy cortical samples were obtained in dry ice. Cortical samples were of both genders, different ages, ApoE stati and Brack stages. Isolation of total RNA: ,50 ug of total cortex tissue were scraped off on dry ice three times for biological triplicates of each sample. TRIzol reagent (Invitrogen Carlsbad, USA) was used according to manufacturer's protocol for total RNA isolation. RNA was dissolved in water, which was followed by ND-1000 spectrophotometer (Nanodrop Technologies, Wilmington, USA) quantification and pico chip quality control analysis (6000 Nanochip assay Agilent Technologies, Meno Park, USA). Dataset S1 EST clones with increased transcription under enhanced c-scretase activity compared to inhibited c-secretase activity. By applying a cut-off with a p value of 0.005 based on the false discovery rate (FDR, i.e. the probability to wrongly accept a difference between the two conditions), we found 2658 EST clones to be differentially expressed, with 1241 EST clones of increased with enhanced c-secretase activity compared to inhibited csecretase activity. FC = Fold change; adj,P,Val = adjusted P-value Found at: doi:10.1371/journal.pone.0006952.s001 (0.20 MB XLS) Dataset S2 EST clones with decreased transcription under enhanced c-secretase compared to inhibited c-secretase activity. By applying a cut-off with a p value of 0.005 based on the false discovery rate (FDR, i.e. the probability to wrongly accept a difference between the two conditions), we found 2658 EST clones to be differentially expressed, with 1417 EST clones of decreased transcription with enhanced c-secretase activity compared to inhibited c-secretase activity. FC = Fold change; adj,P,Val = adjusted P-value Dataset S4 EST clones of transcriptional relevance differentially transcribed under enhanced c-secretase compared to inhibited csecretase activity. By applying a cut-off with a p value of 0.005 based on the false discovery rate (FDR, i.e. the probability to wrongly accept a difference between the two conditions), we found 2658 EST clones to be differentially expressed with enhanced c-secretase activity compared to inhibited c-secretase activity. Among them 56 imply transcriptional relevance. FC = Fold change; adj,P,Val = adjusted P-value. Found at: doi:10.1371/journal.pone.0006952.s004 (0.03 MB XLS)

Bayesian Phylogeography Finds Its Roots

As a key factor in endemic and epidemic dynamics, the geographical distribution of viruses has been frequently interpreted in the light of their genetic histories. Unfortunately, inference of historical dispersal or migration patterns of viruses has mainly been restricted to model-free heuristic approaches that provide little insight into the temporal setting of the spatial dynamics. The introduction of probabilistic models of evolution, however, offers unique opportunities to engage in this statistical endeavor. Here we introduce a Bayesian framework for inference, visualization and hypothesis testing of phylogeographic history. By implementing character mapping in a Bayesian software that samples time-scaled phylogenies, we enable the reconstruction of timed viral dispersal patterns while accommodating phylogenetic uncertainty. Standard Markov model inference is extended with a stochastic search variable selection procedure that identifies the parsimonious descriptions of the diffusion process. In addition, we propose priors that can incorporate geographical sampling distributions or characterize alternative hypotheses about the spatial dynamics. To visualize the spatial and temporal information, we summarize inferences using virtual globe software. We describe how Bayesian phylogeography compares with previous parsimony analysis in the investigation of the influenza A H5N1 origin and H5N1 epidemiological linkage among sampling localities. Analysis of rabies in West African dog populations reveals how virus diffusion may enable endemic maintenance through continuous epidemic cycles. From these analyses, we conclude that our phylogeographic framework will make an important asset in molecular epidemiology that can be easily generalized to infer biogeogeography from genetic data for many organisms.

Phylogenetic inference from molecular sequences is becoming an increasingly popular tool to trace the patterns of pathogen dispersal. The time-scale of epidemic spread usually provides ample time for rapidly evolving viruses to accumulate informative mutations in their genomes [1] . As a consequence, spatial diffusion-among other processes-can leave a measurable footprint in sampled gene sequences from these viruses [1] . Reconstructing both the evolutionary history and spatial process from these sequences provides fundamental understanding of the evolutionary dynamics underlying epidemics, e.g. [2, 3] . It is also hoped that these insights can be translated to effective intervention and prevention strategies [4] and elucidating the key factors in viral transmission and gene flow over larger distances is central in formulating such strategies, e.g. [5] . Phylogeographic analyses are a common approach in molecular ecology, connecting historical processes in evolution with spatial distributions that traditionally scale over millions of years [6] . Many popular phylogeographic approaches [7, 8] can be remiss in ignoring the interaction between evolutionary processes and spatial-temporal domains. One first reconstructs a phylogeny omitting spatial information and then conditions the phylogeo-graphic inferences on this reconstruction [1, 9] , exploiting nonparametric tests to evaluate the significance of this conditional structure, e.g. [7, 10, 11] . To draw conclusions about the epidemic origin or epidemiological linkage between locations, however, we require a reconstruction of the dispersal patterns and process throughout the evolutionary history. Considering locations as discrete states, this boils down to the well-known problem of ancestral state inference [7] . Parsimony is a popular heuristic approach to map characters onto a single phylogenetic tree [12] . Unfortunately, parsimony reconstructions ignore important sources of model uncertainty, including both uncertainty in the dispersal process as well as in the unknown phylogeny [13] . In addition, minimizing the number of state exchanges over a phylogeny is misleading when rates of evolution are rapid and when the state exchange probabilities are unequal [14] . Probabilistic methods draw on an explicit model of state evolution, permitting the ability to glimpse the complete state history over the entire phylogeny and conveniently draw statistical inferences [15] [16] [17] . These analyses typically employ continuoustime Markov chain models for discrete state evolution analogous to common nucleotide, codon or amino acid substitution models [18] . In contrast to parsimony, maximum likelihood-based reconstructions incorporate branch length differences in calculat-ing the conditional probability of each ancestral state given the observed states at the phylogeny tips [14] . Bayesian reconstruction methods enable further generalization of this conditional probability analysis by removing the necessity to fix the Markov model parameters to obtain ancestral states and the necessity to specify a fixed tree topology with known branch lengths. Bayesian inference integrates conclusions over all possible parameter values but to achieve this, however, requires prior probability distributions for all aspects of the model. While probabilistic methods have been previously presented in a bio-or phylogeographic context, in particular Bayesian methods that integrate over phylogenetic uncertainty and Markov model parameter uncertainty [19] , viral phylogeography studies have rarely made use of these developments. This may be a consequence of low awareness of existing software implementations for arbitrary continuous-time Markov chain models [20, 21] or a lack of appreciation for the uncertainty intrinsic in these reconstructions and the ease with which one can formally access epidemiological linkage through probabilistic approaches. A recent phylogeographic study of influenza A H5N1 introduces a heuristic non-parametric test to evaluate whether parsimonyinferred migration events between two particular locations occur at significantly high frequency [22] . Null distributions for these frequencies arise from randomizing tip localities after false discovery rate correction to control for simultaneous testing issues. Although this procedure addresses concerns about statistical inference on sparse frequency matrices, the multiple comparison correction still results in a conservative estimate of significant migration events. Fully probabilistic approaches may further ease statistical inference, yet similar tests remain lacking for likelihoodbased phylogeographic models. Advances in evolutionary inference methodology have frequently demonstrated how novel approaches can be appended to a sequence of analyses, in many cases starting from alignment to parameter estimation conditional on tree reconstructions. For example, demographic inference has involved genealogy reconstruction, estimating a time scale for the evolutionary history, and coalescent theory to quantify the demographic impact on this tree shape [23] . It is well acknowledged that such sequential procedures ignore important sources of uncertainty because they generally purge error associated with each intermediate estimate. With the advent of novel computational techniques like Markov chain Monte Carlo (MCMC) sampling, it has become feasible to integrate many of the models involved and simultaneously estimate parameters of interest. Demographic inference is a wellknown example of genealogy-based population genetics that benefited from these advances [24, 25] . Bayesian MCMC methods also enable ancestral state reconstruction while simultaneously accounting for both phylogenetic and mapping uncertainty. Although this adds much needed credibility to ancestral reconstruction [13] , phylogeographic analysis would benefit even more from fully integrating spatial, temporal and demographic inference. Here, we implement ancestral reconstruction of discrete states in a Bayesian statistical framework for evolutionary hypothesis testing that is geared towards rooted, time-measured phylogenies. This allows character mapping in natural time scales, calibrated under a strict or relaxed molecular clock, in combination with several models of population size change. We use this full probabilistic approach to study viral phylogeography and extend the Bayesian implementation to a mixture model in which exchange rates in the Markov model are allowed to be zero with some probability. This Bayesian stochastic search variable selection (BSSVS) enables us to construct a Bayes factor test that identifies the most parsimonious description of the phylogeographic diffusion process. We also demonstrate how the geographical distribution of the sampling locations can be incorporated as prior specifications. Through feature-rich visual summaries of the space-time process, we demonstrate how this approach can offer insights into the spatial epidemic history of Avian influenza A-H5N1 and rabies viruses in Africa. The highly pathogenic avian influenza A-H5N1 viruses have been present for over a decade in Southern China and spread in multiple waves to different types of poultry in countries across Asia, Africa and Europe [26] . As a result, highly pathogenic A-H5N1 is now a panzootic disease and represents a continuous threat for human spill-over. Strong surveillance has been in place since these viruses caused extensive outbreaks, but the source and early dissemination pathways have remained uncertain. Because parsimony analysis has attempted to shed light on the latter [22] , A-H5N1 provides an ideal example for comparison with Bayesian phylogeographic inference. Rabies is endemic in Asia and Africa, where the primary reservoir and vector for rabies virus (RABV) is the domestic dog. Phylogenetic analysis has revealed several genotypes of lyssaviruses (family Rhabdoviridae); genotype 1 has been found responsible for classical rabies, a fatal disease in terrestrial mammals throughout the world [27, 28] . Here, we explore the phylogeographic history of RABV in domestic dogs in West Central Africa, using recently obtained sequence data, and evaluate the role of viral dispersal in maintaining RABV epidemic cycles. We examine the evolution and spatial dispersion of two viral pathogens, Avian influenza A-H5N1 and rabies, to demonstrate the strengths and limitations of our discretized stochastic model for phylogeography. To reconstruct the spatial dispersion patterns of Avian influenza A-H5N1, we analyze the hemagglutinin (HA) and neuraminidase (NA) gene datasets previously compiled by [22] . Both datasets contain whole gene sequences from 192 A-H5N1 strains sampled from 20 localities across Eurasia. [22] explore these genes individually, as well as concatenated together, through a strictly parsimony-based ancestral reconstruction method. Our Bayesian approach builds upon stochastic models and naturally affords Spreading in time and space, rapidly evolving viruses can accumulate a considerable amount of genetic variation. As a consequence, viral genomes become valuable resources to reconstruct the spatial and temporal processes that are shaping epidemic or endemic dynamics. In molecular epidemiology, spatial inference is often limited to the interpretation of evolutionary histories with respect to the sampling locations of the pathogens. To test hypotheses about the spatial diffusion patterns of viruses, analytical techniques are required that enable us to reconstruct how viruses migrated in the past. Here, we develop a model to infer diffusion processes among discrete locations in timed evolutionary histories in a statistically efficient fashion. Applications to Avian Influenza A H5N1 and Rabies virus in Central and West African dogs demonstrate several advantages of simultaneously inferring spatial and temporal processes from gene sequences. quantification of uncertainty in both the ancestral state reconstructions and the underlying phylogeographic process. Further, as we are able to infer plausible root positions unlike the original analysis, we are not required to include outgroup sequences. To model sequence evolution, we employ the [29] (HKY85) CTMC model of nucleotide substitution; we include discrete gamma-distributed rate variation [30] and assume an unknown, constant population-size coalescent process prior over the unknown phylogeny [31] . Exploratory analyses using the less restrictive Bayesian skyline plot model indicate that the demographic prior has little influence on the phylogeographic inference (data not shown). Figure 1 summarizes the Bayesian maximum clade credibility (MCC) trees for the A-H5N1 HA and NA segments. An MCC tree is a point-estimate characterizing the posterior distribution of trees and represents the tree topology yielding the highest product of individual clade probabilities in their posterior sample [2] ; branch lengths in these MCC trees are posterior median estimates. We further annotated the tree nodes with their most probable In combination with other topological differences between the trees, this difference strongly suggests past reassortment events between both segments, with the progenitor virus of the basal Hong Kong clade and a chicken strain from Hebei having acquired an NA segment from different lineages. Such events are not surprising given frequent reports of A-H5N1 reassortment in China, e.g. [26] , and the particular reassortment event for the basal Hong Kong clade has very recently been confirmed [32] . Despite different time scales for HA and NA, most probable location states agree on Guangdong as the predominant location of these sequences throughout the majority of their evolutionary history. As an indication of the A-H5N1 epidemic origin, we consider the inferred location at the root of the trees (Figure 1 ). In the HA tree, Guangdong and Hong Kong share a vast majority of the posterior mass, neighboring locations in which surveillance efforts report early Avian influenza cases [33, 34] . In the NA tree, although Hong Kong and Guangdong still obtain marginally higher support than other locations, all posterior root state probabilities are much closer to their prior probability. The substantially deeper NA root explains this difference as the depth greatly increases uncertainty on the root state. Table 1 quantifies differences in ancestral state reconstruction uncertainty between the HA and NA trees using the Kullback-Leibler (KL) divergence measure (see Methods). The NA tree results in considerably lower KL divergence than the HA tree, signifying a much smaller deviation of the posterior distribution of the root location from the prior. However, lack of phylogeographic structure in the data does not contribute to this difference because the NA trees return a lower association index (AI). This measure of spatial admixture is based on a sum across all nodes in the tree of the complement of the frequency of the most abundant location among all descendent taxa weighted by the depth of the node in the tree [35] , and thus bears some relationship with an entropy value for descendent taxa locations. The AI rescales this sum by its expectation for randomized location assignments and results in low values for relatively strong phylogeny-locality correlation whereas AI values close to one reflect complete spatial admixture. If the basal Hong Kong clade and a chicken strain from Hebei have indeed acquired a different NA through reassortment, the root state might be difficult to interpret for NA and is not necessarily the same as that for HA. Therefore, we also list uncertainty measures for the marginal posterior distribution of the most recent common ancestor (MRCA) of the Gs/GD lineage, named after the A/ goose/Guangdong/1/96 strain very close to this node (indicated in Figure 1 ). KL divergence is again lower for this node in the NA phylogeny, but the difference is not as pronounced as for the root node. Table 1 also explores the effects of distance-informed priors and BSSVS on location reconstruction. In general, the distance- Figure 1 . Maximum clade credibility (MCC) phylogenies for hemagglutinin (HA) and neuraminidase (NA) genes of Avian influenza A-H5N1. We color branches according to the most probable location state of their descendent nodes. We use the same color coding as [22] . To the upper left of both phylogenies are their root location state posterior probability distributions. A white arrow indicates the A/goose/Guangdong/1/96 sequence; a filled white circle identifies the most recent common ancestor of the Gs/GD lineage named after this strain. doi:10.1371/journal.pcbi.1000520.g001 informed priors furnish little advantage while inferring the root locations for both the HA and NA trees. If anything, KL divergences are slightly smaller for models involving distanceinformed priors than those with flat priors. For these data, this finding is unsurprising as physical distances can be poor proxies for inverse-diffusion rates when dispersal results from a heterogenous mix of migratory birds, transport of poultry and poultry products, and trade of wild birds [36] . Finally, we also investigated the uncertainty that is accommodated by averaging over plausible trees by analyzing the HA data using a fixed tree topology and branch lengths ( Table 1 ). The state reconstructions for the Gs/GD node in the fixed tree topology appear to ignore some uncertainty in comparison to integrating trees, which is not that evident for the root node. Although state reconstruction uncertainty is expected to be correlated among nodes, we also compared the KL divergence summed over all internal nodes, indicating much higher KL divergences using a fixed tree topology, e.g. for HA, m jk~C : 292 vs 523 for integrating trees and a fixed tree respectively. Under BSSVS, we assume a truncated Poisson prior that assigns 50% prior probability on the minimal rate configuration, comprising 19 non-zero rates connecting the 20 locations. This model strongly favors reduced parameterizations. A sensitivity analysis with respect to larger Poisson prior means reinforces that the data prefer a minimal number of rates, as increasing the mean leads to lower overall marginal likelihoods ( Table 2 ). BSSVS has a strong impact on root location reconstruction ( Figure 2 ). Many localities that are weakly supported as the root location without BSSVS obtain negligible posterior probability under BSSVS. Consequentially, BSSVS leads to larger KL divergences for both the HA and NA root nodes ( Table 1 ), suggesting that these reduced models more efficiently exploit the information content of the data. Interestingly, the posterior support for Guangxi increases under BSSVS at the expense of Guangdong in the HA phylogeny ( Figure 2 ). This may be an artifact of the reversible CTMC assumption we enforce. Specifically, at the tips of the phylogeny, several pathways of migration into Guangxi are highly likely. Assuming reversibility dictates that migration out of this location occurs as well; placing these emigrations deeper in the phylogeny is most consistent with the data. Because many locations already receive very low posterior probabilities at the GsGD node, the increase in posterior probability for a few locations now seems to outweigh the marginal reductions in posterior probabilities for most other locations and results in lower KL divergences at this node. By specifying a prior on the number of non-zero rates, we are able to construct Bayes factor (BF) tests for significance of individual rates ( Figure 3 ). To visualize the epidemiological linkage that this test establishes, we employ Google Earth to display all rates with a non-zero expectancy that results in a BF larger than three. The majority of well-supported rates (16 out of 25 for both genes) are concordant between HA and NA. Some variation in support for different migration pathways between HA and NA was also noted in the original parsimony analysis [22] . Importantly, Guangdong presents as an end-point in three wellsupported epidemiological links in HA as well in NA. For HA, four migration links previously identified using the parsimony sFDR test (Guangdong to Fujian, Bangkok to Vietnam, Uthai Thani to Phitsanulok, and Qinghai to Novosibirsk) are also present in our well-supported symmetric rates. We can, however, not confirm epidemiological linkage directly between Guangdong and Indonesia. Despite having more supported rates by this Bayes factor test as compared to the parsimony sFDR test, it remains difficult to We report the Kullback-Leibler divergence between the posterior and prior location distributions of the root and the GsGD most recent common ancestor (MRCA), as well as a phylogeographic association index. We analyze genes independently, assuming equal phylogeographic models (Shared) and by fixing the HA phylogeny through phylogeographic models with prior rates proportional to a constant (C) or distance-informed (DI) and using Bayesian stochastic search variable selection (BSSVS). doi:10.1371/journal.pcbi.1000520.t001 univocally identify the pathways seeding remote localities as Japan and Indonesia, and to connect the eastern diffusion network with the Chinese/Russian inlands. Distance-informed priors do not have strong influence on the Bayes factor test for significant rates. The presence of reassortment amongst the gene segments obfuscates phylogenetic inference for concatenated HA/NA sequence data. In this respect, it is interesting to note that previous parsimony reconstructions on a phylogeny for the concatenated HA and NA segments result in fewer significant diffusion rates compared to the separate analyses; [22] found 2 for the concatenated alignment vs. 5 and 10 for HA and NA separately. The Bayesian framework enables a flexible combination of the data without having to specify a single phylogeny for both segments. To this end, we share the instantaneous rate matrix L between both segment phylogenies and sample all parameters in a single MCMC analysis. Without BSSVS, sharing the rate matrix results in slightly higher KL divergences for both the root node and the Gs/GD node in the HA and NA phylogenies (Table 1 ). Figure 4 illustrates the well-supported rates based on the Bayes factor test of the shared rate matrix with a distance-informed prior. The shared data bring to light two possible pathways seeding the remote localities of Japan and Indonesia; these pathways suggest Guangxi and Hunan as possible source for Indonesia, and Hunan and Hebei as possible source for Japan. A major advantage of the current phylogeography implementation is the ability to infer the migration process in natural time scales. The panels in Figure 5 summarize the temporal dynamics of A-H5N1 spatial diffusion inferred using the shared rate matrix (KML files, Dataset S1 and S2 for HA and NA respectively, which enable visualizing the spread over time in Google Earth are available as supporting information). The lines connecting different locations represent branches in the MCC tree on which state exchanges occur and circle areas reflect the number of branches maintaining a particular state at that time point. By May 1997, Avian influenza lineages have accumulated in Guangdong, where the virus was originally isolated from a farmed goose [33] , and to a large extent in Hong Kong Southeast Asia resulting in severe A-H5N1 outbreaks in 2003. Finally, A-H5N1 virus also spreads to the west in a second major transmission wave. Since this occurs after a major outbreak in migratory waterfowl at Qinghai Lake in Northern China, migratory birds could play a prominent role in this dissemination pathway [37] . We investigate the ''Africa 2'' lineage of rabies transmitted by African dogs. This lineage forms one of the most divergent African rabies virus clades [28, 38] . The data set we analyze here comprises 101 complete nucleoprotein (N) gene sequences sampled across 12 African countries including Chad, Niger, Cameroon, the Central African Republic, Benin, Sierra Leone, Mali, Mauritania, Guinea, Ivory Coast and Burkina Faso [39] . Figure 6A illustrates the location-annotated MCC phylogeny and demographic history for the African dog rabies lineages. We make this initial inference without either BSSVS or a distanceinformed prior. To allow for variation in the underlying coalescent process giving rise to the phylogeny, we assume a piece-wise constant multiple change-point model on the effective populationsize with 20 coalecent-interval groups [25] . As generally observed for rabies viruses [28] , there exists strong signal of phylogenetic clustering according to sampling location. This observation is also These two locations are geographically distant from each other, but they both host viruses from the most basal lineage in the phylogeny ( Figure 6A ). Root inference is somewhat different using BSSVS and a distance-informed prior on the rates ( Figure 6B ). In this case, a more central location, Niger, obtains the highest posterior probability (0.144) but the KL divergence for the root state reconstruction is only marginally greater than zero (0.0645). The exploitation of BSSVS contributes to this effect; as for Avian influenza A-H5N1, distance-informed priors, alone, on the rates have little impact (data not shown). Although geographic origins remain elusive, we are able to identify locations that are epidemiologically linked using the BF test under BSSVS (Figure 7) . Panel A in the figure highlights the rates yielding a BFw3. The resulting migration graph is markedly parsimonious with a distinctive East-West axis running along the Southern border of the Saharan desert. To glean how this graph reflects the migratory process acting along the rabies phylogeny, panel B projects each of the branches of the MCC phylogeny onto the geographic map. In this projection, we translate each branch into a geographic link that connects the branch's most probable starting and ending location states. The height of a link represents the relative length of the time elapsed on the link's corresponding branch, while the color gradient reflects the relative age of the migration. Many recent (magenta) migration events that occur in a relatively short time contribute to the well-supported East-West axis. Although the best supported rates mainly form an East-West axis, many transitions along this axis occur in the last three decades; this suggests that the axis is not representative of a relatively slow unidirectional migration wave. Figure 8 reports the migration pathways over the last thirty years. These migration events accumulating over time, contingent on the estimated time of the branches on which they occur, demonstrate that RABV diffusion in West Africa is characterized by different simultaneous migration events in various directions rather than a unidirectional pattern, and that most of these migrations are short-range, occurring between neighboring countries. The Bayesian phylogeographic inference framework we present here incorporates the spatial and temporal dynamics of gene flow. In this study, we focus on pathogen diffusion because viral sequence sampling on a time-scale commensurate with the rate of substitution permits the inference of spatial patterns in real-time units. In addition, elucidating the phylodynamics of viral epidemics has important implications for public health management. We selected the Avian influenza A-H5N1 example to allow a convenient comparison of Bayesian ancestral state inference with the previous parsimony analysis; on the other hand, statistical analysis of the rabies migration in Africa up to this point has been largely unexplored. Both zoonoses represent a clear threat to human health. The frequent transmission of A-H5N1 from poultry or wild birds to humans suggest that the virus could emerge as or contribute genetically to the next human flu pandemic. Although the lack of a human-to-human transmission mechanism means that rabies will not emerge as a purely human disease, rabies infection causes a fatal neurological disease and at least 55,000 people die from this disease every year, mainly in the developing world [40] . A Bayesian statistical approach presents many advantages over parsimony inference of ancestral states. First, MCMC offers a computational technique to integrate over an unknown phylogeny and unknown migration process as the former is not directly observable in nature and the latter is poorly understood. Accommodating this lack of knowledge protects against potentially severe bias, but can reduce the power to make inferences; phylogeographic analyses are no exception to this. One can regard this uncertainty itself as a 'mixed blessing' because whilst it can hamper drawing definitive conclusions [13] , it protect us from making overstated conclusions. For example, parsimony analysis of the influenza data establishes an epidemiological link between Guangdong and Indonesia [22] . Bayesian inference does not confirm this conclusion and phylogenetic analysis of more recently obtained sequence data now identifies the progenitors of Indonesian strains in the Chinese province of Hunan [26] , a site which our shared analysis also identified as a possible source. Further, unlike parsimony, likelihood-based probabilistic methods consider branch lengths in ancestral reconstructions. The impact of the tree depth on root state reconstructions for the A-H5N1 genome segments clearly illustrates the importance of branch lengths. Moreover, probabilistic methods allow for estimating the relative posterior probability of each location state at any position along the phylogeny; this ability is indispensable in a hypothesis testing framework. As introduced by [19] in a different setting, we also demonstrate how phylogeographic parameters can be estimated from different genomic segments without assuming the same evolutionary history. H5N1 reassortment, however, will have not have fully unlinked HA and NA evolutionary histories and partially shared ancestry may lead to overstated credibility in some aspects of the phylogeographic inference. Bayesian inference also proffers particular benefits within the class of likelihood-based methods, for example, by allowing for straightforward approaches to control model complexity. BSSVS naturally provides a BF test to identify significant non-zero migration rates. Further prior specification easily incorporates geographical detail of the sequence data. Although distanceinformed priors appear to have little impact on the phylogeographic analyses presented here, both BSSVS and informed priors furnish new opportunities for hypothesis testing when comparing competing prior scenarios of the diffusion process. Examples include ''gravity models'' [41] in which infinitesimal rates become functions of the host population-sizes at the end-point locations and a priori structurally-fixed graphs [19] . Finally, it has been recognized that an MCMC-based Bayesian framework is wellsuited to bring together information of different kinds [42] . The BEAST software, which has a strong focus on calibrated phylogenies and genealogies, elegantly illustrates this by offering a large number of complementary evolutionary models including substitution models, demographic and relaxed clock models that can be combined into a full probabilistic model [43] . By adding spatial reconstruction to this arsenal of evolutionary models, the full probabilistic inference now brings us much closer to biogeographical history reconstruction from genetic data. Our primary motivation for exploiting BSSVS to select among all possible migration graphs is to elucidate the limited number of epidemiological links that appropriately explain the viral diffusion process. This parsimonious set both informs major modes of migration and reduces the high statistical variance that burdens estimation of all pairwise transition rates. Following this argument, less uncertainty on node state reconstructions would be expected when focusing on a parsimonious parameterization of the instantaneous rate matrix. The A-H5N1 analysis indeed indicates lower uncertainty of root state reconstructions. However, for some other internal nodes, we note the opposite behavior. We attribute this to the reversibility assumption in the rate matrix. Selection of reversible rates by BSSVS imposes more balanced transitions in the phylogeny among locations that could have unidirectional links in reality. Therefore, work is in progress to develop non-reversible models that may better fit a spatially expanding epidemic like A-H5N1 or recurring epidemic influenza emergence through sourcesink dynamics [2] . Considerable technical hurdles remain to incorporate BSSVS procedures under such models. Because BSSVS places non-negligible probability on structural zeros in the rate matrix, we can not guarantee that all resulting rate matrices are diagonalizable, challenging stable computation. Bearing in mind the reversibility assumption, we pass no judgement on the origin of the A-H5N1 epidemic based on the frequency by which a location is present in well-supported rates, as was previously done in the parsimony analysis [22] . Instead, we focus on node location state reconstructions throughout the phylogeny and their posterior probabilities. Figure 1 suggests that, although Hong Kong and Guangdong both receive posterior support as root location states, the dominant location throughout the phylogeny and hence the more likely hub of diffusion is Guangdong. An inherent assumption of the discrete model of location change is that ancestral viruses necessarily reside at only the sampled locations of the extant viruses. In this respect, it is important to realize that the CTMC process should describe the underlying spatial dynamics more accurately as the sampling density increases. E.g., for A-H5N1 [26] , provide more recently obtained sequence data across a larger set of geographic locations; the data could inform further pathways seeding remote localities that remain elusive in our present analysis. In addition to tackling the reversibility assumption, it may also prove necessary to relax constant diffusion rates through time to realistically model phylogeographic processes in many situations. Covarion-like models [44] and allowing different diffusion matrices across different time intervals in the phylogeny may help achieve these aims. Our rabies phylogeographic analysis confirms a longstanding presence of this viral lineage in West Africa [28] . The virus appears to have a constant population size for about 150 years during which, extrapolating from the more recent spatial dynamics, diffusion occurs continuously with no particular directionality ( Figure 6A ). These continuous dynamics explain why we can not achieve precise root state location inference based only on samples from the last 20 years. In the light of the constant population dynamics, however, the location of the MRCA may be epidemiologically irrelevant as the location probably does not necessarily represent the ultimate source of the rabies endemic. We note that our analysis does not include all currently available strains originating from Chad, which may add to the weak East-to-West dispersal signal revealed by a recent parsimony analysis [39] . Our analysis confirms the model proposed for dog RABV in general; that is, of a series of spatially distinct clusters that experience relatively little contact among them [28, 39] . By providing a time scale for the seeding of these spatial clusters, we again demonstrate a clear advantage of the Bayesian inference over parsimony analysis. The ability to draw migrations over time also promotes a more precise dissection of local and temporal RABV movement on smaller geographical scales. After migration, the virus appears to establish local populations maintaining the viral lineage for at least a limited amount of time. These dynamics are reminiscent of a metapopulation model with continuous turnover of locally established viral populations. A long-standing rabies presence in West Africa is not surprising; already recognized since the late 60s, the territory plays a major role in the rabiescanid ecological balance [45] . It remains a remarkable feat that an acute and mainly fatal disease achieves prolonged endemicity. Because disease-induced mortality can rapidly deplete the number of susceptibles in a population, one expects epidemic cycles with oscillatory dynamics to occur. Rabies cycles and traveling waves have been well documented in wildlife across Europe and North America, e.g. [46] , and more recently [47] , demonstrate such cycles in African dogs. Because their periodicity is notably shorter than expected from epidemiological models, the authors argue that intervention responses also impact the epidemic cycles [47] . Importantly, there is also a remarkable phase synchrony in rabies outbreaks across southern and eastern Africa, most pronounced for distances up to 1,000 km [47] . For oscillating systems in particular, it is well known that dispersal can generate population synchrony [48] . Because previous studies illustrate that even limited amounts of relatively local dispersal can generate synchrony in cyclical dynamics over large spatial scales [48] , and that the resulting synchrony tends to decline as distance increases and varies through time [49] , [47] argue that dispersal could enforce synchrony in dog rabies epidemics across different countries. Our analysis clearly reveals rabies dispersal as a continuous dynamic process that could indeed be essential in maintaining epidemic cycles. As argued by [39] , however, the rate of dispersal is probably not sufficiently high to explain the short epidemic cycles as suggested by [47] . Nevertheless, we underscore that sustained and coordinated responses across political boundaries are necessary to control domestic dog rabies in Africa. Many questions in evolutionary biology require a biogeographical perspective on the population under investigation. We hope to have demonstrated that Bayesian phylogeographic framework can contribute significantly to evolutionary hypothesis testing, and, although we have focused on viral phylodynamics, this approach is generally applicable in molecular evolution. Employing geographically-informed priors delivers a first step in incorporating GIS information. Future developments like irreversible CTMC processes may offer even more biological realism. For many spatial scales and problems, geography can naturally be partitioned into a finite number of discrete sites fS k g for k~1, . . . , K. Examples of these situations include individual cities, islands or countries. Starting from the observed data, at the tips of the phylogeny F we record discretized locations X~(X 1 , . . . , X N ), where X i [fS k g pin-points the sampling site of taxon i. Unobserved in the spatial process are the locations of the most common recent ancestor X root drawing from root distribution p root , the times at which the descendent taxa move and amongst which discrete sites these moves occur, a process which ultimately gives rise to X. Conditioning on X root and the unobserved locations realized at each internal node (X Nz1 , . . . , X 2N{2 ) [13, 16] , suggest modeling the instantaneous locations X (t) for taxa along each branch in F as independent continuous-time Markov chains (CTMCs). CTMCs are the same processes one commonly exploits to model sequence character evolution [15, 50] . Although many readers are familiar with CTMCs, we here highlight several chain properties to which we turn later when discussing CTMC modeling limitations. CTMCs are the simplest stochastic processes that emit discrete outcomes as a continuous function of time. The processes are memoryless, in that the probability of transitioning to a new location only depends on the current location and not the past history. A K|K infinitesimal rate matrix L~fl jk g completely characterizes the CTMC process. Rate matrix L contains non-negative off-diagonal entries and all rows sum to 0, yielding a stochastic matrix upon exponentiation. Solving the Chapman-Kolmogorov equation that specifies the behavior of the chain yields the finite-time transition probabilities p jk (t)~Pr X (t)~S k j X (0)~S j À Á . In matrix form, Determining the finite-time transition probabilities involves matrix exponentiation, generally accomplished through an eigen-decomposition of L. Here, we restrict ourselves to infinitesimal rate matrices that yield only real eigen-values and eigen-vectors. Any matrix similar to a symmetric matrix ensures a real eigen-decomposition; consequentially, we formulate where m is an overall rate scalar, S~fs jk g is a K|K symmetric matrix and P~diag(p 1 , . . . , p K ). Infinitesimal rate matrices of this form generate reversible Markov chains, such that p j l jk~pk l kj and p j p jk (t)~p k p kj (t), ð3Þ placing many restrictions on the underlying geographic process. We discuss these limitations and modeling extensions that allow for irreversible chains in the Discussion. In its most general timereversibile (GTR) form, L contains (Kz2)(K{1)=2 free parameters, with P donating K{1 together with mS's K(K{1)=2 off-diagonal entries. Following standard practice, we normalize entries in S such that m measures the expected (with respect to P) number of transitions per unit time t. One illuminating perspective from which to view a CTMC is that of a random walk on a graph G. From this perspective, the possible realizations of the chain fS k g correspond to the vertex set of G. Between the vertices lie edges that record the infinitesimal transition rates. For example, between S j and S k sits l jk . As a continuous-time random walk, a particle, starting at vertex S j at time 0, first waits an Exponential amount of time with rate {l jj and then randomly decides to which neighboring vertix S k to move with probability {l jk =l jj . Now on S k , the process repeats. Neighboring vertices are those for which a single edge connects them. For character evolution, ''complete'' graphs find almost exclusive use, such that edges exist between all pairs of vertices. At a minimum, however, the graph must remain ''connected'', such that it remains possible to walk between any two vertices on G. Bayesian stochastic search variable selection. When GTR models find use modeling nucleotide substitution, most of the K(K{1)~12 possible transitions have non-neglible probability of occurring and are observed over the evolutionary history. Such is unlikely to be the case for geographical locations; given that there may be many sites and each taxon only has one location (the equivalent of just one single alignment site), we expect most transitions to rarely, if ever, occur. Consequentially, we suspect a priori that many infinitesimal rates are zero. From a statistical perspective, so many degrees of freedom fit to the limited data lead to extremely high variance estimates. These poor estimates arise not only for L, but, more critically, for inference of the unobserved ancestral locations and X root . We circumvent this sparse data problem by invoking BSSVS to select a parsimonious parameterization of L. BSSVS enables us to simultaneously determine which infinitesimal rates are zero depending on the evidence in the data and efficiently infer the ancestral locations. As a beneficial by-product of BSSVS, directly quantifying the evidence about which rates are non-zero furnishes both the most likely migration patterns and the ability to test between competing migratory hypotheses. BSSVS is traditionally applied to model selection problems in a linear regression framework, in which statisticians start with a large number of potential predictors X 1 , . . . , X P and ask which among these associate linearly with an N-dimensional outcome variable Y. For example, the full model becomes Y~½X 1 , . . . , X P bze, where b is a P-dimensional vector of regression coefficients and e is an N-dimensional vector of normally distributed errors with mean 0. When b p for p~1, . . . , P differs significantly from 0, X p helps predict Y, otherwise X p contributes little additional information and warrants removal from the model via forcing b p~0 . Given potentially high correlation between the predictors, deterministic model search strategies tend not to find the optimal set of predictors unless one explores all possible subsets. This exploration is generally computationally impractical as there exist 2 P such subsets and completely fails for PwN. Recent work in BSSVS [51, 52] efficiently performs the exploration in two steps. In the first step, the approach augments the model state-space with a set of P binary indicator variables d~(d 1 , . . . , d P ) and imposes a prior p b ð Þ on the regression coefficients that has expectation 0 and variance proportional to a P|P diagonal matrix with its entries equal to d. If d p~0 , then the prior variance on b p shrinks to 0 and enforces b p~0 in the posterior. In the second step, MCMC explores the joint space of (d, b) simultaneously. To map BSSVS into the phylogeography setting, we consider selection among the 2 random graphs in which each of the K(K{1)=2 edges either exists or does not exist in G. Let d jk be the binary indicator that an edge exists connecting S j and S k . An equivalent parameterization specifies that l jk~0 when d jk~0 and l jk w0 otherwise. So, rate matrix L plays an analogous role to the regression coefficients in BSSVS. An important difference is that l jk [ (0, ?) while b k [ ({?, ?), mandating alternative prior formulations. Prior specification. To specify a prior distribution over d~fd jk g, we assume that each indicator acts a priori as an independent Bernoulli random variable (RV) with small success probability x. The sum of independent Bernoulli RVs yields a Binomial distribution over their sum W~X jvk d jk . In the limit that x%K(K{1)=2, this prior conveniently collapses to where g~x|K(K{1)=2 is the prior expected number of edges in graph G. We entertain two prior choices for P. Diagonal vector (p 1 , . . . , p K ) is the stationary distribution for the CTMC when all edges are included in the graph G. For this complete graph, as the length of the random walk t??, one expects that X (t)~S k with probability p k . One natural choice says that there exists no spatial preference over the long-run and fixes p k~1 =K for all k. However, sites may expound spatial preference over the long-run; for example, such preference can relate to known or unknown quantities such as population-size or geographic size of the site. In these situations, we estimate P simultaneously with the rest of the model by imposing the flat prior (p 1 , . . . , p K )*Dirichlet(1, . . . , 1). Also non-informatively for small values, we take m*Exponential (1) . To complete the CTMC specification, we assume that all unnormalized rates in S are a priori independent and Gammadistributed with prior expectation m jk and variance d jk |v jk , following in the vein of Bayesian SSVS. However, little previous work on prior formulation helps inform our choices of m jk and v jk . This represents a critically important area of research. A common, yet arbitrary choice in the Bayesian phylogenetic literature assumes that rates draw from Exponential distributions, forcing m jk~vjk . We follow this practice in light of there being no obvious way to elicit information on the variance of these rates. Finally, we explore two choices for setting the means. The first assumes no preference over rates, setting all m jk~C , where C is an arbitrary constant; as, after normalization, only ratios of infinitesimal rates participate in the data likelihood, the actual value of C has no influence on the likelihood. The second is informed by the geographical distance between sites. prior. Considerable additional information exists about the sites fS k g and remains unused. Most notably, the geographic distances d jk between (the centroids) of sites is readily measurable. A priori we may believe that more distantly separated sites have the smallest infinitesimal migration rates, yielding Other information is also surely helpful and application-specific. One example involving human hosts quantifies the availability of motorized transportation, such as air flights, between sites. We explore the utility BSSVS and distance-informed priors in our phylogeographic models. Bayes factor test of significant diffusion rates. The Bayes factor (BF) for a particular rate k contributing to the migration graph is the posterior odds that rate k is non-zero divided by the equivalent prior odds, where p k is the posterior probability that rate k is non-zero, in this case the posterior expectation of indicator d k . Since we employ a truncated Poisson prior with mean g~log 2, that assigns 50% prior probability on the minimal rate configuration (K{1), the prior probability q k reduces to We consider rates yielding a BFw3 as well supported diffusion rates constituting the migration graph. A strength of the Bayesian approach we exploit in this paper is the ability to integrate together into a joint model of spatial locations X and aligned molecular sequence data Y~(Y 1 , . . . , Y N ) collected from the N taxa. The joint model affords a natural way to incorporate uncertainty about the unobserved phylogeny F and the character substitution process giving rise to Y. We take a standard statistical phylogenetic approach and assume that a separate CTMC characterized by w generates Y. While we discuss specific choices about this process in the Results sections, we do assume that the sequence and location CTMCs are independent given F, enabling us to write the joint model posterior distribution as Likelihoods Pr X jF, L ð Þ and Pr Yj F, w ð Þ follow directly from Felsenstein's pruning algorithm [15] , efficiently integrating over all possible locations and sequences at the root and internal nodes in F. We approximate the joint posterior (8) and its marginalizations using MCMC implemented in the software package BEAST [43] . We employ standard transition kernels over the parameter spaces of F and w. To sample realizations of L, we consider random-walk operators on the continuous portions and a specialized ''bit-flip'' operator on the Bernoulli rate indicators d jk . [53] discuss this transition kernel further. Finally, in many situations, inference on the posterior distribution of the root and internal node states is of paramount interest. We implement a pre-order, tree-traversal algorithm in BEAST that allows researchers to generate realizations of the root and internal node states following [20] and produce posterior summaries. Importantly, this procedure is not limited to phylogeographic models, making general ancestral state reconstruction available in BEAST for the first time. Summarizing posterior location uncertainty. An important statistical question asks to what extent the data inform our inference when fitting different phylogeographic models. A model of low statistical power makes poor use of the information in the data, while a successful model exploits this information to generate posterior distributions that are maximally different from prior beliefs. One primary outcome of a Bayesian phylogeographic study is the marginal posterior distribution of the root location Pr X root jX, Y ð Þ . We calculate the Kullback-Leibler (KL) divergence [54] from the root location prior P to summarize this information gain, where 0| log 0~0. When the posterior and prior distributions are equal, d KL~0 . In the examples in this paper, we fix p k~1 =K and d KL achieves its maximum log K when the posterior places all estimable mass on a single location. From this perspective, log K{d KL plays the role of a measure of dispersion [55] or uncertainty. As a simple numerical summary, we also use d KL to explore the utility of BSSVS and distance-informed priors on drawing inference from phylogeographic models. Larger values signify that the model extracts more information from the data. To calculate KL divergence, we employ a uniform discrete distribution as reference distribution. Association index. Following existing phylogeographic approaches, we finally score the degree of spatial admixture using a modified association index (AI) [35] . For a given phylogeny F and tip locations X, we obtain the association value d AI by summing over each internal node n, where c n counts the number of sampled locations descendent to n and f n is the frequency of the highest frequency location amongst these descendents. Similar to [10] , we report the posterior distributions Pr d AI jX, Y ð Þ and the AI compares these distributions to those obtained by random permutation of the tip locations X. Deviation from these permuted distributions reflected in low AI values suggests phylogeographic structure whereas AI values close to 1 suggest spatial admixture. Visualizing phylogeographic diffusion. To summarize the posterior distribution of ancestral location states, we annotate nodes in the MCC tree with the modal location state for each node using TreeAnnotator, and visualize this tree using FigTree (available at http://tree.bio.ed.ac.uk/software). To provide a spatial projection, we convert the tree into a keyhole markup language (KML) file suitable for viewing with Google Earth (http://earth.google.com). We introduce the temporal information on the marked-up tree using the TimeSpan KML-function to animate viral dispersal over the time. Example KML files for the Avian Influenza A HA and NA genes are included as supplementary files and software to convert annotated trees to KML is available from the authors on request. Dataset S1 KML file for H5N1 diffusion over time as inferred from HA Found at: doi: 10

The Model Repository of the Models of Infectious Disease Agent Study

The model repository (MREP) is a relational database management system (RDBMS) developed under the auspices of models of infectious disease agent study (MIDAS). The purpose of the MREP is to organize and catalog the models, results, and suggestions for using the MIDAS and to store them in a way to allow users to run models from an access-controlled disease MREP. The MREP contains source and object code of disease models developed by infectious disease modelers and tested in a production environment. Different versions of models used to describe various aspects of the same disease are housed in the repository. Models are linked to their developers and different versions of the codes are tied to Subversion, a version control tool. An additional element of the MREP will be to house, manage, and control access to a disease model results warehouse, which consists of output generated by the models contained in the MREP. The result tables and files are linked to the version of the model and the input parameters that collectively generated the results. The result tables are warehoused in a relational database that permits them to be easily identified, categorized, and downloaded.

T HIS PAPER describes the model repository (MREP) of models of infectious disease agent study (MIDAS), which is a tool developed to organize and catalog the models, results, and suggestions for using results of the MIDAS and store them in a relational database for future use and reference. The database is a repository of epidemiological-based infectious disease models and their derivatives (e.g., inputs, parameters, and outputs). These products will be stored to allow easy retrieval via a query method. This section gives a brief explanation of the MIDAS and the rationale for the MREP, and identifies other related tools in the literature. Section II presents the architecture and design of the MREP. The MREP's key design elements include the ability to link all related model components together and to include a version control tool as part of the relational data base management system (RDBMS) data model. Section III describes the content of the MREP, including the information about the MIDAS models and the studies they comprise. Section IV summarizes key features of the MREP and discusses future enhancements. The MIDAS is a research partnership between the National Institutes of Health (NIH) and the scientific community to develop computational models for policymakers, public health workers, and researchers to help them make better-informed decisions about emerging infectious diseases-both man made and naturally occurring. The MIDAS researchers are working to develop models that can assist the public health community understand how best to respond during outbreaks and epidemics. The MIDAS consists of seven research groups and one centralized informatics group. 1) MIDAS Objectives: The MIDAS Research Groups develop epidemiological models that represent host-pathogen relationships, disease epidemiology and forecasting systems, and pandemic response strategies. They also focus on informationdriven research rather than hypothesis-driven investigations. The MIDAS model developers use real or simulated data available through the MIDAS Web site. As a collaborative network of scientists, the MIDAS investigates using computational and mathematical models that will prepare the nation to respond to outbreaks of infectious diseases by providing policymakers and public health officials with reliable and timely information that can be used to prepare for infectious disease outbreaks. 2) Epidemiological Models: Epidemic models represent powerful tools for gaining insight into how the dynamics of an epidemic are affected by interventions. In order to understand and control the spread of pathogens, it is essential to establish some of the key parameters associated with disease transmission. Fundamental to the dynamics of an epidemic are two quantities: the basic reproduction number (R0) and the generation time (Tg) of the pathogen [1] . The R0 is the average number of secondary cases produced by each primary case at the start of an epidemic in a previously unaffected population. The Tg is the average time between infection of index cases and the secondary cases they produce. The R0 is a measure of the transmissibility of the strain in the population, and largely determines the proportion of the population that will be infected in a pandemic. The ratio R0/Tg is a measure of an epidemic's rate of growth. Many epidemiological models are based on a compartmental, Sampling Importance Resampling (SIR) framework; the host population is partitioned into those that are susceptible, infected, or immune to a particular pathogen. These models assume that the rate at which new infections are acquired is proportional to the number of encounters between susceptible and infected individuals, and leads to an effective reproductive ratio that depends on a threshold density of susceptibles. Thus, the models depend not only on parameters intrinsic to the disease such as latent and infectious periods, but also on contacts between infectious and susceptible hosts. Historically, structuring the population at risk into compartments permits subpopulations of varying risks to be represented. Compartmental models of this kind implicitly assume that the host population is well mixed, such that the probability of infection is equal for all. However, social network structures are clearly not always well mixed, and the complexities of host interactions may have profound implications for the interpretation of epidemiological models and clinical data. To overcome the simplifying assumption of classical transmission models, a new type of model is gaining recognition. Agent-based models (ABMs) represent an important new approach for describing interacting heterogeneous agents. The heterogeneous aspect of agents enables more sophisticated and complex environments to be described by ABMs approaches. Also, it is typical to introduce a geospatial dimension into the framework so that both time and geographical patterns are represented. Thus, every ABM run identifies infected persons where they live and work in general, and their movements within groups (referred to as social networks) that influence disease spread. ABMs have been used to describe phenomena from social systems to immune systems, both of which are distributed collections of interacting entities (agents) that function without a leader. Simple agents interact locally according to simple rules of behavior, responding in appropriate ways to environmental cues and not necessarily striving to achieve an overall goal. An ABM consists of a set of agents that encapsulate the behaviors of the individuals that make up the system, and model execution consists of emulating these behaviors [2] . 3) Models of Influenza Transmission: The major focus of the MIDAS research partnership has been pandemic influenza. In response to the MIDAS mission (see http://grants.nih.gov/ grants/guide/notice-files/NOT-GM-06-106.html), a number of large ABMs describing influenza transmission were developed. The main purpose of the models was to examine possible intervention strategies that would protect the general public from morbidity and mortality should an influenza pandemic strike. One of the initial objectives in all disease transmission studies is to determine the basic R0 value. The goal of intervention is to reduce R0 below the self-sustaining threshold of R0 = 1. The R0 of a future newly emergent influenza strain is unknown, but estimates for previous pandemics are available. For example, an estimate of 1.89 was obtained for the pandemic of 1968 in Hong Kong [3] , and the pandemic of 1957 in Great Britain (GB) was estimated to be between 1.5-1.7 [4] . Also, the reproductive number of the first wave of the 1918 pandemic in the United States was estimated as 2-3 [5] . There is historical evidence that these three influenza pandemics were explosive. The results reported by the MIDAS also suggest that the pandemics can be controlled to some degree. By comparison, childhood diseases such as rubella, pertussis, and measles have R0 values in most populations in the range 7-15 [6] , and consequently, are much less controllable. To estimate the effect of various interventions on the spread of pandemic influenza, the effect of specific interventions on transmission rates needs to be quantified. Generally, estimates of the proportion of infections that occur in the various social contexts such as households, schools, workplaces, and communities have been reported. Then, estimates of relative effect of intervention measures on transmission in each context will need to be established. It would also be useful to know how implementation of a particular measure might disrupt contact patterns in other social contexts. For example, we would like to know the extent to which household and community contacts are increased when schools are closed. There is scant information on the proportion of transmission that occurs in different social contexts. The best data available only allow the proportion of transmission in households to be quantified [7] , [8] . Therefore, while models can give some insight into the likely benefit of single or combined interventions, that insight is somewhat limited by this lack of data. To some extent, the results depend on the assumptions made by the modelers about transmission in the different contexts. The degree of uncertainty does depend on the specific control measures. Modelers can arguably better project the possible effects of antiviral and vaccine use, case isolation, and household quarantine than the effects of school closure, mask use, banning of mass-gatherings, or nonspecific social distancing measures. The most recent MIDAS study examined the effectiveness of a set of proposed targeted, layered containment strategies that combined a number of transmission interventions currently available to public health planners in the United States [9] . These include nonpharmaceutical social distancing measures and antiviral treatment and prophylaxis. All three MIDAS models examined the same set of interventions, although each of the three implemented the interventions using different approaches. The three sets of models also examined the sensitivity of the effectiveness of the intervention combinations to thresholds for triggering the interventions, levels of case ascertainment and compliance, and the transmissibility of the circulating pandemic strain. The intervention scenarios examined reflect those being considered now by the United States Homeland Security Council and Department of Health and Human Services (DHHS). An important goal of MIDAS was creating a repository for storing and managing the computerized models, model results, model parameters, and the specifications used to develop the models. The MREP was developed to house the code being developed by the MIDAS research groups into a professional, organized, and controlled environment. The guiding premise is that responding to an emergency requires a process that can be activated in a controlled, orchestrated manner; this premise also allows legacy code to be identified and past results reproduced. The MREP needs to fit into and support the emergency response process. To support ease of access, the MREP is implemented using relational database management technology and a Webbased interface. In addition, there are other compelling reasons for developing the MREP, including promoting quality assurance and enhancing productivity during day-to-day activities. Losing track of the exact versions of the model that generated specific model results is easy. By imposing a structure that links together all model components, houses these connected components in a centralized database, and annotates those components, we can preserve and reuse essential linkages. In summary, the MREP provides the following capabilities to the MIDAS: first, a process for responding to an emergency event. Characteristics of MIDAS models across all modeling groups can be quickly identified and linked to relevant documentation. Second, a quality assurance mechanism for registering models into the MREP. For a model to be part of the MREP, it must: 1) have a name; 2) be linked to a readily available description; 3) be linked to a contact; 4) identify the date and time of creation; and 5) specify the terms of distribution. The registration system is consistent with the process identified by Le Novère et al. [10] . Third, a tracking mechanism that catalogs, locates, and identifies the different versions of the models involved in different experiments that comprise MIDAS studies. Fourth, the MREP has productivity features that allow modelers to efficiently locate previous model versions and reuse the code in new models. Fifth, the MREP maintains inventories of work developed by the research groups in a locatable form. 1) MREP Scope: A repository is a collection of resources that can be accessed to retrieve information. Repositories often consist of several databases tied together by a common search engine. The MREP consists of a collection of infectious disease models and pertinent information about those models including: model specifications, inputs (static), parameters, results, source code, object code, scripts for compiling the object code, scripts for executing the model, user manuals, and other model documentation including published manuscripts. Each element in the MREP is part of a relational database. This permits them to be linked to each other so that the inputs to a specific model and the results generated by running that model using those inputs are connected. Therefore, one major feature of the MREP is tracking and connecting all of the components of a model, which allows researchers to revisit previously generated results. A second important feature is that the linked components can be unified as part of the information retrieval process. This enables a query engine to present information to a user without having to browse irrelevant information. In summary, the MREP is important to the MIDAS scientific research environment because it links specific model runs with the explicit model code and input data with the corresponding output data. It also provides a link to the specifications that guided model development. If researchers wish to identify how specific interventions were implemented and rerun a particular model or analyze it in any way after the run is completed, the MREP provides all the information necessary. The MIDAS MREP appears to be unique among computerized epidemiological model repositories. However, a number of quantitative biology-based model repositories exist and provide a useful point of comparison. We performed a literature search and identified other model repositories similar in some degree to the MREP. Specifically, we sought applications that maintained a database where models (model code, data inputs, and outputs) were shared by users in a controlled environment (i.e., an environment that connects model components). We identified a number of existing relevant model repositories and we contrast five of those that catalog and share information about a specific set of models and model runs. A number of online archives and/or data repositories from a number of nonmodeling applications were also identified. The repositories described later offer information about a specific class of models to their user communities, and, in this context, they are similar in their capabilities to the MREP. None of them, however, catalogs infectious disease models and none of them attempts to maintain a version-controlled environment that offers code to the users from a stable, documented environment provided by a version-managed system. Nevertheless, all of the examples offer their users annotated models, and all are concerned about providing reliable programs with usable documentation. 1) Biomodels Database: The BioModels.net project describes itself as an international effort to: 1) define standards for model curation; 2) define vocabularies for annotating models with connections to biological data resources; and 3) provide a free, centralized, publicly accessible database of annotated computational models in Systems Biology Markup Language (SBML) and other structured formats [11] (for detail, see http:// www.ebi.ac.uk/biomodels/). The database component of BioModels.net is especially designed for working with annotated computational models: each model is carefully reviewed and augmented by human annotators on the BioModels.net team to add metadata linking the model elements to other biological databases and resources. The BioModels database at the European Bioinformatics Institute (EBI) system is a true database, featuring browsing, crossreferencing, searching, and facilities for visualization, exporting models in different formats, and remote application programming interface (API) access. The BioModels Database is a data resource that allows biologists to store, search, and retrieve published mathematical models of biological content. Models present in the BioModels Database are peer reviewed, annotated, and linked to relevant data resources, such as publications, databases of compounds and pathways, controlled vocabularies, and similar items. All models use SBML as their standard form of representation. The premise of this tool is that researchers must be able to exchange and share their results. The development and broad acceptance of common model representation formats such as SBML is a crucial step in that direction, allowing researchers to exchange and build upon each other's work with greater ease and accuracy. To make assembling useful collections of quantitative models of biological phenomena easier, establishing standards for the vocabularies used in model annotations as well as criteria for minimum quality levels of those models is crucial. The BioModels.net project aims to bring together a community of interested researchers to address these issues. 2) CellML-A Biological Model-Based Repository: This application is similar to the BioModels Database application and uses a markup language called CellML developed specifically for describing biological processes contained in CellML [12] . The repository is a Web site that stores and exchanges computer-based mathematical models. This site allows scientists to share models described by the CellML markup language. It also enables them to reuse components from one model in another, thus accelerating model building (for detail, see http://www.cellml.org/examples/repository/index.html). The CellML language is an open standard based on the Extensible Markup Language (XML) and is designed to describe biological models. The CellML also includes information about model structure (how the parts of a model are related to one another), model mathematics (equations that describe the underlying biological processes), and metadata (additional information about the model that allows scientists to search for specific models or model components in a database or other repository). 3 The premise behind this MREP is taken from the AnyBody Web site [13] . The site describes how the cost of musculo-skeletal injuries is rapidly increasing, while fundamental understanding of the mechanical functions of the body is increasing at a dramatic rate. However, it notes that there are still many unknown questions and problems researchers are addressing. For example, the AnyBody Web site indicates that the number of ergonomic-based injuries caused by excessive use of the computer mouse is exploding, yet the actual cause of many of these injuries remains a mystery. Policymakers and others, therefore, find it difficult to issue guidelines to reduce the problems. Also, research into human locomotion has historically been relegated to experimental studies (for detail, see http://anybody.auc. dk/). The stated purpose of the AnyBody project is to develop mechanical models of different elements of the human body, and then, perform detailed studies of the behavior of these models. Typical model examples include the analysis and optimization of tools and workplace layout, and designing sports equipment and hand tools for maximum efficiency. A unique software system called the AnyBody Modeling System was developed to conduct necessary research into causes and treatments of musculo-skeletal injuries. There are four major aims of the AnyBody project. The first is to develop methods for analyzing movement strategies and tendon, muscle, and joint forces in humans performing specific manual tasks. The second is to investigate what numerical simulation can teach us about the function of the human body. The third is to use the analysis for ergonomic optimization of tools, workplaces, and man/machine interfaces. The fourth is to provide an MREP to enable interested researchers to share the models. Overall, AnyBody identifies useful information and models that can be shared by researchers interested in studying musculo-skeletal injuries. This repository contains software for manipulating and learning probabilistic-logical models [14] . The aim is to construct an MREP that will allow dissemination of software for probabilistic-logical models and facilitate comparisons among competing approaches (for detail, see http://www.informatik.uni-freiburg.de/∼kersting/plmr/). The site notes that probability models are important methods for representing uncertainty, and mentions that various probabilistic frameworks include Bayesian networks, hidden Markov models, and stochastic context-free languages, along with other popular tools for describing appropriate scenarios exhibiting uncertainty. It notes that these types of models have been applied to problems in diagnosis, forecasting, automated vision, sensor fusion, manufacturing control, speech recognition, and computational biology. However, traditional approaches have a major drawback-they have a rigid structure, and therefore, lack of versatility in representing complex models. To overcome these limitations, the site states that various researchers have recently proposed logical extensions of classical probabilistic models incorporating the notions of objects and object interconnections. This repository consists of a database of software, documentation, and links to developers. The models are not specifically designed to be general purpose; rather, they are instructional in nature and represent a basis for discussion. The potential scope for the models in this repository is huge, with many different methods for defining and describing probabilistic-logical models. The cancer intervention and surveillance modeling network (CISNET) is a consortium of investigators sponsored by the National Cancer Institute "whose focus is using modeling to improve our understanding of the impact of cancer control interventions (e.g., prevention, screening treatment, etc.) on population trends in incidence and mortality" [15] . They use models to project future trends and help determine optimal cancer control strategies (for detail, see http://cisnet.cancer.gov/about/). The CISNET also describes using a comparative modeling approach in which each modeler focuses on an individual area. However, whenever possible, they develop a common "base" question that allows comparison across all models using a set of common population inputs. Then, they develop a common set of intermediate and final outputs. To aid in this process of model description and comparison, the CISNET has developed the Model Profiler, an Internet-based application. Each CISNET team has a private model profile Web site on which it maintains model profile information and controls what parts of the profile are shared with other teams. By using a core documentation format that is the same for each group, the published profile information can be compared among models. The system allows modelers to describe their models, and allows interested readers to read about, compare, and contrast simulation models. The sites described earlier we believed were fairly representative of the repositories available to researchers. Our literature search indicated that other sites are available. SigPath is described by Campagne et al. [16] as an information management system that stores both quantitative information on cellular components and their interactions, and the basic reactions governing those interactions; EcoCyc, which is described as a comprehensive database resource for Escherichia coli by Keseler et al. [17] ; JWS Online, described by Olivier and Snoep as a repository of kinetic models describing biological systems that can be interactively run and interrogated over the Internet [18] ; and the Database of Quantitative Cellular Signaling, which Sivakumaran et al. describe as a repository of models of signaling pathways [19] . A number of model repositories are identified in the literature. Many of these models support computational biology applications, and certainly, by some measures are more sophisticated than the MREP we describe here. One manifestation of sophistication is using a markup language based on XML concepts that imposes a standard method for describing and representing models common to a repository. We assert that ABMs are conceptually less mature than computational biology models, and broader in their scope (i.e., agents can represent genes, proteins, and cis-regulatory elements at one end of the spectrum, and represent people, states, and countries at the other end). We chose not to grapple with the notion of trying to develop a common taxonomy to describe agent-based epidemiology models. Rather, we focused on building a repository that captures ABMs in whatever form they were developed and creates a common set of documentation and annotations, so that the model can be understood (at some level) and reused should the need arise. The MREP is designed to house, manage, and allow users to run infectious disease models from an access-controlled disease MREP. The MREP contains source code of disease models that have been developed by external developers and tested in a production environment. Different versions of models used to describe various aspects of the same disease are housed in the repository. During registration, models are linked to their developers, to a paper or PubMed reference that describes the model, to the name and contact information of model creators, to the date and time of creation, and to the terms of distribution. In addition, a code, available from the versioning software described later, is used to distinguish between different model versions. The annotations captured during the model registration process also identify a model's purpose, specifications, and relevant features. The MREP also houses, manages, and controls access to a disease model results warehouse, which consists of output generated by the models contained in the MREP database. The results, tables, and files will be linked to the version of the model and the input parameters that collectively generated them. They will also be stored in a relational database to permit them to be easily identified, categorized, and downloaded. The MREP includes a version control system (also referred to in the literature as a configuration control system) as one of its core elements. The system manages the source code, documents, graphics, and related files. Version-control software provides a database that is used to keep track of the revisions made to a program by all the programmers and developers involved in it. The MREP uses Subversion as its version control system [20] . Subversion is a free, open-source application that is licensed under the Creative Commons Attribution License (see http://svnbook.red-bean. com/en/1.0/). The MREP also includes a Graphical User Interface (GUI) application that allows the user community to use the MIDAS models more easily. The GUI interfaces with Subversion, the database management system, and the computer environment. The GUI allows the user access to the model source code, provides an interface to input data sets, permits results to be viewed directly or downloaded to a user workstation, and provides a mechanism to submit the models for rerunning. The MREP is comprised of the following major components: 1) a source code repository and version control system; 2) a model documentation tree; 3) a data warehouse (input and output data sets); and 4) an application interface (API) consisting of a database, browser, and graphics user interface components that allow the model user to develop input data sets, run the codes, and browse output results. A standard system development approach was used creating the initial version of the MREP using approximately two fulltime equivalents (FTEs) over a 13-month period. B. Architecture 1) Overview: There were two features that were considered extremely important by the developers to include in the MREP design. These features include: 1) a version control system that provides configuration control over model source code and 2) complete flexibility regarding output formats. The first of these features was a response to an absence of model standards (such as SBML) in the storage of source code. The most common model development languages represented in the MREP are anticipated to be C, C++, Java, and Matlab. To maintain control in a language-free environment, using a version control element was deemed essential. The second feature was the ability to support a common output format. All of the models generate results in different formats, including Concurrent Versions System (CVS), text, and Portable Document Format (PDF), among others. The MREP offers these results to users in the received format. Users can view results directly or by downloading a file and using a viewer that they supply. Fig. 1 represents a high-level logical data model for the MREP. Five primary components comprise the architecture, including the MIDAS Compute Server; a File System; a relational database management system (RDBMS); a Version Control System; and an External Systems Gateway. A description of each is provided later. 2) MIDAS Compute Server: The MIDAS cluster is the central computational resource for the MIDAS research groups and is referenced in Fig. 1 as the cluster. The primary interaction that an analyst using the system will have with the cluster is through interacting with the job queuing system. Additionally, code maintainers/developers will have full access to all of the Linux system's features as needed to make changes to the code. The MIDAS cluster is managed by Cluster Resource's MOAB (see http://www.clusterresources.com/pages/products/ moab-cluster-suite.php) [21] , an advanced cluster scheduler capable of optimizing scheduling and node allocations. MOAB allows site administrators to control job scheduling, priority, and where jobs are run. 3) File System: The file system is part of the MIDAS cluster resource and is an integral part of the cluster system. A second system element is primarily used to archive study results in the MREP warehouse. At present, the archival file system is configured with 2 terabytes (TB) of disk storage that will eventually be expanded to 14 TB. Users interact with the file system to associate simulation run I/O data locations with metadata that will be used to identify and tag achievable simulation results. The configuration control system also interacts with this system. 4) Relational Database System: The database management system platform will be housed by the MIDAS Web portal, which serves as the interface for electronic information exchange for the MIDAS network. The MIDAS portal is accessible to the public, but only registered users can access and provide information to the private section of the portal. The MIDAS portal runs on RTI's Oracle Application Server 10 g (v. 9.0.4.1.0) server farm and uses Oracle technology [22] to manage the information within the repository (see www.oracle.com/appserver/index.html). The database system tags metadata that reference specific simulation run results with the input and output data sets and source code associated with those runs. The database system will allow users to perform keyword searches to identify repository elements assigned to those metadata. a) Data model-hierarchal design: The database design will maintain tables of metadata that describe the following entities: Projects-a collection of studies with a common set of objectives; Studies-a collection of runs that were produced by one or more models; Models-the core code that is designed to describe computer environments that generate the runs; Model versions-a specific instance of a model to handle the production of runs having a specific set of attributes; and Runs-a set of information referred to as results that describe a single realization of a simulated epidemic. Each realization is associated with a unique set of parameter values or alternatively is associated with a repeated set of parameter values. In the last situation, the results are referred to as a run replicate. Each model run produced by a specific model version is housed in the results data warehouse and consists of information that is part of the experiment. Fig. 2 defines the interconnections between the entities. A study is either linked to a published manuscript that defines project aims and study results or to a document that describes an MIDAS study that is part of a larger MIDAS project effort to examine specific hypotheses about disease spread and containment. For example, the DHHS project examined the problem of whether influenza could be stopped at its source (i.e., South East [SE] Asia), and if so, what methods of containment are important. Two studies from this project are part of the MREP: the Emory SE Asia Study and the Imperial SE Asia Study. In a second example, the combined project focus is to address the problem of "what can be done to mitigate pandemic influenza if it gets established in the U.S." The specific objectives of the combined study are to assess the feasibility and effectiveness of different types of interventions strategies. The types of interventions specifically exclude prepandemic vaccines and limit the available quantity of a partially efficacious vaccine, but utilize as much antiviral treatment as required. Two paradigms are the focus of this study: the entire United States and a large city (Chicago). Six studies (and corresponding models) examine this problem. Two of these models simulate transmission in the continental United States, three others represent transmission in the city of Chicago, and the sixth presents the results of a historical review of the 1918 Spanish influenza epidemic. The last study helped justify the design of the intervention strategies. Each model is composed of one or more model versions. Usually, a modeler attempts to describe the complete set of simulations through parameter manipulation. However, in many instances, it is likely that new and novel interventions will not be anticipated by the code developer. In these instances, it is common for the developer to create a different version of the model to handle a subset of the simulated interventions. One of the explicit objectives of the MREP is to track the different model versions and to link them to the results they produce. The run is the lowest unit of analysis of the MREP. Each run constitutes a single replicate of a single set of parameters or a summary run over all replicates having a common set of parameters. b) Query engine: The overall plan is to populate the MREP with models that are initially in accordance with MIDAS goals, which are currently limited to infectious disease models. Eventually, we anticipate including models from a broader perspective; namely, models that are relevant to any pathogen and appropriate transmission environment. This will potentially necessitate using a query tool to readily identify models of interest. For example, the MIDAS ABMs use synthetic population data that describe a specific geographical region. The model results identify individuals that influence disease spread within that region. The query tool enables users to identify a set of model results that pertain to a region of interest. Then, by using the geospatial identifiers that are tagged to individual model results, users can drill down into subregions and neighborhoods that are affected by epidemics. Please note that it is possible to use the search keys to identify the studies that focus on a particular type of intervention strategy, and then, by examining Model Version details, determine the implementation details to decide a target computer to rerun the model. This could be done to replicate earlier results or to begin the process of modifying parameters to assess new interventions. 5) Version Control: The version control system maintains the various model version source codes and executables. Each model version in the MREP is maintained in Subversion, the free, open-source version control system that manages files and directories over time. A tree of files is placed into a central database. This database is similar to an ordinary file server, except that it remembers every change made to files and directories. This allows the user to recover older versions of data, results, and/or code and to examine their change history. Subversion is a distributed application; therefore, it can access its database across networks. This allows people to use Subversion on different computers and fosters collaboration by allowing various people to modify and manage the same set of data from their respective locations. Furthermore, progress can occur more quickly because there is no single conduit through which all modifications must occur. Because the work is versioned, we prevent the possibility of losing that conduit if an incorrect change is made to the data, in that the change can be easily undone. Subversion is a full version control system. The MREP only uses a small subset of Subversion's functionality. The MREP user interface allows registered users to check out model executables that the user may then run on the RTI cluster. Another MREP interface allows the user to browse the source code tree for any model of interest. The MREP Subversion server is hosted on a virtual Linux host using VMWare Server software. 6) External Applications: An important component of Fig. 1 , referred to as External Systems, is a general set of tools that are available outside of the MREP. These tools are used to visualize, process, and analyze results from the MREP. The data from the MREP are served up via a HyperText Markup Language (HTML) portal. These data can be downloaded to the user's workstation or can be visualized directly from the MREP. For example, many of the data files in the MREP are stored as PDF, text, or.xls files, and can be viewed directly from the repository using Adobe Acrobat Reader, a text editor, or Microsoft Excel, respectively. Other files can be downloaded and imported into external systems available to specific users. The results/outputs of production runs will also be housed in the MREP. The model results will only include outputs from registered models. For a model to be registered, it must be placed under version control. When output from the model is used in a paper submitted for publication or otherwise presented publicly, or when the model's code is declared stable (by the developer), the model is a candidate for inclusion in the repository. If it has not been developed under version control, it will be moved to Subversion, tested, and moved to the MREP data warehouse. The model code will be annotated with the following set of metadata: name of model, link to model description, contact information, date of model creation, distribution terms, model specifications, model implementation of those specifications, disease, region of analysis, types of intervention strategies, computer resource requirements, user's manual (link), validation measures (link to supporting manuscript), and compile and/or run scripts. The information about model specification and how those specifications were implemented is particularly important for explaining model differences. For example, if an implementation strategy calls for a reduction of 50% in model contacts, it is plausible to implement the strategy by halving the number of people contacted or alternatively maintaining the same number of persons in the contact network while halving the number of contacts with each person. The model results for each of the implementations could vary significantly. 2) Model Results: Model results are also captured in a model results warehouse and linked to the model that generated them. Each result unit is tagged to a second set of metadata that includes model name, version number, developer, intervention, parameter file, fixed input file, and script used to generate results. 3) Model Inputs: Model inputs are also placed in the repository and linked to the model that uses the inputs and the corresponding results that are generated. Each set of inputs is tagged to a set of metadata defined at the model version level: model name, version number, developer, intervention, and parameter location (including name, type, and range of values). Either a query tool is used to locate repository results or a complete list of models is displayed, and the user selects from the list. These results are then available for download and display. The process proceeds according to the following five steps. In step 1, query keys are specified: the study and/or model and/or results are selected that meet the user-specified search criteria. For example, disease, geographic description, objective of study, model name and version number, and model developer. In step 2, the query tool identifies the model in the repository with the specified attributes and displays the information. At this point, the user can either access the annotations or drill down to lower-level (run-level) linked results (results are displayed and/or downloaded if desired), or identify input data files and scripts that run the model. In step 3, the model identified by the earlier steps can be checked out of Subversion. Please note that all version updates are performed by the MREP administrator and that model development occurs under version control. When a model is modified, debugged, and tested, it can be entered back into Subversion, but only as new model version. The modified model is then entered into the repository. In step 4, the model is loaded into the MREP. This step annotates the model with both descriptive text and model metadata as part of the check-in process. An HTML file is created that connects the model, results, and input data; it also creates directories for source code, object code, inputs, results, and scripts. In step 5, model results are loaded into the model result warehouse and metadata specified at the run level, including model name, version number, developer, result category, replicate, link to parameter list, and location of results. Currently, four projects, 12 studies, five models, six model versions, and 538 runs are loaded into the MREP. The vaccine project is a single study project that uses a single influenza-based model of a medium-sized city in the United States. It was developed by the Emory Group headed by Ira Longini. The study consists of six runs generated from a single epidemiological model of disease spread. The main hypothesis behind the vaccine distribution study is to assess whether targeted antiviral prophylaxis (TAP), taken prophylactically, is effective in containing influenza. The authors conclude that TAP is nearly as effective as vaccinating 80% of the population, and further, that vaccinating 80% of children less than 19 years of age is almost as effective as vaccinating 80% of the entire study population. The Influenza Containment-SE Asia was developed by the MIDAS network and was completed in September 2005. It consists of two studies and two models. Both studies simulated disease transmission in regions that included some part of Thailand. The Imperial SE Asia model was developed by Ferguson et al. [23] and the Emory SE Asia model was developed by Longini et al. [24] . The overall hypothesis of the project was to examine if it is possible to contain Avian Influenza in the place of origin before it becomes a pandemic. The Emory SE Asia study represents a region of rural Thailand and consists of one model and 18 runs. Sixteen of the runs were produced by a single version of the model. However, the developers used a special version (the second) of the model that simulates the impact of geographically targeted antiviral prophylaxis (GTAP) to produce two other runs. All runs are loaded in the MREP. The Imperial model represented all of Thailand plus a perimeter region around its border that extended into its four border countries. Only the baseline (no intervention) run, produced by the Imperial SE Asia model, is loaded into the MREP at this time. The Influenza Containment-United States and Great Britain project examined whether pandemic flu could be mitigated in the United States, assuming containment in SE Asia failed. This project was developed in collaboration with the DHHS and consisted of three studies: the Imperial Assessment study, the Epicast assessment, and the EpiSims assessment. The Imperial Assessment study consisted of two models: one describing disease transmission in the United States and the second describing transmission in GB. The GB model served as a counterpoint for the United States model, suggesting some interesting influences of geography and national patterns of behavior on disease spread. The Epicast DHHS experiment described disease in the United States and examined influences on the spread of disease on a population derived from the U.S. 2000 Census data. The principal investigator of the Imperial assessment study is Ferguson et al. [4] . The MIDAS principal investigator of the Epicast assessment study is Ira Longini [25] . The EpiSims assessment study simulated disease behavior in a mid-size city in the United States and used the same type of containment strategies that were used by the other two U.S. experiments. The EpiSims DHHS experiment consisted of a single model and generated 516 runs, involving a complete factorial design of nine binary variables and a partial design that examined the influence of three more variables. A complete set of 516 runs is loaded in the MREP for this study. The baseline GB run is also loaded into the MREP. Four national U.S. Epicast runs are currently loaded in the MREP. The Combined project is also a U.S.-based study. It represents a refinement of the Influenza Containment-U.S. and GB project. Specifically, it examines a more complete set of social distancing interventions with the goal of determining practical strategies for implementation at the state and local level. This is an ongoing study that consists of six sets of experiments: two U.S. experiments, three Chicago-based experiments, and a historic study that examined certain characteristics of the 1918 influenza pandemic. Each study is associated with a single model with a single version per model. The objective of the combined study is to simulate a set of scenarios at the city and the national level. The scenarios are designed to address specific concerns by various federal agencies. The entire study amounts to about 150 distinct scenarios, each with multiple (thousands of) runs. This is an ongoing study that will be loaded into the MREP. Currently, only the principle results are loaded into the MREP; the code that generated the results has not yet been loaded. A set of economic runs has also been generated as part of this study. The economic models involve assessments of cost-effective interventions with respect to containing disease spread. The MREP represents a one-of-a-kind resource for housing and cataloging infectious disease models. The strength of the MREP's design is its data model hierarchy that accurately portrays the stages of a study and its derivatives, at least from the MIDAS perspective. This data model begins with a high-level study as represented by an overall set of study objectives and design specifications and culminates at a low level with a set of runs (results) that contribute to assessing those objectives. In between the studies and the runs linked to those studies are the experiments that represent the different and independent points of view characterized by different research groups and the models those groups used to generate their results. The final data model design element is based on the assumption that containment strategies are not always accommodated within a single model; a change in the model code is sometimes the favored approach for representing simulated behavior differences, that is, the containment strategy responses. A second important element of the MREP is its use of a recognized code versioning application to house different model versions. Using Subversion fosters a highly controlled environment that promotes quality assurance/quality control (QA/QC) activities through a rigorous association among different model versions-their inputs and the results the input data set and model version generate. A final feature of the MREP is the absence of standards associated with including models results. The MREP permits Statistical Analysis Software (SAS), text, Excel, images, and virtually any output format for which a reader exists. Results can be left in the source environment or downloaded onto a user's workstation. In the future, we plan to add an epidemiology-based ontology and develop a separate query tool. This is in part a substitute for a markup language. This tool will identify all models within the MREP (using the ontology information) that reference specific model parameters and connect those parameters to the values assigned by the study developers. This will provide a convenient mechanism for identifying and comparing assumptions across models within the MREP.

Escape from Autologous Neutralizing Antibodies in Acute/Early Subtype C HIV-1 Infection Requires Multiple Pathways

One aim for an HIV vaccine is to elicit neutralizing antibodies (Nab) that can limit replication of genetically diverse viruses and prevent establishment of a new infection. Thus, identifying the strengths and weaknesses of Nab during the early stages of natural infection could prove useful in achieving this goal. Here we demonstrate that viral escape readily occurred despite the development of high titer autologous Nab in two subjects with acute/early subtype C infection. To provide a detailed portrayal of the escape pathways, Nab resistant variants identified at multiple time points were used to create a series of envelope (Env) glycoprotein chimeras and mutants within the background of a corresponding newly transmitted Env. In one subject, Nab escape was driven predominantly by changes in the region of gp120 that extends from the beginning of the V3 domain to the end of the V5 domain (V3V5). However, Nab escape pathways in this subject oscillated and at times required cooperation between V1V2 and the gp41 ectodomain. In the second subject, escape was driven by changes in V1V2. This V1V2-dependent escape pathway was retained over time, and its utility was reflected in the virus's ability to escape from two distinct monoclonal antibodies (Mabs) derived from this same patient via introduction of a single potential N-linked glycosylation site in V2. Spatial representation of the sequence changes in gp120 suggested that selective pressure acted upon the same regions of Env in these two subjects, even though the Env domains that drove escape were different. Together the findings argue that a single mutational pathway is not sufficient to confer escape in early subtype C HIV-1 infection, and support a model in which multiple strategies, including potential glycan shifts, direct alteration of an epitope sequence, and cooperative Env domain conformational masking, are used to evade neutralization.

The current AIDS pandemic is the result of genetically diverse viral subtypes and circulating recombinant forms (CRFs) of HIV-1 group M, of which subtypes A, C, and D account for a majority of infections worldwide [1, 2] . A key source of this genetic diversity is the viral env gene, which encodes the envelope (Env) glycoproteins, gp120 and gp41 (reviewed in [3] ). On the virion, monomers of non-covalently associated gp120 and gp41 subunits trimerize to form 'spikes', and together these facilitate entry into a target cell. Env has a complex conformation and undergoes substantial rearrangements in both subunits upon gp120 binding to CD4 and coreceptor [4, 5, 6] . Env also contains the principal targets for neutralizing antibodies (Nab) [7, 8] , and epitopes are targeted in both Env subunits [9] . However, many potential neutralization targets are transiently or not exposed on the trimeric form of virion-associated Env, including the V3 domain, CD4-induced epitopes, and the CD4 binding site [10, 11, 12] . Despite these limitations, most HIV-1 infected patients develop robust Nab responses against their autologous virus, particularly those infected with subtype C [13, 14, 15, 16, 17, 18, 19] . To confer potent and broad neutralization, it is expected that an epitope will need to possess at least four properties: (i) exposure on the virion-associated native Env trimer, (ii) conservation across diverse HIV-1 variants, (iii) immunogenicity, and (iv) lack of autoreactivity. To date, there are no epitopes that meet these criteria. However, our knowledge of the epitopes that are recognized by Nab during natural infection with diverse HIV-1 is somewhat limited. It is not known which or how many epitopes are targeted by the initial autologous Nab response, what proportion of these epitopes is strain-specific or shared, how antigenicity differs between patients or viral subtypes, or what the predominant escape mechanisms are. Furthermore, there is mounting evidence that at least some of these parameters differ between HIV-1 subtypes (reviewed in [20] ). We, and others, have characterized the autologous Nab response as it first develops, and have observed high titers of autologous Nab activity against the infecting strain in most (but not all) patients [13, 14, 15, 16, 17, 18, 19] . In addition we have reported higher Nab titers in subtype C infected patients compared to subtype B infected patients that were evaluated in parallel, prompting our group and others to propose that there are differences in antigenicity between subtype B and C Envs [15, 20, 21, 22] . Recent studies honed in on regions that could be involved in early subtype C autologous Nab responses, and these included the V1V2 hyper-variable domain and the C3 to V4 subregion of gp120 [23] . However, these regions could not account for all of the Nab activity present in patient plasma, suggesting the involvement of additional determinants. Furthermore, we have demonstrated that both V1V2-dependent and -independent pathways are utilized for escape from Nab during chronic subtype C infection [24] . In addition, we have shown a strong association between mutations in the a2 helix region of C3 and neutralization resistance, although these mutations did not directly alter neutralization sensitivity when transferred between sensitive and resistant Envs from linked transmission partners [25] . Temporal studies of HIV-1 have demonstrated that HIV-1 undergoes recurrent cycles of escape from autologous Nab [17, 19, 26, 27] , and escape also occurred during infection with a chimeric SIV-HIV-1 (SHIV) virus in response to vaccine-induced Nab [28] . Yet, our knowledge of the specific molecular events that lead to escape remains incomplete. Shifting carbohydrate moieties in and around the outer surface of gp120, as well as changes in the hyper-variable domains, have been proposed as general mechanisms used by HIV-1 and SHIV to alter neutralization epitopes, although most of these studies are based on subtype B Envs. Nab escape in SHIV-infected macaques has been shown to involve glycan changes in the V1, V2, and V3 domains, perhaps by shielding conserved epitopes such as the CD4 binding site [29, 30, 31] . Consistent with this finding, a recent study identified V1V2 as the major determinant of strain-specific autologous Nab in macaques infected with two different strains of SHIV [32] , and this domain was also shown to be the principal determinant of inherent Nab resistance for HIV-1 strain JRFL [10] . Subtype B HIV-1 can also escape from autologous Nab by shifting N-linked carbohydrates on the outer domain of gp120 with little involvement of V1V2 [17] . Others have demonstrated that subtype B viral escape could also occur in the absence of frank changes in glycosylation, with no clear mutational pattern emerging [33] . In experimental SIVmac infection, the emergence of N-and O-linked carbohydrates in the V1 and V4 hypervariable domains has been shown to confer escape from autologous neutralization [34, 35, 36] . Furthermore, the presence of specific glycans in V1 reduced the immunogenicity of SIVmac in the context of an experimental infection [37] . Taken together, these studies hint at the complexity of HIV-1 neutralization and escape, but also suggest that common themes may exist. Thus, unlike cytotoxic T lymphocyte (CTL) epitopes and their escape mutations, which are frequently predicted by the association of viral sequence polymorphism and HLA alleles, Nab epitopes and escape pathways in Env can be inherently difficult to identify based on sequence alone. We have therefore undertaken a molecular approach to define these in subtype C HIV-1 Env during early infection. Using a pseudovirus-based assay that facilitates evaluation of individual, patient-derived Envs, we analyzed the neutralizing ability of longitudinal plasma samples against contemporaneously-derived, autologous Envs from two subtype C seroconvertors who generated potent Nab against the infecting Env. Sequential neutralization escape variants emerged in both patients, and we used Env domain exchange and sitedirected mutagenesis approaches to map the pathways involved in Nab escape at multiple time points throughout the first two years of infection. Nab escape in these two patients was driven by V1V2dependent and -independent pathways, and these pathways exhibited different levels of stability over time. Yet, spatial representation of the sequence changes in gp120 indicated that immune pressure was directed at the same Env regions in both subjects. The derivation of autologous monoclonal antibodies (Mabs) from one patient demonstrated how a single potential glycan change in V1V2 afforded simultaneous resistance against multiple antibodies. These studies therefore provide a detailed look at Nab escape in subjects recently infected with the most predominant subtype worldwide, and demonstrate that the flexibility of Env facilitates the use of multiple mechanisms. We previously demonstrated that 9 out of 11 subtype C infected subjects from the ZEHRP cohort developed robust autologous Nab responses against the infecting Envs, with 50% inhibitory (IC50) titers often exceeding 1:3,000 within the first few months of infection [15] . For two of the subjects who developed potent autologous Nab and were identified as viral p24 antigen positive (Table 1) , we sampled the emerging quasispecies by single genome PCR amplification, cloning, and sequencing of biologically functional env genes from longitudinal plasma and PBMC DNA samples. For both subjects, the 0-month Envs were cloned at the first seropositive time point estimated to be within 48 days of infection, and longitudinal timing was calculated in months from A significant obstacle to developing an HIV vaccine is the potential for the virus to escape from the immune response induced by immunization. We previously showed that subjects in a Zambian cohort developed potent neutralizing antibody responses shortly after becoming infected by subtype C HIV-1, and here we have extended those findings to demonstrate that cycles of viral escape occurred in two of these subjects despite a potent immune response. We investigated the determinants of immune escape, and found that a single common mutational pathway was not sufficient to facilitate viral escape. Instead, we demonstrate that multiple strategies, including potential changes in glycosylation pattern, direct alteration of an epitope sequence, and cooperative envelope interactions, were used independently or together to evade neutralization. We also recovered individual monoclonal antibodies from one of the subjects and found that a single mutation can confer escape from different neutralizing antibody specificities. The studies demonstrate the remarkable flexibility of subtype C HIV-1, and suggest that the envelope glycoproteins are uniquely equipped to adjust to the specific properties of the immune response in each newly infected host. this point forward [38] . Samples from five subsequent time points over the first two years of infection were evaluated. A subset of Envs was chosen to represent the diversity of the circulating quasispecies at each time point (see arrows in Fig. S1A and B) , and was evaluated for sensitivity to neutralization by each contemporaneous (simultaneously collected) plasma sample using the JC53-BL (Tzm-bl) pseudovirus assay [15, 24, 25] . The IC50 Nab titer for each plasma-Env combination was calculated from each virus infectivity curve using a growth function. Fig. 1 shows that the median IC50 titer of the 0-month Envs (designated according to the first seropositive time point; see Table 1 ) was higher than the contemporaneous Envs at each time point, indicating repeating cycles of neutralization resistance. This difference in median IC50 titer was statistically significant at all time points for 185F. However, for 205F, the median IC50 titer differed significantly at only two time points, probably due to the wide range of Nab sensitivities observed for the contemporaneous Envs of this subject (Fig. 1B) . Nevertheless, Nab-resistant variants were present at each time point and these were neutralized by subsequent plasma samples, indicating continued induction of a de novo Nab response (data not shown). To gauge whether breadth developed within the window of evaluation, cross-neutralizing activity of a single plasma sample from subject 185F (23-months) and subject 205F (20-months) was measured against heterologous 0-month subtype C Envs from six other subjects in the same cohort (Fig. 2) . Consistent with previous studies of early subtype C infection [15, 16] , Nab in 185F and 205F was mostly strain-specific. However, each of these plasma samples neutralized 2 of the 6 heterologous Envs with an IC50 of greater than 1:100, and in two cases approaching 1:1000. Interestingly, 205F plasma potently neutralized the 185F Env (red dot on right point plot), but the reciprocal was not observed (green dot on left point plot). Thus, Nab in these plasma samples was directed against predominantly strain-specific targets, but were also capable of recognizing some common epitopes by approximately two years after infection. For 185F, the median Nab IC50 titer for the 0-month Envs was significantly greater than the contemporaneous Envs at every time point using a Mann-whitney test (Fig. 1A) . Using the criteria of at least a 100-fold decrease in sensitivity to neutralization compared to the median of the 0-month Envs, a resistant Env (highlighted in green) from each time point was selected for in-depth investigations into Nab escape. At 28-months, three different Nab resistant Envs were selected because the phylogenetic tree indicated that multiple lineages of resistance were circulating at this time point (Fig. S1A ). For 205F, the difference between 0-month and contemporaneous Envs reached significance at only two of the five time points, although there were Envs at each time point for which Nab activity was undetectable at the highest dilution of plasma tested (1:20, Fig. 1B ). These Envs, which were often 1000fold less sensitive to Nab than the median of the 0-month Envs, were selected for detailed studies of Nab escape. Two genetically diverse Envs from distinct lineages were also selected to represent the 20-month time point in 205F (Fig. S1B ). We next investigated the adaptations that were responsible for escape from contemporaneous Nab in 185F and 205F. The 0month Envs were potently neutralized by plasma from all subsequent time points ( Fig. 1A and B) and were used to provide a neutralization sensitive background, which remained more than 95% conserved at the amino acid level with subsequent variants and could be used to investigate the molecular determinants of escape for each Nab resistant variant. To do this, two approaches were used: (i) where sequence changes were limited in the Nab resistant Env, site-directed mutagenesis was used to introduce potential escape mutations into the 0-month Env and (ii) where multiple sequence changes were present in the Nab resistant Env, larger Env subregions (i.e. V1 to V5, V3 to V5, V1V2, etc.) were transferred from the Nab resistant Env into the 0-month Env. The neutralization sensitivity of the chimeric and parental Envs was then evaluated using plasma contemporaneous with the Nab resistant Env. For 185F, the 5-month Nab resistant EnvPB1.1 was chosen to determine which sequence adaptations were responsible for early Nab escape. Fig. 3A demonstrates that the chimeras containing either the region spanning the V1 loop through the end of the V5 domain (V1V5) or the V3 loop through the V5 domain (V3V5) from the 5-month Env displayed a level of resistance similar to the parental Env. However, the Nab sensitivity of the chimera containing the entire V1V2 domain (V1V2) was unchanged compared to the 0-month Env, despite a K192Q change in V2 (Fig. 3C) . We therefore surmised that Nab resistance was heavily dependent upon the V3V5 subregion, in which there were 3 residues that differed between 0-month EnvPB3.1 and 5-month EnvPB1.1. These changes were an E335A in the first position of the a2 helix, and two changes in the V5 region: I459T, which may also impact CD4 binding, and S463N ( Fig. 3C ; based on HXB2 numbering). None of these changes altered any of the predicted Nlinked glycosylation sites. To assess its individual contribution to Nab resistance, each amino acid change was introduced into the 0month EnvPB3.1. The V5 mutations I459T and S463N each independently produced a decrease in neutralization sensitivity, while these mutations combined recapitulated the Nab resistance level of the V3V5 chimera ( Fig. 3B ). In contrast, the E335A change in the a2 helix did not decrease neutralization sensitivity when introduced by itself into the 0-month Env (Fig. 3B ). Together, these findings indicate that the combined V5 mutations facilitated neutralization resistance at 5 months, while the changes in the a2 helix and the V2 loop did not contribute to a detectable level, at least within the context of the 0-month Env. A different scenario was observed for early escape in 205F. For the 2-month escape variant EnvPB2.3, the V1V5 region contained determinants for Nab resistance (Fig 3D) , and this entire region differed from 0-month EnvPL6.3 by only 3 amino acids (Fig. 3E ). Two changes were located within V1 (an N134S substitution that introduced a potential N-linked glycosylation site near the Nterminal V1V2 stem and a Y140P substitution; numbered according to Fig. S2 ). In contrast to 185F, introduction of the two changes in V1 decreased Nab sensitivity by more than 10-fold in the context of the 0-month Env (Fig. 3D) . A third change, A453T, was located within the b23 region and has the potential to impact CD4 binding (Fig. 3E ). This residue by itself reduced Nab sensitivity by about 3-fold ( Fig. 3D ). Thus, for 205F, complete Nab resistance at 2-months could be achieved by sequence changes in V1, including addition of a potential N-linked glycan, and a change in a region involved in CD4 contact [39] . Although the mutagenesis studies strongly suggested that the changes in V5 at 5-months were Nab escape mutations, these specific residues were not maintained in the subsequent Nab resistant Envs (Fig. S3) . However, the sequence of the V5 domain continued to evolve over time, suggesting ongoing selective pressure from Nab. At the last time point analyzed, 28-months, genetically distinct lineages of Nab escape variants were circulating (Fig. S1A ). The chimera-mapping approach revealed that these different Env variants had acquired resistance through at least two distinct mutational pathways (Fig. S4 , see bottom 3 panels). More detailed mapping revealed that for 28-month EnvPL5.1, the V5 domain continued to contribute to Nab resistance (Fig. 4A) , retaining the major escape pathway operative at 5-months. By contrast, 28-month EnvPL3.1 achieved a similar level of resistance through an escape pathway that required the gp41 ectodomain in addition to the cognate V1V5 domain (Fig. 4B ). For this Env, the V1V5 region from the 28-month Env independently conferred only partial escape onto the 0-month Env, while insertion of the 28-month ectodomain alone had no effect on Nab sensitivity. Further dissection revealed somewhat surprisingly that the V1V2 domain was a major contributor to Nab resistance in the context of the gp41 ectodomain, while the V3V5 region in this context did not appear to contribute to escape. Thus, a cooperative interaction between the gp41 ectodomain and V1V2 appeared to be conferring Nab resistance in this Env. Nab escape in 205F is primarily determined by V1V2 over time Having found that changes in the V1V2 and b23 regions drove early escape from Nab in 205F, we investigated whether this pattern was maintained over time. Env chimeras were created for 205F using Nab resistant Envs from five time points over a 26month follow-up period (Fig. 1B) and evaluated using the same approach as for 185F. V1V2 was the major determinant of Nab resistance at all time points analyzed (see Fig. S5 ). By contrast, the V3V5 region alone had little effect on resistance, although in combination with V1V2 it clearly contributed to escape. In an effort to more precisely define Nab targets and escape pathways in 205F, B cell hybridomas were generated from viably frozen PBMC samples collected at 49 months after infection, which were the earliest available sample of this type. A 0-month Env (clone PB1.1) was used to screen for neutralizing activity in the hybridoma supernatants, and two hybridomas produced monoclonal antibodies (Mabs; 6.4C and 13.6A) that neutralized this and other 0-month Envs ( Fig. 5A and B, respectively) . Surprisingly, the 0-month Envs were not equally sensitive to neutralization by the two Mabs, despite being very homogeneous in sequence and potently neutralized by patient plasma (Fig. 1B) . For Mab 6.4C, 0-month EnvPL6.3 was moderately more sensitive to neutralization than the other two 0-month Envs (Fig. 5A ). In contrast, neutralizing activity for 13.6A was not detectable against EnvPL6.3, but the other two 0-month Envs were neutralized at levels similar to those observed with Mab 6.4C (Fig. 5B ). The 2month EnvPB2.3 was neutralized by both Mabs (Fig. 5A and B). Neutralizing activity against Envs cloned at 8-months or beyond, however, was undetectable for both Mabs ( Fig. 5A and B), providing strong evidence that these Mabs could be representative of those elicited during early infection and that the later Env variants had developed resistance mutations that protected against both specificities. An identical pattern was observed for the purified 6.4C and 13.6A Mabs, with a mean IC50 against the sensitive 0-months Envs of 39 and 156 ng/ml, respectively (data not shown). Envs from 8-months and beyond were not neutralized at 10 mg/ml of either purified Mab (data not shown). Resistance against Mabs 13.6A and 6.4C involves loss and gain of predicted glycan addition sites in V1V2 Neutralization of the 205F chimera panel by each Mab localized differences in sensitivity to the V1V2 domain (data not shown). Examination of the V1V2 sequences of 0-to 26-month 205F Nab resistant Envs revealed that each one differed in length, pattern of predicted glycosylation sites, and sequence (Fig. S2) . However, all of the Nab resistant Envs from 8-months and beyond had acquired a mutation that created a potential glycosylation site in V2 at position 197 (highlighted in yellow in Fig. S2 ). Furthermore, 0-month EnvPL6.3 was the only 0-month Env that was resistant to 13.6A, and it lacked a potential N-gly site in V1 relative to the other Envs (highlighted in yellow in Fig. S2 ). Thus, Figure 2 . Moderate neutralization breadth of plasma from 185F and 205F against heterologous 0-month Envs. A single plasma sample from 185F (23-months) and 205F (20-months) was evaluated for neutralizing activity against heterologous Envs acquired during acute/early subtype C infection of six subjects. The Nab IC50 titer for each plasma-Env combination is shown on the vertical axis on a log10 scale. Each data point represents a single plasma-Env combination. Arrows indicate the Nab IC50 titer for the autologous plasma-Env combination; all other data points represent heterologous Envs. The plasma sample is indicated below each point plot. doi:10.1371/journal.ppat.1000594.g002 A and B) , and 205F 2-month PB2.3 (D) and chimeras or mutants from each of these Envs in the corresponding 0-month Env was evaluated using contemporaneous plasma. Pseudoviruses were created by expressing each Env with an HIV-1 env-deficient backbone, and their infectivity for JC53-BL13 (Tzm-bl) cells was evaluated in the absence or presence of serially-diluted patient plasma with luciferase as a quantitative measure. Percent virus infectivity relative to no test plasma is plotted on the vertical axis; the reciprocal of the plasma dilution is plotted along the horizontal axis on a log10 scale. Each curve represents one Env against serial plasma dilutions, and error bars represent the standard deviation of at least two independent experiments using duplicate wells. The legends list the parental Nab resistant Env followed by the chimeric and mutant Envs created in the 0-month Env background. Amino acid alignments are shown to indicate the sequence differences between the different Envs for 185F (C) and 205F (E). The color of the text corresponds to the curves on the graph. Only regions that contained differences are shown. doi:10.1371/journal.ppat.1000594.g003 . 185F Nab resistant variants at 28-months utilize distinct escape pathways. Neutralization of 28-month EnvPL5.1 (A) and 28month EnvPL3.1 (B) and the chimeric Env pseudoviruses generated from each of these Envs in the 0-month Env background was evaluated using the 28-month plasma sample in JC53-BL13 cells with luciferase as a quantitative measure. Percent virus infectivity is plotted against the reciprocal of the log10 plasma dilution. Error bars represent the standard deviation of at least two independent experiments using duplicate wells. The panels below each graph indicate the region that was transferred from the 28-month Nab resistant Env (PL5.1 in A and PL3.1 in B; gray boxes) into 0-month EnvPL3.1 (white boxes). (C) Amino acid alignment of the V5 region for the 0-month and 28-month Envs. doi:10.1371/journal.ppat.1000594.g004 we hypothesized that a different array of potential glycan addition sites determined the pattern of sensitivity to the two Mabs. Fig. 6A shows the naturally occurring patterns of these predicted glycan sites that were detected in the early 205F Envs with the positions of the sites of interest indicated in red (V1) and blue (V2). To define the effects of these potential glycan addition sites in V1 and V2 on sensitivity to the two Mabs, both sites were introduced into 0-month EnvPL6.3, which carried neither (Fig. 6B) . For Mab 13.6A, introduction of the V1 predicted glycan site into 0-month EnvPL6.3 resulted in a dramatic increase in neutralization sensitivity (Fig. 6C ). In contrast, for 6.4C, introduction of the V1 predicted glycan produced a moderate decrease in sensitivity (Fig. 6D) . Introduction of the V2 predicted glycan site into EnvPL6.3, with or without the V1 predicted glycan site, resulted in strong protection against both Mabs (Fig. 6C and D). Thus, predicted glycosylation at this site in V2 potentially tracked with protection against both Mabs (for a summary of longitudinal Envs and glycan sites see Table 2 ). These results also demonstrate that while both Mabs target a V1V2-dependent epitope, they recognize distinct structures. The detection of a 0-month Env that was resistant to one of the Mabs was unexpected given the early timing and high sensitivity of these Envs to patient plasma. Therefore, the frequency of this predicted glycan site in V1 during acute/early infection was investigated using 21 uncloned single genome amplified V1V4 sequences from a p24-positive, antibody-negative sample (1-Mar-03) and 31 from the 0-month sample (27-Mar-03), which was antibody positive [38] (see Table 1 ). These sequences were combined with the five cloned 0-month Envs from this study, and a highlighter plot was created using the HIV Database (Fig. S6) . Forty-four out of 57 sequences (77%) were identical throughout V1V4, and all of these contained the predicted glycan addition site in V1. Two variants (including EnvPL6.3), both from the antibody positive time point, carried an identical G to A mutation that abrogated the predicted V1 glycan addition site. Taken together, these observations provided strong evidence that the founder virus, like the 0-month Envs PB1.1 and PL4.1, carried the predicted glycan site in V1, and that the mutation in 0-month EnvPL6.3 arose shortly after transmission, but circulated only transiently. In a previous study, we demonstrated that subtype C infected seroconvertors mount robust Nab responses against their autologous viruses during the early stages of infection [15] . Here we have extended those findings to demonstrate that cycles of viral escape occur despite potent Nab and that these cycles involve multiple mechanisms and regions of Env. In subject 185F, early Nab escape required amino acid substitutions in V5 that were independent of glycosylation. It is possible that these changes directly altered an epitope in V5, as this region may be accessible to Nab on the Env trimer. However, attempts to remove Nab activity with a V5 peptide were unsuccessful, and Env chimeras in which unrelated Envs were engineered to carry the 185F 0-month V5 sequence lacked biological activity (data not shown). Thus neither of these approaches allowed definitive identification of a V5 epitope, and the latter suggested that the V5 domain itself, or the proximal region of gp120, likely evolved in concert with adjacent regions of the protein. Another possibility is that the early changes in V5 created conformational changes that protected a distinct target. The N-terminal region of V5 has been shown to contain contact sites for both CD4 and Mab b12 [39] , and the escape mutations could therefore have influenced exposure of epitopes such as the CD4 binding site. These two alternatives, epitope mutation or masking, are not mutually exclusive, and it is conceivable that V5 changes could protect from more than one antibody specificity. This is clearly the case for 205F, where a single amino acid change in V2 creating a potential glycan addition site resulted in resistance against two distinct Mabs. At later time points in subject 185F, the flexibility of the Env structure provided alternative mutational pathways to resist neutralization. Escape pathways in 185F oscillated between changes localized to the gp120 outer domain (V3V5), and conformational masking strategies that required interaction between spatially separated Env subregions ( Fig. 7A and B) . V5 or V3V5 was the major determinant at 5-, 14-, and 17-months, and also in one of the 28-month Nab resistant Envs (see Fig. S4 for neutralization curves). However, in Envs from two other time points (11-and 23-months), V3V5 did not independently confer resistance, but appeared to require contributions from V1V2. In Nab resistant Envs from three time points (20-, 26-, and 28months), the gp120 V1V5 region and the gp41 ectodomain were both required for resistance. Further mapping for one of these Envs demonstrated that the V1V2 domain contributed in large part to this phenotype, but the V3V5 domain in this context did not. Interestingly, some of the V1V2 domains in subject 185F contained changes in the predicted glycosylation pattern relative to the 0-month Env, while others had sequence changes that would not alter the original glycosylation pattern (Fig. 7A : 5-PB1.1, 11-PL5.1, 23-PL5.1, and 28-PL5.1). A novel finding is that Nab resistant Envs at 28-months utilized distinct Env sub-regions to block the same Nab pool. This provided a striking example of convergent, intra-patient evolution during early infection. One pathway was heavily dependent on the V5 domain, while the other exhibited V1V2 and gp41 codependence. The V5 domains of these two Envs contained the same predicted glycosylation shift (Fig. 7A : green and white spheres in 28-PL3.1 and 28-PL5.1) but differed in primary amino acid sequence (Fig. 4C) . This raises the possibility that the V5dependent Env contained mutations that directly confer epitope escape, while the V1V2-dependent Env retained the target but escaped through indirect mechanisms. In addition, both V1V2 and the regions flanking V5 are proximal to the CD4 binding site and could therefore alter its exposure, as has been proposed for changes in V2 and V5 in the context of a SHIV infection [30] . Thus, in examining a single subject in great detail, we have uncovered remarkable flexibility in the pathways of viral escape during early infection. These results further highlight how the plasticity of the Env hyper-variable domains coupled with complex conformational interactions could provide numerous options for escape. In contrast to subject 185F, Nab escape in 205F was driven predominantly by changes in the V1V2 domain ( Fig. 8A and B) . A preference for potential glycan shifts in V1V2 became evident from the spatial representation of gp120, where in the 14-month Env, four predicted glycosylation site changes are observed in V1 alone (Fig. 8A : green and white spheres in Env 14-PB5.4). The importance of V1V2 for escape was further illustrated by the demonstration that two predicted glycan sites, one in V1 and one in V2, influenced neutralization by two Mabs derived from this same subject. Thus, this study is the first to identify specific mutations that confer autologous Nab resistance at the single antibody level. This made possible several observations that were not apparent from polyclonal plasma. First, a single substitution can confer resistance against multiple antibody specificities within an individual. While we did not formally show that glycosylation at the substituted site was responsible for resistance against Mabs 6.4C and 13.6A, there is strong evidence from other studies to support that this is the case. Second, different pathways of escape also operate at the single antibody level. Resistance against 13.6A could be achieved either by addition of the predicted glycan site in V2 or by loss of the predicted glycan site in V1. Interestingly, only the modification of V2 was retained in subsequent escape variants, suggesting that it could have been more advantageous in terms of escape and or maintenance of replication fitness. Third, mutations that confer escape from multiple monoclonal antibody specificities do not necessarily confer escape from the entire polyclonal Nab milieu in plasma. The V2 modification in the 8-month Env conferred complete resistance against 6.4C and 13.6A at 10 mg/ ml, but only partial escape from patient plasma ( Fig. S5 and data not shown). This finding suggests that escape determinants mapped against plasma will only reflect the dominant Nab specificities that are present at relatively high concentration (able to inhibit virus infectivity at greater than a 1:100 dilution in our assay), but other lower titer Nab specificities could also drive escape mutations. The relative contributions of different antibody specificities in plasma will undoubtedly vary among subjects, and potentially even within a subject over time, resulting in the need for customized escape pathways that are driven by each dominant Nab response. Thus, to derive a complete picture of autologous Nab and escape, it will be necessary to recover and characterize individual Mabs with different specificities, as was done here and recently by others to dissect the B cell response in subjects with neutralization breadth [9] . These studies demonstrate how the HIV-1 subtype C Env is uniquely equipped to respond to the current immune response of each individual host by adjusting its pathways of escape. The V2-based mutation that conferred resistance against 6.4C and 13.6A appeared during the first eight months of infection; however, 13.6A and 6.4C were recovered from memory B cells circulating 41 months later. As such, it was not possible to determine whether related B cells were circulating during early infection in 205F. The V1-based change that conferred resistance against 13.6A was present at 0-months (at which time the subject was seropositive), but only transiently. If this mutation occurred in response to immune pressure from 13.6A, then this antibody must have been present within ,48 days from the calculated time of infection. Indeed we have observed very low level neutralizing activity (IC50 = ,1:40) in the 0-month plasma of 205F against 0-month Envs [15] and (data not shown), consistent with this concept. While others have observed that the very early antibody response (within the first ,40 days after infection) lacks neutralizing activity and is directed predominantly against gp41 [40] , our findings raise the possibility that viral neutralization and escape could occur earlier than previously thought. Individual Mabs, derived from B cells early in infection The blue gp120 backbones were generated by homology modeling of the 0-month Env sequence onto the CD4-liganded HIV-1 YU-2 gp120 structure [54] , with modeled V1V2 and V3 loops as described previously [25] . Red indicates amino acid changes relative to the 0-month Env sequence. Spheres indicate changes in potential N-linked glycosylation sites (green = loss, white = gain and tested against founder virus Envs, may be required to detect this initial Nab activity. It will therefore be important to determine whether Nab activity and viral escape is present in other subtype C infected subjects at very early time points, as well as whether early escape mutations are associated with decreased viral fitness. The blue gp120 backbones were generated by homology modeling of the 0-month Env sequence onto the CD4-liganded HIV-1 YU-2 gp120 structure [54] , with modeled V1V2 and V3 loops as described previously [25] . Red indicates amino acid changes relative to the 0-month Env sequence. Spheres indicate changes in a potential N-linked glycosylation sites (green = loss, white = gain Spatial properties of escape pathways Importantly in this study, not all sequence changes were linked directly with Nab escape. For example, one of the first sequence changes that was present in the 185F 5-month Env was a substitution in the first position of the a2 helix (E335A); however, this change did not alter Nab sensitivity to contemporaneous plasma when introduced into the 0-month Env. Reversal of this mutation (A335E) in the 5month escape variant also did not increase its sensitivity to autologous Nab (Murphy et al., in preparation) . The 205F Nab escape variants also exhibited variation in the a2 helix beginning at 14-months, but again this region did not appear to contribute independently to Nab escape. These findings support that a2 was not targeted directly by Nab in these instances, despite ongoing sequence evolution. The a2 helix has been linked to autologous and heterologous Nab sensitivity of subtype C Envs from Zambia, South Africa, and India by our group and others [23, 25, 41] . However, its exact role(s) in Nab sensitivity or escape remains undetermined [23, 25] . We have speculated that the a2 helix plays an ancillary role in Nab escape, and perhaps participates in maintenance of the tertiary structure of the gp120 outer domain or the quaternary structure of the trimer [25, 42] . The findings presented here support our earlier findings, but do not rule out the possibility that the a2 helix is targeted by Nab in some instances, or by low titer Nab specificities. Studies are ongoing in our laboratory to more precisely define the role of changes in the a2 helix in the context of autologous Nab. In addition to the a2 helix, sequence variation was observed in V1V2, V5, and other regions of the outer domain in both 185F and 205F Envs ( Fig. 9A and B, respectively) . However, in 185F, the V3V5 region had the strongest effect on Nab resistance, while in 205F V1V2 was the major determinant. A companion study of four subtype C infected seroconvertors and our own previous study of subtype C chronically infected subjects reported consistent findings, in that V1V2 was commonly involved in Nab escape, but to varying degrees [24] and (Moore et al., in press). Importantly, the combined biological results and spatial analysis of these two subjects demonstrate how the perpetual flexibility of the V1V2 and V5 domains provides a formidable defense against Nab. This could be due in part to their ability to simultaneously mask multiple epitopes through limited changes, but may also involve direct escape. Although these studies were conducted on a small number of subjects, it is still beneficial to work toward developing a mechanistic model that can explain the underlying complexity of escape pathways between and within subjects. The results presented here provide the basis for such an endeavor. First, the escape pathways observed here appear to define Nab resistance through a combination of direct and indirect mechanisms. Direct epitope changes may be sufficient in the setting of limited antibody specificities, such as during the early phase of infection, while indirect 'masking' or cooperative mechanisms may be required later when multiple antibody specificities are circulating. However, it will be important to expand and confirm these studies by characterizing Nab escape in additional subjects. The frequency, timing, and underlying basis for convergent escape pathways within a subject will require further investigation, as will the significance of regions of subtype C Env that appear to be under positive selection but do not contribute directly to escape from plasma Nab. Lastly, it will be important to determine if different escape pathways share any common conformational basis or point to specific regions of the Env that should be incorporated into or excluded from vaccine immunogens. Informed consent and human subjects protocols were approved by the Emory University Institutional Review Board, and the Figure 9 . Sequence variation occurs in similar regions of Env in 185F and 205F. A 3-dimensional representation of the sequence variation in gp120 over time in 185F and 205F is shown in panels (A and B) , respectively. The gp120 backbones were generated by homology modeling of the 0-month Env sequences of 185F and 205F onto the CD4-liganded HIV-1 YU-2 gp120 structure [54] , with modeled V1V2 and V3 loops as described previously [25] . Blue to green to red indicates degree of sequence conservation (high to low) within the alignment. doi:10.1371/journal.ppat.1000594.g009 University of Zambia School of Medicine Research Ethics Committee. Written Informed consent was obtained from human subjects. The Zambia Emory HIV Research Project (ZEHRP) was established in Lusaka in 1994 to provide voluntary HIV-1 testing and counseling, long-term monitoring, and health care to cohabiting heterosexual couples. Details of the cohort have been described elsewhere [43] . Briefly, HIV-discordant couples enrolled in studies of transmission are monitored for seroconversion of the negative partner at three-month intervals, at which time the participants also receive preventative counseling and condoms. Banked plasma samples from seronegative partners are tested for p24 antigen by ELISA to identify individuals with acute infection [38] . The two subtype C infected seroconvertors studied here were participants in this cohort and were identified as p24 antigen positive and seropositive by rapid test and western blot, as described in [38, 44] . Plasma viral loads were determined using the Roche Amplicor HIV-1 assay. None of the subjects received antiretroviral therapy during the evaluation period. Conditions for single genome PCR amplification of full-length gp160 (plus Rev, Vpu, and partial Nef coding sequences) from the genomic DNA of uncultured peripheral blood mononuclear cells and cDNA from plasma have been described previously [38, 44] . The viral env amplicons were directionally T/A cloned into the CMV-driven expression plasmid pcDNA3.1-V5HisTOPO-TA and screened for biological function as pseudoviruses following co-transfection with an Env-deficient subtype B proviral plasmid (SG3Denv) into 293T cells [45] . Seventy-two hours later, supernatant was collected and used to infect JC53-BL13 (Tzmbl) cells. At 48 hours post-infection, b-gal staining was performed and each well was scored positive or negative for blue foci. DNA sequencing of env genes was carried out by Lone Star Labs, Inc. (Houston, TX) utilizing the ABI Prism H Automated DNA sequencer 377XL and Big Dye TM Terminator Ready Reaction Cycle Sequencing Kit. Nucleotide sequences were edited and assembled using Sequencher v4.7, translated using Se-Al v2.0all, and nucleotide or amino acid alignments were created using Clustal W v1.83. Neighbor joining phylogenetic trees were generated by Clustal W v1.83 using gap-stripped nucleotide sequences of the complete env gene, and reliability of branching orders was assessed by bootstrap analysis using 1,000 replicates. Trees were visualized using NJ Plot. Aligned sequences were imported into the Highlighter tool to analyze viral diversity (http://www.hiv.lanl.gov/content/ sequence/HIGHLIGHT/highlighter.html). Sequences have been deposited into Genbank under the accession numbers GQ485312-GQ485447. Patient plasma samples were evaluated for neutralizing antibody activity against virions pseudotyped with autologous patientderived viral Envs using a single round reporter assay described previously [17, 45] . Briefly, JC53BL-13 (Tzm-bl) cells were plated and cultured overnight. Two thousand infectious units of each pseudovirus was combined with five-fold dilutions of heatinactivated patient plasma and incubated for 1 hour at 37uC. Normal heat-inactivated human plasma was added as necessary to maintain a constant overall concentration. The virus-Ab mixture was then added to JC53BL-13 cells, and after two days, the cells were lysed, and the luciferase activity of each well was measured using a luminometer. Background luminescence was determined in uninfected wells and subtracted from all experimental wells. Percent infectivity was calculated by dividing the number of luciferase units at each plasma dilution by the value in the well containing no test plasma. The dilution of patient plasma that inhibited 50% of virus infectivity (IC50 titer) was determined using the Microsoft Excel 2004 for Mac Growth Function. Each experiment was performed independently at least twice with duplicate wells. Chimeric Envs were constructed using a domain exchange strategy that has been described previously [24, 25] . The primer sequences and their HXB2 locations are shown below. A PCR screen was performed to identify transformants in which the fragments ligated together in the correct orientation using forward primer EnvA and the reverse primer that was used to amplify the exchanged domain. Colonies that were positive by PCR screen were inoculated into LB-Ampicillin broth for overnight cultures, and the plasmid was prepared using the QIAprep Spin Miniprep Kit. Env chimeras were then screened for biological function as described above. For chimeras that produced functional Env pseudotypes, the plasmids were re-transfected into 293T cells on a larger scale to produce a working pseudotype virus stock. Transfection supernatants were collected at 72 hours post-transfection, clarified by low speed centrifugation, aliquoted into 0.5 ml or less portions, and stored at 280uC. The titer of each pseudovirus stock was determined by infecting JC53-BL13 cells with 5-fold serial dilutions of virus as described previously. All Env chimeras were confirmed by nucleotide sequencing of the entire env gene. The 0-month Env backbones (minus the target domains) with pcDNA3.1 vector sequences were PCR amplified using primers that anneal to conserved regions adjacent to the target domain primers. These primers amplify away from the target domains. A 59 phosphate group (Phos) was added to these primer sets during synthesis to facilitate ligation to the target domain amplicons. The primer sequences and their HXB2 locations were as follows: For subject 185F, V1V5 domain backbone, forward primer 59-Phos-atgagggacaattggagaagtg-39 (HXB2 nt 7647 to 7668) and reverse primer 59-Phos-gctttaagctttgatcccataaac-39 (HXB2 nt 6576 to 6553); V1V2 domain backbone, forward primer 59-Phoscaagcctgtccaaaggtctct-39 (HXB2 nt 6831 to 6851) and reverse primer 59-Phos-gctttaagctttgatcccataaac-39 (HXB2 nt 6576 to 6553); V3V5 domain backbone, forward primer 59-Phos-atgagggacaattggagaagtg-39 (HXB2 nt 7647 to 7668) and reverse primer 59-Phos-ccttacacacacaatttctac-39 (HXB2 nt 7118 to 7098); ectodomain backbone, forward primer 59-Phos-aagatatttataatgatagta-39 (HXB2 nt 8271 to 8291) and reverse primer 59-Phostttttctctctccaccactctcc-39 (HXB2 nt 7754 to 7732); V5 domain backbone, forward primer 59-Phos-atgagggacaattggagaagtg-39 (HXB2 nt 7647 to 7668) and reverse primer 59-Phos-catgttatgtttcctgcaatg-39 (HXB2 nt 4527 to 4507). The PCR amplification conditions for the 0-month Env backbones were 1 cycle of 95uC for 3 min; 35 cycles of 95uC for 1 min, 50uC to 60uC for 30 s (the optimal annealing temperature was determined for each primer set), 72uC for 10 min; 1 cycle of 72uC for 15 min; and storage at 4uC. The amplification conditions for the target domains were the same, except the extension time at 72uC was reduced to 30 s. The 25 ml PCR mixtures contained 50 ng of each primer, 10 ng of the plasmid template, 2.5 mM MgCl 2 , 0.2 mM deoxynucleoside triphosphate, and 16 reaction buffer. PfuUltra II DNA polymerase (Stratagene) was used to generate the blunt-ended PCR amplicons, which were digested with DpnI to remove contaminating template DNA and gel purified from an agarose gel using the QIAquick Gel Extraction Kit (QIAGEN) prior to ligation. Each target domain DNA fragment was then ligated to the purified 0-month env backbone to produce a chimera using T4 DNA ligase (5 U/ml; Roche) at 4uC overnight. The ligation reaction mixture (usually one-third of the volume) was transformed into maximum-efficiency XL2-Blue Ultracompetent cells (1610 9 CFU/mg DNA; stratagene) so that the DNA volume did not exceed 5% of the cell volume. The entire transformation was plated onto LB-ampicillin agar plates, generally resulting in 10 to 50 colonies per ligation reaction. To investigate whether individual amino acid sequence differences contributed to the neutralization resistant phenotype, PCR-based site-directed mutagenesis was used to introduce an amino acid change as described previously [24] . Briefly, a set of primers was used that each had either the wildtype or mutated sequence. The primer sequences and their HXB2 locations are shown below, where the substituted nucleotides are underlined. All mutants were confirmed by sequencing the entire env gene. The PCR amplification conditions were 1 cycle of 95uC for 1 min; 18 cycles of 95uC for 50 s, 60uC for 50 s, 68uC for 8 min; and 1 cycle of 68uC for 7 min. The 25-ml PCR mixtures contained 63 ng of each primer, 5 ng of the plasmid template, 0.2 mM deoxynucleoside triphosphate, and 16reaction buffer. PfuUltra HF DNA polymerase (Stratagene) was used, amplicons were digested with DpnI to remove contaminating template DNA, and 2 ml was transformed into maximum-efficiency XL10-Gold ultracompetent cells (5610 9 CFU/mg DNA; Stratagene). Half of the transformation was plated onto LB-ampicillin agar plates, generally resulting in 10 to 50 colonies per reaction. All mutants were confirmed by sequencing the entire env gene. 185F Human B cell hybridomas were generated from viable frozen PBMC samples from subject 205F based on a protocol of EBV immortalization of peripheral blood B cells as described previously [46, 47, 48] . This protocol includes immortalization of pre-selected (CD22+, IgM2, IgD2, IgA2) memory B cell populations that are cultured in medium supplemented with immunostimulatory CpG sequences and irradiated allogeneic PBMC, as described in [49, 50, 51] . EBV-transformed B cell cultures were screened for Mabs that neutralized the 205F 0-month Env and that bound envelope glycoproteins by ELISA [52, 53] . The method used to screen for neutralizing Mabs is a modification of the Tzm-bl (JC53-BL13) luciferase reporter cell assay originally developed by Wei et al. [15, 17, 24, 45] . EBV-transformed B cell hybridomas with neutralizing activity were cloned in the presence of CpG and irradiated PBMC. Culture supernatant was collected from the two hybridomas, 6.4C and 13.6A, and used in the neutralization studies to map targets and escape. Figure S1 Phylogenetic tree of longitudinal 185F and 205F Envs. A. A neighbor-joining tree was generated using Clustal W v. Figure S2 Amino acid sequence alignment for 205F Envs. Three 0-month Nab sensitive Envs and six Nab resistant Envs from subsequent time points were selected for study from 32 Envs. Env clones are indicated by the time point (in months), source (PB = PBMC DNA or PL = plasma), and clone number. Sequences are shown in reference to the 0-month EnvPB1.1, with amino acid differences indicated by the letter, and deleted residues indicated by a dot. Domains that were transferred into the 0-month Env to create chimeras are as follows: V1V5 (blue, gray and green; HXB2 nt 6557 to 7634), V1V2 (blue; HXB2 nt 6557 to 6876), V3V5 (green; HXB2 nt 7110 to 7634). Major Env domains are indicated above the region, and the a2 helix is underlined. Two potential N-linked glycan addition sites of interest in V1 and V2 (NXS or NXT where X is any residue but proline) are highlighted yellow. Found at: doi:10.1371/journal.ppat.1000594.s002 (0.37 MB PDF) Figure S3 Amino acid sequence alignment for 185F Envs. Two 0-month Nab sensitive Envs and ten Nab resistant Envs from subsequent time points were selected for study from 58 Envs. Env clones are indicated by time point (in months), the source (PB = PBMC DNA or PL = plasma), and clone number. Sequences are shown in reference to the 0-month Env PL3.1, with amino acid differences indicated by the letter, and deleted residues indicated by a dot. Domains that were transferred into the 0month Env to create chimeras are as follows: V1V5 (blue, gray, and green; HXB2 nt 6577 to 7646), V1V2 (blue; HXB2 nt 6577 to 6810), V3V5 (green; HXB2 nt 7119 to 7646), gp41 ectodomain (yellow; HXB2 nt 7755 to 8270). Found at: doi:10.1371/journal.ppat.1000594.s003 (0.50 MB PDF) Figure S4 The major determinants of Nab resistance in 185F Envs change over time. Neutralization of 185F parental and chimeric Env pseudoviruses was evaluated using longitudinal plasma samples that are indicated in each panel. Each plasma sample was contemporaneous with the Nab resistant Env, and neutralization sensitivity was evaluated in JC53-BL cells using luciferase as a quantitative measure. Percent virus infectivity is plotted against the reciprocal of the log10 reciprocal plasma dilution. Error bars represent the standard deviation of at least two independent experiments using duplicate wells. All chimeras were created in the 0-month EnvPL3.1 background (red lines). In the legend, the parental Env clones are indicated, followed by each chimera that contains the indicated region from the Nab resistant Env in the 0-month Env. 'ecto' stands for the gp41 ectodomain. Found at: doi:10.1371/journal.ppat.1000594.s004 (0.22 MB PDF) Figure S5 The major determinants of Nab resistance in 205F Envs are in V1V2. Neutralization of 205F parental and chimeric Env pseudoviruses was evaluated using longitudinal plasma samples that are indicated in each panel. Each plasma sample was contemporaneous with the Nab resistant Env, and neutralization sensitivity was evaluated in JC53-BL cells using luciferase as a quantitative measure. Percent virus infectivity is plotted against the reciprocal of the log10 reciprocal plasma dilution. Error bars represent the standard deviation of at least two independent experiments using duplicate wells. All chimeras were created in the 0-month EnvPL6.3 background (red lines). In the legend, the parental Env clones are indicated, followed by each chimera that contains the indicated region from the Nab resistant Env in the 0-month Env. Found at: doi:10.1371/journal.ppat.1000594.s005 (0.14 MB PDF) Figure S6 Highlighter plot showing nucleotide mismatches in the gp120 V1V4 region for 205F Envs. Single genome amplified, uncloned V1V4 sequences derived from the 1-Mar-03 sample (n = 21) or the 27-Mar-03 sample (n = 31) from 205F and V1V4 sequences from single genome amplified cloned Envs from the 27-Mar-03 sample (n = 5) were subjected to Highlighter analysis (www.hiv.lanl.gov). Ticks represent mismatched bases compared to the master sequence listed at the top. The three 0-month Envs that were analyzed for neutralization by 6.4C and 13.6A in Fig. 11 are boxed in red. PB = uncultured PBMC DNA; PL = plasma. The G.A mutation (diamonds) located near the beginning of the sequence represents the loss of a potential N-gly site in V1 that tracks with resistance against 13.6A. Found at: doi:10.1371/journal.ppat.1000594.s006 (0.11 MB PDF)

Crystal Structure of the N-Acetylmannosamine Kinase Domain of GNE

BACKGROUND: UDP-GlcNAc 2-epimerase/ManNAc 6-kinase, GNE, is a bi-functional enzyme that plays a key role in sialic acid biosynthesis. Mutations of the GNE protein cause sialurea or autosomal recessive inclusion body myopathy/Nonaka myopathy. GNE is the only human protein that contains a kinase domain belonging to the ROK (repressor, ORF, kinase) family. PRINCIPAL FINDINGS: We solved the structure of the GNE kinase domain in the ligand-free state. The protein exists predominantly as a dimer in solution, with small populations of monomer and higher-order oligomer in equilibrium with the dimer. Crystal packing analysis reveals the existence of a crystallographic hexamer, and that the kinase domain dimerizes through the C-lobe subdomain. Mapping of disease-related missense mutations onto the kinase domain structure revealed that the mutation sites could be classified into four different groups based on the location – dimer interface, interlobar helices, protein surface, or within other secondary structural elements. CONCLUSIONS: The crystal structure of the kinase domain of GNE provides a structural basis for understanding disease-causing mutations and a model of hexameric wild type full length enzyme. ENHANCED VERSION: This article can also be viewed as an enhanced version in which the text of the article is integrated with interactive 3D representations and animated transitions. Please note that a web plugin is required to access this enhanced functionality. Instructions for the installation and use of the web plugin are available in Text S1.

Sialic acids are Nor Osubstituted terminal monosaccharides with a nine-carbon backbone highly expressed on eukaryotic cell surfaces [1] . Sialylation of glycoproteins and glycolipids modulates a wide range of biological and pathological events including early development [2] , tumorigenesis [3] , viral and bacterial infection, and immunity [4, 5] . In vertebrate systems, N-acetylneuraminic acid (Neu5Ac) is the metabolic precursor of all known naturally occurring sialic acids [6] . Neu5Ac is synthesized in the cytosol from UDP-N-acetylglucosamine (UDP-GlcNAc) by four consecutive reactions; and UDP-GlcNAc is a derivative of fructose-6phosphate and the end-product of the hexosamine biosynthesis pathway (Figure 1 ). The first two steps of the biosynthesis of Neu5Ac from UDP-GlcNAc are catalyzed by the bi-functional enzyme UDP-GlcNAc 2-epimerase/N-acetylmannosamine kinase (GNE). GNE contains an N-terminal epimerase domain and a C-terminal kinase domain [7] . The epimerase domain converts UDP-GlcNAc to Nacetylmannosamine (ManNAc), which is then phosphorylated at the 6 position by the kinase domain. GNE is feedback-inhibited by the activated form of Neu5Ac, i.e., cytidine-monophosphate Nacetylneuraminic acid (CMP-Neu5Ac). The kinase domain belongs to the ROK (Repressor, ORF, Kinase) family. The ROK family consists of a set of bacterial proteins that include repressors for sugar catabolic operons, and sugar kinases [8] . Gne is the only known gene in the entire human genome that encodes a ROK domain-containing protein. Three protein isoforms have been described for human GNE, where isoform 1 is ubiquitously expressed and is believed to be responsible for the basic supply of sialic acids. Isoforms 2 and 3 are generated by alternative splicing and show tissue specific expression patterns. Isoforms 2 and 3 have reduced epimerase activities but almost intact kinase activities and may fine-tune the production of sialic acids [9] . Wild type GNE forms homohexamer in solution [10] , and allosteric regulation of the epimerase and kinase activities of GNE is important for the normal function of the protein [10, 11] . Mutations in the epimerase domain lead to the rare congenital metabolism disorder sialurea, which results in the production of high levels of Neu5Ac due to loss of the allosteric feedback control of the UDP-GlcNAc 2-epimerase activity by CMP-Neu5Ac [12] . Late onset autosomal recessive inclusion body myopathy, which is also known as hereditary inclusion body myopathy (hereinafter referred to as HIBM), and allelic Nonaka myopathy are neuromuscular disorders that are caused by a number of different mutations within the gne gene. The mutations are located at either the epimerase domain or the kinase domain [13] and lead to hypoactivity of the enzyme [11] . Mutagenesis and enzymatic activity analysis revealed that the activities of the epimerase domain and the kinase domain are interrelated such that a single mutation in one domain could affect the activities of both domains [11] . Here, we solved the structure of the dimeric GNE kinase domain in the ligand-free state. The structure reveals the dimerization interface of the kinase domain and also suggests a possible hexameric assembly of the protein. Furthermore, the structure provides insights into the relationship between GNE mutations and GNE-related metabolism disorders. Overview of the GNE kinase domain monomer The overall structure adopts a typical bi-lobal kinase architecture. Both the N-lobe and the C-lobe have the a/b fold. Each lobe consists of a central b-sheet flanked by a-helices on both sides of the sheet. The last helix C-terminal to the C-lobe is part of the Nlobe and perpendicular to the interfacial helix of the C-lobe. Residues 475-498 of the N-lobe are invisible in the electron density map (Figure 2 ). The GNE kinase domain contains a type I zinc-binding motif GHx 9-11 CxCGx 2 G(C/H)xE, which forms an HC3 type zincfinger with residues H569, C579, C581, C586. The zincbinding motif is a characteristic feature for all ROK family members [14] . The kinase domain also contains a DxGxT type ATP-binding motif [15, 16] . The side chains of this ATPbinding motif residues point toward the cleft between the Nlobe and the C-lobe. Comparison with the actin/hexokinase/ hsp70 ATPase domains suggests that the disordered residues 475-498 form part of the binding pocket for the adenosine moiety of ATP [17] and are located near the DxGxT ATPbinding motif. Taken together, these findings suggest that the ATP binding pocket of the GNE kinase domain is located in the cleft between the two lobes. Previous deletion mutations study has suggested that the GNE kinase domain is responsible for dimerization, while a segment of residues between the epimerase and kinase domains, residues 360-382, is a potential site for trimerization [18] . Our gel filtration data ( Figure 3) show that the kinase domain exists predominantly as a dimer in the solution, with small amounts of monomer and a higher order oligomer. The apparent molecular weight of the oligomer fits a hexamer of the kinase domain (see also below). However, the possibility of a tetramer [19] cannot be completely ruled out due to the low resolution of the gel-filtration column at this molecular size The GNE kinase domain was crystallized in space group P3 1 21 with three molecules in the asymmetric unit. Protein interface and assembly analysis using the PISA server [20] suggests that two of the three molecules dimerize through the C-lobe with an average buried surface area of 1587 Å 2 per molecule ( Figure 4 ) whereas the third molecule dimerizes with a two-fold symmetry related molecule through the same C-lobe ( Figure 5a ). The solvation free energy gain upon formation of the interface, D i G, is 224.2 kcal?mol 21 , indicating that the dimer interface is very stable and may not simply be a crystal packing artifact. A crystallographic hexamer can be produced when a two-fold rotational symmetry operation is applied to the three molecules (one and a half dimers) in the asymmetric unit ( Figure 5 ). In this hexamer, the N-lobes of three kinase molecules are pointing to the same side of the ''hexamerization plane'', while the N-lobes of the other three molecules are pointing to the opposite side of the plane (Figure 5b ). This assembly mode of the kinase domain allows locating the epimerase domain further away from the hexamerization plane and is consistent with the proposition that the interdomain segment (residues 360-382) is the site of trimerization [18] . Structural homology search of the GNE kinase domain using the FATCAT (Flexible structure AlignmenT by Chaining Aligned fragment pairs allowing Twists) server [21] revealed the top four non-redundant hits to be PDB codes 2aa4, 1xc3, 1z05 and 1z6r. All these structures contain the signature zinc-binding motif of the ROK family. The structure of E. coli putative ManNAc kinase (PDB 2aa4) was the top hit with twist-adjusted r.m.s. deviation (opt-rmsd value) of 1.94 Å . The structure of a putative sugar kinase from Bacillus subtilis was the second best hit (PDB 1xc3). The other two homologous structures were transcription repressors that belong to the ROK family (PDB: 1z05 and 1z6r). Vibrio cholerae transcriptional regulator (PDB: 1z05) is a homolog of the E. coli Mlc protein (PDB: 1z6r). The latter is a transcriptional repressor that controls the expression of malT, the central transcription activator of the E. coli maltose system [22] . The structure of the GNE kinase domain aligns well with the E. coli Mlc structure: the N-lobe of GNE kinase domain aligns to the E-domain of Mlc and the C-lobe aligns to the Odomain of Mlc. It is interesting to note that the Mlc O-domain is responsible for the oligomerization of Mlc protein [22] in a way similar to the dimerization of GNE kinase through the C-lobe. However, these four structures do not contain sugar ligands that would help inform on a substrate binding mode for GNE. To evaluate the putative sugar binding site, the sequence of the GNE kinase domain was aligned with that of E. coli glucokinase complexed with glucose (PDB: 1sz2) [23] , which is the closest homologous structure containing a bound substrate currently available in the PDB data bank. The five residues involved in sugar binding are conserved in GNE (N516, D517, E566, H569, E588, GNE numbering). These five residues are arranged to accommodate the sugar substrate ( Figure 6 ). Two residues, H569 and E588, are located in the ROK family zinc-binding signature motif and H569 directly coordinates the zinc ion. This finding suggests that zinc may play a catalytic role in sugar substrate binding, as well as a structural role. Since the identification of the relationship between gne mutations and HIBM [13] , more than 60 mutations have been found to be associated with HIBM [24] . Among these mutations, 25 missense mutations at 23 unique sites are located in the kinase domain of the GNE protein. These 23 mutation sites can be classified into 4 different groups based on their solvent accessibility, and their locations ( Table 1 ). The first group of residues I557, G559, V572, and G576 is located at the dimerization interface of the C-lobe and mutation of these residues may interfere with dimerization of the kinase domain. It is noteworthy that kinase domain dimerization does not affect the solvent accessibility of G576 (Table 1 ), indicating that G576 is not directly involved in dimerization. The amino acid side chain of a mutant at this position would point into a hydrophobic niche that also accommodates the side chain of L574 from another monomer. The G576E mutation would exert both charge and space hindrances on the side chain of L574 and thus disrupt the dimerization (Figure 7) , consistent with the previous observation that the G576E mutant of the full length GNE remains as a trimer [11] . The location of this group of residues is also close to the residues involved in sugar substrate binding (Figure 6b ). Residues V572 and G576 are located on the zinc-binding signature motif of the ROK family (Figure 6b, 8b) , which could play both a functional and a structural role. Mutations of these residues could thus also affect the sugar substrate binding affinity of the kinase domain indirectly. The second group of residues includes those located at the interfacial helices between the N-lobe and the C-lobe, i.e. N519, A524, F528, G708, and M712. Since the interlobar cleft is the site of ATP and carbohydrate binding as well as where phosphoryl transfer occurs, mutation of these residues could change the interlobar movement during catalysis and thus affect the kinase activity of the protein. For example, the first identified HIBM-related mutation, M712T [13] , would likely abolish the hydrophobic interaction of the side chain of M712 with that of L523 from the C-lobe helix (Figure 8a ). In the previous study [11] , the M712T mutation has been shown to cause a 30% reduction in the kinase activity without affecting the epimerase activity of full length GNE. On the contrary, mutations of other residues in this second group reduce not only the kinase activity but also the epimerase activity of the full length protein ( [11] , Table 1 ). This suggests that the kinase domain is allosterically coupled to the epimerase domain. The structure of the full length GNE is needed to fully understand the coupled effects of the kinase and epimerase domains. The third group currently includes residue P511. P511 has the highest relative solvent accessibility (.40%) among the 23 mutation sites and is located on a loop region of the structure. The underlying mechanism for the association of P511H and P511L mutations with HIBM is elusive without further data, but mutation of a proline to any other residue type will inevitably change the flexibility of the loop region around this residue and could thus change the allostery of full length GNE in the higher-order oligomeric state. The fourth group of residues includes all the rest of mutation sites in Table 1 . All residues have hydrophobic side chains and low solvent accessibilities, and are located within secondary structural elements. Mutations of these residues may disrupt the secondary structural elements at given mutation sites, and could interfere with the hydrophobic interactions of the secondary structure elements that stabilize the protein quaternary structure. We show here the 3D structure of the N-acetylmannosamine kinase domain of GNE, the only ROK family kinase encoded in the human genome. The kinase domain dimerizes through an Data of missense mutations were extracted from reference [11] and [24] . b UDP-GlcNAc epimerase and ManNAc kinase activities are percentage values relative to the corresponding activities of the full length wild type GNE. Data extracted from reference [11] . c Oligomeric state of the full length mutant GNE. Data extracted from reference [11] . d Relative solvent accessibility of the residue calculated using the DSSP program [30] and normalized according to values in reference [31] . A value of 1 means full exposure of the residue while a value of 0 means the residue is fully buried. e The type of the secondary structure element the residue is located at was assigned using the DSSP program [30] . doi:10.1371/journal.pone.0007165.t001 interface at the C-lobe. This is consistent with mutagenesis data from other groups on the full length GNE protein [11, 18] . The crystallographic hexamer, which consists of a trimer of kinase dimers, could serve as a prototype of a proposed full length GNE hexamer. Structure comparison of the GNE kinase domain with previously studied proteins revealed potential substrate binding sites at the interlobar cleft and also the structural and functional importance of the signature zinc-binding motif of the ROK family. Four groups of missense mutations associated with hereditary inclusion body myopathy are classified and their effects on the enzymatic activity can mostly be explained by the structure model. The cDNA template encoding the kinase domain of GNE was codon optimized for overexpression in E. coli and synthesized commercially (Codon Devices, Inc.) The DNA fragment encoding GNE residues 406-720 was PCR amplified and subcloned into the pET28-MHL vector (gi:134105571) using an In-Fusion dry-down PCR cloning kit (ClonTech). Protein was overexpressed in E. coli BL21(DE3) CodonPlus-RIL cells (Stratagene) grown in terrific broth medium. The culture was grown at 37uC in a LEX bubbling system (Harbinger Biotech. & Engineering Corp.) until OD 600 reached 3.0. The temperature of the culture was then lowered to 15uC and the cells were induced with 0.5 mM isopropyl 1-thio-b-D-galactopyranoside and allowed to grow further overnight. Cells were harvested by centrifugation and flash frozen in liquid nitrogen and stored at 280uC until purification. Frozen cells were thawed and resuspended in 10 mM HEPES buffer (pH 7.5) containing 500 mM sodium chloride, 5% glycerol, 2 mM b-mercaptoethanol, and supplemented with 5 mM imidazole, and mechanically lysed using a microfluidizer (Microfluidics, model M-110EH) at 1,000 bar pressure. The lysate was clarified by centrifugation. GNE protein was bound with nickel-nitrilotriacetic acid (Ni-NTA) beads (Qiagen) at a ratio of 2.5 mL 50% Ni-NTA flurry per litre of cell culture. The bound protein was washed twice with the same HEPES buffer containing 30 mM or 75 mM imidazole, and finally eluted with the HEPES buffer supplemented with 300 mM imidazole. The elutant containing the GNE protein was further purified by Supderdex-75 size exclusion chromatography (GE Healthcare). The eluted fractions were pooled, concentrated to a final concentration of 40 mg per mL, and stored in a buffer containing 10 mM HEPES, pH 7.5, 500 mM sodium chloride, 5% glycerol and 5 mM dithiothreitol. The purity of the protein was better than 95% judging from SDS-PAGE gel. Selenomethionine (SeMet) labelling of the protein was carried out using prepacked M9 SeMet growth media kit (Medicilon) following manufacturer's instructions. The ligand-free form crystals were grown at room temperature in sitting drops. A final concentration of 5 mM ADP, 1:100 chymotrypsin (w/w) were added into the protein stock solution and 0.5 mL protein solution was mixed immediately with 0.5 mL well solution containing 15% polyethylene glycol (PEG) 4000, 0.2 M ammonium acetate, 0.1 M sodium citrate, pH 5.6 and set up for vapour diffusion crystallization. The SeMet crystal used for structure determination was grown in 14.55% PEG4000, 0.2 M ammonium acetate, 0.1 M sodium citrate, pH 6.0, with 1:100 chymotrypsin (w/w) and 5 mM ADP in a sitting drop setup. Crystals grew to a mountable size within 24 hours. Paratone oil was used to cryo-protect the crystals. Diffraction data of a selenomethionyl derivative of the GNE kinase domain were collected at beamline 19ID of the Advanced Photon Source (APS) at a wavelength of 0.9792 Å . Initial phases were obtained by single wavelength anomalous diffraction with SOLVE and density modification with RESOLVE [25] . For model building, the phases were combined with data collected at APS beamline 23ID-B at a wavelength of 0.9793 Å (see Table 2 ). The refined model of the target resulted from iterative application of density modification with DM and RESOLVE, interactive model building with COOT [26] , coordinate and B-factor refinement with REFMAC [27] and PHENIX [28] , and geometry validation with MOLPROBITY [29] . Diffraction data and refinement statistics are summarized in Table 2 . The current model was deposited at the Protein Data Bank with PDB ID 3EO3. Datapack S1 Standalone iSee datapack -contains the enhanced version of this article for use offline. This file can be opened using free software available for download at http://www.molsoft.com/ icm_browser.html. Found at: doi:10.1371/journal.pone.0007165.s001 (ICB) Text S1 Instructions for installation and use of the required web plugin (to access the online enhanced version of this article). Found at: doi:10.1371/journal.pone.0007165.s002 (PDF)

Increased Host Species Diversity and Decreased Prevalence of Sin Nombre Virus

Emerging outbreaks of zoonotic diseases are affecting humans at an alarming rate. Until the ecological factors associated with zoonoses are better understood, disease emergence will continue. For Lyme disease, disease suppression has been demonstrated by a dilution effect, whereby increasing species diversity decreases disease prevalence in host populations. To test the dilution effect in another disease, we examined 17 ecological variables associated with prevalence of the directly transmitted Sin Nombre virus (genus Hantavirus, etiologic agent of hantavirus pulmonary syndrome) in its wildlife host, the deer mouse (Peromyscus maniculatus). Only species diversity was statistically linked to infection prevalence: as species diversity decreased, infection prevalence increased. The increase was moderate, but prevalence increased exponentially at low levels of diversity, a phenomenon described as zoonotic release. The results suggest that species diversity affects disease emergence.

D uring the past 60 years, the number of emerging pathogens affecting humans has substantially increased (1) . Of these emerging infectious diseases, 62% are zoonotic (2) , meaning they are naturally hosted by, and persist in, wildlife but also affect human populations. The ecological factors associated with zoonotic disease emergence are likely complex and are poorly understood. Most often, because of limited time, resources, and the exigencies of the situation, outbreak investigations of emerging diseases seek only to discover the pathogen responsible for the disease in humans. But ecological studies are of critical importance to long-term containment of zoonotic disease emergence; they are the only way to ascertain the wildlife source of the disease, the dynamics of the host-pathogen relationship, and the ecological factors associated with an outbreak. Knowledge of all these factors is needed to proactively protect the public from zoonotic diseases; without this knowledge, new diseases will continue to emerge. The worldwide distribution of these largely zoonotic diseases suggests a globally distributed mechanism for their emergence. Anthropogenic factors-including pollution, land-use conversions, and climate change-likely contribute to disease emergence by several mechanisms (3), one of which has been hypothesized to be decreased species diversity. The number of species currently being lost, as well as the rate of species loss, is unprecedented (4); these losses generally have negative effects on ecosystem functioning (5, 6) . It likely is not coincidental that areas where many zoonoses are emerging among humans are the same areas where loss of species is accelerating, e.g., Central Africa (Ebola, monkeypox, Marburg virus), West Africa (Lassa virus, HIV-2), Southeast Asia (Nipah virus, severe acute respiratory syndrome, avian influenza), and South America (dozens of strains of hantaviruses and arenaviruses). Lyme disease, a vector-borne zoonosis, is affected by loss of species by a process known as the dilution effect (7), whereby increasing species diversity decreases disease prevalence by diluting the availability of competent hosts with increased numbers of noncompetent hosts. Little research on the dilution effect has been carried out beyond its effect on Lyme disease (8) , yet the global implications of the phenomenon-if the effects are applicable to other types of diseases and transmission dynamics-could have substantial and enduring effects on human health and conservation. Hantaviruses provide a model system in which to test the dilution effect in directly transmitted zoonoses. Since their initial discovery in the Western Hemisphere in 1982, several dozen hantavirus strains have been found, each hosted by a unique rodent species (9); novel hantaviruses have recently been discovered in shrews (10, 11) . Natural hosts are asymptomatic and chronically infected; intraspecies spread is hypothesized to be through bites (12) . Humans become infected with hantavirus by inhaling aerosolized excreta from infected rodents (13) . Occasionally hantavirus pulmonary syndrome (14) develops; this syndrome has a mortality rate of almost 40% and no prophylaxis, treatment, or cure (15) . Most of the 506 confirmed cases in the United States have been caused by Sin Nombre virus (SNV). Studies have found that low diversity ecosystems dominated by the rodent hosts for 3 distinct hantaviruses had high infection prevalence in the host (16, 17) , suggesting a role for species diversity. Although the mechanism of disease dilution would differ in directly transmitted zoonoses (e.g., hantaviruses), as opposed to vector-borne diseases, a dilution effect could occur if 1) individuals of the host species remain as species diversity decreases, 2) the disease is spread within the host species through direct encounters (such as biting), and 3) presence of other species causes encounters among the host species to decrease. Other ecological factors could affect the number of intraspecific deer mouse (Peromyscus maniculatus) encounters, including increased density of deer mice and vegetative factors that lead to variation in population numbers (e.g., available cover and forage) ( Table 1 ). Some studies have found high SNV prevalence in host populations when deer mice densities were high (18, 19) . However, although the concept of density-dependent transmission is not unique to hantaviruses, its applicability to the deer mouse-SNV system has been elusive. SNV prevalence also has been shown to vary with habitat characteristics and quality (15, 18, 19) , although interpretation of this variation has been difficult because SNV prevalence varies as much within as among habitat types (20) . In this study we examined small mammal populations in 5 forested sites over a 3-year period, October 2002 through September 2005. We monitored mammal species diversity, deer mouse densities, and SNV infection prevalence in the mammals to test the hypotheses that 1) areas of higher mammal species diversity would exhibit lower prevalence of SNV infection in host populations, 2) areas of higher host density would contain higher infection prevalence of SNV in the host populations, and 3) vegetative factors could be related to prevalence of SNV infection among deer mice. We sampled small mammals at 5 (21) . To sample as many different mammal species as possible, we set up a trapping web 200 m in diameter (22) at each site and used 4 trap types: Sherman (H.B. Sherman Traps, Tallahassee, FL, USA), handmade wire mesh, Tomahawk (Tomahawk Live Trap Co., Tomahawk, WI, USA), and pitfall. Each station included an aluminum folding Sherman live trap and a custom-built mesh live trap (23) of similar dimensions (7.6 cm × 8.9 cm × 22.9 cm). Two sizes of Tomahawk live traps were used to trap larger animals; a 61 cm × 17.8 cm × 17.8 cm trap was placed at each 50-m trap station, and a 91.4 cm × 25.4 cm × 30.5 cm trap was placed at each 100-m trap station. Pitfall traps were made by using a 19-L bucket (30-cm All captured animals were treated as if they were infected with SNV, and standard precautionary methods were implemented (25) . After point of capture was recorded, animals were transferred from traps to sealable plastic bags or, if too large, left in the trap and brought to the center of the web, where they were weighed and measured and examined for age, sex, reproductive status, scarring, or other notable characteristics. Retroorbital blood samples were collected by using heparinized microcapillary tubes and either placed in cryovials and frozen in liquid nitrogen or placed in serum separator tubes and refrigerated for no more than 1 week before testing. Infection prevalence was determined by ELISA (26) . Infected deer mice were counted 1 time (time of first capture). During the first 2 years of the study, to obtain tissue samples for a companion study, deer mice were euthanized in a chloroform chamber (25) . The resulting specimens were tagged and stored at the Museum of Vertebrate Biology at Portland State University. All other animals captured were marked with ear tags and released at the point of capture. During the last year of the study, deer mice were also tagged and released. To determine whether removal affected subsequent capture rates within the same trapping period, the differences between the number of captures on the first and last day of the trapping period were calculated and averaged, then compared between removal and replacement sampling with the Welch 2-sample t-test. Because no significant differences were found between the first 2 years and the last year of the study (t = 0.50, p = 0.63, df = 8), data from all 3 years were analyzed together. This research was conducted under the auspices of federal, state, and city permits, and it complied with the American Society of Mammalogists' guidelines for animal care and use (27) . Deer mouse density was calculated by using the Distance program (28) . Mammal species diversity was measured by using the Simpson diversity index (D S ) (29) , which takes into account both richness (number of species) and evenness (number of individuals within each species) and ranges from 0 (least diversity) to 1 (maximal diversity). D S further represents the probability of interspecies encounters (30) . Pairwise comparisons of D S values among parks was conducted by using the Student t test; differences of D S were divided by the square root of their variances (30) . To minimize the possibility of type 2 errors resulting from multiple comparisons, a statistically conservative Bonferroni correction was made (α = 0.05/10 comparisons, or 0.005) (31) . Deer mouse densities were compared pairwise by using the Welch 2-sample t-test. Logistic regression with binomial errors was initially used to assess the association between infection prevalence and deer mouse density and species diversity. However, the resulting models showed such extensive overdispersion that we considered logistic regression to be an unsuitable statistical method for these data (32) . Accordingly, we used nonlinear regression analysis. An analysis of similarity returns a statistic (R) based on a Bray-Curtis dissimilarity measure, which considers the difference of the mean ranks between and within groups. Most values fall between 0 and 1; 1 is the most dissimilar. Significance is assessed by comparing the observed value of R to the permutation distribution of R (33). Again, because of multiple comparisons, a Bonferroni correction was made such that α = 0.005 (31) . We then used stepwise (backward) logistic regression with binomial errors to assess the association between infection prevalence and vegetative characteristics. Although only 5 sites were examined, the intensity of the sampling yielded a total of 5,057 individuals from 21 species, resulting in a thorough species inventory over a gradient of diversity in small mammal ecological communities. Deer mice averaged 62% of all captures (Table 2) and were the dominant species at all sites. Mammal species diversity differed significantly among sites (p<0.001; Table 2 ), except sites 3 and 4 (p = 0.1). Densities varied spatially and temporally; all parks exhibited the highest densities during year 2 (Table 3) . Interannual variances of densities were large due to seasonal differences in capture rates, such that no statistical differences in densities were found either within or among parks. Infection prevalence also varied, although it remained consistently low at 4 of the 5 sites. During year 1, infection prevalence was significantly higher at site 1 than at sites 2 and 3 (p<0.001) but not different than at sites 4 and 5 (p = 0.20 and 0.32, respectively). Site 1 was the only site where infection prevalence significantly increased between years 1 and 2 (p = 0.005); thus, infection prevalence at this site was significantly higher than at any of the other parks during year 2 (p<0.001). High infection prevalence was maintained at site 1 during year 3. Although the rate for site 2 increased significantly between years 2 and 3 (p = 0.035), prevalence remained significantly higher at site 1 than at any other park during year 3 (p<0.01). Using nonlinear regression, we found a significant negative relationship between infection prevalence and mammal species diversity. Infection prevalence increased as diversity decreased, up to an inflection point where the rate of infection increased exponentially (Figure) . No regression model was able to account for the association between infection prevalence and density of deer mice, either alone or with species diversity in the model. A pairwise analysis of similarity was used to compare sites floristically; all parks differed significantly from each other (p<0.001). Stepwise backward logistic regression with binomial errors found no association between infection prevalence and any vegetative factors alone or in combination with other vegetative factors. Population densities fluctuated synchronously at all sites, yet infection prevalence increased significantly at only 1 site, which suggests that factors other than density alone are involved in disease transmission. If, as hypothesized, transmission were through aggressive encounters (12) , SNV would spread most efficiently in an ecosystem composed solely of deer mice, where every encounter would be a potential disease-transmitting encounter. As more spe-cies, and more individuals within those species, are added to the community, the number of potential disease-transmitting encounters decreases because species other than deer mice are nonhost (not competent, or nonamplifying) species. This type of decreased intraspecies interaction has been termed "encounter reduction" (34) and would occur if increasing species diversity increases the number of competitors in an ecosystem, thereby increasing the amount of time a host species has to spend securing limited resources (food, nest sites), in turn decreasing the time spent on intraspecies encounters. An increase in species diversity, in combination with an increase in the densities of individuals within those species, as we observed in this study, should also mean an increase in the number of predators of the rodent host species. It is reasonable to hypothesize that predators keep rodent numbers under control, in turn limiting pathogen spread both among rodents and into human populations, although it has been difficult to empirically support this hypothesis (35) . Our results suggest that predators control infection prevalence not by controlling the density of host species but instead by an unrelated mechanism, possibly encounter reduction. When predators are present in the ecosystem, host species should spend more time in the nest, in hiding, or within the familiarity of their territory, all to avoid predation and all likely to decrease intraspecies encounters. This hypothesis is supported by the fact that capture ratebut not density-was highest at site 1 during year 2 relative to all other parks (p<0.01), which means that deer mice were moving about and encountering traps more often. We hypothesize that when predation and competition are decreased or absent, for this small mammal community at a Simpson diversity index ≈0.43, a zoonotic release of predatory and competitive controls appears to have occurred, in which SNV infection prevalence increased drastically. This hypothesis would account for the lack of differences in infection prevalence rates at sites 2-5; although the Simpson index for these sites varied significantly; the threshold for zoonotic release had not been breached at any of those sites. Above the threshold level, sites would maintain a low level of infection, or perhaps locally lose infection altogether. In this study, SNV infection prevalence was either so low during some seasons at some sites as to be virtually undetectable by traditional trapping techniques or ephemerally absent. In particular, SNV was undetected or absent most often at the most diverse site (no SNV was detected in 8 of 12 seasons at site 5, in 6-7 seasons at sites 2-4, and in 1 season at site 1). Our system differs from the Lyme disease system, which depends on a vector that is not host specific (black-legged tick) to transmit the disease. Here, in contrast, presence of nonhost species in the small mammal community will not directly affect the transmission of SNV; instead, the behavior of members of the natural host species will be affected, decreasing SNV transmission rates among competent hosts through encounter reduction. Increased diversity in both the Lyme disease and SNV systems appears to lead to decreased disease prevalence, although the mechanisms differ. Another difference between the 2 systems is the threshold relationship between species diversity and SNV prevalence, which suggests that the shape of the dilution curve may be mechanism dependent and is the reason we proposed the term "zoonotic release." Given that many hantaviruses are hosted by generalist rodent species (i.e., those able to exploit a broad variety of ecological resources) that dominate ecosystems as species diversity decreases (e. Host density should likewise be considered a factor in this phenomenon because density increased before infection prevalence increased. However, the result of the logistic regression between density and infection prevalence, although significant, was marked by considerable overdispersion, suggesting that this was the wrong model, and its significance was greatly overestimated (32) . Additionally, at all parks deer mouse density increased but infection prevalence did not, clearly indicating that density is not the sole driver of infection prevalence in this system. A logistic regression with both density and mammal species diversity in the model showed similar overdispersion. Our results suggest that dependence on both density and frequency play a role in SNV transmission, which may be one of the reasons it has been so hard to determine their respective roles in the transmission of hantaviruses (36) . More extensive studies should therefore be undertaken wherein species diversity, density, and frequency of encounters are carefully measured to determine their respective roles in disease transmission. The finding that infection prevalence of a directly transmitted zoonosis may be inversely related to species diversity has implications for human health. The toll in illness and death from emerging zoonotic diseases is high, and outbreak investigations are costly (37) . These investigations often fail to identify the source of a pathogen, let alone answer the question of why an outbreak occurred at a given time and place. If the host species or vector is found, eradication usually is neither possible nor desirable, particularly when the species are as ubiquitous as deer mice. Prophylaxis is difficult when transmission is airborne, as in hantaviruses, for which potentially everyone in a region is at risk. Ecosystem-level control may be the best way to protect the public from the increasing threat of many zoonotic diseases. Wildlife also are at risk for infection with novel pathogens, and the factors underlying wildlife disease emergence are similar to those in humans (38) ; a dilution effect may therefore help protect wildlife as well. For example, a study of West Nile virus suggested that increased bird species richness depressed the prevalence of the virus in ecosystems (39) . Thus, wildlife could be protected in 2 ways: first, from dilution of diseases that are potentially harmful to them and second, from maintenance of healthy ecosystems. Extension of a dilution effect to directly transmitted diseases has implications for conservation as well. Although protecting species diversity is a cause that would seem universal in its appeal, conservationists often are perceived as being overly biocentric and having little concern 1016 Emerging Infectious Diseases • www.cdc.gov/eid • Vol. 15 for human welfare. In addition, many benefits derived from maintaining diverse ecosystems are difficult for the layperson to decode and seem far removed from daily life such that despite scientific research, unparalleled loss of species caused by anthropogenic factors continues at an unabated rate. Conservation likely will not succeed without the support of the general public, who in turn influence the environmental policies our society embraces. To gain support of the general public, tangible human benefits from conservation should outweigh the immediate-usually economicgains of nonconservation land use (40) . Linking human health to biodiversity could be just the benefit for gaining the public's support of conserving biodiverse ecosystems. Protection from disease is a tangible objective; it is easily understood and translated and it has direct benefits for all. As a consequence, extension of a dilution effect to directly transmitted diseases could have broad conservation implications by raising the public's concern about conservation in a manner that has yet to be emphasized.

Chinese-like Strain of Porcine Epidemic Diarrhea Virus, Thailand

Since late 2007, several outbreaks of porcine epidemic diarrhea virus (PEDV) infection have emerged in Thailand. Phylogenetic analysis places all Thai PEDV isolates during the outbreaks in the same clade as the Chinese strain JS-2004-2. This new genotype PEDV is prevailing and currently causing sporadic outbreaks in Thailand.

, is an enveloped, single-stranded RNA virus belonging to the family Coronaviridae. The PEDV genome contains genes for the following proteins: pol1 (P1), spike (S) (180-220 kDa), envelope (E), membrane (M) (27-32 kDa), and nucleocapsid (N) (55-58 kDa) (2) . The M protein is a structural membrane glycoprotein, which plays an important role in the assembly process; the S surface glycoprotein harbors the specific host cell receptor binding sites (3) . During late 2007, the PED outbreak appeared first in Nakornpathom province before spreading throughout the country. Pig losses from the recent PED outbreaks were extensive. Obvious clinical signs were severe diarrhea (Figure 1 , panel A) and dehydration with milk curd vomitus in suckling piglets. Most of the affected farms reported the disease first in farrowing barns and subsequently lost 100% of newborn piglets. Pigs of all ages were affected and exhibited degrees of diarrhea and inappetite, which varied by their ages. Boars and sows had mild diarrhea and anorexia for a few days and recovered within a week. In piglets that died, the small intestinal wall was congested and intestinal contents were watery with undigested milk curd ( Figure 1, panel B). Segmental enteritis was indicated by segmental disappearance of intestinal lacteal caused by malabsorption in affected intestinal parts ( Figure 1 , panels C and D). Atrophic enteritis, characterized by blunting of the intestinal villi and sloughing of intestinal epithelium, occurred in all affected piglets (Figure 2 , panel A). Immunohistochemical tests, performed by using monoclonal anti-PEDV S protein (JBT Biotechnology Laboratory, Seoul, South Korea), demonstrated dark brown staining in intestinal epithelial cells (Figure 2, panel B) . Massive feedback of piglet feces and minced piglet guts to gestating sows was recommended by local veterinary practitioners to prime the sow's immune response and pass protective immunity to the piglets. At affected farms, the outbreak lasted <3 weeks. Samples from 8 provinces (24 farms) in Thailand from December 2007 through March 2008 were submitted to the veterinary diagnostic laboratories of Kasetsart University and Chulalongkorn University. A total of 33 porcine samples were confirmed as positive for PEDV by reverse transcription-PCR (RT-PCR) (4) before virus isolation (Table) . Published primers (5) were used for generating the PEDV 651-bp partial S gene. Primers were designed to amplify the PEDV M gene and yielded the amplified product of 715 bp on the basis of CV777 and Br1/87. Products were purified by using a QIAquick Gel Extraction Kit (QIAGEN, Hilden, Germany) and were sequenced by 1st BASE Pte Ltd (Singapore). Nucleotide and deduced amino acid sequences of the 33 PEDV isolates were aligned, edited, and analyzed with ClustalX version 1.83, Bioedit version 7.0.5.2, and MegAlign software (DNAStar Inc., Madison, WI, USA), respectively. Phylogenetic trees were generated by using partial S and full-length M genes, including the deduced amino acid sequences with selected reference PEDV strains, by applying the Jotun Hein method in the MegAlign software. To assess the relative support for each clade, bootstrap values were calculated from 1,000 replicate analyses. The M gene sequence analysis of 31 PEDV isolates obtained in Thailand indicated that the nucleotide sequence of the entire M gene was highly conserved. All recent PEDV isolates in Thailand had 99.3%-100% nucleotide homology. The lowest sequence identity (96.5%) was with the Chinese strain, EF185992/LZC, and the highest sequence identity (99.2%-99.7%) was with the Chinese strain, JS-2004-2, and concurrent isolates from the National Institute of Animal Health, Thailand, M_NIAH 07-08 (data not shown). All 33 PEDV isolates had 97.0%-98.8% DNA sequence identities of the S gene with each other. Our findings demonstrated that the recent PEDV isolates in Thailand were genetically diverse in their S genes either within Our results indicated that the recent Thai PEDV isolates clustered in the same group were highly homologous with the Chinese strains, JS-2004-2 and LJB/03. They were responsible for the recent PED outbreak in Thailand and able to produce pathologic effects similar to the Chinese isolates (6) . Notably, 08NP04, isolated 4 months after the first outbreak, had the highest identity to JS-2004-2. Also, 08NP04 and 07NP01 (the first isolates in 2007) originated in the same geographic area. The Chinese-like strain of the virus might have gained entry into Thailand via unknown routes as early as December 2007. In addition, rendering trucks traveling from farm to farm might have encouraged widespread transmission of the disease. Structural differences in the partial S gene could help elucidate pathogen- esis and antigenic structures of the recent PEDV isolates because S glycoproteins are responsible for inducing the virus neutralizing antibodies and known to be highly conserved in PEDV strains (7) . Continuing investigation of PEDV isolates will contribute to the prevention and control of this virus. The phylogenetic relationship of the Thai PEDV strain indicated that the recent Thai PEDV isolates differed genetically from previous Thai isolates. Despite precautions, sporadic outbreaks continue to occur. In addition, disease transmission frequently occurs due to the purchase of new stock with improper gilt acclimatization and biosecurity. Immunity induced through vaccination, currently unavailable in Thailand, does not provide lifelong protection from this virus. However, vaccination is recommended to encourage specific immunity to PEDV in all stock when an acute outbreak occurs. Undoubtedly, effective biosecurity is a key management tool for PED prevention and control.

Optimism/pessimism and health-related quality of life during pregnancy across three continents: a matched cohort study in China, Ghana, and the United States

BACKGROUND: Little is known about how optimism/pessimism and health-related quality of life compare across cultures. METHODS: Three samples of pregnant women in their final trimester were recruited from China, Ghana, and the United States (U.S.). Participants completed a survey that included the Life Orientation Test - Revised (LOT-R, an optimism/pessimism measure), the Short Form 12 (SF-12, a quality of life measure), and questions addressing health and demographic factors. A three-country set was created for analysis by matching women on age, gestational age at enrollment, and number of previous pregnancies. Anovas with post-hoc pairwise comparisons were used to compare results across the cohorts. Multivariate regression analysis was used to create a model to identify those variables most strongly associated with optimism/pessimism. RESULTS: LOT-R scores varied significantly across cultures in these samples, with Ghanaian pregnant women being the most optimistic and least pessimistic and Chinese pregnant women being the least optimistic overall and the least pessimistic in subscale analysis. Four key variables predicted approximately 20% of the variance in overall optimism scores: country of origin (p = .006), working for money (p = .05); level of education (p = .002), and ever being treated for emotional issues with medication (p < .001). Quality of life scores also varied by country in these samples, with the most pronounced difference occurring in the vitality measure. U.S. pregnant women reported far lower vitality scores than both Chinese and Ghanaian pregnant women in our sample. CONCLUSION: This research raises important questions regarding what it is about country of origin that so strongly influences optimism/pessimism among pregnant women. Further research is warranted exploring underlying conceptualization of optimism/pessimism and health related quality of life across countries.

The psychosocial constructs of optimism and pessimism have been under study for several decades. Optimism is associated with more active coping strategies, lower levels of psychological distress [1] [2] [3] [4] [5] , health-enhancing behavior [6] , higher immune functioning [7] , better health outcomes [8] and even lower mortality. [9] [10] [11] On the other hand, pessimism has been shown to have prophylactic effects in certain circumstances. In particular, pessimism can insulate people from the psychological consequences of failure, including anxiety, depression, and diminished self-esteem. [12] Thus, the impact of optimism and pessimism is potentially enormous yet still very unclear. In this context, even less is known about differences across cultures. While several studies have shown levels of optimism/pessimism to vary across cultures, [1, [13] [14] [15] [16] [17] [18] [19] findings have been inconsistent in terms of which cultural groups are more or less optimistic. No research to date has compared Asian, African, and Western cultures in the same study. Recent research among pregnant women has yielded interesting, albeit similarly inconsistent, findings. Lobel et al. [8] found that women who were the least optimistic had babies of the lowest birthweight, even when controlling for gestational age. Moyer et al. [20] found that optimism/pessimism among Ghanaian pregnant women was inversely associated with knowledge of HIV and previous HIV testing. In other words, those who were not tested prior to their pregnancy and had the least knowledge of HIV were the most optimistic. The authors also found that, when compared to a similarly aged sample of nonpregnant women in the United States (U.S.), the Ghanaian women were significantly more optimistic. [20] The construct of health-related quality of life is one variable that has been linked to optimism/pessimism in past research. [21] In one study, researchers found that, even when health status was controlled, pessimists had significantly worse health-related quality of life (HRQOL) scores than optimists or so-called "realists." [21] In that study, pessimists were those who expected disproportionately negative outcomes associated with their Hepatitis C diagnosis, optimists were those who expected few negative outcomes associated with their Hepatitis C diagnosis, and realists were those who had a fairly accurate perception of the impact Hepatitis C was going to have on their lives. Optimists' HRQOL scores in this study mirrored the scores of the general U.S. population, even though the population being studied (chronic hepatitis C patients) has been shown to have significantly lower QOL than the general population. The results of the research to date suggest that further examination of the cross-cultural issues in optimism/pes-simism and health-related quality of life among pregnant women is warranted. This research was undertaken to explore the differences among a three-sample matched cohort of pregnant women at the same stage in their pregnancies in Ghana, China, and the U.S. The specific aims of this research were to 1) identify if and to what extent optimism and pessimism vary across similar populations of pregnant women in three different countries; 2) determine if and to what extent self-assessed quality of life scores vary among similar populations of pregnant women in three different countries, and 3) determine if optimism and/or pessimism is predictive of or associated with current self-perceived health status and/or selfassessed Health-Related Quality of Life (HRQOL) and how that might vary by culture. China Data were collected from women presenting for prenatal care at the obstetric outpatient clinic at the Peking University First Hospital between May and July 2006. As one of the largest and most well-known academic medical centers in Beijing, Peking University First Hospital draws both public and private patients from in and around Bejing. Clinics average 600 pregnant women per week and 3000-3500 deliveries per year. Data were collected from women presenting for prenatal care at the Obstetrics and Gynecology Clinic at the Noguchi Research Institute/Medical School at the University of Ghana in Accra, Ghana, between May and July 2005. This facility is housed in a public hospital that is also the largest government hospital in Ghana. Patients from all over Ghana travel to this clinic to receive their care. Clinics average 500-600 pregnant women per week and 10,000 -12,000 deliveries per year. At all three sites, pregnant women in their last trimester of pregnancy who were 18 years old or older were asked to participate in this research. Women facing an imminent health crisis or those in active labor were excluded. At all three sites, research assistants talked patients through an informed consent form. Translators were used when nec-essary. Surveys were administered verbally to all participants in Ghana as part of a larger study. [20] In China and in the U.S., surveys were designed to be self-administered, but women were given the option to have the survey administered verbally. (None chose this option.) The instruments used for the survey in each study location varied slightly, but for the purposes of this analysis, each study site used three key instruments -a demographic questionnaire, the Life Orientation Test (LOT-R), and the Short form 12, or SF-12. The instruments were pilot tested separately in each location, and minor modifications were made to ensure comprehension and comparability across sites. In China and Ghana, the instruments were translated into the dominant language of the region and backtranslated into English by native bi-lingual speakers. The original and back-translated versions were compared for consistency, and any inconsistencies were resolved by discussion and consensus among the research team. A Demographic and Health Questionnaire was used to measure patient characteristics that may be associated with optimism, pessimism, HRQOL, or pregnancy outcomes. These include age, number of pregnancies, other medical conditions, previous treatment for mental health problems such as depressed mood or anxiety, previous use of anti-depressants, and self-perceived health status. Women were also asked to rate their perception of the difficulty of their pregnancy on a scale of 1 to 4, with 1 being "extremely easy" and 4 being "extremely difficult." The Life Orientation Test -Revised (LOT-R) [22] is a revision of the original Life Orientation Test [23] . It assesses optimism/pessimism using a series of questions that inquire about an individual's attitudes in daily life. This instrument has been widely validated [24] and used in both Ghana [25] and China. [14] Its 10 items generate an overall score, as well as two possible subscales: affirmation of optimism and affirmation of pessimism. The participant answers each item based on a 5-point scale, with response options ranging from strongly disagree to strongly agree. A higher score relates to a greater level of optimism. The Short Form 12, or SF-12, [26] is a 12-item quality of life instrument derived from the Short Form 36, or SF-36, an instrument used and validated around the world to determine self-assessed health-related quality of life (HRQOL). The SF-36 and shortened SF-12 generate not only summary scales of mental and physical functioning (MCS and PCS), but also a profile of patients' HRQOL across eight domains: physical functioning (PF), role physical (RP), bodily pain (BP), general health (GH), vitality (VT), social functioning (SF), role emotional (RE), and mental health (MH). According to SF-36 developers, 12 of the original SF-36 items accounted for at least 90% of the variance in PCS-36 and MCS-36 in both general and patient populations, and those same 12 items reproduced the profile of the eight SF-36 health concepts sufficiently for studies in which the length of the instrument may be prohibitive. All research protocols and survey instruments were reviewed and approved by the institutional review boards at the U.S. and foreign institutions participating. Pregnant women presenting for prenatal care were approached and asked to participate in a research study. If they expressed an interest, women were asked how far along they were in their pregnancies. Those in their final trimester were taken through a consent form. A research assistant (and translator, when necessary) answered any questions the participant might have and made sure women had a copy of the consent form to keep. Data were gathered using paper and pencil forms (China, U.S.) and verbal interviews (Ghana). No identifying information was collected from Ghanaian participants, given that collecting post-delivery follow-up data was anticipated to be extraordinarily difficult and thus was not attempted. Hospital registration numbers were collected from Chinese and U.S. patients to allow for post-delivery follow-up for a separate research protocol than that described here. Hospital registration numbers were removed from the original survey and replaced with a unique ID number once the registration number was recorded in a separate location for follow-up purposes. Responses from the hard copies of the self-administered surveys (China and the U.S.) and the interviewer-administered surveys (Ghana) were entered into an Excel spreadsheet and cleaned. Cleaned data from each site (China, N = 251; Ghana, N = 101; U.S., N = 311) was combined into one large dataset. A data subset was created by matching women on three key variables: maternal age, number of weeks pregnant at the time of enrollment, and number of previous pregnancies. Women were matched within 3 years of age (average age difference across matched sets = 0.85 years); within 1 pregnancy (average difference in number of pregnancies across matched sets = 0.48); and within 5 weeks of gestational age (average difference in gestational age across matched sets = 2.66 weeks). This matching schema was undertaken to attempt to create a sample that was as similar as possible on key variables that could influence optimism/pessimism. Previous research has suggested that age can be related to optimism/pessimism [27] , as well as number of previous deliveries [20] . We also postulated that women at different stages of pregnancy might feel differently in terms of optimism/pessimism, and thus wanted to reduce the impact of stage of pregnancy on our results. Given the one child policy in China, "number of previous pregnancies" was used as a key variable rather than number of previous deliveries. This resulted in a matched sample of 168 women, 56 in each country, that were predominantly nulliparous. We refer to these threeway matches as "matched sets" throughout the manuscript. Frequencies and descriptive statistics were calculated. Means across the three groups were compared using ANOVA with post-hoc pairwise comparisons. In addition, multivariate linear regression analysis was used to create a model to identify the factors most strongly contributing to the overall LOT-R scores. Variables that were significantly associated with the LOT-R in univariate analysis were entered into the multivariate model. The model was reworked using only those variables that maintained significance until the best-fit model was identified. Interactions between key variables were examined as well. A pvalue of .05 was taken to be statistically significant. Table 1 illustrates the demographic characteristics of women in the sample. Note that two of the three variables upon which women were matched (age and number of previous pregnancies) were not significantly different (see Table 2 ). In fact, the average difference in age across the matched sets was 0.857 years, with a maximum difference of 3 years. The average difference in number of previous pregnancies was .482, with the highest difference being 2.0, reflected in only 4 of the 56 sets. Women from China were enrolled significantly later in their pregnancies than women in the U.S. (p = .04, mean difference of 1.52 weeks), yet the average difference in gestational age across the matched sets was 2.66 weeks. All of the women in the Ghanaian sample were married, as were a majority from the China sample (98.2%), and in the U.S. sample, a smaller yet significant percentage of the women (76.8%) were married. Educational variables were assessed slightly differently in Ghana than in China and the U.S., making it difficult to compare across countries. That said, it is clear that the U.S. sample included women with higher levels of education than both China and Ghana (p < .001). Ghanaian women in our sample had Table 2 illustrates the health-related variables assessed across the three samples. Note that while the number of pregnancies was not significantly different across the three groups, the number of previous deliveries was significantly different at p < .001. Women from our sample in China reported having significantly easier pregnancies than women from our Ghanaian and U.S. samples (p < .001), and far more women in the U.S. sample report ever seeing a healthcare professional for help with emotional issues (p < .001). Table 3 ). Both the LOT-R optimism subscale and the LOT-R pessimism subscale indicate significant differences across country samples as well. (See Figure 1 .) Figure 2 illustrates differences in self-assessed healthrelated quality of life across the samples. Physical Functioning (PF) was significantly different across the country samples (ANOVA, p = .049), but the difference was primarily between Ghanaian and U.S. samples (post-hoc pairwise comparisons, p = .038). General Health (GH) was significantly different as well (ANOVA, p = .001), and here the difference was primarily between Chinese and Ghanaian samples (p = .001). Vitality (VT) showed significant differences across country samples as well (ANOVA, p < .001), with pronounced differences between Chinese and U.S. samples (p < .001) and Ghanaian and U.S. samples *Ns may total more than 56. Two separate questions were asked: "Have you ever been treated... yes/no" and "Are you currently being treated...yes/no " Thus the number "currently being treated" is a subset of the number who have "ever been treated." ANOVAs were conducted using the "ever treated" and "never treated" groups for comparison. (p < .001), but not Chinese and Ghanaian samples. Finally, Role Emotional (RE) scores were also significantly different across country samples (ANOVA, p < .001), but the difference was most pronounced between Ghanaian and U.S. samples (p < .001). Among all the women across the three samples, univariate analysis indicated that overall LOT-R optimism/pessimism scores were significantly associated with country of origin (p = .001; Ghanaian women sampled had the highest scores, U.S. women sampled had scores in the middle, Chinese women sampled had the lowest scores), educational attainment (p = .046; optimism increased with education level), whether women worked for money (p = .006; working women were more optimistic than nonworking women), and number of previous deliveries (p = .004; those with the fewest and the most deliveries were the most optimistic -those with 2 previous deliveries were the least optimistic). Women who had never seen nor were currently seeing a health care provider for emotional issues were more optimistic than those who had ever seen or were currently seeing a health care provider for emotional issues. (p = .002, p = .001 respectively) Treatment for mental health issues with medication was another variable that was associated with LOT-R scores: those ever treated and those currently being treated with medication were significantly less optimistic than those who had never been treated. (p < .001, p = .005 respectively) Women's self-reported experience of the ease or difficulty of the current pregnancy was also associated with LOT-R scores (p = .007). Those who reported having the easiest pregnancies had the highest levels of optimism. LOT-R scores were positively correlated with SF-12 mental health summary scores (p = .001), vitality subscale scores (p = .041), and mental health subscale scores (p < .001). The LOT-R optimism subscale was significantly associated with the same scales (MCS, p = .018; VT, p < .001; MH, p = .001), as well as the general health subscale (GH, p = .002). The LOT-R pessimism subscale was associated with the mental health summary score (p = .013), the role physical subscale (p = .032), the role emotional subscale (p = .034), and the mental health subscale (p = .038). Note that LOT-R scores were not associated with the selfreported presence of any ongoing health issues. Multivariate linear regression analysis indicated four key variables that predicted approximately 20% of the variance (adjusted R square = .199) in overall LOT-R scores: country of origin (p = .015), working for money (p = .025), level of education (p = .001), and ever being treated for emotional issues with medication (p = .002). No significant interactions were identified. These data suggest that in a three-country cohort of pregnant women matched on age, number of weeks pregnant, and number of previous pregnancies, significant differences existed with regard to both optimism/pessimism and health-related quality of life. According to the LOT-R, Ghanaian pregnant women in our sample were the most optimistic and Chinese women in our sample were the least optimistic of the matched sets. Interestingly, Ghanaian women sampled also reported higher levels of pessimism than Chinese and U.S. women sampled. This is consistent with research that suggests optimism and pessimism may not be mutually exclusive constructs. [28] Physical functioning, general health, vitality, and role emotional subscales of the SF-12 also indicate significant differences by country per our samples -U.S. women sampled had the lowest physical functioning and vitality scores, Ghanaian women sampled had the highest physical functioning and general health scores, and Chinese women sampled had the lowest general health scores and highest vitality scores. LOT-R Optimism/Pessimism Scores by Country (*ANOVA: p < 0.001) Figure 1 LOT-R Optimism/Pessimism Scores by Country (*ANOVA: p < 0.001). LOT-R = overall optimism; higher score = greater level of optimism LOT-OPT = optimism subscale; higher score = greater level of optimism LOT-PESS = pessimism subscale; higher score = greater level of pessimism Correlates of optimism/pessimism included country of origin, educational attainment, working for money, number of previous deliveries, and having ever seen a healthcare provider for emotional issues. Self-reported ease or difficulty of the pregnancy was also associated with LOT-R scores. Perhaps most striking in these findings is that we were able to create a multivariate linear regression model that predicted approximately 20% of the variance in overall LOT-R scores. The key variables were country of origin, working for money, level of education, and ever being treated for emotional issues with medication. What is particularly fascinating about this model is that country of origin somehow remains a strong factor in the modeleven with variables that might be seen as proxies for country of origin (education, working for money, ever being treated for emotional issues). It raises a question about the potential differences that underlie being Ghanaian, Chinese, or American that yield differential LOT-R scores. It also raises questions about what accounts for the remaining 80% of the variance. Further research with more comprehensive instrumentation is necessary to begin to answer these questions. There are several limitations to this research. First, the study design does not allow for the definitive determination of a causal relationship -we can only report observed associations. Second, in each setting, the women recruited represented a convenience sample from one hospital in one city, minimizing the ability to generalize based upon our findings. Third, we created our matched sets based on three variables we believed were most important to control for at the outset: maternal age, gestational age, and number of previous pregnancies. Given the one child policy in China, we did not match on number of previous deliveries -as that would have eliminated all but the nulliparous women in Ghana and the U.S. and reduced our sample size to an untenably low level. However, upon analysis it was discovered that the number of previous deliveries was significantly different across countries. In China, women were more likely to have had a previous pregnancy that did not result in a delivery. Whether these were induced or spontaneous terminations is not known, although previous research indicates that more than half of women in China have had at least one abortion. [29] It is possible that Chinese women who have had the same number of pregnancies as their African and North American counterparts but have not had as many deliveries are in some way inherently different. Our univariate analysis did indicate that number of deliveries was associated with overall LOT-R scores (p = .004), but this significance was not sustainable in the multivariate logistic regression model. Note as well that the samples differ significantly on a few key variables. First is in terms of educational attainment. In Ghana and China, approximately a third or respondents were high school graduates or less, compared to only seven percent in the U.S. Our data suggest that educational attainment is associated with optimism/pessimism and health-related quality of life. What is less clear is through what mechanism. Further research is needed to disentangle these associations. The second key difference in the samples is related to their reports of current and previous treatment for emotional issues. While nearly 40% of U.S. women in our sample reported ever seeking care for emotional issues, only 5% of Chinese women in our sample said the same. While this may reflect true differences in need for mental health care, more likely it reflects social norms and access issues surrounding treatment for mental health issues. Nonetheless, such differences are indicative of the likely many other variables that would differ across cultures that were not measured in this study and that might have an effect on optimism/pessimism or selfassessed health-related quality of life. Finally, the samples were likely to be very different in terms of standard of living. Women in Ghana were recruited from a public hospital, suggesting lower income participants, whereas women in China and the U.S. were either 'public' or 'private,' depending upon the type of insurance being used. We attempted to assess income to sort out these differences in a meaningful way, but there were numerous challenges. The first is the distinction between family and individual income. In China, family income may or may not include extended family. There may be four -or even six -adults in one household, awaiting the arrival of a single child. In Africa, family income was challenging for some respondents to answer, whether due to lack of knowledge or the challenge of translating non-currency income into survey response options. In addition, assuming for a moment that our survey instrument accurately collected and recorded a representative gradient of incomes in each culture, how does one compare being 'low income' in the U.S. versus being 'low income' in sub-Saharan Africa? The combination of missing data and uncertain interpretation of our income-related items made meaningful comparisons challenging at best, misleading at worst, thus income data were removed from analysis. Another limitation to this research is that we were unable to suitably control for or verify "difficult" pregnancies. We did not have independent corroboration of women's reported symptoms or reported ratings of the ease or dif-ficulty of their pregnancies. Our sole measure of the ease or difficulty of women's pregnancies was their self-report. That said, the LOT-R measures what is referred to as dispositional (trait) optimism/pessimism -or something that is seen as relatively stable over time. Other measures of situational (state) optimism may have been more responsive to changes over time and may have been influenced by easy or difficult pregnancies. However, for the purposes of this research, we believe that a measure of dispositional optimism/pessimism is the most appropriate and that it should not have been differentially influenced by women's pregnancies. This is corroborated by the finding that despite reporting the easiest pregnancies of the three groups of women, Chinese women in our sample reported the lowest levels of optimism. Another limitation to this research is the uncertainty surrounding the comparability of the constructs of optimism/pessimism across cultures. In a review of Chinese folk wisdom of behavioral health, researchers cite that for Chinese subjects, being optimistic means to be able to accept one's current life conditions positively rather than to expect good things to occur in one's life. [30] This is a more present-focused interpretation of optimism, rather than the future-focused interpretation most commonly used in the West. In the West, optimism is defined as the expectation that good things will happen to you in the future. Given these potentially contradictory definitions, future research needs to identify and clarify definitions of optimism and pessimism as commonly understood in different areas of the world. Similarly, it is unclear how Ghanaians would articulate the concept of optimism. Research in Africa has shown that optimism has been inversely correlated to income and a number of other indicators of higher standards of living and positively correlated with a number of variables associated with deep poverty. [31] This finding -if realraises several critical questions. Are those who are optimistic in the midst of poverty the ones that survive? Is optimism a result of the perception that if one is extremely poor it can't get much worse, so it must be getting better? Are those with higher incomes and better standards of living so afraid of losing it all that they steel themselves against that potential by assuming the worst? Clearly more research is needed to begin to tease out some of these issues. One final limitation is the inconsistency of data collection methodology across sites. Data were colleted verbally in Ghana, whereas data were collected via self-reported surveys in China and the U.S. We do not believe this difference significantly impacted the outcome variables in question, however it is possible there may be differences attributable to the method of data collection. In conclusion, these results raise some critical questions worthy of further exploration. Rigorous studies of crosscultural differences are not only methodologically, but also logistically, very challenging. That said, research needs to address the question of whether the underlying constructs of optimism/pessimism and health related quality of life are conceptualized similarly across cultures. In addition, further research is warranted that explores the potential link between psychosocial constructs such as these and measurable health outcomes. Only then can researchers and practitioners explore the possibility of interventions to influence what have typically been seen as stable constructs.

Emerging Technologies for the Detection of Rabies Virus: Challenges and Hopes in the 21st Century

The diagnosis of rabies is routinely based on clinical and epidemiological information, especially when exposures are reported in rabies-endemic countries. Diagnostic tests using conventional assays that appear to be negative, even when undertaken late in the disease and despite the clinical diagnosis, have a tendency, at times, to be unreliable. These tests are rarely optimal and entirely dependent on the nature and quality of the sample supplied. In the course of the past three decades, the application of molecular biology has aided in the development of tests that result in a more rapid detection of rabies virus. These tests enable viral strain identification from clinical specimens. Currently, there are a number of molecular tests that can be used to complement conventional tests in rabies diagnosis. Indeed the challenges in the 21st century for the development of rabies diagnostics are not of a technical nature; these tests are available now. The challenges in the 21st century for diagnostic test developers are two-fold: firstly, to achieve internationally accepted validation of a test that will then lead to its acceptance by organisations globally. Secondly, the areas of the world where such tests are needed are mainly in developing regions where financial and logistical barriers prevent their implementation. Although developing countries with a poor healthcare infrastructure recognise that molecular-based diagnostic assays will be unaffordable for routine use, the cost/benefit ratio should still be measured. Adoption of rapid and affordable rabies diagnostic tests for use in developing countries highlights the importance of sharing and transferring technology through laboratory twinning between the developed and the developing countries. Importantly for developing countries, the benefit of molecular methods as tools is the capability for a differential diagnosis of human diseases that present with similar clinical symptoms. Antemortem testing for human rabies is now possible using molecular techniques. These barriers are not insurmountable and it is our expectation that if such tests are accepted and implemented where they are most needed, they will provide substantial improvements for rabies diagnosis and surveillance. The advent of molecular biology and new technological initiatives that combine advances in biology with other disciplines will support the development of techniques capable of high throughput testing with a low turnaround time for rabies diagnosis.

clinical and epidemiological information, especially when exposures are reported in rabies-endemic countries. Diagnostic tests using conventional assays that appear to be negative, even when undertaken late in the disease and despite the clinical diagnosis, have a tendency, at times, to be unreliable. These tests are rarely optimal and entirely dependent on the nature and quality of the sample supplied. In the course of the past three decades, the application of molecular biology has aided in the development of tests that result in a more rapid detection of rabies virus. These tests enable viral strain identification from clinical specimens. Currently, there are a number of molecular tests that can be used to complement conventional tests in rabies diagnosis. Indeed the challenges in the 21st century for the development of rabies diagnostics are not of a technical nature; these tests are available now. The challenges in the 21st century for diagnostic test developers are two-fold: firstly, to achieve internationally accepted validation of a test that will then lead to its acceptance by organisations globally. Secondly, the areas of the world where such tests are needed are mainly in developing regions where financial and logistical barriers prevent their implementation. Although developing countries with a poor healthcare infrastructure recognise that molecularbased diagnostic assays will be unaffordable for routine use, the cost/benefit ratio should still be measured. Adoption of rapid and affordable rabies diagnostic tests for use in developing countries highlights the importance of sharing and transferring technology through laboratory twinning between the developed and the developing countries. Importantly for developing countries, the benefit of molecular methods as tools is the capability for a differential diagnosis of human diseases that present with similar clinical symptoms. Antemortem testing for human rabies is now possible using molecular techniques. These barriers are not insurmountable and it is our expectation that if such tests are accepted and implemented where they are most needed, they will provide substantial improvements for rabies diagnosis and surveillance. The advent of molecular biology and new technological initiatives that combine advances in biology with other disciplines will support the development of techniques capable of high throughput testing with a low turnaround time for rabies diagnosis. Validated diagnostic tests that confirm the presence of rabies virus or a lyssavirus variant have been the foundation of rabies control strategies in many countries. Historically, histopathological techniques such as the Sellers Stain technique [1] were used to determine the presence of Negri bodies as rabies virus-specific antigen, however due to poor sensitivity and specificity this technique is no longer recommended by the World Health organization (WHO). The Fluorescent Antibody test (FAT) [2] relies on the ability of a detector molecule (usually fluorescein isothiocyanate) coupled with a rabies specific antibody forming a conjugate to bind to and allow the visualisation of rabies antigen using fluorescent microscopy techniques. Microscopic analysis of samples is the only direct method that allows for the identification of rabies virus-specific antigen in a short time and at a reduced cost, irrespective of geographical origin and status of the host. It has to be regarded as the first step in diagnostic procedures for all laboratories. Autolysed samples can, however, reduce the sensitivity and specificity of the FAT. The Rabies Tissue Culture Infection Test (RTCIT) [3] and the Mouse Inoculation Test (MIT) [4] are based on the propagation and isolation of the virus. These diagnostic tests are used to detect virus particles either directly in tissue samples (FAT) or indirectly in animals and in tissue culture (MIT and RTCIT, respectively). The rationale for the use of virus isolation (RTCIT/MIT) from a sample where there is a suspicion of infection with rabies virus is always recommended, especially when Koch's postulates are likely to be met. Such amplification of the viral pathogen facilitates additional molecular analysis to be undertaken, including sequencing of the viral isolate and subsequent phylogenetic analysis. Conventional diagnostic tests for rabies (FAT, RTCIT, MIT) are not labour intensive and rely upon low throughput. The FAT can be completed in less than two hours. In contrast, both the RTCIT and MIT require longer turnaround times (4-days and 28-days, respectively). The fluorescent antibody virus neutralisation (FAVN) test [5] and the Rapid Fluorescent Focus Inhibition Test (RFFIT) [6] utilise a similar principle, to measure the level of virus neutralising antibody in vaccinated individuals. 'Indirect' serological methods, including the FAVN and RFFIT measure the host response to infection/vaccination only and do not detect the presence of infectious virus/antigen directly. However, host antibody detection (FAVN/RFFIT) is an indirect tool to measure the presence of rabies virus in a non-immunised individual by evaluating the host response to infection. The test may lack sensitivity and specificity, and the interpretation of the test results may be difficult as the host response to infection varies substantially between individuals. As such, the negative predictive value of serological tests for rabies diagnosis is considered poor. Therefore, serological assays are not suitable as diagnostic tools for routine rabies testing. These internationally approved methods have provided accurate and timely information of animal and human rabies cases thereby supporting surveillance for rabies and providing a greater understanding of the epidemiology of this disease (Box 1). For numerous laboratories in rabies-endemic regions in the developing world, cost and simplicity are critical factors in the delivery of disease diagnosis and cannot be neglected, even when the principal consideration is for rapid diagnosis. Therefore, cost and simplicity need to be considered if new technologies are to be adopted in the regions of the world where they are most needed. Molecular tools based on the detection and manipulation of the genetic information of the virus are becoming more widely accepted and accessible for the diagnosis of rabies. The advent of molecular biology is changing the face of diagnostic virology generally enabling high throughput and short turnaround-time analysis of samples. In the 21 st century, it is expected that diagnostic virology techniques for high throughput rabies virus detection will progress rapidly towards the use of molecular diagnostics replacing more conventional testing techniques such as virus isolation and histopathology. It is also possible that immunological tests, measuring 'surrogate' markers such as cytokines and electrolytes, will augment the standard diagnostic approach; nevertheless they will continue to remain oddities outside the realms of the routine diagnostic laboratory and be confined to a few reference laboratories. Semi-automated or automated instruments and robotics-based techniques are useful when large numbers of the same test are undertaken and these tests will continue to increase in popularity and use, especially in central reference laboratories rather than in each local or regional facility. New technological advances will undoubtedly be faster, more accurate and may, in time, offer a cost-effective alternative to traditional rabies diagnostic tests. These paradigm shifts including modern advances in technology will lead to the effective control of rabies in animals and wildlife [7] (Box 2). This review provides information on some of the latest developments and diagnostic techniques for determining the presence of rabies virus or nucleic acid in diagnostic samples. The principal focus of this review is to highlight the new developments in virology related to techniques for the diagnosis and surveillance of rabies. Literature reviews were identified through Web of Science, PubMed and Scopus using various search phrases. This review also drew on information provided to international organisations, mainly WHO and OIE, funded by the UK Department for Environment, Food and Rural Affairs (Defra) in an advisory context on diagnostic and surveillance strategies for rabies. This review however, does not reflect the views of Defra, WHO or OIE. This review provides information on some of the latest developments and diagnostic techniques for determining the presence of rabies virus in diagnostic samples. Our aim is to provide a viewpoint on the current thinking in diagnostic virology for rabies, reflecting the 'neglected' nature of this tropical disease and the contrasting needs of diagnostic laboratories in developed and developing countries. Development of Rapid Immunohistochemical Test (dRIT) in the evaluation of suspect rabies tissue samples A direct Rapid Immunohistochemical Test (dRIT) for the postmortem detection of rabies virus antigen in brain smears has been developed [8] . Using a cocktail of highly concentrated and purified biotinylated monoclonal antibodies, rabies antigen can be detected by direct staining of fresh brain impressions within 1 hour. This test employs anti-rabies monoclonal antibodies specific for the nucleoprotein, a viral protein produced in abundance during productive infection. The FAT is based on antibodies specific for the same protein but, being conjugated to fluorescein isothiocyanate, requires a fluorescent microscope to visualise any specific antibody bound to viral protein within the test sample [2] . In contrast, the new Box 1. Key Learning Points dRIT antibody cocktail is biotinylated such that following a short incubation with a streptavidin-peroxidase complex, antibody-antigen binding complexes can be visualised through the addition of the substrate, 3-amino-9-ethylcarbazol. Performed on brain tissues, the dRIT has proven as sensitive as the FAT for fresh specimens [9, 10] . Brain impressions stained using the dRIT technique can be read within one hour and the antibody cocktail used has been shown to detect classical rabies virus strains (genotype 1) that have been assessed [11] . Currently, the FAT is routinely used to detect virus antigen in badly decomposed sample material. For the purpose of testing samples in the developing world where suitable cold storage for samples is often unavailable, this factor is important in the development of new tests. This obstacle has been overcome through evaluating sample preservation in phosphate buffered 50% glycerol at a range of temperatures for different time periods prior to testing for virus antigen. Glycerol saline solutions have been previously recognised as suitable storage media for tissue samples in the absence of cold storage [12] (Box 2). Using the dRIT in field studies in Tanzania, viral antigen could be detected in samples after considerable time periods post collection regardless of the regimen of glycerol preservative used [11] . Applications of the dRIT to analyse field samples in other rabies endemic regions have also proven highly successful. Field trials in Chad sought to study the dRIT in direct comparison to the FAT to attempt to confirm previous studies as to the incidence of rabies within a district known to be endemic. In this study, results between the two tests were 100% in agreement [9] and the only issue regarding use of the dRIT over the FAT was the need for the dRIT kit to be stored refrigerated prior to use. The dRIT will enable developing countries to perform routine rabies surveillance at greatly reduced cost and without the need for prohibitively expensive microscopic equipment along with the expertise and financial input needed to maintain them. The cost effectiveness of the dRIT will allow knowledge and technology transfer to areas of the developing world that currently are unable to diagnose rabies cases. Another method for the detection of rabies virus antigen from postmortem samples is a recently developed rapid immunodiag- www.plosntds.org nostic test (RIDT) based on the principles of immunochromatography [13] . The immunochromatographic lateral flow strip test is a one-step test that facilitates low-cost, rapid identification of various analytes including viruses [14] . Briefly, suspect material is subjected to a test device similar to a lateral flow device. Conjugated detector antibodies attached to two different zones on a membrane indicate the presence of viral antigen. Preliminary validation studies with a limited number of samples showed that the RIDT might have great potential as a useful method for rabies diagnosis without the need for laboratory equipment [13] . However, thorough validation including various circulating variants of RABV and other lyssaviruses is still needed before this test could be relied upon and be used as an alternative for the gold standard FAT. Reverse-transcriptase polymerase chain reaction (RT-PCR) Various conventional RT-PCR protocols for the diagnostic amplification of lyssavirus genome fragments have been published (Tables 1-3 ). Since primers were selected from conserved regions of the genome, most assays amplify parts of the nucleoprotein (N-) gene as earlier proposed [15] . In generic approaches intended to detect all lyssaviruses either hemi-nested or fully nested amplifications are used and have applications for both antemortem (saliva, CSF, brain) and postmortem samples (principally brain tissue) ( Table 2 ). Some of these diagnostic procedures are also applied for further virus characterization, including sequencing reactions [16] or restriction fragment length polymorphism (RFLP) [17] . Also, strain-specific RT-PCRs have been developed to distinguish various rabies virus (RABV) strains in a particular region [18] . Classical RT-PCR assays proved to be a sensitive and specific tool for routine diagnostic purposes [19, 15] , particularly in decomposed samples [20, 21, 22] or archival specimens [23, 24] . The sole detection of amplified RT-PCR products by gel-based systems however, especially when using hemi-nested RT-PCRs generates the risk of cross-contamination, does not allow an exact quantification of genome copies and does not include tests for specificity [25] . Hybridisation methods [26] and PCR-ELISA methods were established to overcome these difficulties [27] , although these techniques have not become universally accepted. Additionally, many laboratories now use partial sequencing to confirm the detection of a lyssavirus and obtain data that can be used in a phylogenetic analysis of viruses circulating in a specific region. The importance of sequencing the PCR products was highlighted in an experimental study [28] . This study demonstrated that although the nested RT-PCR was shown to be the most sensitive of the diagnostic techniques employed, host genomic amplicons of the same size as the target amplicons were observed on the agarose gels, which were subsequently confirmed as false positives following direct sequencing [28] . With the introduction of fluorogenic probes, detection of sequence specific templates can be achieved in real-time. Specificity is ensured by an inherent hybridization reaction, and cross-contamination is avoided due to the closed tube nature of the test [29, 30] . Consequently, for RABV and other lyssaviruses, various PCR assays using TaqMan technology have been described (Tables 4-5) . A generic real-time TaqMan-PCR for the detection and differentiation of lyssavirus genotypes 1, 5, and 6 has also been developed [31] . This assay utilises a pan-lyssavirus primer set, which has been shown to amplify a large panel of representative lyssaviruses, with probes specifically designed to discriminate between classical rabies virus and the European Bat Lyssaviruses type-1 and -2 (EBLV-1 and EBLV-2). PCR assays using TaqMan technology have applications for antemortem and postmortem samples. The pan-lyssavirus primer www.plosntds.org set can also be used in conjunction with a specific dye such as SYBR Green to allow for rapid detection of the amplicons. Validation of probe based assays relies on the availability of representative viruses or nucleic acid. However, for some lyssavirus genotypes only a limited number of viruses or sequences are available for primer/probe design, and they may not represent the genetic diversity of all current variants that are circulating. Single mutations for the North American RABV strains [32] in the region of the primers or the probe can alter the sensitivity of the PCR. Thus the genetic diversity among lyssaviruses may hamper the use of a single assay for all lyssaviruses. As such scanning surveillance may benefit more from the use of a pan-lyssavirus primer SYBR green assay rather than a strain or specific based assay. The use of a rapid automated NASBA technique was successfully applied to the ante-mortem saliva and cerebrospinal fluid (CSF) of four rabies patients in Thailand and shown to have a ten-fold increase in sensitivity compared to RT-PCR [33] . The assay detected rabies viral RNA as early as two days after onset of symptoms. The NASBA technique involves the use of three enzymes (reverse transcriptase, RNase H and T7 RNA polymerase) to synthesise multiple copies of target RNA under isothermal conditions. Briefly, a large number of RNA copies are generated using a pair of specific primers, one of which contains the T7 RNA polymerase binding site, and the other which has an electrochemiluminescence detection region attached to the 59 end. The Table 3 . Conventional, gel-based PCR-assays for the detection of lyssavirus species other than RABV. www.plosntds.org amplified RNA is detected using an automated reader, enabling rapid throughput testing. It is relatively easy to use and the whole process from extraction to detection can take as little as four hours. This technology has already been applied for point of care testing of bacterial pathogens [34] and viral pathogens [35, 36] . The NASBA technique has also been adapted to investigate rabies virus replication in situ, whereby the relatively lower isothermal temperatures of NASBA compared to in-situ RT-PCR ensure that cell integrity is maintained [37] . LAMP offers an alternative DNA amplification method to the polymerase chain reaction for applications to the ante-mortem saliva and CSF testing. The originators of the technique suggest that it amplifies with high specificity, efficiency and without the need for thermal cycling [38] . Amplification is achieved through the specific binding of two inner and two outer primers to the target sequence. The inner primers initiate strand synthesis whilst the outer primers displace the inner primers, allowing them to self-anneal to the nascent strand. This creates hairpin structures that trigger further strand synthesis that in turn lead to concatenation of the target sequence [38] . Polymerisation and strand displacement are achieved using a single enzyme, Bst 1 DNA polymerase. The technique is rapid, generating large quantities of target sequence within minutes. For the amplification of RNA viruses, a reverse transcription step is undertaken prior to the LAMP reaction. Primer sets have been successfully developed to detect a range of pathogenic viruses including West Nile virus [39] , Japanese Encephalitis virus [40] , Foot and Mouth Disease virus [41] and Chikungunya virus [42] . To assess the applicability of LAMP to the detection of rabies virus we designed a primer set using PrimerExplorer V4 software (Eiken Chemical Company Ltd., Japan) that can detect the Challenge Virus Standard (CVS) fixed strain of rabies virus (Table 6 ). In addition to the standard set of four primers, two further loop-binding primers have been added to increase the rate of strand displacement and synthesis [43] . The reverse transcription and LAMP reactions were undertaken simultaneously (RT-LAMP) in a single tube at 65uC using a www.plosntds.org thermostable reverse transcriptase, hence avoiding the step process inherent in an RT-PCR. Target amplification was monitored by the incorporation of the double stranded DNA binding fluorophore picogreen (Figure 1a) or by separation on a 1% agarose gel (Figure 1b) . At a constant temperature of 65uC, CVS RNA could be detected within 30 minutes. Development of RT-LAMP assays for use in diagnosis and surveillance is challenged by the considerable sequence variation observed within the rabies virus genome [44] that can frustrate specific primer design. Preliminary attempts at this suggest that multiple combinations of primers (up to 12 different primers) can lead to sensitive, rapid amplification of RABV genomes from a wide range of geographical locations. The use of isothermal amplification has the benefit of reducing the technological requirements of thermal cycling used in RT-PCR. This in turn offers the opportunity, when linked with lateral flow devices, to develop surveillance protocols where testing can take place in the field or in less sophisticated laboratories. Microarray linked to sequence independent PCR amplification offers the ability to rapidly identify pathogenic viruses from postmortem samples [45, 46, 47] . We have undertaken a study that has demonstrated the ability of a microarray to detect each of the seven lyssavirus genotypes (VLA Weybridge, unpublished data). The microarray is composed of oligonucleotide probes 70 nucleotides in length and includes probe sets for each of the seven classified genotypes and sets for the unclassified lyssaviruses. The www.plosntds.org diagnostic capability of the array was illustrated showing the ability of the array to detect RABV in a human case of rabies as the amplified RNA bound specifically to the classical rabies virus (genotype 1) probe set ( Figure 2 ). Development of a novel ultrasensitive and stable potentiometric immunosensor A stable potentiometric immunosensor for the detection of various analytes of interest from complex real world samples such as blood, serum and milk has been described [48] . The biosensor detects enzyme labelled immunocomplexes formed at the surface of polypyrrole coated, screen-printed gold or silver electrodes. Detection is through a secondary reaction that produces charged products with a shift in potential, being measured by local changes in redox state, pH and/or ionic strength. The magnitude of the change in potential is directly related to the level of target in the matrix such that the assays are quantitative and the numerical output is rapidly transmissible. The bioassays produce very rapid results (5-15 minutes), are highly reproducible (%CV,5%) and are ultra-sensitive (,50 fM). Thus this sort of biosensor offers the rapidity of production of signal of a lateral flow device with the sensitivity of third generation ELISAs. Immunoassays can be developed in a sandwich or competitive format for small (e.g. Digoxin; MW 780 Da) or large (e.g. hepatitis B surface antigen; .300 kDa) molecules [49] . In addition to immunoassays, it is also possible to detect specific nucleic acids and whole cells using the technology. Due to the method of production of the electrode strips, the base assays are inexpensive (,US$1 per strip) as is the detection hardware (US$2,000). In addition, most immunoassays can be readily adapted to this format with minimal additional optimisation. Potentiometric immunosensors for detecting the rabies virus nucleoprotein are in progress and offer the ability to rapidly screen complex non-clarified matrices, possibly at pen-side, in a cost-effective manner. Serological assays are not suitable as diagnostic tools for routine rabies testing as the host response to infection varies substantially between individuals. However, serology is still useful, particularly to monitor the development of the immune response. We would suggest that detection of rabies antibodies in serum and CSF, early after presentation and in the absence of a history of vaccination may be a positive indicator for a therapeutic intervention. Viral pseudotypes, the core of one virus coated with envelope protein derived from a second virus, offers a safe alternative to the use of pathogenic viruses in neutralisation assays. Using pseudotypes expressing genotype 1 CVS-11 glycoprotein, high titre stocks (1.3-3.2610 5 infectious units/ml) were produced that proved 100% specific and highly sensitive compared with neutralisation titres achieved using the FAVN [50] . A high correlation was also observed (r = 0.89). Using pseudotypes expressing EBLV-1 (genotype 5) and EBLV-2 (genotype 6) G-proteins, neutralising antibody titres broadly correlated with the degree of G-protein diversity. A vaccine study in Tanzania compared the two assays with pseudotypes showing 100% specificity and 94.4% sensitivity to the FAVN with a high correlation of antibody titres (r = 0.92). Incorporation of Lagos bat virus (genotype 2), Mokola virus (genotype 3) and Duvenhage virus (genotype 4) G-proteins, as well as lacZ as a reporter gene, makes the pseudotype assay an attractive option for serosurveillance in Africa and other resource limited countries. In addition, as the pseudotype assay uses substantially less input serum (10 ml) compared to FAVN and RFFIT, multiple tests can be undertaken on samples where collection volumes are limited or valuable e.g. bat sera. Due to the neurotropic nature of rabies virus, infection results in enormous viral replication in the CNS in the final stage of the disease that leads to massive antigen and viral genome concentrations. This makes detection of viral antigen in brain tissue by tests such as the FAT or the dRIT [7] very robust and relatively simple to perform, and these have become rapid gold standard tests. As for detection of viral genome, approaches are now available which process multiple specimens from nucleic acid extraction through to genetic typing, with significantly reduced risks of contamination. In addition, the use of TaqMan RT-PCR or similar technologies on robotics platforms, allow for rapid largescale rabies detection, typing and quantification in real time [32, 31, 51] . The development of PCR-based methods (Box 2) provided an alternative method for post mortem rabies diagnosis [26] , and the possibility of ante mortem diagnosis of human rabies [52] . RT-PCR methods invariably involve multiple transfers of nucleic acids between different tubes. Coupled with the high sensitivity of PCR methodologies, any small amount of contam- Figure 1 . Amplification of rabies virus RNA by reverse transcription loop-mediated isothermal amplification (RT-LAMP). For each reaction 1 mg RNA, prepared using TriZol (Invitrogen) from normal mouse brain (1A, circles; 1B 2) or infected mouse brain (1A, triangles; 1B, +) was added to a reaction mixture containing all six primers at concentrations indicated in Table 1 ination will undoubtedly produce false-positive results. Attempts have been made to adapt RT-PCR to reduce manipulations thereby reducing contamination risks. The visualisation of PCR products by gel electrophoresis exposes facilities and operators to large quantities of amplified material and thus many adaptations have been directed at replacing this step. New and improved rapid diagnostic tools for rabies using Taqman technology have been developed that avoid cross-contamination due to the closed tube nature of the test [29, 30] . A further benefit of RT-PCR has been to enable practical molecular characterization of rabies viruses [26] that has added significantly to the understanding of virus evolution and epidemiology. This approach has superseded the use of monoclonal antibodies for typing and characterising new strains of rabies virus. This has provided the evidence to support the delineation of lyssaviruses into genotypes [53] (Box 2) and was used for the classification of another four putative members of the genus [54, 55] . Also, this technique was a prerequisite for the understanding of the molecular biology of lyssaviruses [56] and underpinned future developments in rabies diagnosis and prevention. It is likely that in the future microarray techniques in combination with sophisticated bioinformatics and arrays with a hierarchical set of probes will provide an alternative to rapid virus discovery and characterisation. The development of 'real-time' RT-PCR techniques allows the quantification of this RNA in 'real-time,' giving a relatively quick and reliable method for the measuring levels of viral RNA. PCR based techniques are not currently recommended by the WHO for routine post-mortem diagnosis of rabies. However, in laboratories with strict quality control procedures in place and demonstrable experience and expertise, these molecular techniques have been successfully applied for confirmatory diagnosis and epidemiological surveys. For these reasons, it is likely that international bodies will accept their use in the future for routine rabies diagnosis. Reverse transcription PCR has been reported to confirm rabies diagnosis intra-vitam in suspect human cases, when conventional diagnostic methods have failed and post-mortem material is not available (Box 1) [57] . Rabies virus RNA can be detected in a range of biological fluids and samples (e.g. saliva, CSF, tears, skin biopsy sample and urine). Owing to the intermittent shedding of virus, serial samples of fluids such as saliva and urine should be tested but negative results should not be used to exclude a diagnosis of rabies. All positive PCR results should be sequenced to confirm the origin of the virus and rule out possible contamination. In terms of the RNA concentrations in the brain, the sensitivity especially of nested or real-time PCRs may be beyond the threshold needed for routine post-mortem testing. Also, contamination of negative samples could lead to an unjustifiable administration of a high number of costly post exposure prophylaxis and would produce false data for the rabies surveillance. However, with the introduction of accreditation for laboratories, quality control measures are being implemented in a growing number of laboratories worldwide. Such quality controls for diagnostic rabies PCRs should encompass several measures, including the inclusion of appropriate positive, negative, and inhibition controls in assay runs. The consistency and the interassay reproducibility should also be ensured over time by monitoring performance. Only if laboratories meet the required standard [58] , can PCR fulfil its full potential. The use of PCR should not be restricted only as a confirmatory diagnostic test for decomposed samples but also as a powerful tool for virus typing Figure 2 . Microarray identification of rabies virus RNA prepared from a brain sample from a confirmed case of human rabies. Total RNA was extracted using TriZol (Invitrogen) and treated with DNase I prior to conversion to double stranded DNA [45] . Non-specific amplification was achieved using a DNA polymerase (Klen Taq, Sigma) and the products were labelled through binding of Alexa Fluor 555 reactive dye (Invitrogen) to amplicons. Labelled target DNA was denatured at 95uC and chilled on ice before dilution in hybridization buffer and addition to the microarray slide. Hybridization occurred at 50uC overnight. Slides were washed and the target-probe binding was captured using GenePix Pro 6.1 software (Molecular Devices). Statistical analysis of the data was conducted using DetectiV software [64] . doi:10.1371/journal.pntd.0000530.g002 and molecular epidemiology studies. The lack of standardization is a major obstacle to the general use of PCR for rabies diagnosis, especially in developing countries. It is evident that the RT-PCR dominates genetic detection of rabies virus and it seems probable that this technique will dominate rabies diagnosis in the 21 st century. However, we should not discount alternatives that have the benefit of isothermal amplification that will enable implementation in laboratories where access to thermal cyclers is an obstacle. A NASBA technique was successfully applied to the saliva and CSF of four living patients with rabies and detected rabies viral RNA 2-days after the onset of symptoms. This technique has also been adapted to investigate rabies virus replication in-situ. LAMP also falls into this category and can be adapted for use with lateral flow devices thus making its application very simple. Existing assays for rabies virus antibody prevalence studies either require high containment facilities or do not distinguish between neutralising and non-neutralising antibodies [59] [60] . Recently however, a neutralisation assay using retroviral pseudotypes was described [50] , not bound by the restrictions listed above and also allowing a choice of endpoint reporter proteins (bgalactosidase, green fluorescent protein or luciferase) [61] . A further benefit of this technique is its adaption to using small volumes of sera thus making them useful for surveillance. Currently, high-throughput rabies virus molecular detection methods augment standard diagnostic tests or are in the process of development and refinement for use alone. As we progress through the 21 st century, it is possible that these techniques will replace conventional tests (Box 1). Obstacles to adoption include cost, complexity and local acceptance of their use. It is also possible that immunological tests by measuring 'indirect' markers such as cytokines and electrolytes will increase in use. These tests however, will probably remain in the realm of large reference laboratories where resources allow the development of novel assays. As far as semi-automated or automated instruments and robotics-based techniques are concerned, they are useful when large numbers of the same test are undertaken such as surveillance and companion animal testing and these tests will continue to increase in popularity and use, especially in central reference facilities. There is a clear need to simplify molecular diagnostic techniques so these tests can be applied universally in developing and developed countries. It is likely that new developments will focus on generating low volume and yet affordable diagnostic tests for rabies. More use will be made of point-of-care (POC) diagnostic testing using portable extraction techniques linked to PCR machines with the use of lyophilised reagents to overcome coldchain dependencies in tropical countries. In the 21 st century, these technologies will have a demonstrable impact on people living in developing countries, especially where rabies is still considered a 'neglected' disease. By contrast in the developed world, these new technological advances will undoubtedly be faster, more accurate and cost-effective leading to a 'Theragnostics Approach' that combines therapeutics with diagnostics for the human treatment of rabies. Interest in treating human rabies aggressively is gaining momentum, largely due to the reported success in treating a 15year-old girl, in whom clinical rabies developed one month after she was bitten by a bat, using a combination of therapeutic coma with antiviral drugs whilst allowing for the host immune system to confer immunity -The 'Milwaukee Protocol' (Box 2) [62, 63] . Bioinformatics, genomics, proteomics, and functional genomics are the molecular biology tools that are essential for the progress of molecular 'theragnostics', where both early diagnosis and monitoring of serology are critical factors for the successful treatment of a rabies patient. In addition, theragnostics could eliminate the unnecessary treatment of patients for whom rabies immunotherapy is not appropriate i.e. immunosuppressed patients, resulting in substantial drug cost savings for the healthcare system.

Ducks: The “Trojan Horses” of H5N1 influenza

Abstract Wild ducks are the main reservoir of influenza A viruses that can be transmitted to domestic poultry and mammals, including humans. Of the 16 hemagglutinin (HA) subtypes of influenza A viruses, only the H5 and H7 subtypes cause highly pathogenic (HP) influenza in the natural hosts. Several duck species are naturally resistant to HP Asian H5N1 influenza viruses. These duck species can shed and spread virus from both the respiratory and intestinal tracts while showing few or no disease signs. While the HP Asian H5N1 viruses are 100% lethal for chickens and other gallinaceous poultry, the absence of disease signs in some duck species has led to the concept that ducks are the “Trojan horses” of H5N1 in their surreptitious spread of virus. An important unresolved issue is whether the HP H5N1 viruses are maintained in the wild duck population of the world. Here, we review the ecology and pathobiology of ducks infected with influenza A viruses and ducks’ role in the maintenance and spread of HP H5N1 viruses. We also identify the key questions about the role of ducks that must be resolved in order to understand the emergence and control of pandemic influenza. It is generally accepted that wild duck species can spread HP H5N1 viruses, but there is insufficient evidence to show that ducks maintain these viruses and transfer them from one generation to the next.

Avian influenza is caused by type A viruses of the family Orthomyxoviridae. The influenza A viruses infect primarily free-living aquatic birds, and they are classified by their hemagglutinin (HA) and neuraminidase (NA) surface glycoproteins. All 16 HA and 9 NA subtypes have been isolated from aquatic birds; wild ducks are the main reservoir. The viruses cause asymptomatic or low pathogenic infection in these natural hosts. 1 However, certain strains of influenza A virus have crossed the host range barrier and infected other species, including humans. These viruses are the source of the influenza pandemics that emerge at irregular intervals. 1, 2 The H5 and H7 subtypes are of particular concern because they can become highly pathogenic (HP), causing systemic illness and death in both avian and mammalian species, including humans. 2 The H5N1 virus that emerged in Asia in 1996 is unique among the HP avian influenza (HPAI) viruses in that it has continued to circulate in avian species for more than a decade and has spread to more than 60 countries in Eurasia (http://www.who.int/csr/ disease/avian_influenza/en/). While the H5N1 HPAI viruses are 100% lethal to chickens and gallinaceous poultry, they often cause asymptomatic infection in some species of domestic and wild ducks. These ''silent spreaders'' of H5N1 HPAI viruses are therefore referred to as ''Trojan horses''. [3] [4] [5] Clearly, ducks play a complex and vital role in the biology and the overall natural history of influenza, including H5N1 HPAI viruses. Ducks are members of the subfamily Anatinae, which contains most species of anserine birds. This subfamily is nearly cosmopolitan in distribution, and its members occupy almost all aquatic habitats. The ecology of these birds, summarized in Figure 1 , facilitates the maintenance and spread of avian influenza viruses. Although human influenza A isolates and the currently circulating H5N1 HPAI viruses typically infect the upper respiratory tract, the primary site of infection in ducks is the intestine. 6 Avian influenza viruses enter the environment when the host defecates or drools, and they then infect susceptible hosts as they feed and drink. Avian influenza virus replication has been observed in the respiratory tract, 6 but the contribution of this site to maintenance of infection in the population is unresolved. Specifically, fecal shedding of H4N7, H7N3, and H11N9 virions from experimentally infected mallard ducks persists longer and at higher titers than tracheal shedding. 6 When a large number of birds roost on a small pond (for example, in the staging ⁄ marshalling areas), as many as 10 10 EID 50 •g )1 •d )1 infectious virions are estimated to enter the environment in the fecal matter of each infected duck. 6 Further, avian influenza viruses are stable in water 1,7 and have been isolated from the surface of ponds containing a large number of waterfowl. 8, 9 Although aerosol transmission cannot be dismissed, the larger number of positive cloacal than tracheal swabs, the high fecal virus titer, and the stability of the virions in water suggest that lowpathogenic avian influenza (LPAI) viruses persist in duck populations through fecal-oral transmission. 1 This mechanism could partially explain the higher prevalence of infection in surface-feeding (dabbling) ducks than in diving ducks that typically feed in deeper water. 10 Surveillance data suggest year-round transmission of avian influenza viruses within duck populations. The prevalence of infection exhibits an annual cyclical pattern in both North American 1,11 and Eurasian 12 duck populations (Figure 1 ), peaking before and during the fall migration as a result of the influx of immunologically naïve juveniles. 1, 9, 10, 13 Experi-mentally infected white Pekin ducks have shed virus for more than 3 weeks after inoculation. 3, 14 Coupled with limited morbidity and serum antibody response, 3 infected birds are likely to shed virus during the first few weeks of the fall migration, dispersing it along their numerous migration corridors. However, the prevalence of infection is much lower along the migration routes and at the wintering grounds than at the marshalling areas. 9, 12, 15, 16 This disparity may reflect the development of immunity to circulating virus subtypes within the duck population or a decline in transmission because of population dispersal. 13 In general, prevalence of infection is higher at the wintering grounds and spring nesting sites in duck populations from Europe than in North American populations (Figure 1 ). The most likely explanation for this difference is random variation, since surveillance studies from multiple areas in North America and in Europe often obtain slightly different prevalence values in the duck populations. Many factors can affect prevalence including, but not limited to, the size of the duck population, sampling location, and time of collection. Thus, the few multi-year studies that exist likely exemplify the variation that one would observe if additional sampling sites were included in the studies, and not the differences in geography between Europe and North America. 16 Prevalence is at its lowest Figure 1 . Overview of the annual movement and behavior of migratory ducks and their role in interspecies transmission. During spring and fall migration, the ducks rest and feed for a few days to weeks at numerous stopover sites (wetlands, lakes, or ponds) along the migration route. The length of stay and the aquatic habitat allows the transmission of influenza viruses to and from the domestic duck populations. Domestic ducks that become infected are likely to maintain the virus locally and increase the probability of its spread to other species. In the diagram, solid arrows indicate confirmed routes of transmission of LPAI and ⁄ or HPAI viruses between species. The dashed line represents a probable but unconfirmed route of transmission. The graphs indicate the average prevalence of low-pathogenic avian influenza in North American and European duck populations during 3 stages of the annual migration. 10, 16 during the spring migration but increases again after the breeding season, when the ducks have moved to the molting and staging areas. 11, 13, 16 It is not clear how the duck population acquires avian influenza viruses during the spring of every year. Infectious virions may persist through the winter in the frozen waters of the breeding areas and reinfect the ducks when they return in the spring. 1, 6 Alternatively, the duck populations may carry the viruses throughout the entire migration. Year-round prevalence in the ducks supports the latter, although persistence in the frozen habitats could play a role in the perpetuation of the viruses. 10 Some virus subtypes are isolated more frequently than others. 10, 11 Three HA subtypes, H3, H4, and H6, are common in both North American and European ducks, 11, 12 and the most prevalent subtype combinations in both areas are H4N6 and H6N2. 16 Explanations vary for why certain HA and NA subtypes (and combinations) are common or rare in wild birds. The general hypothesis is that these subtypes likely have the highest fitness, with replication rates balanced by a level of virulence that sufficiently increases transmission probability to the level where infection in the next cohort of birds is almost guaranteed. It is speculated that this could be largely influenced by the functional balance between HA binding affinity and NA enzymatic activity. 17 Although the H6 gene is of Eurasian origin and is widely distributed in North American ducks, genomic analysis of viruses suggests limited intercontinental exchange between Eurasia and the Americas. 18 Therefore, novel viral genotypes must arise via mutation and reassortment of the genomes in circulation within a specific geographic area. The marshalling areas provide such an opportunity by attracting populations of ducks from various breeding areas and migration corridors, with each population harboring a potentially different combination of subtypes. 16 Coinfection of ducks with two or more virus subtypes is common, 19 as is reassortment, 14 and emergent strains that are most virulent to gallinaceous poultry can have low pathogenicity in the duck hosts. 20 However, the role of ducks in the maintenance and spread of influenza viruses, and especially in the emergence of novel genotypes, appears to depend on their migratory behavior. Specifically, ducks that migrate annually are likely to spread influenza viruses along the migration routes, primarily by exposing the resident and domestic duck populations at the numerous stopover sites. 10, 16, 21 In contrast, domestic and resident ducks maintain the viruses in close proximity to other species and have been implicated in the spread of both LPAI and HPAI viruses to domestic poultry and terrestrial birds. 20, 22, 23 Low-pathogenic avian influenza LPAI viruses can pass through the upper digestive tract of ducks and replicate in the lower intestinal tract without causing clinical manifestations of disease. 3, 6 Further evidence that the intestinal tract is the target organ of LPAI viruses in ducks includes the replication of virus in the lower intestinal tract, but not in the lungs, after direct inoculation into the crop and rectum 6 and high fecal virus titers after intravenous inoculation. 3 The specific site of LPAI virus replication is believed to be the crypts of Lieberkühn in the large intestine. 3 The other potential target organ for LPAI viruses in ducks is the respiratory tract. Lungs of mallard ducks intranasally inoculated with LPAI viruses showed mild pneumonia, and lymphocyte and macrophage infiltration within 2 days. Immunostaining for viral nucleoprotein revealed intermittent staining of respiratory epithelial cells but no viral replication in the lung tissue. 24 This evidence indicates that the respiratory tract and not the lung tissue itself is the primary target of infection. The species diversity of ducks may also play an important role in the pathogenicity of influenza viruses. Mallard duck embryos inoculated with LPAI virus have significantly lower mortality rates than inoculated Muscovy duck embryos; however, in regards to replication, the LPAI viruses behave in a different way. Viral antigens were found in the internal organs (nasal sinuses, pharynx, trachea, bronchi, lung, and air sacs) of the mallard duck embryos, but not in those of the Muscovy duck embryos. The reason for this mortality ⁄ virus replication paradox in mallard ducks is unclear but is in keeping with the evidence that mallard ducks are considered to be the main reservoirs of LPAI viruses in nature. 25 Although several studies have examined the serum antibody response in both naturally and experimentally infected ducks, knowledge of the avian immune response to influenza viruses is very limited. 26 White Pekin ducks inoculated with an H7N2 LPAI virus developed negligible serum hemagglutination inhibition (HI) antibody titers despite fecal shedding of virus until day 7 post-inoculation. Animals reinoculated 46 days later with the same virus strain had a marked antibody response, but virus could not be isolated from any of the organs. These results and the lack of a secondary immune response after inoculation with formalin-inactivated virus suggested that the rapid immune response in re-infected birds may restrict influenza infection to short time scales. 3 It is noteworthy that prior infection does not protect ducks against subsequent infection with other virus subtypes. For example, ducks infected with an H4N6 subtype are protected from reinfection with the same virus, but they shed virions for 8 days after challenge with an H11N3 isolate. 27 These data have applications in the field, where isolation of influenza virus from migratory waterfowl is infrequent during the winter, potentially indicating the existence of a significant level of immunity in wintering ducks acquired from previous influenza infections. Further illustration of this is seen from a study of wild waterfowl in Italy over six winter seasons, in which 17 of the 20 viruses isolated were of the H1N1 subtype, suggesting that the wintering ducks had some degree of immunity to the other subtypes of circulating influenza strains. 28 Highly pathogenic avian influenza Several experimental studies have investigated the pathogenicity of H5N1 HPAI viruses (isolated since 2002) in ducks. Cherry Valley Pekin ducks inoculated with a 2003 HPAI H5N1 strain isolated from duck meat at a quarantine inspection station in China developed neurologic signs, including blindness and head shaking, although none died. High virus titers were detected in the respiratory organs (lung and trachea), brain, liver, kidneys, and colon, and microscopic lesions were observed in the brain (viral encephalitis), heart (myocarditis with degeneration and necrosis of myocytes), and bursa (mild lymphoid follicular hyperplasia). 29 Viral neurotropism and pancreatotropism have been observed in multiple other studies of recent HPAI virus isolates. Ducks lethally challenged with these H5N1 HPAI viruses showed severe neurologic signs, including torticollis, incoordination, tremors, and seizures. 30, 31 Immunohistochemistry positivity was recorded in the cerebrum and brain stem, and in situ hybridization detected virus in the neurons and glial cells of the cerebral gray matter, further confirming the strong neurotropism of post-2002 isolates. 30, 31 Although the route of entry of virus into the central nervous system has not been determined, at least two different hypotheses have been proposed, including ascending transmission of virus via vagal, olfactory, and trigeminal nerve fibers, and penetration of the blood-brain barrier. 32, 33 Another recurring characteristic of recent H5N1 HPAI viruses in ducks is that virus titers are frequently higher in oropharyngeal swabs than in cloacal swabs. 30, 34 Pharyngeal excretion of H5N1 HPAI viruses has been suggested to originate from the lung and ⁄ or air sac, as only these tissues have shown immunohistochemical evidence of virus replication. Preferential pharyngeal excretion suggests that pharyngeal swabs, as well as the customary cloacal swabs, should be taken when conducting surveillance of avian influenza viruses in wild ducks. 34 Otherwise, the prevalence of H5N1 HPAI may be underestimated. Additional studies of the role and rates of respiratory transmission of H5N1 HPAI viruses in ducks are needed, especially as they relate to cloacal excretion. In 1996, the parental virus (A ⁄ Goose ⁄ Guangdong ⁄ 1 ⁄ 96; A ⁄ Gs ⁄ GD ⁄ 1 ⁄ 96) of currently circulating H5N1 HPAI viruses emerged in southern China. Genetic evidence revealed that this virus originated from H5 LPAI viruses carried from northern Japan by wild ducks or other migratory birds. 15, 35 A reassortant H5N1 HPAI virus subsequently emerged in poultry at farms and live animal markets in Hong Kong in 1997. Genetic analyses showed that the H5 HA gene of the reassortant virus was derived from an A ⁄ Gs ⁄ GD ⁄ 1 ⁄ 96-like virus, while the remaining gene segments were derived from low-pathogenic viruses: the N1 NA gene came from A ⁄ Teal ⁄ Hong Kong ⁄ W312 ⁄ 97 (H6N1) virus, and the internal genes from A ⁄ Quail ⁄ Hong Kong ⁄ G1 ⁄ 97 (H9N2) virus. 36 The reassortant virus caused the first lethal infection in humans (6 deaths among 18 known cases) by direct bird-to-human transmission. 37 Between 1999 and 2002, H5N1 HPAI viruses with the H5 HA gene of A ⁄ Gs ⁄ GD ⁄ 1 ⁄ 96-like viruses but with a diversity of genotypes in the other genes, re-emerged multiple times in Hong Kong. 37 The first indication of the spread of H5N1 HPAI viruses from domestic to wild species of aquatic birds occurred in Kowloon and Penfold Park in Hong Kong, 38 where 19 different duck species were infected. Some species, including the Red-Crested Pochard (Netta rufina), were highly susceptible (19 ⁄ 20 died), whereas others, including the Bahama Pintail (Anas bahamensis), were less susceptible (4 ⁄ 21 died). The next major event in nature was the massive die-off of waterfowl at Qinghai Lake in China. [39] [40] [41] Four different genotypes of H5N1 HPAI viruses co-circulated in the waterfowl there; one of these became dominant and spread westward to India, Europe, and Africa. Notable features of the dominant Qinghai Lake H5N1 HPAI isolates were the acquisition of a lys627 mutation in the PB2 gene and the absence of pathogenicity in mallard ducks. 42 The lys627 mutation has been found to be associated with pathogenicity in mammalian species, 43, 44 suggesting that it may have been generated while the virus was replicating in a mammal. The virus was likely transmitted from domestic ducks to wild ducks en route to Qinghai Lake. The role of duck species in the westward spread of the Qinghai-H5N1 virus remains controversial. Circumstantial evidence from global wildlife surveillance supports the hypothesis that migratory birds, including wild ducks, have contributed to the current Eurasian endemic of H5N1 HPAI viruses. 45 Surveillance studies in Thailand in 2004 showed that most domestic grazing ducks infected with H5N1 HPAI viruses were asymptomatic 4 and that the initial spread of H5N1 HPAI viruses to chickens and humans corresponded to the movement of grazing ducks. 4, 46 In fact, infected domestic ducks grazing on man-made wetlands (e.g., harvested rice fields and irrigation canals) may maintain the infection and spread it to wild birds that feed at the same sites. If these wild birds are migratory and experience limited morbidity, they in turn can disperse HPAI viruses widely (Figure 2) , as suggested by the high genetic similarity of HPAI isolates from Africa, Europe, and the Middle East to the Qinghai-H5N1 virus. 37 The evolution of H5N1 HPAI viruses by reassortment with LPAI viruses in the aquatic bird reservoir played an important part in the genesis of the multiple genotypes, clades, and subclades of Asian H5N1 HPAI viruses and is ongoing. 37 However, it is the ever-increasing poultry industry that provides the reassortment interface between wild and domestic avian species. The number of domestic ducks, chickens, and other poultry continues to increase, but biosecurity and separation of species is not always taken into account. Ducks raised in a closed high-biosecurity system in Thailand were protected from infection while H5N1 HPAI viruses were circulating among backyard ducks, open house ducks, and grazing ducks. 4 Therefore, biosecurity can prevent the spread of influenza viruses from wild to domestic ducks. Live poultry markets (wet markets) have been identified as a risk factor in the genesis of novel influenza viruses 49 and were identified as the source of the human outbreak of HP H5N1 viruses in Hong Kong. The ban on ducks, geese, and later, quail, together with improved biosecurity (clean days), have markedly reduced the influenza virus diversity in the Hong Kong wet markets. 37 Live poultry markets are being phased out in Hong Kong, and in the interim no live poultry can be carried over from 1 day to the next. Taiwan plans to close all live poultry markets by 2009, and Shanghai authorities are reducing the number of wet markets. Overall, however, the role of live poultry markets in the emergence and control of pandemic influenza has been largely ignored. Universal closure of live markets would make biological sense but is difficult in regions where refrigeration is not widely available. Vaccination has been accepted as an option for the control of HPAI by the Food and Agriculture Organization of the United Nations and the World Organization for Animal Health. Emphasis is placed on the use of vaccine to facilitate eradication. The continued use of poultry influenza vaccines without an eradication plan has immediate benefits but also long-term consequences. The difficulty with continued vaccine usage is that it promotes genetic variation and allows shedding of virus in the absence of disease signs, thus creating the potential for epicenters of virus spread. Further, while both inactivated oil emulsion whole-virus H5 vaccines and recombinant NDV vaccines containing H5 HAs are highly efficacious in chickens, the recombinant NDV vaccines are less efficacious in ducks. The experience in Vietnam illustrates these points. In 2005, after 61 human cases of HP H5N1 virus infection and 19 deaths, universal poultry vaccination and reduction of duck populations were implemented, with dramatic results. In 2006, there were no human cases of H5N1 influenza. However, in 2007-2008, there were 14 human cases of HP H5N1 virus infection and 10 deaths (http://www.who.int/ csr/disease/avian_influenza/country/cases_table_2009_03_11/ en/index.html). The difficulty of controlling H5N1 HPAI viruses in ducks by vaccination and the enormous task of vaccinating sufficient poultry to maintain ''herd immunity'' remain daunting obstacles. Influenza in humans is considered a non-eradicable disease due to periodic introduction of viruses from their natural reservoir, wild migratory birds -mainly ducks. The culling of wild birds is not an option. The sole current option is biosecurity and eradication of HP influenza from domestic poultry. The longer-term goal will be to understand the genetics of natural resistance in ducks and to introduce these traits into domestic animals. There is consensus that the migratory waterfowl of the world (predominantly wild ducks) serve as the natural reservoirs of all influenza A viruses, which cause asymptomatic infection in these birds. Influenza viruses have probably co-evolved with ducks over millennia, establishing an equilibrium between hosts and parasites so that neither suffers a significant loss of biological fitness; the evidence being minimal signs of disease in the hosts and the annual isolation of common subtypes. The unanswered question is whether these migratory bird species are the reservoirs of the currently circulating H5N1 HPAI viruses. Until the emergence of the Asian H5N1 HPAI viruses, the available data indicated that each new outbreak of HP H5 or H7 virus died out or was stamped out and that subsequent HP strains emerged from the low-pathogenic H5 and H7 virus reservoir. All species of birds tested to date support replication of some HP H5N1 strains and, providing they are not killed rapidly, could contribute to the spread of H5N1 HPAI viruses. The present review has concentrated on ducks, some species of which are susceptible to H5N1 HPAI virus-caused disease and death while others (e.g., mallards) are quite resistant. 34, 50 Therefore, ducks that are infected but are naturally resistant to disease could have contributed to the spread of H5N1 HPAI viruses westward from Qinghai Lake in 2005 to Europe, Africa, India, and the Middle East. An unanswered question is whether the H5N1 HPAI viruses are being carried back to the duck breeding areas and are infecting the next generation. Extensive surveillance in the migratory pathways in Europe and Asia has provided no evidence of H5N1 HPAI viruses in the new generation of birds after the breeding season. While all duck species tested to date are susceptible to lethal infection with H5N1 HPAI viruses, some species, including the mallard and Pintail ducks, are less susceptible and many survive. It is in these relatively resistant species that the H5N1 HPAI viruses could be maintained. However, surveillance studies to date in these species have detected no H5N1 HPAI viruses in breeding or juvenile birds. Experimental studies show that some mallard ducks continue to shed virus for up to 17 days, allowing the development of humoral immunity and subsequent selection of antigenic variants within the same bird. 23 If this occurs, it could be argued that a limited number of ducks would be sufficient to maintain the virus in nature. Continued surveillance is needed to determine whether H5N1 HPAI viruses are maintained in nature by a small number of naturally resistant ducks that are long-term virus shedders. While naturally resistant ducks can be argued to have been involved in the spread of H5N1 HPAI viruses from Qinghai Lake to the rest of Eurasia, it is difficult to explain why H5N1 HPAI viruses have not spread to susceptible species in the Philippines, Australia, and the Americas, which are on the direct flyways of migratory waterfowl. More than 6AE6 million birds migrate from Eastern Asia to Alaska yearly (alaska.usgs.gov ⁄ science ⁄ biology ⁄ avian_influenza ⁄ migrants_tables.html). Despite intense surveillance in Alaska, no H5N1 HPAI viruses have been detected to date, and influenza viruses of Eurasian lineage are introduced into the Americas only rarely. 18 The major spreaders of influenza in domestic poultry are humans. As described by , from the molecular epidemiology data, transmission of H5N1 influenza in domestic poultry is perpetuated largely through movement of poultry and poultry products rather than continued reintroduction of viruses from migrating birds. 42 The alternative reservoir, the domestic duck population, has a higher likelihood of perpetuating H5N1 HPAI viruses. Prospective surveillance continued to isolate H5N1 HPAI viruses from apparently healthy ducks, geese, and chickens in Southeast Asian poultry markets during 2004-2006. Naturally resistant ducks might not be expected to show disease signs, but the absence of morbidity in highly susceptible geese and chickens is surprising. The widespread use of vaccine in chickens may explain this observation, but vaccine has been used less in geese and ducks. An alternative possibility is that the susceptible poultry had cross immunity as the result of exposure to co-circulating influenza viruses. Experimental studies have demonstrated that chickens previously infected with H9N2 virus and then inoculated with H5N1 HPAI virus become infected and shed virus but do not show disease signs. 51 The continuing co-circulation of multiple subtypes of LPAI viruses in domestic poultry could explain why a small percentage of susceptible domestic species can appear healthy while shedding transmissible levels of H5N1 HPAI virus. To provide answers to these unresolved questions about the role of domestic species, it will be necessary to establish long-term prospective surveillance in domestic poultry in the hypothetical ''epicenter zones,'' including China, Indonesia, Vietnam, Egypt, and Nigeria. It is noteworthy that in these regions, control of H5N1 HPAI virus is attempted by the continuing use of vaccines. An area that has been surprisingly neglected is the genetics of ducks, the ultimate reservoir species of all influenza A subtypes. The wild duck reservoir contributes some or all of the genes of future pandemic strains in humans and future panzootic strains in domestic poultry. Immune mechanisms in ducks are currently understudied, and the molecular basis of resistance of some duck species to lethal infection is unresolved. Sequencing of the genome of the mallard duck is warranted, as it could provide insight into the factors that contribute to markedly reduced influenza virus pathogenicity. Because wild ducks are the main reservoir of all influenza A viruses and the ultimate source of future pandemics, members of the scientific community who are interested in understanding the emergence and control of pandemic influenza should direct their attention to the following questions: • Do ducks (wild or domestic) serve as the reservoirs of the Asian H5N1 HPAI viruses? • What genomic characteristics of ducks are associated with natural resistance in some species? • Is antigenic diversity driven naturally in ducks or is it the consequence of vaccine usage? • What dose of vaccine antigen is required to prevent transmissible levels of excretion of H5N1 HPAI viruses by ducks of different species (and geese and swans)? • Is eradication of Asian H5N1 HPAI viruses achievable? • Can the use of transgenic animals containing the natural resistance gene(s) of mallard ducks prevent pathogenic influenza virus infection?

Non-Hemagglutinating Flaviviruses: Molecular Mechanisms for the Emergence of New Strains via Adaptation to European Ticks

Tick-borne encephalitis virus (TBEV) causes human epidemics across Eurasia. Clinical manifestations range from inapparent infections and fevers to fatal encephalitis but the factors that determine disease severity are currently undefined. TBEV is characteristically a hemagglutinating (HA) virus; the ability to agglutinate erythrocytes tentatively reflects virion receptor/fusion activity. However, for the past few years many atypical HA-deficient strains have been isolated from patients and also from the natural European host tick, Ixodes persulcatus. By analysing the sequences of HA-deficient strains we have identified 3 unique amino acid substitutions (D67G, E122G or D277A) in the envelope protein, each of which increases the net charge and hydrophobicity of the virion surface. Therefore, we genetically engineered virus mutants each containing one of these 3 substitutions; they all exhibited HA-deficiency. Unexpectedly, each genetically modified non-HA virus demonstrated increased TBEV reproduction in feeding Ixodes ricinus, not the recognised tick host for these strains. Moreover, virus transmission efficiency between infected and uninfected ticks co-feeding on mice was also intensified by each substitution. Retrospectively, the mutation D67G was identified in viruses isolated from patients with encephalitis. We propose that the emergence of atypical Siberian HA-deficient TBEV strains in Europe is linked to their molecular adaptation to local ticks. This process appears to be driven by the selection of single mutations that change the virion surface thus enhancing receptor/fusion function essential for TBEV entry into the unfamiliar tick species. As the consequence of this adaptive mutagenesis, some of these mutations also appear to enhance the ability of TBEV to cross the human blood-brain barrier, a likely explanation for fatal encephalitis. Future research will reveal if these emerging Siberian TBEV strains continue to disperse westwards across Europe by adaptation to the indigenous tick species and if they are associated with severe forms of TBE.

Tick-borne encephalitis virus (TBEV) causes up to 14,000 human cases of tick-borne encephalitis (TBE) across Eurasia annually [1, 2] . TBE outbreaks are now registered in about 30 European countries with a recorded morbidity increase of about 400% during the past 30 years [3] . TBEV is a member of the tick-borne flavivirus (TBFV) group that, together with mosquito-borne and no-known vector virus groups comprise the genus Flavivirus within the family Flaviviridae. Human pathogens within the genus Flavivirus include Japanese encephalitis virus, Dengue virus and Yellow fever virus that cause annual epidemics of fever, encephalitis and hemorrhagic fever in the tropics and some sub-tropical regions [4, 5] . In its natural habitat, TBEV is maintained by transmission between infected and non-infected ticks when they co-feed on small forest animals [6] [7] [8] . Humans are incidental hosts for ticks and may become infected by a feeding infected tick. The clinical manifestations caused by TBEV range from inapparent infections and fevers, with complete recovery of patients, to debilitating or fatal encephalitis. The proportion of fatal human infections varies widely in different regions and in different years. The factors that determine disease severity are poorly defined but correlations between viral subtype and disease severity have been described. TBEV strains are currently divided into 3 closely related subtypes, i.e. western-European (WE), Siberian (SIB) and Far Eastern (FE) [9] . FE TBEV is recognised as the most virulent pathogen with a 20-40% case fatality rate. The SIB subtype is considered less virulent (7-8% case fatality rate) but chronic disease occurs more frequently (1-3%) . Western European strains are the least virulent with case fatality rates lower than 2%. However, a range of clinical manifestations, from asymptomatic to encephalitic is observed for all TBEV subtypes [1, 2] and the underlying basis for this has not yet been adequately explained. Conventionally, each TBEV subtype has been associated with distinct geographic ranges within the Old World region of the northern hemisphere, hence the groupings Far East, Siberia and Western Europe [9] . However during recent decades the epidemiology of the TBFV appears to have been changing, with SIB TBEV becoming the dominant subtype apparently gradually replacing the WE or FE subtypes that previously appeared to monopolise many regions [10] [11] [12] [13] [14] [15] [16] . Moreover, the SIB subtype is being isolated more frequently from patients who develop the most severe forms of encephalitis, with the virus invading the entire brain in contrast with the more focal virus localization observed previously. Over a period of time, the most severe cases of TBE have been more frequently associated with the SIB strains than with the FE strains [17] , indicating that this is not an artifact of increased surveillance. Whilst these reports are disturbing they have not as yet been addressed at the molecular virological level. TBEV virions are spherical particles with an ,11 kb-RNA genome embedded in a capsid that is surrounded by a lipid envelope mainly containing a virus envelope (E) glycoprotein. This E protein plays a key role in many stages of the virus life cycle; it mediates virus binding to receptors on the cell surface (adsorption) which triggers receptor-mediated endocytosis. Exposure of the endocytosed virus to the acid pH converts the native E protein dimers into fusogenic trimers [18, 19] ; the latter promote fusion of virion and endosomal membranes thus releasing viral RNA into the cytoplasm. The E protein also plays the major role in inducing the host immune response and mediates hemagglutination (HA), i.e. the ability of virions to agglutinate avian erythrocytes; for decades HA has been used in routine diagnosis [20] . Whilst most strains of TBEV show HA activity, during the past 10 years atypical HA-deficient strains have been isolated with increasing frequency in Europe from both ticks and infected patients. More than 40 HA-deficient strains are now recognized and they all exhibit reduced pathogenicity for mice when compared with HA-competent strains. They are also called ''antigenically deficient'' (AD) strains, in contrast to the more common antigenically competent (AC) strains; the term ''AD'' is derived from the strict correlation between loss of HA and immunoprecipitating activities [21] . The AD-viruses were also deficient in complement-fixation and neutralization tests when analysed using either hyperimmune antisera or sera obtained from patients recovered from TBE [21] . Here, using molecular methods of analysis, we show that HA-deficiency is linked with the adaptation of SIB TBEV strains to western European Ixodes ricinus ticks reflecting altered, E protein-mediated, receptor/fusion functions. We also illustrate how this might result in continued westward dispersal and emergence of new highly pathogenic virus variants (see Fig. S1 ). Yar-strains of TBEV i.e. Yar71, Yar 114, Yar 46-2, and Yar 48 were isolated in the European part of Russia (Yaroslavl' region) between 1999-2001 and details for their isolation are listed in Table 1 . All 4 Yar viruses had equivalent infectivities when compared with the Vasilchenko (Vs) strain of TBEV (see Methods) used as the positive control virus (Table 1 ) and produced comparable concentrations of E protein when analysed by Western blot (data not shown). However, they were completely negative in HA tests over a range of pH 5.75-7, regardless of whether they were prepared in newborn mice or in PS cells. The control Vs virus and pGGVs virus, recovered from the infectious clone (see Methods), produced high positive HA titres (1:640-1:1280) at pH 6.2 and this was therefore the pH of choice for all subsequent tests (Table 1) . Thus, the 4 Yar isolates can be defined as HA-or AD-deficient in common with the other 40 strains that have been isolated in Europe and were also identified as HA-and AD-deficient [21] . Yar-viruses belong to the ''Baltic'' group of the Siberian TBEV subtype Phylogenetic analyses based on a 1110-bp fragment of the E gene (positions 1114-2224 in Vs virus L40361) showed that the Yar-viruses belong to the SIB subtype of TBEV [9, [22] [23] [24] [25] [26] , which includes the control Vs virus (see Methods). Fig. 1 illustrates the overall branching pattern in agreement with previously published results [24, 25] . Three Siberian sub-clusters I, II and III were clearly identifiable and supported by high bootstrap values. Yarviruses were grouped with strains of sub-cluster III designated ''Baltic'' [27] , isolated only in Europe, in contrast to strains of clusters I and II that were found across Europe and Asia. Separate phylogenetic analyses based on the C, prM, E and NS1 genes produced trees that were congruent with the one presented (data not shown). To identify amino acid(s) responsible for the loss of HA-activity, we sequenced Yar viruses and aligned them with 290 available TBEV E sequences ( Fig. 2A) . One amino acid 175N in the E protein was common to all HA-deficient Yar-viruses, in contrast with the highly conserved 175T. Therefore, the substitution T for N at amino acid position 175 was introduced into a TBEV infectious clone (IC) designated pGGVs [28, 29] to generate mutant virus IC-T175N that produced HA titres similar to parent Vs virus and also control pGGVs virus rescued from the infectious clone (Table 1 ). Since no other common amino acids were identified that distinguished Yar-viruses from the other strains, we hypothesised that individual non-shared amino acid substitutions might be responsible for the HA-deficient phenotype. The alignment in Fig. 2A revealed 3 non-conserved mutations, each of which was unique to one of the Yar strains, i.e. D67G (Yar 46-2), E122G (Yar 71 and Yar 114) and D277A (Yar 48). When compared with the parent pGGVs virus, each of these mutations increased net charge and hydrophobicity of the E protein and was surface orientated, mapping on the most protruding loops of the E protein in its native dimeric conformation (Fig. 2B) . The loss of HA activity in TBEV has only previously been reported in relation to selective adaptation of TBEV to ticks. One substitution, E87K was generated during the propagation of a WE Figure 1 . Phylogenetic analysis of Yar strains. MEGA version 4 [43] was used to align E genes (between nucleotide positions 1114-2223 of Vs virus genome) of SIB TBEV strains (accession numbers are specified). Tree topology was reconstructed by Neighbor-Joining. The Tamura-Nei model was used for estimation of evolutionary distances [44] . Bootstraps were based on 1000 replications; values below 90% are hidden. The scale bar shows the number of nucleotide substitutions per site. Geographic origins of SIB strains and clusters I, II and III are shown on the right hand side of the tree. Yar strains are highlighted using triangles. KFDV was used as the outgroup. doi:10.1371/journal.pone.0007295.g001 strain to I. ricinus ticks [30] and two substitutions, E122G and T426I respectively, followed a few SIB TBEV passages in H. marginatum ticks [31] . In support of our observations, the E protein surface in these independently reported HA deficient viruses, is predicted to be of either positive or neutral charge, as we have described for the substitutions, D67G, E122G and D277A respectively. Whilst the molecular details of interactions between virions and erythrocytes remain unknown it has been suggested that HA activity might be mediated by the trimeric E protein in its postfusion conformation rather than by native E dimers [32] . Five of these trimers form a fusion pore that enables fusion between the viral and cellular endosomal membrane thus releasing the viral RNA into the cellular cytoplasm. Therefore we mapped the chimaeric Yar-simulated mutants onto the crystal structure of the trimeric (post-fusion) conformation of the E protein [18] . This demonstrated that the 3 Yar-virus mutations are located along the most protruding parts of the lateral surface of the trimer (Fig. 2C ). Therefore they are likely to be able to make direct contact with the erythrocyte surface and/or with each other. Thus, we hypothesised that each of these three amino acid substitutions could individually abolish HA deficiency in TBEV. To test this hypothesis experimentally we used a TBEV infectious clone to engineer three mutant viruses IC-D67G, IC-E122G and IC-D277A (See Methods) that simulate the wild-type viruses Yar 46-2, Yar 71/Yar 114 and Yar 48 ( Table 1 ). The introduction of any one of the three mutations into the parent HA positive pGGVs virus rendered it HA negative (Fig. 3 ). All three engineered mutants demonstrated delayed growth in PS cells, most visible during the first 24 hours (Fig. 4) . Nevertheless, they all subsequently reached titres similar to pGGVs HA-positive virus by day 3 (Table 1) . Although, mutant IC-D67G exhibited better growth characteristics than IC-E122G and IC-D277A it was reproducibly slightly lower than the control pGGVs virus. All 3 natural HA-negative isolates (Yar71, Yar 46-2, and Yar 48) displayed small-plaque phenotype (1 mm, Fig. 4B ), in . Abbreviated comparative alignment of TBEV E protein sequences (complete version is available on request). TBEV strains are specified by subtype (FE-, SIB-or WE-subtypes) and GenBank accession numbers. HA-disabling mutations are encircled. The ''tick-specific'' amino acids that differentiate I. persulcatus-transmitted viruses (shadowed) from the I. ricinus-transmitted viruses are highlighted in yellow; those that increase hydrophobicity are boxed and surface-faced amino acids are marked with asterisks (*). The fusion peptide is underlined. (B) Mapping of HA-disabling residues onto native dimeric conformation of E protein crystal structure [45] (1SVB.pdb); the monomer is shown as it lies on the virion membrane. The ''persulcatus'' and ''ricinus'' residues are highlighted in orange; fusion peptide is in green and HAdeficient amino acids (arrows) are red. The position of Yar substitution D277A (red coloured) coincides with the position of ''ricinus/persulcatus'' substitution (orange colour is masked). (C) Residues 67, 122 and 277 (purple spheres) are mapped onto the E protein in trimeric post-fusion conformation (1URZ.pdb) [18] . The fusion peptide is highlighted in green and can be seen protruding out of the virion membrane towards the endosomal membrane. Different subunits of the trimer are coloured in red, blue and yellow. doi:10.1371/journal.pone.0007295.g002 comparison with control Vs strain which produced much larger plaques (3 mm) but only two corresponding mutations, E122G and D277A caused plaque size reduction in the engineered viruses (0.7 to 1.5 mm as shown in Fig. 4B ). Clearly, as yet unidentified additional mutations contributed to the small plaque phenotype of Yar 46-2 (D67G). Similarly Yar 48 virus formed smaller plaques (1 mm) than the corresponding IC-D277A mutant (1.5 mm) demonstrating that HA activity and plaque phenotype are not always determined by the same amino acid. Vs virus is different from many other laboratory-maintained TBEV strains; in cell culture it develops cytopathic effect (cpe) relatively slowly [22] . All tested mutants showed no increase or decrease of cpe in comparison with the control infectious clone pGGVs, even at high multiplicity of infection (10 PFU/cell). After ip inoculation, the control virus pGGVs produced a relatively high morbidity rate (65%) in correspondence with previous results [29] . In contrast, all three engineered virus mutants exhibited lower neuroinvasiveness as determined by morbidity rates (Fig. 4C ). Mice observed for 21 days following ip inoculation with IC-D67G produced antibodies against TBEV, indicating that they had been infected (data not shown). HA-deficiency correlates with increased TBEV reproduction in feeding ticks and tick-to-tick transmission efficiency Studies on the reproduction and tick-to-tick transmission efficiency of IC-D67G, IC-E122G and IC-D277A were carried out using a novel tick/mouse laboratory model initially developed by Labuda et al [7] . In the first set of experiments, 30 ticks were infected by injection in the leg with each engineered virus as described in Methods and TBEV titres in salivary glands were measured in each of 6 ticks at each time point, i.e. on 2 nd , 7 th , 14 th and 21 st day following infection. The reproduction characteristics of IC-E122G and IC-D277A in fasting ticks were similar to those of control pGGVs virus whereas the titres of IC-D67G were significantly lower (Fig. 5A) . On the 21 st day post-infection, fasting infected ticks were allowed to feed on mice and virus titres were measured in each of 6 individual ticks. This analysis was employed for each engineered virus, 3 days after feeding. Fig. 4A shows that infectivity of IC-E122G had increased approximately 1000-fold and IC-D277A and IC-D67G had increased approximately 300-fold, whereas for control pGGVs virus the increase of virus titre was about 10-fold. Since 6 ticks were used for each tested virus for each time-point, the differences in titres between different virus mutants were statistically significant, based on Student t-tests (p,0.05). In the second set of experiments, the efficiency of virus transmission from infected adult female ticks to uninfected nymphs during co-feeding on mice (tick-to-tick transmission; see Methods) was evaluated by estimating the proportion of infected nymphs. In addition, the titres of virus in each recipient nymph were estimated in a plaque assay. The results show in each case, that HA deficiency directly correlated with increased TBEV titres in nymphs, following feeding and also tick-to-tick transmission efficiency (Fig. 5B) . We analysed 290 TBEV E protein sequences for the presence of amino acid substitutions similar to those that resulted in the loss of HA in TBEV, i.e. acidic (positively charged) residues that were replaced by hydrophobic and/or neutral amino acids (i.e. glycine) and localized on the virion envelope protein surface ( Table 2 ). In total, 5.8% (including Yar-viruses) of the strains exhibited similar mutations. Strains with potentially increased charge and/or surface hydrophobicity were identified in all three TBEV subtypes, i.e. FE, SIB and WE; they were isolated from different geographical regions and a variety of hosts including ticks, rodents and humans. Notably, the substitution D67G was detected in 7 viruses isolated only from mammalian hosts and human patients. No other correlation between isolation source, geography or virus subtype specificity was observed. A significant number of TBEV E proteins have been sequenced partially (227 viruses of 290 available from GenBank), therefore it could not be excluded that TBEV strains with increased surface charge/hydrophobicity are quite common in nature. Indeed among 63 completely sequenced E proteins those with increased charge/hydrophobicity on the virion surface comprise ,20% (including Yar viruses). We also compared E proteins of TBEV strains isolated from I. ricinus (WE-strains) with those isolated from I. persulcatus (FE-and SIB-strains), to identify group amino acids that might be involved in TBEV adaptation to these two different tick species (Fig. 2A ). Ten amino acid differences were revealed that have previously been localised in hypervariable clusters of the envelope protein [33] and five showed overall increased E protein hydrophobicity in the I. ricinus-transmitted WE-strains in contrast with FE-and SIBstrains ( Fig. 2A) . Four of these ten substitutions were localised on the virion surface, with two (S47A and S88G) increasing the surface hydrophobicity of ''ricinus'' strains and two (D178E and D277E) being conserved (Fig. 2B) . Remarkably, the position of the ''tick''-specific (i.e. I. ricinus or I. persulcatus) D277E amino acid substitution overlapped with the D277A substitution of Yar 48 ( Figs. 2A and 2B) . Two substitutions that also increased the hydrophobicity of strains adapted to I. ricinus were localised either inside the E protein (T115A) or on the membrane-oriented side (S267A), i.e. under the virion surface. Four substitutions (T427A, T431S, V433I and L437V) were located in the transmembrane domain of the E protein, with one (T427A) more hydrophobic for ''ricinus'' strains. Louping ill virus (LIV), a TBEV-related virus which is transmitted by I. ricinus in the UK, also showed the same more hydrophobic pattern and localisation of ''tick-specific'' amino acids as the I. ricinus-transmitted WE TBEV strains ( Fig. 2A) . Discussion TBE is currently reported in more than 30 countries of Eurasia and causes significant outbreaks of encephalitis in 16 countries, including 13 EU Member States (Austria, the Czech Republic, Estonia, Finland, Germany, Greece, Hungary, Latvia, Lithuania, Poland, Slovak Republic, Slovenia, Sweden) and three non-EU Member States (Norway, Russia and Switzerland) [3] . The clinical manifestations of TBE in endemic regions vary widely, from inapparent and febrile infections, with recovery of patients, to debilitating or fatal encephalitis, even in some vaccinated individuals [1, 2] and no adequate explanations have as yet been produced. Here, we have investigated the molecular mechanisms and epidemiological implications of the emergence of unusual TBEV strains originally identified as HA-or AD-deficient [21] . In contrast with most TBEV strains, these novel viruses fail to agglutinate avian erythrocytes, show reduced antigenic characteristics and replicate relatively poorly in mammalian cells. We demonstrated that these strains display increased hydrophobicity and positive charge on their virion surface. By engineering genetically modified viruses we proved that these distinct surface characteristics, including HA-deficiency, are caused by any one of three single amino acid substitutions D67G, E122G or D277A in the E protein. These mutations significantly increased virus reproduction in feeding ticks and increased the efficiency of tickto-tick virus transmission when infected and uninfected ticks co-fed on the same animal. Thus mutations leading to HA-and ADdeficiency are directly associated with selection for enhanced virus transmission between ticks, a process that facilitates virus survival in the natural habitat [7] . Although I. ricinus ticks can be routinely maintained in laboratories, SIB TBEV is normally associated with transmission by I. persulcatus, which is not as readily available for laboratory experiments. Nevertheless, WE TBEV strains show ,100% transmission efficiency in their natural tick vector I. ricinus (manuscript in preparation). Thus, logically, Yar viruses (i.e. SIB TBEV strains) should be transmitted efficiently in their natural vector I. persulcatus. Geographically, I. ricinus and I. persulcatus overlap in Europe and numerous reports describe the isolation of Siberian strains from I. ricinus which is now recognised as a second vector for SIB TBEVs [11, 34] . We therefore propose that the driving force behind the westward dispersal of these HA-deficient strains is their adaptation to newly-encountered European I. ricinus ticks. This hypothesis is also supported by our analysis of E protein comparative alignments between FE, SIB and WE subtypes that revealed more hydrophobic ''ricinus'' amino acid patterns, compared with ''persulcatus'', in correspondence with tick preference of WE or FE/SIB strains respectively. Louping ill virus, which is transmitted by I. ricinus in the UK, shares this ''more hydrophobic'' pattern with the WE TBEVs. For many viruses, HA-activity is recognised as a reflection of receptor [35] [36] [37] or low-pH dependent fusion activity [38, 39] . It was suggested that HA activity of flaviviruses is mediated by fusion activity of the E protein [32] because it is optimal at pH 6.2 which promotes conversion of native E protein dimers into fusion-active trimers [18, 19] . This implies that the Yar mutations identified herein destabilise trimer-trimer contacts or contact between trimers and erythrocyte membranes, thus preventing HA activity. Alternatively, these mutations may impact on both E protein functions, ie virus adsorption to the cell surface and pH-dependent fusion of virions with endosomal membranes. Therefore, depending on charge and hydrophobicity, tick cell receptors may restrict ''easy'' entry of WE-strains into I. persulcatus or SIB-strains into I. ricinus. Clearly these barriers to infection are not absolute since Vs virus (SIB TBEV) has a limited capacity to replicate in I. ricinus (Fig. 5 ). Phylogenetic analysis also supports our hypothesis; it was proposed that WE-strains diverged from ancestral FE-and SIB lineages [40] , implying that the emergent WE subtype adapted to I. ricinus from I. persulcatus by evolving a more hydrophobic E protein (Fig. 2) . The coincidence of increased hydrophobicity of the virion surface for I. ricinus-adapted WEstrains and SIB Yar viruses presumably reflects a similar molecular requirement for different viruses to adapt to the same host. Alternatively, the emergence of the atypical Yar viruses may result from adaptation of TBEV to both tick species. Indeed SIB TBEV strains have been isolated from both I. ricinus and I. persulcatus on numerous occasions [11, 34] and regular switching between them cannot be excluded. These data might explain the apparently increasing dissemination of SIB TBEV in Europe; in a few decades this virus could reach more western territories, possibly even the UK where I. ricinus is the vector for the Louping ill virus that is closely related to TBEV. The mutant IC-D67G was distinct from other HA-deficient TBEV strains since there was no obvious correlation between loss of HA activity and significantly reduced growth in mammalian cells. The Yar 71 virus, with the corresponding D67G mutation was isolated from a fatally infected individual (Table 1 ) and ominously, similar substitutions have also been detected in other TBEV isolates from hospitalised patients with encephalitis (Table 2 ). It is possible that due to altered surface charge and hydrophobicity, strains with D67G might be more able to penetrate the human blood-brain barrier (neuroinvasiveness) or more rapidly spread between human neurones, with no correlation to reduced mouse neuroinvasiveness. This would explain the recently discovered association of the most severe form of human encephalitis with the SIB strains [17] . The molecular basis of antigenic deficiency of Yar viruses has not been elucidated but it might be related to the increased incidence of the most severe forms of TBE having been associated with SIB subtype [17] . Clearly more studies are required before we can understand, at the molecular level, the implications of the phenomenon of HA-and AD-deficiency in terms of the development of TBE from fever to encephalitis as an interplay between virus neuroinvasiveness and ability to evade the immune response. Thus, the emergence of HA-deficient TBEV mediated by adaptation to different tick species might represent a mechanism for the westward dissemination of SIB TBEV, increased TBE incidence in Europe, and might also be the reason for encephalitis in humans (see Fig. S1 ). To confirm and develop these ideas, future research should focus on large-scale genomics and transmission studies of TBEV isolates recovered from patients and ticks. Porcine embryo kidney cells (PS) were used to produce TBEV stocks, to recover mutant viruses, for plaque assay and studies of cytopathogenicity. Yar-strains of TBEV i.e. Yar71, Yar 114, Yar 46-2, and Yar 48 were isolated in the European part of Russia (Yaroslavl' region) between 1999-2001 (Table 1 ) and stored as 10% mouse brain suspensions. SIB TBEV strain Vasilchenko (Vs) and its infectious clone (pGGVs) used as control viruses have been described previously [22, 28] . I. ricinus ticks were bred in the Institute of Zoology, Slovak Academy of Science, Bratislava [7] . The RNA of each Yar virus was extracted from 200 ml of 10% infected mouse brain suspension or infected PS cell supernatant using Total RNA Isolation System (PROMEGA). The RT-PCR was used to amplify the 59-C-prM-E gene region of Yar viruses as described [28, 41] . PCR products were directly sequenced using a Taq BigDye Terminator v3.1 Cycle Sequencing Kit (Applied Biosystems). Sequences of the C-prM-E region of Yar-viruses were deposited in GenBank with accession numbers EU444077, EU444078, EU444079 and EU444080 (Table 1 ). Nucleotide and deduced amino acid sequences were aligned using program package BioEdit [42] . Phylogenetic analyses were conducted using MEGA4 [43] . Tree topology was reconstructed by the Neighbor-Joining method and the Tamura-Nei model was used for estimation of genetic distances [44] . The reliability of the tree was evaluated by bootstrapping based on 1000 replications. The unique amino acid substitutions in the E protein of each Yar virus were mapped onto the crystal structure of the native dimeric [45] or low-pH induced trimeric [18] Cytopathic effect (cpe), plaque assay and virus growth cycles in PS cells The routine protocols for cpe, plaque assay and growth curve experiments were described in detail previously [25, 28, 29, 41] . Briefly, for cpe and growth curve experiments PS cells in 24-well plates were infected with viruses at a multiplicity of infection (moi) of 1 PFU/cell, in four replicates. To estimate cpe, the inoculum was replaced with RPMI medium after virus adsorption for 1 h at 37C. Infected cell monolayers were initially examined by microscopy and then stained with 0.1%naphthalene black at 24, 48, 72, 96 or 120 h post-infection for further examination of the extent of cpe. For plaque assay, original virus stocks were 10-fold serially diluted and after 1-h of virus adsorption at 37uC infected monolayers were overlaid with 1% SeaPlaque Agarose (Cambrex, USA). After incubation at 37uC for 5 days monolayers were fixed with 10% formol saline and stained with 0.1% naphthalene black. For virus growth curve experiments, after virus adsorption for 1-h at 37uC, monolayers were washed 5 times with serum-free RPMI medium and overlaid with 1 ml of medium containing 2% FCS. The supernatant medium from infected cells was collected at 2, 10, 12, 16, 20 and 24 hours pi and frozen at 270uC. The titres of infectious virus were determined by plaque assay. Hemagglutination assay PS cells were infected at a moi of 0.1 PFU/cell in 500 ml culture flasks and infectious supernatant medium was collected at day 5pi. The TBEV virions with estimated initial virus titers of 228610 6 PFU/ml were concentrated 100 times using 7% polyethylene glycol (PEG) in the presence of 2.4% NaCl overnight at 4uC and precipitated by centrifugation at 8000 rpm for 3 h. The resulting pellet was resuspended in 500 ml of PBS. For routine hemagglutination test [20, 46] , 50 ml of the concentrated virus sample (10 8 PFU/ml) were placed in the first well of 96-well plates and diluted two-fold in borate buffer (pH 9.0). Then 100 ml of a suspension of 0.5% newborn chick erythrocytes was added to each well. Results are recorded as the reciprocal of the maximum virus dilution that produced agglutination. The ligation in vitro of two overlapping plasmids, pGGVs H and pGGVs 660-1982del produces full-length infectious clone of TBEV strain Vs [28, 29] . The megaprimer-mediated domain swapping mutagenesis technique [47] was utilized to introduce point mutations in pGGVs 660-1982H (Fig. 6 ). For the Hi-Fi PCR, a pair of primers was used to amplify regions of about 200 nucleotides (megaprimer); one primer contained the appropriate point mutation within the E gene. Subsequently, the megaprimer was used to amplify the pGGVs 660-1982 H template in 15 cycles of circular PCR at 95uC for 30 sec, at 60uC for 30 sec and at 72uC for 5 min. Each cycle of the PCR produced nicked dsDNA molecules, with nascent (circular) DNA strand originating from template bacterial dsDNA and the other one (nicked) from newly amplified PCR product. Following 15 cycles, the accumulation of amplified linear complementary ssDNA strands resulted in the formation of annealed circular (twice nicked) molecules. To facilitate screening, the dammethylated bacterial (template) DNA was digested with 40 U of DpnI (New England Biolabs) at 37uC for 1 h. Following this, PCR products were electroporated into AbleK bacterial cells (Stratagene) and selected clones were completely sequenced. To recover engineered viruses, the mutated plasmids, constructed on the basis of pGGVs 660-1982 H, were ligated with pGGVs 660-1982del to restore the full-length cDNA of TBEV as described previously [29] . Each mutant full-length clone was subsequently linearised by Sma I and used for SP6 transcription to produce full-length RNA. The SP6-transcribed RNA was transfected into PS cells using Lipofectin reagent (Invitrogen) according to the manufacturer's protocols. Infectious supernatant medium was collected on day 5 pi. The presence of virus in infected cells was confirmed by immunofluorescence microscopy using monoclonal antibodies specific for flavivirus E proteins [46] and by RT-PCR. The entire E-gene was sequenced to ensure no additional substitutions appeared during the genetic manipulation or initial virus replication in PS cells. Unfed adult females of I. ricinus ticks were inoculated with virus under a stereo zoom microscope (Wild M 400, Wild Heerbrugg AG, Switzerland) into the coaxial plate of the second pair of legs using a digital microinjector TM system (MINJ-D-CE; Tritech Research, Inc.; USA). Clean nitrogen served as a gas source to produce an injection pressure of 20 psi ( = app. 1.38bar). The injection interval was set to 1.0 sec. Hollow glass needles with a microscopically fine tip were prepared using a P-30 Micropipette puller (Sutter Instrument Company, USA). To investigate virus reproduction in fasting ticks, groups of 45 female ticks were infected with 500 PFU/tick of one TBEV strain. Infected ticks were incubated at room temperature (2464uC) and 85-90% RH in a desiccator for 21 days. At 2, 7, 14 and 21 days pi salivary glands of 6 ticks were dissected, individually homogenized and the concentration of infectious virus was estimated by plaque titration. Virus tick-to-tick transmission experiments were carried out essentially as described previously [6] [7] [8] . Two of 45 infected adult female ticks were allowed to feed simultaneously on Balb/C mice for 3 days with 15 uninfected I. ricinus nymphs that were attached in close proximity (1-1.5 cm) to the feeding donor females. Surviving nymphs and salivary glands of donor females were used to determine the titres of infectious virus using plaque assays. The co-feeding transmission rate was estimated as the proportion of nymphs (%) that became infected. Experimental animal procedures were performed in accordance with the guidelines for care and maintenance of animals (Act of the Government of the Slovak Republic 2003 regulating the use of experimental animals). All animal experiments were approved by the State Veterinary and Food Administration of the Slovak Republic (permission numbers 12284/03-220 and 2362/06-221). The ethical permission to carry out the work with mice was obtained from the Ethical Review Committee of the Institute of Virology, Slovak Academy of Sciences''. Ten adult ICR mice were inoculated intraperitoneally (ip) with 2000 PFU/mouse. Mice were observed for 21 days and morbidity rate was estimated as the proportion of animals that showed clinical symptoms that included hind-leg paralysis. Sick and healthy mice were tested for the presence of anti-TBEV antibodies by HA-inhibition test [20] . Statistical analysis was performed on the data obtained from virus replication studies in PS cells and ticks, neutralization and neuroinvasiveness test using EXCEL and SigmaPlot 11 software (Systat Software Inc., USA). Standard errors of mean (SEM) were estimated for each dataset. Between-groups comparisons were performed using unpaired, two-tailed Student's t-test. Values of p,0.05 were considered as significant. Conceived and designed the experiments: MAK VVP ML BK EAG TSG.

Diagnosis and management of drug-associated interstitial lung disease

Symptoms of drug-associated interstitial lung disease (ILD) are nonspecific and can be difficult to distinguish from a number of illnesses that commonly occur in patients with non-small-cell lung cancer (NSCLC) on therapy. Identification of drug involvement and differentiation from other illnesses is problematic, although radiological manifestations and clinical tests enable many of the alternative causes of symptoms in advanced NSCLC to be excluded. In lung cancer patients, high-resolution computed tomography (HRCT) is more sensitive than a chest radiograph in evaluating the severity and progression of parenchymal lung disease. Indeed, the use of HRCT imaging has led to the recognition of many distinct patterns of lung involvement and, along with clinical signs and symptoms, helps to predict both outcome and response to treatment. This manuscript outlines the radiology of drug-associated ILD and its differential diagnosis in NSCLC. An algorithm that uses clinical tests to exclude alternative diagnoses is also described.

The diagnosis of drug-associated interstitial lung disease (ILD) involves three elements: clinical suspicion, differentiation from other parenchymal lung diseases using computed tomography (CT) and other clinical tests for alternative disease, and a compatible histological pattern. This review article will discuss the radiological evaluation of a patient with non-small-cell lung cancer (NSCLC) suspected of suffering from drug-associated ILD, the diagnosis and management of drug-associated ILD, and the development of a diagnostic algorithm designed to distinguish gefitinib ('Iressa')-associated ILD from other forms of parenchymal disease. The onset of drug-associated ILD during therapy for advanced NSCLC usually occurs within a few weeks of the start of treatment (Thomas et al, 2000; Gupta et al, 2002; Kudrik et al, 2002; Read et al, 2002) . Indeed, retrospective analysis of the first 152 patients in Japan to experience gefitinib-associated ILD showed that 475% of cases occurred within 3 months, with the majority of these occurring within 4 weeks. The symptoms of drug-associated ILD, as with all forms of the condition, include rapidly developing breathlessness and a dry and unproductive cough, together with fever (Thomas et al, 2000; Gupta et al, 2002; Kudrik et al, 2002; Read et al, 2002) . Such symptoms are nonspecific and can occur with a large number of common illnesses often associated with NSCLC, or they may be due to cancer progression or lung cancer therapy. These patients are also prone to pneumonia and many have radiation-related injury as a result of prior treatment. Cardiovascular causes of the symptoms, such as fluid overload, congestive heart failure and pulmonary embolus, are not uncommon. Differentiation of drug-associated ILD from these illnesses is difficult and the diagnosis is usually made by exclusion. Highresolution CT (HRCT) is recommended on first suspicion of ILD to provide an assessment of the parenchymal nature of the cause of symptoms. The radiological manifestations of drug-associated ILD, although heterogeneous and nonspecific, enable many of the alternative causes of symptoms in advanced NSCLC to be excluded. There is no specific radiological pattern of parenchymal change connected with drug-associated ILD. Furthermore, in the early stages of disease, patients with symptoms secondary to drug reaction may have a normal chest radiograph. High-resolution CT allows a more precise assessment of the presence, pattern and distribution of parenchymal and airway abnormalities than a chest radiograph. It has the advantage over lung biopsy of providing an overall view of the extent and pattern of parenchymal involvement rather than being limited to a small region, which may not be representative of the overall pattern of disease. However, there is limited information on the correlation between the findings on HRCT and histological patterns in drug-associated lung disease. Data based on a small number of cases suggest that the different histological patterns of drug reaction are not reflected by distinctive HRCT findings (Cleverley et al, 2002) . *Correspondence: Professor NL Müller; E-mail: nmuller@vanhosp.bc.ca Despite these limitations, it seems reasonable to approach the radiological manifestations of drug-associated lung disease by the use of the underlying histological pattern (Ellis et al, 2000; Myers et al, 2003) . Using such an approach, the most common manifestations can be classified into diffuse alveolar damage, hypersensitivity pneumonitis, organising pneumonia, nonspecific interstitial pneumonia (NSIP) and eosinophilic pneumonia (other less common drug reactions are not discussed here). Any one histological pattern can be caused by a number of different drugs. Furthermore, similar histology is found in other conditions that are not associated with drug use, such as idiopathic interstitial pneumonias, viral, bacterial and fungal pneumonia, pulmonary haemorrhage or leukaemia and collagen vascular disease. Occasionally, HRCT may demonstrate findings that are highly specific for the diagnosis, including increased attenuation in amiodarone lung and areas of fat attenuation in lipoid pneumonia. Diffuse alveolar damage is characterised histologically by the presence of alveolar airspace and interstitial oedema, hyaline membrane formation and proliferation of type II pneumocytes (Rossi et al, 2000; Cleverley et al, 2002) . In relation to drug-associated pulmonary disease, it occurs most commonly with cytotoxic agents such as bleomycin and, less commonly, with aspirin and narcotics (Rossi et al, 2000; Cleverley et al, 2002) . The corresponding radiological features are also found in adult respiratory distress syndrome (ARDS). The chest radiograph shows bilateral patchy or homogeneous airspace consolidation involving mainly the middle and lower lung zones (Rossi et al, 2000; Cleverley et al, 2002) . High-resolution CT demonstrates extensive bilateral ground-glass opacities and dependent areas of airspace consolidation (Rossi et al, 2000; Erasmus et al, 2002) ( Figure 1A ). A number of drugs may result in hypersensitivity pneumonitis, including methotrexate, cyclophosphamide and antidepressants such as fluoxetine and amitriptyline (Ellis et al, 2000) . The radiological and HRCT findings are identical to those seen in hypersensitivity pneumonitis secondary to the inhalation of organic dust and consist of bilateral ground-glass opacities and/ or small poorly defined centrilobular nodular opacities (Ellis et al, 2000; Cleverley et al, 2002) . The majority of patients also demonstrate lobular areas of air trapping, although this is less Figure 1 High-resolution CT images demonstrating radiology of drug-associated ILD. (A) A 77-year-old man with diffuse alveolar damage secondary to amidarone; note the extensive bilateral ground-glass opacities, airspace consolidation and bilateral pleural effusions. (B) A 36-year-old woman with hypersensitivity pneumonitis secondary to sertraline; note the extensive bilateral ground-glass opacities and lobular areas of air trapping (arrows). (C) A 69-year-old man with BOOP-like reaction to amiodarone; note the mild reticulation and bilateral areas of consolidation and ground-glass opacities in a predominantly peribronchial distribution. (D) A 47-year-old man with NSIP reaction to bleomycin; note the extensive bilateral ground-glass opacities with mild superimposed reticulation. (E) A 47-year-old man with eosinophilic pneumonia reaction to dilantin; note the patchy bilateral areas of consolidation involving the peripheral regions of the upper lobes. apparent than in extrinsic allergic alveolitis to airborne agents (Ellis et al, 2000) ( Figure 1B ). Organising pneumonia, also known as bronchiolitis obliterans organising pneumonia (BOOP)-like reaction, has been reported most frequently in association with methotrexate, cyclophosphamide, gold, nitrofurantoin, amiodarone, bleomycin and busulphan (Cleverley et al, 2002; Erasmus et al, 2002) . The chest radiograph shows patchy bilateral areas of consolidation, masses or nodules, which may be symmetric or asymmetric (Müller et al, 1990; Ellis et al, 2000) . In a few patients, the disease manifests with a lone mass. On HRCT, areas of consolidation often have a predominantly peripheral or peribronchial distribution (Müller et al, 1990; Ellis et al, 2000) ( Figure 1C ). Nonspecific interstitial pneumonia is one of the most common forms of drug-associated pneumonitis. Nonspecific interstitial pneumonia is characterised histologically by homogeneous alveolar wall thickening by fibrous tissue and mononuclear inflammatory cells. The reaction is seen in association with a variety of drugs, the most common being methotrexate, amiodarone and carmustine (Erasmus et al, 2002) . The corresponding radiographical and HRCT findings usually consist of patchy or diffuse ground-glass opacities (Rossi et al, 2000; Erasmus et al, 2002) ( Figure 1D ). On disease progression, there may be evidence of fibrosis with development of a reticular pattern and traction bronchiectasis. In some patients, the fibrosis is patchy in distribution and predominantly peribronchovascular, a pattern most commonly seen in patients receiving nitrofurantoin. Late chemotherapy lung may predominate in the upper and lateral parts of the lung. Eosinophilic pneumonia is characterised histologically by the accumulation of eosinophils in the alveolar airspaces and infiltration of the adjacent interstitial space by eosinophils and variable numbers of lymphocytes and plasma cells. Peripheral blood eosinophilia is present in p40% of patients. Eosinophilic pneumonia secondary to drug reaction is seen most commonly in association with methotrexate, sulphasalazine, para-aminosalicylic acid, nitrofurantoin and nonsteroidal anti-inflammatory drugs. Chest radiography and HRCT show bilateral airspace consolidation, which tends to involve mainly the peripheral lung regions and the upper lobes (Rossi et al, 2000; Cleverley et al, 2002 ) ( Figure 1E ). Alternative diagnoses to drug-associated ILD in NSCLC include progression of the cancer, infection, radiation-related lung injury, fluid overload, congestive heart failure and pulmonary embolus. Additionally, some lung cancer patients may develop BOOP or other steroid-responsive inflammatory disorders that cannot be clearly related to drug therapy. Other possible causes of dyspnoea that do not give infiltrates include comorbid diseases such as chronic obstructive pulmonary disease (COPD) and aspiration of food and saliva (particularly in patients with vocal cord paralysis or brain metastases). In the USA, pulmonary embolism is particularly common in patients with lung cancer, with as many as 20% of patients estimated to develop a deep vein thrombosis or pulmonary embolism during the course of their disease (Lieberman et al, 1961; Sack et al, 1977; Lee and Levine, 2003) . Most episodes of pneumonia in patients with lung cancer are due to bacteria. This is particularly the case when risk factors of neutropenia, endobronchial lesions, underlying COPD and aspiration are present. Although opportunistic infections are not common, fungal infections should be considered if the patient has received a high dose of corticosteroids. Viral infections with herpes simplex, cytomegalovirus or respiratory syncytial virus may also rarely result in pneumonia in patients who have received highdose corticosteroids or very intensive chemotherapy. The development of lung fibrosis following radiation therapy is well documented and is usually confined to the radiation port (Abid et al, 2001; Aviram et al, 2001) . With the use of threedimensional radiation portals, however, resulting infiltrates from radiation may not result in the traditional straight-edged infiltrate and may be more difficult to distinguish from other entities. Patients with lung cancer who present with respiratory failure should undergo systematic investigation. Pulmonary function tests, such as measurement of forced expiratory volume in 1 s, carbon monoxide diffusing capacity and measurement of arterial oxygen saturation with pulse oximetry, are commonly used. These tests ascertain the type of defect, for example, obstructive or restrictive ventilatory defects, which helps establish the cause, such as an exacerbation of any underlying obstructive airways disease versus interstitial disease. They also indicate the severity of the disease that helps determine the need for further assessment and treatment. A standard chest CT scan is commonly performed to exclude a diagnosis of disease progression or pulmonary embolism. However, as discussed previously, obtaining high-resolution cuts is very helpful (HRCT) if drug-associated ILD is suspected. Bronchoscopy is useful to evaluate some NSCLC patients with dyspnoea to assess extension of the cancer and to exclude an opportunistic infection using bronchioalveolar lavage. However, the value of a bronchoscopy in establishing the diagnosis of drugassociated ILD is less clear, since bronchioalveolar lavage is not specific for drug-associated disease, and biopsies obtained by transbronchial biopsy are small and often do not yield enough tissue to make this distinction. Open lung biopsy is rarely performed in NSCLC patients with respiratory distress since most patients have advanced disease with a limited prognosis. Furthermore, procedures requiring surgery are not usually believed to arrive at a definitive diagnosis of drug-associated ILD. Patients with mild symptoms or pulmonary function abnormalities (such as a decrease in diffusing capacity of o20% from baseline or no change in oxygen desaturation during exercise), or with transitory or slight radiographical infiltrates are monitored using pulmonary function tests, symptoms and usually CT scans. Diagnostic evaluation and treatment for drug-related lung disease is considered in those patients who experience dyspnoea at rest or on mild exertion, have a X20% decrease in carbon monoxide diffusing capacity, or experience oxygen desaturation at rest or during exercise. Patients whose radiographical infiltrates are extensive or progressive are also considered for therapy. Retrospective analysis of the adverse-event reports from patients diagnosed with ILD following gefitinib treatment is also difficult as there is often limited or heterogeneous clinical information, no pathology result and no access to the results of radiological investigations. A diagnostic algorithm has therefore been developed to assess the accuracy of the reports of drug-associated ILD among Japanese patients receiving gefitinib. This approach to differential diagnosis used an algorithm developed to aid early diagnosis of drugassociated ILD in clinical practice. The algorithm used both radiological and clinical evidence to exclude alternative diagnoses, such as infection, tumour progression, heart failure and pulmonary embolism, to arrive at a presumptive diagnosis of drug-associated ILD (Figure 2 ; Table 1) . Patients were categorised based on the strength of evidence supporting the diagnosis of drug-associated ILD (category 1, good supporting evidence for ILD; category 2, limited supporting evidence for ILD; category 3, no supporting evidence for ILD). In the first assessment, adverse-event reports from Japanese patients receiving gefitinib for NSCLC were evaluated using the algorithm, with available clinical information and radiographical reports but without access to chest radiographs or HRCT films. These findings were then compared with those of a second assessment, conducted by an independent panel of expert radiologists and physicians. The panel assessed the radiological and clinical findings for the same patient population using a set of standardised criteria (Cleverley et al, 2002) . The initial 152 reported patients with NSCLC in Japan who had experienced adverse events involving the lungs while receiving treatment with gefitinib were included in the first assessment of the algorithm. Of these, 135 were included in the second assessment because radiological examinations, including 47 with CT imaging, were available. Approximately 20% of these patients (23 out of 135) were considered by the expert panel not to have drug-associated ILD, highlighting the difficulty in diagnosing drug-associated ILD in patients with NSCLC. These results were then compared with those obtained using the algorithm. A large proportion (17 out of 23) of patients not considered by the panel to have drug-associated ILD had been categorised as having 'good' or 'limited' supporting evidence for drug-associated ILD. The initial algorithm based on the terminology used in radiology reports, without considering differential diagnosis criteria, was not adequate. Infection, heart failure and tumour progression can be differentially diagnosed and excluded using additional radiological clinical data. Therefore, the algorithm is now used to enable the clinician to make a diagnosis having first undertaken all the necessary steps in the clinical examination and investigation. This approach is applied to the prospective nested case -control study to investigate the relative risk and risk factors for ILD in NSCLC patients in Japan treated with and without gefitinib. An independent case review board will review each reported case of ILD using the information gathered from the algorithm. General principles in the management of acute respiratory distress in patients with lung cancer are influenced by the multiple causes of respiratory failure that are associated with cancer, lung cancer therapy and the presence of comorbid disease. In addition, the diagnosis may remain uncertain, even when invasive procedures are performed, and empirical therapy for the likely causes is frequently given. Finally, respiratory failure in patients with cancer results in high mortality (Groeger et al, 1999) requiring aggressive assessment and therapy. There are no firm guidelines for the treatment of drug-associated ILD and therapy tends to be on an empirical basis. Withdrawal of the drug suspected of causing the ILD is the first step in treatment. For patients in respiratory failure, high-dose methylprednisolone (250 mg four times a day i.v.) for several days is commonly used. If the patient responds, then the dose is reduced (0.5 -1 mg kg À1 day À1 orally) for several weeks before being gradually tapered. For patients in respiratory distress, methylprednisolone (1 mg kg À1 day À1 or 60 mg day À1 ) is commonly used, again with gradual dose reduction. Low-dose methylprednisolone (10 -20 mg) is prescribed for patients with mild radiographical or pulmonary function abnormalities, particularly if oral corticosteroids are contraindicated (White and Stover, 1984; Baughman et al, 1994; American Thoracic Society, 2002) . Immunosuppressive agents such as azathioprine have been used as steroid-sparing agents in the treatment of drug-associated ILD, particularly in chronic cases of bleomycin-associated ILD (Maher and Daly, 1993) . These agents are useful for patients in whom corticosteroids cannot be tapered or who cannot tolerate corticosteroids. It is also advisable to avoid combining agents associated with ILD, such as bleomycin and possibly mitomycin, with other agents that cause lung damage, such as oxygen and radiation, as this may result in worsening of the lung damage. If radiotherapy is indicated in a patient who has experienced mild bleomycin toxicity, then concomitant low-dose corticosteroid may minimise any further lung damage. Following resolution of the drug-associated ILD, some patients are susceptible to exacerbations on subsequent insults (e.g. during a respiratory infection) and may require further treatment. In cancer patients, drug-associated ILD is most commonly observed during mitomycin, paclitaxel, docetaxel or gemcitabine therapy. Table 2 outlines the types of injury reported with these agents and their response to therapy (Buzdar et al, 1980; Chang et al, 1986; Goldberg and Vannice, 1995; Rivera et al, 1995; Ramanathan and Belani, 1996; Bookman et al, 1997; Merad et al, 1997; Piccart et al, 1997; Semb et al, 1998; Vander Els and Miller, 1998; Dunsford et al, 1999; Thomas et al, 2000; Fogarty et al, 2001; Read et al, 2002; Barlesi et al, 2004) . Acute pneumonitis, both interstitial and noncardiac pulmonary oedema pattern, has been observed with mitomycin. Response to high-dose methylprednisolone has been reported within 24 -48 h of therapy; however, approximately 60% of patients developed ongoing and persistent lung disease (Rivera et al, 1995) . Chronic pneumonitis, similar to bleomycin-type pneumonitis, has also been reported during mitomycin therapy, which responded to prednisone therapy (Buzdar et al, 1980; Chang et al, 1986) . As with bleomycin, it has been suggested that oxygen therapy be avoided during mitomycin therapy, although the evidence for this is not as established as that for bleomycin (Klein and Wilds, 1983) . Infusion hypersensitivity is very common during paclitaxel therapy. Pretreatment with an antihistamine, a corticosteroid and an H2 blocker largely prevents or ameliorates this reaction (Bookman et al, 1997) . Mild interstitial pneumonitis with transitory infiltrates has also been reported; observation or lowdose corticosteroids resulted in a good response, even when this occurred in conjunction with radiotherapy (Goldberg and Vannice, 1995; Ramanathan and Belani, 1996) . Fluid retention syndrome during docetaxel therapy is very common and is associated with dyspnoea. This reaction can be prevented by dexamethasone (Piccart et al, 1997; Semb et al, 1998) . Interstitial pneumonitis and ARDS-like pattern have also been reported with docetaxel, particularly when it is used in combination with radiotherapy or gemcitabine. Both reactions respond well to corticosteroid therapy; however, when docetaxel is used in combination with radiotherapy or gemcitabine, the response to corticosteroid therapy can be variable to poor, with some deaths reported (Merad et al, 1997; Dunsford et al, 1999; Read et al, 2002) . Self-limiting dyspnoea is common during gemcitabine therapy. Additionally, many patients experience a mild capillary leak throughout the lung, similar to that seen in ARDS. However, these patients are asymptomatic and have normal pulmonary function; therefore, gemcitabine therapy can be continued without the need for corticosteroids. Gemcitabine-associated ILD has also been reported and the response to corticosteroid therapy is variable. As with docetaxel, more deaths due to ILD have been reported when gemcitabine is used in combination with other therapies (radiotherapy or chemotherapy) (Vander Els and Miller, 1998; Dunsford et al, 1999; Thomas et al, 2000; Fogarty et al, 2001; Barlesi et al, 2004) . In the management of patients with NSCLC, it is often necessary to treat patients with an agent that has been previously associated with ILD. In this scenario, it is important to consider the patient's previous response to that therapy, the severity of the lung damage and the respiratory distress, the presence of fibrosis and the previous response to corticosteroid therapy. For some patients, the benefits may outweigh the risks and therapy may be re-instituted with concomitant low-dose prednisone (10 -20 mg day À1 ); however, for some patients and/or agents the potential for causing further lung damage may be too great. In summary, lung cancer patients receiving systemic therapy frequently develop dyspnoea and infiltrates. It is difficult to make a specific diagnosis in most cases because of the difficulty of performing invasive procedures in this patient population. Radiological assessment, and HRCT in particular, can play a key role in establishing a diagnosis of drug-associated ILD; however, in the vast majority of cases, the radiological manifestations of drugassociated pulmonary injury are nonspecific, making an accurate diagnosis difficult. Corticosteroids are indicated for suspected drug-associated ILD; however, the outcome is variable unless the patient develops respiratory failure, in which case the mortality is high. As a result, clinicians are reluctant to withhold corticosteroid therapy if there is any indication of drug toxicity, further complicating the diagnosis. However, once a patient responds to corticosteroid therapy, the decision to re-institute a drug suspected of causing ILD is made on an individual basis. In this article we described an algorithm developed to assess the incidence of drugassociated ILD in Japanese patients receiving gefitinib for NSCLC. Such an algorithm, once validated, may be a useful tool in the differential diagnosis of ILD in patients with cancer and help clarify some of the apparent discrepancies in the incidence and reporting of ILD. Buzdar et al (1980) , Chang et al (1986) , Goldberg and Vannice (1995) , Rivera et al (1995) , Ramanathan and Belani (1996) , Bookman et al (1997) , Merad et al (1997) , Piccart et al (1997) , Semb et al (1998), Vander Els and Miller (1998) , Dunsford et al (1999) , Thomas et al (2000) , Fogarty et al (2001) and Read et al (2002) . a Even when paclitaxel is used in conjunction with radiotherapy. b When docetaxel is used in combination with radiotherapy or gemcitabine, the response is variable to poor.

Primary biliary cirrhosis

Primary biliary cirrhosis (PBC) is an immune-mediated chronic cholestatic liver disease with a slowly progressive course. Without treatment, most patients eventually develop fibrosis and cirrhosis of the liver and may need liver transplantation in the late stage of disease. PBC primarily affects women (female preponderance 9–10:1) with a prevalence of up to 1 in 1,000 women over 40 years of age. Common symptoms of the disease are fatigue and pruritus, but most patients are asymptomatic at first presentation. The diagnosis is based on sustained elevation of serum markers of cholestasis, i.e., alkaline phosphatase and gamma-glutamyl transferase, and the presence of serum antimitochondrial antibodies directed against the E2 subunit of the pyruvate dehydrogenase complex. Histologically, PBC is characterized by florid bile duct lesions with damage to biliary epithelial cells, an often dense portal inflammatory infiltrate and progressive loss of small intrahepatic bile ducts. Although the insight into pathogenetic aspects of PBC has grown enormously during the recent decade and numerous genetic, environmental, and infectious factors have been disclosed which may contribute to the development of PBC, the precise pathogenesis remains enigmatic. Ursodeoxycholic acid (UDCA) is currently the only FDA-approved medical treatment for PBC. When administered at adequate doses of 13–15 mg/kg/day, up to two out of three patients with PBC may have a normal life expectancy without additional therapeutic measures. The mode of action of UDCA is still under discussion, but stimulation of impaired hepatocellular and cholangiocellular secretion, detoxification of bile, and antiapoptotic effects may represent key mechanisms. One out of three patients does not adequately respond to UDCA therapy and may need additional medical therapy and/or liver transplantation. This review summarizes current knowledge on the clinical, diagnostic, pathogenetic, and therapeutic aspects of PBC.

Primary biliary cirrhosis (PBC) [1] is an immune-mediated chronic progressive inflammatory liver disease that leads to the destruction of small interlobular bile ducts, progressive cholestasis, and, eventually, fibrosis and cirrhosis of the liver without medical treatment commonly necessitating liver transplantation. Addison and Gull [2] have first described a disease with a PBC-like picture in 1851, but the term "primary biliary cirrhosis" was coined in 1949 when a cohort of 18 patients with characteristic features of PBC was published [3] . PBC, predominantly affecting middle-aged women, is characterized by biochemical markers of cholestasis, serum antimitochondrial autoantibodies (AMA), and lymphocytic infiltration of the portal tracts of the liver [4] . Histologically, the hallmark of the disease is damage to biliary epithelial cells (BEC) and loss of small intrahepatic bile ducts accompanied by significant portal tract infiltration with CD4 and CD8 T cells, B cells, macrophages, eosinophils, and natural killer cells [5, 6] . Being one of the first conditions in which specific autoantibodies were recognized, PBC is regarded as a "model autoimmune disease." Both environmental factors and inherited genetic predisposition appear to contribute to its pathogenesis [7] . Although there has been tremendous progress in unraveling potential pathophysiologic factors in PBC over the past years [7] , the actual impact of each of the identified genetic and environmental associations is still controversial. It is the aim of this article to review the current knowledge of major pathologic features in PBC and to try to combine these findings to an overall picture of the pathogenesis of PBC. The most frequent symptoms in PBC are fatigue and pruritus, occurring in up to 85% and 70% of patients, respectively [8, 9] . Median survival in untreated individuals has been reported to be 7.5 to 16 years [1, 10] , but has largely improved since the introduction of ursodeoxycholic acid (UDCA) therapy and liver transplantation. Patients that are treated with UDCA at an early stage of the disease and respond well to therapy may reach normal life expectancy [11] [12] [13] [14] . However, the beneficial mechanisms of UDCA treatment are still incompletely understood, and about one third of patients fail to adequately respond to UDCA monotherapy. In the second part of this review, we will therefore summarize the rationale behind UDCA therapy and give an overview on future therapeutic options currently under study. PBC occurs in individuals of all ethnic origins and accounts for up to 2.0% of deaths from cirrhosis [15] . It primarily affects women with a peak incidence in the fifth decade of life, and it is uncommon in persons under 25 years of age. Incidence and prevalence vary strikingly in different geographic regions (as does the quality of epidemiological studies related to PBC), ranging from 0.7 to 49 and 6.7 to 402 per million, respectively [16] [17] [18] [19] [20] [21] [22] [23] . The highest incidence and prevalence rates are reported from the UK [16, 21] , Scandinavia [17] , Canada [18] , and the USA [19, 22] , all in the northern hemisphere, whereas the lowest was found in Australia [20] . There is no clear evidence to support or exclude the concept of "a polar-equatorial gradient," as it has been reported for other autoimmune conditions [24] . Increased awareness of the condition and the increasing availability of diagnostic tools, in particular serological testing, have led to a more frequent and earlier diagnosis of PBC [25] . More than half of patients diagnosed today with PBC are asymptomatic at presentation [26, 27] . They generally attract attention by findings of elevated serum alkaline phosphatase (AP) and/or total serum cholesterol, especially in asymptomatic patients often during routine checkup. A diagnosis of PBC is made "with confidence" when biochemical markers of cholestasis, particularly alkaline phosphatase, are elevated persistently for more than 6 months in the presence of serum AMA and in the absence of an alternative explanation [28, 29] . In PBC patients, AMA is generally present in high titer. Low-titer AMA may not be specific and may disappear on retesting [30] . Compatible histological findings confirm the diagnosis and allow staging before therapeutic intervention, but in many cases, histological workup is not necessary to diagnose PBC [29] . Serum AP and γGT are commonly elevated and define, together with AMA, the diagnosis of PBC. Mildly elevated serum aminotransferases (ALT, AST) are usually observed in PBC but are not diagnostic. Increased serum levels of (conjugated) bilirubin as well as alterations in prothrombin time and serum albumin are late phenomena in PBC like in other cirrhotic states and unusual at diagnosis. However, serum bilirubin is a strong and independent predictor of survival [31] with a high impact on all established models for prognosis. Serum cholesterol is commonly elevated in patients with PBC, alike other cholestatic conditions. The increased cholesterol level in PBC is largely caused by the presence of LpX [32] . LpX is an abnormal lipid particle which is characteristic for cholestatic liver disease and that is directly derived from biliary lipids that regurgitate into the blood [33] . Upon lipoprotein fractionation, LpX is usually found in the very low-density lipoprotein fraction but is very different from other lipoproteins. In contrast to normal lipoproteins, which have a core filled with neutral lipid, LpX consists of liposomes (with an aqueous lumen) of phospholipids and free cholesterol. LpX is not taken up in atherosclerotic plaques and may reduce the atherogenicity of low-density lipoprotein (LDL) cholesterol by preventing LDL oxidation [34] . Accordingly, the increased serum cholesterol in PBC patients is not associated with increased risk for cardiovascular disease. In contrast, long-lasting PBC is associated with the occurrence of xanthomata and xanthelasma. Hypercholesterolemia in PBC patients is responsive to statin treatment. Long-term treatment with UDCA also reduces serum cholesterol [35] . Though not diagnostic, thyroid-stimulating hormone levels should be assessed in any patient who is believed to have PBC due to the high association of PBC with thyroid dysfunction, mainly caused by Hashimoto thyroiditis. AMA autoantibodies are pathognomonic for PBC and lead to the diagnosis with a high specificity and sensitivity. AMA-positive individuals, even if no signs of cholestasis and/or liver inflammation are present, are very likely to develop PBC. Mitchison et al. [36] , in a small study, evaluated liver pathology of 29 asymptomatic AMApositive (levels>1:40) individuals lacking AP elevation. At inclusion, all but two had abnormal liver histology, and in 12, findings were diagnostic for PBC. A 10-year followup of these subjects revealed that 24 of the 29 remained AMA-positive; all 24 developed biochemical evidence of cholestasis and 22 became symptomatic [37] , confirming a high positive predictive value of positive AMA testing for the development of PBC. Sensitivity of AMA, however, although high, is limited. Investigators have reported patients who clinically, biochemically, and histologically have all the features of PBC despite consistently negative AMA testing both by immunofluorescence and with the most specific immunoblotting and immunoenzymatic techniques [38] [39] [40] [41] [42] . Overall, AMA seems to be negative in 5% of patients who otherwise have all the features typical for PBC [43] and an identical autoreactive CD4 T cell response to the critical autoantigen, PDC-E2 [44] . This small group of AMA-negative patients, however, may erroneously include patients with PBC-like symptoms induced by causes other than autoimmunity, such as patients with mutations in the ABCB4 (MDR3) gene [45] . These patients secrete reduced amounts of phospholipids into the bile, which is harmful to hepatocytes and cholangiocytes. The pattern of serum immunoglobulin fractions in PBC is characterized by an elevation of serum IgM [46] , possibly due to an abnormal chronic B cell activation by Toll-like receptor-dependent signaling [47] . However, specificity of this finding is limited and IgM levels are not commonly used as a diagnostic criterium. Nonspecific antinuclear antibodies (ANA) and/or smooth muscle antibodies are found in serum of one third of patients with otherwise clear-cut PBC [48] , but are of limited diagnostic value. In contrast, specific ANA directed against nuclear body or envelope proteins such as anti-Sp100, presenting as multiple (6) (7) (8) (9) (10) (11) (12) nuclear dots at indirect immunofluorescence staining and anti-gp210, presenting as perinuclear rims have shown a specificity of >95% for PBC, although their sensitivity is low. These specific ANA can be used as diagnostic markers for PBC in the absence of AMA titers [29] . Ultrasound examination of the liver and biliary tree is obligatory in all cholestatic patients in order to differentiate intrahepatic from extrahepatic cholestasis. When the biliary system appears normal and serum AMA are present, no further radiologic workup is necessary. Abdominal lymphadenopathy, particularly in the hilar region of the liver, is seen in 80% of patients with PBC [49] . Transient elastography (TE) has been introduced as a new, simple and noninvasive imaging technique for determining the degree of fibrosis in patients with chronic liver diseases, mainly chronic hepatitis C [50] . Corpechot et al. [51] compared liver stiffness as determined by TE (Fibroscan R ) to histological findings obtained by liver biopsy in 101 patients with PBC (n = 73) or primary sclerosing cholangitis (PSC; n = 28) and showed a highly significant correlation of liver stiffness with both degree of fibrosis and histological stage. Further studies in independent cohorts of PBC patients are warranted before TE can be regarded as an established alternative to liver biopsy in the staging of chronic cholestatic liver disease. Still, TE appears attractive as a screening tool in future therapeutic trials, as it may help overcome the limited staging accuracy of liver biopsy due to heterogeneous distribution of inflammation and fibrosis in PBC. A liver biopsy is not anymore regarded as mandatory for the diagnosis of PBC in patients with elevated serum markers of cholestasis and positive serum AMA [28, 29] , but may be helpful in excluding other potential causes of cholestatic disease and in assessing disease activity and stage. A liver biopsy may also be helpful in the presence of disproportionally elevated serum transaminases and/or serum IgG levels to identify additional or alternative processes. Histological staging of PBC (stage 1 to stage 4) is determined by the degree of (peri)portal inflammation, bile duct damage and proliferation, and the presence of fibrosis/cirrhosis according to Ludwig et al. [52] and Scheuer [53] . Stage 1 disease is characterized by portal inflammation with granulomatous destruction of the bile ducts, although granulomas are often not seen. Stage 2 is characterized by periportal hepatitis and bile duct proliferation. Presence of fibrous septa or bridging necrosis is defined as stage 3 and cirrhosis as stage 4 [52] . Findings of fibrotic or cirrhotic changes (stage 3 or 4) are accompanied by a worse prognosis [54] . Florid duct lesions as defined by focal duct obliteration and granuloma formation are regarded as typical for PBC. The liver is not uniformly involved, and features of all four stages of PBC can be found in one biopsy specimen. The most advanced histological features are used for histological staging. At diagnosis, the majority of patients are asymptomatic and present e.g. for workup of elevated serum levels of AP or cholesterol [55, 56] . In symptomatic patients, fatigue and pruritus are the most common complaints and have been reported in 21% and 19% of patients at presentation, respectively [27, 57] . Unexplained discomfort in the right upper quadrant of the abdomen has been reported in approximately 10% of patients [58] . In the majority of asymptomatic and untreated patients, overt symptoms develop within 2 to 4 years, although one third may remain symptom-free for many years [27, 56] . During the course of the disease, up to 80% of PBC patients complain of chronic fatigue impairing quality of life and interfering with daily life activities [8, 59] . No correlation with the severity of the liver disease could be demonstrated [59] , but there is an association with autonomic dysfunction (in particular orthostatic hypotension) [60] , sleep disturbance and excessive daytime somnolence [60] , and, although weak, depression [61] , which might necessitate treatment for themselves. The exact pathophysiological mechanisms leading to chronic fatigue in PBC and other cholestatic diseases are not unraveled so far. Standard therapy of PBC with UDCA and even liver transplantation may fail to improve this often disabling symptom. PBC is more frequently associated with pruritus than other chronic cholestatic liver diseases. During the course of the disease, pruritus occurs in 20% to 70% of patients and can often be the most distressing symptom [62] . It develops independently of the degree of cholestasis and the stage of disease. Its pathogenesis remains poorly understood and the potential pruritogens in cholestasis are undefined. The therapeutic efficacy of anion exchange resins like cholestyramine, the pregnane X receptor agonist, rifampicin, plasmapheresis, and albumin dialysis as well as nasobiliary drainage led to the conclusion that in cholestasis putative pruritogens accumulate in the circulation, are secreted into bile, and undergo an enterohepatic cycle. Itch could then be induced locally in the skin or in neuronal structures. Bile salt metabolites, progesterone metabolites, histamine, and endogenous opioids, among others, have all been proposed as causative agents. However, evidence for a key role of any of these suggested pruritogens in cholestasis is weak [63] . Variceal hemorrhage or other signs of portal hypertension secondary to fibrosis or cirrhosis are uncommon at first presentation, but have occasionally been described [64] . However, portal hypertension as determined by measurement of the portohepatic pressure gradient (PHG) is common in PBC, and a stable or -under treatmentimproved PHG is a predictor of survival [65, 66] . Patients with PBC have been reported to be at increased risk of osteoporosis in some studies [67] , but reports in the literature remain contradictory. Patients with advanced PBC may be at particular risk for osteoporosis and may present with osteoporosis and yet be otherwise asymptomatic from their liver disease. The development of osteoporosis in PBC patients has been attributed to both, decreased osteoblast activity and increased osteoclast activity [68] . Metabolism of vitamin D is normal in PBC, but malabsorption of both calcium and vitamin D may occur predominantly in latestage disease. Pancreatic insufficiency and celiac disease, which are associated with PBC [69] [70] [71] , may further aggravate malabsorption. When secretion of bile and bile salts is insufficient, i.e., bile salts drop below the critical micellar level in the duodenum, malabsorption of both fat and fat-soluble vitamins may ensue. Serum levels of vitamin A and E have been shown to be low in a minority of patients with primary biliary cirrhosis prior to development of jaundice [72] . Osteomalacia is barely seen in PBC, as a liver transplant is performed in most patients before the development of this complication of prolonged deep jaundice. Though often asymptomatic, recurrent urinary tract infections have been reported in up to 19% of women with PBC [73] , and a potential pathophysiologic relevance of Escherichia coli strains has been suggested. Reports on the rate of breast carcinoma in women with PBC differ and reported either an increase in risk [74, 75] or equal risk [18, 76] compared to a healthy population. Hepatocellular carcinoma in late-stage PBC has been reported to occur at rates similar to other kinds of cirrhosis [77] [78] [79] , but seems to be more frequent in men with PBC: One study reported an average HCC prevalence of 5.9% in advanced PBC (4.1% in women but 20% in men) [77] . A number of mostly immune-mediated diseases are commonly observed in patients with PBC. Thyroid dysfunction is frequently associated with PBC, often predating its diagnosis [80] . Sicca syndrome is seen in up to 70% of patients [81] . Incomplete or complete CREST syndrome (calcinosis cutis, Raynaud syndrome, esophageal motility disorder, sclerodactyly, teleangiectasia) is not uncommon [82] . Celiac disease has been reported in up to 6% of patients [72] and is by far more commonly associated with PBC than inflammatory bowel diseases [83] . A florid bile duct lesion with damage to BEC and subsequent destruction of small bile ducts is the histopathologic hallmark of PBC. The exact pathogenetic mechanisms responsible for BEC damage in PBC remain unknown. Increasing experimental evidence, however, suggests that the florid bile duct lesion in PBC is initiated by environmental trigger(s) acting on a genetically susceptible individual [84] . Genetic factors have an impact on PBC pathogenesis that is stronger than that in nearly any other autoimmune disease [85] [86] [87] . Accordingly, a concordance rate of about 60% was seen in monozygotic twins (as opposed to nearly 0% for dizygotic twins) [88, 89] . A significantly increased incidence of PBC is seen in relatives of PBC patients [89, 90] . The relative risk of a first-degree relative of a PBC patient is 50-to 100-fold higher than for the general population [91] , yielding a prevalence rate up to 5-6% [19, 90, 92] . Interestingly, among affected monozygotic twins, though the age of disease onset is similar, progression and disease severity vary, emphasizing the role of epigenetic and probably environmental factors [89] . It has been difficult to identify distinct susceptibility genes in PBC so far. Genetic associations in PBC were shown with major histocompatibility complex encoded genes. PBC is apparently associated with the DRB1*08 family of alleles, although marked variation is observed between different ethnic groups. Association with the DRB1*0801-containing haplotype is seen in populations of European origin, whereas in populations of Asian origin, an association is seen with the DRB1*0803 allele. A protective association has been described with DRB1*11 and DRB1*13, but once again, significant population differences are observed [93] [94] [95] [96] [97] [98] [99] [100] [101] [102] . Both associations of PBC with the DRB1*08 allele as well as the protective association of DRB1*11 and DRB1*13 were recently confirmed in the largest series ever reported including 664 unrelated patients from Italy [103] . The odds ratio for developing PBC was 3.3 for DRB1*08-positive subjects, whereas it was reduced to 0.3 for subjects positive for DRB1*11, to 0.7 for DRB1*13, and to 0.1 for carriers of both DRB1*11 and DRB1*13. This study again highlighted the relevance of geographic variation, with marked differences in allele association between northern and southern Italy. Associations have also been reported with polymorphisms of genes involved in innate or adaptive immunity. Allelic variations of tumor necrosis factor α (TNFα) and of cytotoxic T lymphocyte antigen 4 (CTLA-4), a key regulator of the adaptive immune system, have been repeatedly associated with susceptibility to different autoimmune diseases (such as type I diabetes mellitus and systemic lupus erythematosus) and were also associated with PBC [104] [105] [106] [107] [108] [109] . Although an association of CTLA-4 variants and susceptibility to PBC could not be demonstrated in all studies [110, 111] , Poupon et al. [112] most recently confirmed a potential role of TNFα and CTLA-4 variants in the pathogenesis of PBC. In 258 PBC patients and two independent control groups of 286 and 269 healthy volunteers, the authors investigated distribution of newly identified htSNPs of 15 selected candidate genes: two related to immunity encoding for CTLA-4 and TNFα, ten genes related to bile formation encoding hepatobiliary transporters, and three related to adaptive response to cholestasis encoding nuclear receptors. In a case-control analysis, only haplotype-tagging single nucleotide polymorphisms (htSNPs) in CTLA-4 and TNFα showed differences in distribution between PBC and controls, confirming their potential role in the pathogenesis of PBC. In contrast, htSNPs of the ten transporter genes as well as the three nuclear receptor genes under study were equally distributed, confirming previous studies of htSNPs in key transporters without major impact [1, 113, 114] . A strong association of the allelic variant TNFα rs 1799724 (C/T) with disease progression was shown. Most interestingly, a strong association with disease progression was also shown for AE2 rs 2303932 (T/A), a gene encoding for the apical anion exchanger 2 (AE2) in cholangiocytes and hepatocytes. In both cases, presence of the variant was associated with delayed disease progression. In a multi-variate Cox regression, the AE2 variant rs2303932 (T/A) was an independent prognostic factor for disease progression in PBC under UDCA treatment, in addition to serum bilirubin, alkaline phosphatase, and serum albumin levels which are established surrogate markers of prognosis in PBC [112] . Most recently, a landmark genetic association study was published [332] . Associations with the risk of disease were unravelled for 13 loci across the HLA class II region, two SNPs at the interleukin 12 alpha (IL12A) locus, one SNP at the interleukin 12 receptor beta 2 locus, and one previously described SNP at the CTLA4 locus. Associations with more than 10 further loci were described. Various associations with other loci have been described in individual populations, mostly of limited size. However, the vast majority have not been confirmed in independent cohorts, and to date, none of the genetic associations described in PBC have been proven sufficiently [84, 88, 108] . Future genetic linkage studies in affected families as well as association studies in large cohorts of unrelated patients may disclose genetic variants conferring susceptibility or influencing progression and severity of disease. Such linkage studies are awaited for PBC [115] . It remains speculative whether the female preponderance (gender ratio up to 10:1) reflects an X-chromosomelinked locus of susceptibility. Alternatively, a protective role of Y-linked genes could be assumed, or just a gender-specific exposure to environmental triggers like cosmetics [116] or nail polishers [117] , as discussed below. However, speculation on a pathomechanistic role for X chromosomal genes was supported by the observation of an increased frequency of X chromosome monosomy in PBC as well as in other autoimmune diseases. This increase in X chromosome monosomy might lead to haploinsufficiency for specific X-linked genes and thereby increase disease predisposition [118] . Case reports of PBC in patients with Turner syndrome (45, X0) also supported this hypothesis [119] . Estrogen signaling has also been proposed to play a role in the homeostatic proliferative response of cholangiocytes in PBC. Accordingly, studies on polymorphisms in estrogen receptor genes revealed associations with the disease, at least in some populations [120] . At the tissue level, cholangiocytes from PBC patients in the earliest disease stages (but not cholangiocytes from normal controls) express estrogen receptors [121] . Agents able to modulate estrogen-receptor-mediated responses (such as tamoxifen) have therefore been proposed as novel, BEC homeostasis targeting therapies, and case reports support this hypothesis [122, 123] , but as yet, this potentially interesting therapeutic approach has not undergone formal assessment in clinical trials. Most recently, altered expression of hepatic microRNA (miRNA) has been described in liver tissue of PBC patients [124] . Certain miRNA negatively regulate protein coding gene expression and may play a critical role in various biological processes. However, a causal link between altered miRNA expression and the development of PBC still remains unproven. Despite strong evidence for a genetic background in PBC, epidemiological studies have early suggested a role for environmental factors in triggering and/or exacerbating PBC [20, 125, 126] . A significant role for environmental factors was supported by the identification of geographic disease "hot spots," as first reported in the northeast of England, using formal cluster analysis. The original UK analysis reported an increased frequency of PBC in former industrial and/or coal mining areas [127] . Another recent study from New York examined the prevalence of PBC and PSC near superfund areas and reported significant clusters of PBC surrounding toxic sites [128] . In synopsis, these observations gave rise to the hypothesis of a chemical environmental factor, potentially associated with contaminated land, which could either trigger disease or cause disease through a direct toxic effect [84] . This hypothesis would also provide one possible explanation for the tissue tropism of PBC if the toxin or toxins are excreted into bile (and thereby concentrated in the biliary tree) [84] . The observation that hormone replacement therapy and frequent use of nail polish are linked to the risk of developing PBC further supports the potential impact of environmental factors in the pathogenesis of PBC [117] . Smoking also seems to be a risk factor for PBC and has been demonstrated to accelerate progression [117, 129] . Associations of exposure to chemical environmental compounds and xenobiotics (including drugs, pesticides, or other organic molecules) with various human autoimmune diseases have been described as summarized in [86] . Xenobiotics may contribute to the pathogenesis of PBC by triggering autoimmune reactions. Different mechanisms for the induction of autoimmunity by xenobiotics have been proposed [86, 130] . A potential direct toxic effect of xenobiotics my cause cell death by apoptosis or oncosis, inducing the generation of immunogenic autoepitopes. In addition, chemical modification of native cellular proteins by removal and/or exchange of a hapten has been shown to change processing in antigen-presenting cells and may lead to the presentation of cryptic, potentially immunogenic peptides. Furthermore, xenobiotics may have the potential to modify host proteins to form neoantigens. Neoantigenspecific T cells and B cells, once primed, may cross-react with the formerly inert native autoantigens. In accordance with this hypothesis, Amano et al. [131] studied a number of xenobiotics with a structure similar to lipoic acid, a residue on the E2 epitope of the pyruvate dehydrogenase complex (PDC-E2), the main autoreactive antigen identified in PBC so far in PBC. Replacement of lipoic acid by certain xenobiotics enhanced the reactivity of PBC sera against the PDC-E2 epitope. Particularly, one of the xenobiotics, 2-nonynoic acid, induced reactivity of PBC sera stronger than that of the native lipoic acid residue. Interestingly, the methyl ester of 2-nonynoic acid has a viol-/peach-like scent and is used as an ingredient in perfumes. It is ranked 2,324th out of 12,945 chemical compounds in terms of occupational exposure with an 80% female preponderance due to its use in cosmetics. Among the environmental factors that have been suggested as potential causative agents in PBC, particularly different bacteria have been discussed. In early histologic lesions in PBC, non-caseating granulomas are observed, as seen in other granulomatous liver diseases including sarcoidosis [1] , drug reactions, and, most interestingly, infections. Furthermore, non-caseating granulomas are unique to PBC when compared to other autoimmune pathologies. This has led to suspicion of a microbial basis for PBC [132] . In support of this hypothesis, certain bacteria were found to contain PDC components fully cross-reactive with the mammalian form. It was proposed that exposure to these homologues could trigger crossreactive immunity. In favor of a bacterial etiology, recent data suggest that Toll-like receptor ligands induce an augmented inflammatory response in PBC. In combination, presence of cross-reactive antigens in a proinflammatory environment would theoretically be able to break tolerance [133, 134] . With this theoretical background in mind, early studies associating various bacteria with PBC re-attract interest: E. coli has been reported to be present in excess in the feces of patients with PBC. In addition, the incidence of urinary tract infections often induced by E. coli is high in PBC patients [73, 135] , and history of urinary tract infections increases the risk of PBC [117] . Another microorganism that has been proposed as a candidate for the induction of PBC is Novosphingobium aromaticivorans [136] . Titers of antibodies against lipoylated bacterial proteins of this ubiquitous organism, which metabolizes organic compounds including estrogens, were 1,000-fold higher compared to those against E. coli in patients with PBC, but no antibodies were observed in a large cohort of healthy subjects. Lactobacilli and Chlamydia, which show some structural homology with the autoantigen (although reactivity against them is considerably less than that against either E. coli or N. aromaticivorans), have also been implicated as putative pathogens, as have Helicobacter pylori and Mycobacterium gordonae [137] [138] [139] [140] . Recently, a case of PBC following lactobacillus vaccination for recurrent vaginitis was reported. The vaccine contained Lactobacillus salivarius, which exerts a high homology to the beta-galactosidase of Lactobacillus delbrueckii (LACDE BGAL [266] [267] [268] [269] [270] [271] [272] [273] [274] [275] [276] [277] [278] [279] [280] , and cross-reactivity of patients' autoantibodies against the human PDC-E2 212-226 epitope and LACDE BGAL [266] [267] [268] [269] [270] [271] [272] [273] [274] [275] [276] [277] [278] [279] [280] was found. Affinity to the Lactobacillus epitope was higher than to the native mammalian, suggesting that antimicrobial reactivity may have preceded that to the self-mimic [141] . However, the AMA status of this patient before repetitive lactobacillus vaccination could not be assessed and causal relation of lactobacillus exposure and development of PBC remains speculative also in this study. Despite these intriguing associations, no compelling data have been provided to show that one individual infectious agent can reproducibly be detected in patients with PBC. Although attractive, the model of bacterial infections as cause of PBC is thus supported by little direct evidence. Further objective data are warranted, obtained either from prospectively followed cohorts or through case-control epidemiological approaches, confirming a role for bacteria in triggering PBC [84] . An alternative infectious agent has recently been proposed as trigger of PBC when a human retrovirus was identified both in liver tissue and hilar lymph nodes from PBC patients. EM analysis of liver tissue obtained from PBC patients revealed retrovirus compatible particles in BECs. In periportal lymph nodes, mouse mammary tumor virus (MMTV) was detected and correlated with aberrant distribution of PDC-E2 in perisinusoidal cells. Homogenates of these periportal lymph nodes also had the capacity to infect BEC cultures inducing marked phenotypic change, and this effect could be abolished by irradiation of the culture media, suggestive of an infectious agent [142, 143] . Retroviral infection hypothetically could cause BEC damage either through a direct viral cytopathic effect, through cross-reactivity between viral protein and self-PDC, a "molecular mimicry" model, or virus-induced apoptosis [84] . Retroviral infection would also provide explanations for some key phenomena in PBC. PBC can recur rapidly after transplantation with all of the clinical manifestations including the detection of AMA in serum [144] , the aberrant expression of the AMA-reactive protein on BEC [145] , and histologic evidence of disease in up to 45% of patients [146] . In this respect, the observation of an association of more potent immunosuppressive therapy following transplant and earlier and more aggressive recurrence of PBC [147] is also of interest. MMTV replication is regulated in part by a progesterone-responsive glucocorticoid regulatory element in the promoter region, offering an alternative explanation for the female preponderance seen with PBC [148] . These findings attracted attention in the field and were acknowledged by other investigators who pointed out the need for clinical trials with antiretroviral therapies [149] . Subsequently, in a small non-randomized pilot study, therapy with Combivir (lamivudine + zidovudine) improved inflammatory scores, normalized AP, and reduced bile duct injury in patients with PBC [150] . These findings await confirmation in a randomized, controlled trial. Unfortunately, major findings of the outlined in vitro studies could not be reproduced by independent groups, and others raised concerns that these findings might mainly reflect contamination or technical artifacts [151] . In an independent study, a large number of sera of PBC patients and healthy controls did not show reactivity against MMTV encoded protein, and no detectable immunohistochemical or molecular evidence for MMTV was found in liver specimens or peripheral blood lymphocytes [152] . It was also speculated that beneficial effects of antiretroviral therapy could be partly explained by anti-apoptotic properties of nucleoside analogs [151, 153] . Furthermore, mechanisms by which human betaretrovirus would enter human cholangiocytes are also not identified. A more recent study, however, strengthened the case for involvement of retroviruses in (immune-mediated) liver disease: Sera of 179 patients with diverse chronic liver diseases and 31 controls were tested for reverse transcriptase activity and presence of human betaretrovirus by polymerase chain reaction. Reverse transcriptase activity was detected in 73% of autoimmune hepatitis patients, 42% of PBC subjects, 35% of patients with viral hepatitis, 22% of liver patients without viral or autoimmune pathogenesis (non-alcoholic fatty liver disease and alcoholic liver disease), and 7% of control subjects. In polymerase chain reactions, 24% of PBC samples were positive for human betaretrovirus compared to 13% in autoimmune hepatitis, 5% in other liver diseases, and 3% in non-liver disease control subjects [154] . If these data can be confirmed, a retroviral compound in the pathogenesis of immunemediated and viral liver disease seems attractive, though not specific for PBC. Thus, despite some intriguing findings, the pathogenetic relevance of retroviruses in the development of PBC remains enigmatic. Appendectomy, other abdominal surgeries, and tonsillectomy were significantly more frequently reported in patients with PBC in an epidemiological study in North America [155] . However, an earlier population-based case control study conducted in England did not show such associations [156] . More recently, another case-control study provided evidence that there was at least no association between PBC and the occurrence of appendectomy and pointed out the selection bias present in the previous study done in North America [157] . The linkage to appendectomy was theoretically attractive, since a PBC-specific immune response to the highly conserved caseinolytic protease P of Yersinia enterocolitica in 40% of patients with PBC was reported [158] . It is noteworthy, that infection with Y. enterocolitica is one of the major causes of acute terminal ileitis mimicking acute appendicitis [159] . Autoimmunity is a phenomenon of dysregulated immune response against self-antigens. If persistent, this can result in inflammatory tissue damage. The immune response to antigens is tightly controlled by various pathways whose deregulation may lead to autoimmune responses. Genetic predisposition and environmental factors affect the susceptibility to such deregulation [86] . Tolerance against self-antigens is achieved in the lymphopoietic differentiation in early life, when highaffinity self-reactive lymphocytes are deleted in the primary lymphoid organs, thymus, and bone marrow. Second, in the periphery, there is activity of a subset of T lymphocytes, T regulatory cells (Tregs), which are dedicated to regulatory function expressing the CD4 and CD25 surface markers and the transcription factor forkhead box P3 (FOXP3). Additional backups along the maturation process of lymphocytes are described, limiting the induction and expression of autoimmunity. These regulatory mechanisms include apoptosis pathways, cytokines and their receptors, chemokine signaling, T cell-T cell interactions, and intracellular signal transduction. Accordingly, loss of self-tolerance could involve multiple faults, most of which are of genetic origin. It is worth mentioning that autoimmunity, defined by the presence of autoantibodies and autoreactive lymphocytes does occur naturally. It appears that such naturally occurring autoantibodies and autoreactive lymphocytes are modulators for the suppression of early infections, clearance of apoptotic bodies, immune surveillance against cancer cells, among others, as reviewed in [160] . In this not yet completely unraveled system regulating immune response, the mechanisms responsible for the development of autoimmunity and autoimmune diseases remain enigmatic. Imbalance of T cell regulation can be sufficient on its own to initiate or propagate autoimmunity in various chronic inflammatory diseases such as inflammatory bowel disease (IBD) or rheumatoid arthritis. In line with this concept, recent transgenic animal models highlight the role of T cell regulation in the development of autoimmune diseases. For instance, IL-2 receptor −/− mice were shown to develop severe anemia and IBD [161] , possibly due to the decreased numbers of Tregs facilitating autoimmune reactivity in the presence of proliferating (and probably, activated) T-cells. Absence of the IL-2 receptor in these animals leads to proliferation of T cells and decreased numbers of Tregs. Another hypothesis is the concept of molecular mimicry based on the similarity of pathogen and host antigenderived epitopes recognized by the immune system [162, 163] , which render bacteria and viruses candidates for the induction of autoimmune disease. This mechanism has first been suggested to be responsible for the development of rheumatic fever, and though this could never be confirmed, there is evidence suggesting associations of infectious triggers for several systemic autoimmune diseases including multiple sclerosis [164, 165] , systemic lupus erythematosus (SLE) [166] , and rheumatoid arthritis [167] . Autoantigens are unable to elicit a primary immune response themselves. However, T cells stimulated by a pathogenic cross-reactive epitope can recognize such targets. As a prerequisite, the pathogen-derived cross-reactive epitope has to be sufficiently different from the host-derived epitope. The role of infectious agents in development of autoimmunity has recently been reviewed elsewhere [168, 169] . Xenobiotics represent another environmental factor foreign to human organisms, and may induce immune reactions or have the potential to modify host proteins and render them more immunogenic. Examples include drugs, pesticides, or other organic molecules. A number of xenobiotics have been associated with several human autoimmune diseases. Chemicals which were linked to autoimmunity include mercury in glomerulonephritis [170] , hydrazines in SLE [171, 172] , iodine in autoimmune thyroditis [173] , and halothane in drug-induced hepatitis [174, 175] . Halothane-induced liver disease occurs when susceptible individuals develop immune response against trifluoroacetylated (TFA) self-proteins upon halothane exposure. Noteworthy is that the lipoylated E2 domain of human PDC is also recognized by anti-TFA [176] . As a further trigger of autoimmunity, it has been speculated that increased cell turnover or, more specifically, increased cell apoptosis may lead to exposure of otherwise rarely exposed antigens and induction of immune response to self. Enhanced apoptosis has been implicated in several autoimmune diseases, including Hashimoto's thyreoiditis [177] . This hypothesis would also provide an appealing explanation for the tissue specificity of most autoimmune reactions despite the often ubiquitous expression of the targeted autoantigen. However, apoptosis is genuinely designed to actually prevent inflammatory reactions to cell death, and ingestion of apoptotic cells by macrophages induces the expression of anti-inflammatory cytokines such as TGF-β and IL-10, both promoting the expression of Tregs, suppressing an autoimmune response during apoptosis [160] . While apoptosis per se is a non-inflammatory process, it can lead to abnormal antigen presentation, especially of previously sequestered antigens. Evidence for a role in the development of autoimmune disease, however, is limited and particularly in organ-specific autoimmune diseases [178] . Whether apoptosis-related mechanisms lead to PBC is unclear [179] , but cholangiocytes in PBC may undergo increased cell turnover, e.g., due to metabolic stress, resulting in an inadequate immune response [179] . Strikingly, BECs in patients with PBC seem to be under significantly increased apoptotic stress compared to healthy controls or patients with other causes of inflammatory reactions in the liver, such as chronic viral hepatitis or PSC [180] [181] [182] . However, it is yet to be elucidated whether this effect is really a cause of autoimmunity in PBC or rather the consequence of increased inflammation. PBC is associated with other autoimmune diseases, both within individuals and among families, reflecting the "clustering" characteristic for autoimmunity [183] . PBC was one of the first conditions in which the presence of autoantibodies in the serum was identified and in which the antigen specificity of this autoreactive response was characterized [5] . It is therefore often referred to as a "model autoimmune disease." The predominant autoreactive antibodies in PBC are AMAs, which, with a high sensitivity and specificity, are virtually diagnostic for PBC when detected in serum. The so far identified targets of AMA are all members of the family of 2-oxo-acid dehydrogenase complexes (2-OADC). This includes the E2 subunits of the pyruvate dehydrogenase complex (PDC-E2), the branched chain 2-oxo-acid dehydrogenase complex (BCOADC-E2), the 2-oxo-glutaric acid dehydrogenase complex (OGDC-E2), and the dihydrolipoamide dehydrogenase binding protein (E3BP) [184] , all localized within the inner mitochondrial matrix, catalyzing oxidative decarboxylation of keto acid substrates. The targeted E2 subunits all have a common N-terminal domain containing single or multiple attachment sites for a lipoic acid cofactor to lysine. Previous studies have demonstrated that the dominant epitopes recognized by AMA are all located within these lipoyl domains of the target antigens [185] [186] [187] [188] . The autoreactive CD4 and CD8 T cells infiltrating the liver in PBC recognize the same PDC-E2 domain (peptides 159-167 and 163-176) [189, 190] , and the same accounts for the dominant autoreactive B cells [191] . CD8 T cells isolated from livers of patients with PBC have been found to exert cytotoxicity against PDC-E2 pulsed autologous cells (peptides 159-167) [192] , supporting the hypothesis of a T cell response contributing to bile duct injury in PBC. It remains a mystery how PDC-E2 and other epitopes localized to the inner membrane of mitochondria become targets of autoimmune injury in PBC. One working hypothesis, largely enforced by the Gershwin group, is that modifications of 2-OADC by xenobiotics may alter these self-proteins to cause a breakdown of tolerance facilitating an autoimmune response. This group identified a 12-amino acid residue peptide of the inner domain of PDC-E2 containing the lipoic acid cofactor carrying 173 lysine to elicit the strongest reactivity of purified sera from PBC patients when compared to a number of other potential epitopes. They subsequently could show that reaction of the same PBC sera was significantly increased by lipoylation of this epitope [193] . Subsequently, the identified PDC-E2 residue was modified by replacing lipoic acid with a series of similar but distinct synthetic structures. Following this modification, PDC-E2-specific autoantibodies from patients with PBC reacted with higher affinity to the modified epitopes than to the native PDC-E2 peptide. The structure inducing maximal affinity was demonstrated to be derived from 2-nonynoic acid, a compound widely used in cosmetics [116] . Xenobiotics, in studies of the same group, were also shown to induce PBC-like features in different animal models. Immunization of rabbits with 6-bromohexanoate (6BH), coupled to bovine serum albumin (BSA), led to break of tolerance to PDC-E2 as judged by the detection of AMA [194] . In guinea pigs, immunization with the compound 6BH-BSA led to the development of histological lesions typical of autoimmune cholangitis with the concurrent appearance of AMA, albeit with a long latency of 18 months [195] . Most recently, these efforts resulted in the development of an inducible animal model of PBC in C57BL/6 mice. In these animals, 2-octynoic acid coupled to BSA after a short follow-up of 8-12 weeks induced manifest autoimmune cholangitis, typical AMA, increased liver lymphoid cell numbers, an increase in CD8 liverinfiltrating cells, and elevated levels of serum tumor necrosis factor α and interferon γ. Remarkably, unlike many other animal models of PBC, the reported immunogenic response was liver-specific, and no inflammation was found in organs other than the liver [196] . However, the model still has disadvantages, e.g., lacking the development of fibrosis. An alternative but complementary concept for loss of self-tolerance in PBC is the idea of underlying immune deficits. This concept is based on clinical and experimental evidence. PBC exhibits clustering with various autoimmune disorders, both within individuals and family. Moreover, in PBC there are reduced levels of Tregs suppressing immune reactions against self [197] . In two genetically manipulated mouse strains, spontaneous occurrence of a PBC-like lymphoid cholangitis together with positivity for anti-PDC-E2 within several weeks of life was reported when either transgenic expression of a dominant negative TGF-β II receptor (dnTGF-β RII) or transgenic disruption of the IL-2 receptor alpha that is highly expressed on Tregs was performed [198, 199] . Interestingly, B cells had a suppressive effect on the inflammatory response in the dnTGF-β RII model of PBC [200] . A role for IL-2 signaling defects in development of PBC was supported by a report of a PBC-like liver disease in a child with inborn deficiency of IL-2 receptor alpha [201] . An additional mouse model of NOD.c3c4 was described to develop autoimmune biliary disease and was found to test AMA-positive [202] . In all models outlined here, the biliary epithelium is infiltrated with CD4 and CD8 T cells, whereas granulomas and eosinophilic infiltration are seen only in NOD.c3c4 mice. Conversely, IL-17 has recently been shown to be involved in various autoimmune disorders. In PBC liver tissues, density of IL-17(+) CD4 lymphocytic infiltration (Th17) was higher than in healthy liver. Although enhanced density of Th17 cells is not specific to PBC, it is in line with the observation of the decreased number of Tregs and may illustrate the counterplay of Tregs and Th17 cells in PBC [203] . There has been impressive progress in unraveling pathogenetic factors of PBC over the past decade. In favor of a genetic background, an outstanding concordance rate among monozygotic twins has been identified, as well as various genetic associations. Clustering of distinct HLA alleles and, among others, allelic variations of TNFα, CTLA-4, and anion exchanger 2 (AE2), respectively, have been described. In addition, there is compelling evidence for an environmental factor required for the development of PBC. Xenobiotics capable of modulating mammalian proteins to form neoantigens have been shown to induce pathologies resembling PBC in animal models. Furthermore, various infectious agents have been associated with PBC in humans, although direct experimental evidence for their role in the pathogenesis is limited. Lastly, disruption of AE2 [204] and genes involved in the regulation of immune response, such as the TGF-β II or IL-2 receptor or the AE2, gives rise to histologic and serologic changes mimicking PBC in different animal models and may be contributing to susceptibility for PBC in human. However, a common causative pathway to PBC, if it exists, could not yet be identified, and it is still controversial which of the outlined factors predispose most to PBC. It may well be speculated that the search for "the cause" of PBC might lead into the dark, and the identified pathogenetic factors may contribute to a different extent in each patient. Why should the cause of PBC not be variable as is the clinical picture? The latter is highly variable: We define PBC today as the presence of cholestasis and AMAs, but only 90-95% of patients are AMA-positive and some AMA-positives hardly develop PBC. Most patients respond to UDCA treatment, but one third does not and disease progression is highly variable, to mention only the most obvious variabilities. The clinical picture that, by convention, we call PBC may well emerge from an individual composition of the described and other pathological factors. Jones [84] suggested a model of PBC development distinguishing "upstream" from "downstream" events. "Upstream" in this model refers to the causes of BEC loss, ductopenia and cholestasis, which are unique to PBC (and probably unique to each individual patient) and include genetic and toxic factors, infectious agents, and immune-mediated events. "Downstream" of these initiating mechanisms, nonspecific pathologic events occur resulting in bile duct damage, hepatocyte injury, inflammation, and fibrosis independently of the primary -individual variablecause. "Downstream" events such as hydrophobic bile salt retention aggravate the underlying injury, further promoting hepatic and cholangiocellular damage [205] . This pathogenetic model provides an explanation for the limited efficacy of immunosuppressive drugs in PBC. Agents such as prednisolone, in the treatment of PBC of limited efficacy, may mainly affect upstream mechanisms. In clinical trials, these agents have, however, largely been evaluated in relatively advanced, symptomatic patients. In these individuals, downstream processes may have become predominant. The balance of evidence in PBC remains strongly in favor of an autoimmune process in which the autoreactive attack is directed at epitopes within self-PDC-E2. The following factors can, in principle, contribute to this breakdown of self-tolerance and have been suggested for PBC: 1. There is strong and compelling evidence to support molecular mechanisms of cross-reactivity between the PDC-E2 lipoic acid co-factor and environmental xenobiotics inducing auto-immunogenity of modified self-PDC-E2. Animal modeling data would argue that a cross-reactive B cell response induced to xenobioticmodified self-PDC can, through a process of epitope spreading driven by antigen-specific cross-reactive B cells, translate into the breakdown of T cell tolerance responsible for the effector T cell mechanisms thought to be directly responsible for BEC loss. The clinical implication of this observation again is that immunomodulatory approaches to therapy should play a role particularly in the earliest stages of PBC. 2. Alternatively or in addition, molecular mimicry mechanisms with cross-reactivity between self-PDC and bacterial or potentially viral structures may support or induce breakdown of tolerance. 3. Most data on genetic associations with disease point to loci or genes involved in immune function. Altered regulation of self-tolerance can possibly support or induce immune reaction against self-PDC-E2, expressed normally or aberrantly by BECs. 4. Primary events of cholangiocellular apoptosis or cell damage have been suggested that could lead to aberrant presentation of self-antigens or create an inflammatory environment potentially triggering immune dysregulation. Suggested mechanisms include metabolic stress, possibly triggered by dysfunction of transporters involved in cell maintenance like AE2. This mechanism was suggested [179] based on the finding that an AE2 variant is a strong and independent prognostic factor for disease progression in PBC under UDCA therapy [112] . Furthermore, secretion of directly toxic environmental compounds into the bile or viral infection would trigger cholangiocellular damage. BECs, either as an associated process to outlined factors or as part of the homeostatic mechanism designed to retain BEC function, may lead to altered self-recognition. Each of these mechanisms may occur simultaneously or sequentially and, to a variable extent, result in the breakdown of tolerance and immune-mediated liver pathology. Once an autoimmune reaction is initiated, various vicious cycles are conceivable. Inflammatory reaction secondary to loss of tolerance will lead to further cholangiocellular damage and apoptosis, increasing presentation of (altered) self-PDC and increase autoreactivity. Once bile duct loss and cholestasis are established, retention of bile salts and other toxic compounds perpetuates damage to BECs and subsequently to the entire liver. Maintenance of these vicious cycles may be supported by an immune system that is genetically dysregulated and insufficiently capable of suppressing autoimmune reactions. The outstanding paradox in PBC pathogenesis remains the tissue tropism of the immune attack on the small intrahepatic bile ducts, although the mitochondrial targets are ubiquitously expressed proteins. An increased vulnerability of the primarily affected BECs therefore is a prerequisite in the pathogenesis of PBC. Staining of BECs from PBC livers with monoclonal antibodies against the mitochondrial PDC-E2 autoantigen showed a specific reaction at the apical surface which was not found in controls [206, 207] . Other non-PBC-related mitochondrial proteins show the expected cytoplasmic pattern [208] . It was subsequently demonstrated that the apical staining is due to a complex between (auto-)antimitochondrial IgA and PDC-E2, giving rise to speculations that IgA might be a player in the immune-mediated destruction of BECs [209, 210] . Apoptosis of BECs has been proposed as a cause of aberrant neoantigen presentation, responsible for activation or attraction of autoreactive T lymphocytes or antibodies. Unlike other cell types [211] for which autoantibody recognition of PDC-E2 is abrogated during apoptosis, probably by glutathiolation of the lysine-lipoyl moiety of PDC-E2, the antigenicity of PDC-E2 persists in the apoptotic BECs in which glutathiolation does not occur [211, 212] . The course of apoptotic markers in patients with PBC peaks in the middle stages of the disease (stages II-III) rather than in earlier stages. This could solely support a role of apoptosis as a trigger for loss of self-tolerance, with probably only a little number of cells being affected rather than being the main initiating event in PBC. In a putative positive feedback loop, apoptosis, autoreactive T cell response, cholestatic bile salt retention, and failed replicative homeostasis may result in the destruction of BECs and bile duct structures. UDCA has been administered in Chinese traditional medicine as a remedy for liver diseases and other disorders in the form of black bear's bile since the time of the T'ang dynasty (618-907AD) and was reestablished in the late 1950s in Japan as a choleretic agent with gallstonedissolving and anticholestatic properties. In the Western literature, the beneficial effects of UDCA on serum liver tests for patients with hepatic disorders were first reported in the 1980s [213, 214] , and UDCA has since been established for the treatment of PBC. Today, it is the only FDA-approved drug and standard therapy for PBC. UDCA was shown to improve serum biochemical markers such as bilirubin, AP, γGT, cholesterol, and IgM levels [215] [216] [217] [218] [219] [220] . UDCA may slow down histologic progression to liver cirrhosis [219, 221] , improve quality of life, survival free of transplant, and overall survival [11] [12] [13] [14] 222] . It is safe and side effects are few [223] . However, the mechanisms of action of UDCA in chronic cholestasis remain enigmatic [224] . About a third of patients is not sufficiently controlled with UDCA monotherapy [12, 13] , which drives the search for additional therapeutic approaches (Fig. 1) . Cholestasis leads to retention of bile salts and other potentially toxic constituents of bile not only systemically but particularly in hepatocytes. This promotes hepatocellular damage and, in a chronic state, can induce the development of fibrosis and cirrhosis of the liver. As recently summarized by Paumgartner and Pusl [225] , the general biochemical and physiologic principles in treating cholestatic liver disease can be broken down to "(1) reduction of hepatocellular uptake of bile salts and other organic anions; (2) stimulation of the metabolism of hydrophobic bile salts and other toxic compounds to more hydrophilic and less toxic metabolites; (3) stimulation of orthograde secretion into bile and (4) stimulation of retrograde secretion of bile salts and other potentially toxic cholephils into the systemic circulation for excretion by the kidney; (5) protection of injured cholangiocytes against toxic effects of bile salts; (6) inhibition of apoptosis caused by elevated levels of bile salts; and (7) inhibition of fibrosis" [225] . Treatment with UDCA addresses most of these postulated principles as recent clinical and experimental data suggest [224, 226, 227] . Still, no relevant effect of UDCA on (1) bile salt uptake and (2) bile salt metabolism [228, 229] has been demonstrated in man. Hepatic secretion When patients with PBC are treated with UDCA, serum levels of bilirubin [216] and endogenous bile salts [230] decrease. This effect of UDCA seems to be mediated by posttranscriptional rather than transcriptional mechanisms, as mRNA levels of key transporters like the conjugate export pump, ABCC2/MRP2, and the bile salt export pump, ABCB11/BSEP, are not affected by UDCA in man. Enhanced hepatic protein levels of BSEP, but not MRP2, have been observed under UDCA treatment in patients with gallstones [228] and may contribute to improved elimination of bile salts [231] . In an animal model of cholestasis, UDCA increases the density of Bsep and Mrp2 in the canalicular membrane of the rat by stimulating transporter targeting and insertion into the membrane [232] by a cooperative protein kinase C α/protein kinase A-dependent mechanism [224, [233] [234] [235] . An integrin-dependent dual signaling pathway involving mitogen-activated protein kinases (MAPK) Erk 1/2 and p38 MAPK has been shown to mediate UDCA-induced canalicular BSEP insertion in normal [236, 237] , but not cholestatic [238] hepatocytes. Whether UDCA-induced kinase-mediated phosphorylation of carriers contributes to enhanced canalicular transporter density and activity needs to be studied in more detail [224, 235, 239] . Dilution of bile and the flushing of bile ducts were postulated as important functions of cholangiocellular bile formation [225] . In PBC, expression of the hepatic AE2 and biliary bicarbonate secretion by AE2 are impaired [240, 241] . UDCA stimulates both AE2 expression [242] and bicarbonate secretion [241] . Retrograde secretion of bile salts and other potentially toxic cholephiles As an adaptive mechanism during cholestasis, basolateral conjugate transporters such as ABCC3/ MRP3 are upregulated and cholephiles such as bilirubin glucuronides, which are not adequately secreted into bile, can leave the liver cell via the basolateral route. In contrast to the postulated increase in retrograde secretion [228] , UDCA diminishes basolateral efflux due to the effective stimulation of orthograde secretion of potentially toxic compounds into the bile. Protection of cholangiocytes Bile, with its high concentration of hydrophobic bile salts, exhibits extracellular cytotoxicity in vitro, but does not cause cholangiocyte injury under physiological conditions. In PBC, inflammatory bile duct injury may be aggravated by hydrophobic bile salts. Two mechanisms have been discussed by which UDCA protects the biliary epithelium against toxic effects of bile, i.e., relative reduction of hydrophobic bile salts and enrichment of phospholipids in bile. When administered at recommended doses of 13 to 15 mg/kg/day, UDCA content may rise up to 50% of total bile salts [243] and to an even higher percentage when the dose of UDCA is increased [244, 245] . In consequence, bile composition is shifted towards less toxic and less hydrophobic bile salts. UDCA administration in patients with gallstones increases expression of MDR3, a phospholipid flippase, in the liver [228] . This might explain the stimulation of biliary phospholipid secretion by UDCA, as described in patients with PSC [246] . Secreted phospholipids form mixed micelles with bile salts, thereby mitigating their toxic effects on cholangiocytes. Inhibition of apoptosis Hydrophobic bile salts induce hepatocellular apoptosis [247] [248] [249] by death receptordependent [250, 251] and -independent [252, 253] mechanisms, an effect that may become relevant when bile salts accumulate in the liver in cholestatic states. UDCA exerts anti-apoptotic effects in experimental in vivo models [247, 249, 252] as well as in vitro in primary human hepatocytes [254] . This anti-apoptotic effect may contribute to the alleviation of liver injury during UDCA treatment. Inhibition of fibrosis Release of chemokines and cytokines by injured cholangiocytes and infiltrating inflammatory cells [255] as well as hepatic stellate cell proliferation induced by bile salts may play a role in fibrogenesis [256] . UDCA has been reported to delay development of severe fibrosis and cirrhosis in PBC [221] . This may be related to the aforementioned anticholestatic and anti-apoptotic effects rather than direct antifibrotic effects of UDCA. The bile salt derivatives 6-ethyl CDCA (6-ECDCA) and nor-UDCA inhibit fibrosis in the bile-duct-ligated mouse. While 6-ECDCA seems to mediate this antifibrotic effect via farnesoid X receptor (FXR) and small heterodimer partner (SHP) [257] , the mechanism of action of nor-UDCA remains yet unclear [258] . Immunomodulatory properties of UDCA have been controversially discussed in the past [259] [260] [261] [262] . The glucocorticoid receptor in rat hepatocytes is activated by UDCA in a ligand-independent way, whereas suppression of IFN-γinduced MHC class II expression was found to be glucocorticoid-receptor-dependent [263] . It remains to be defined whether glucocorticoid receptor activation is unique to UDCA, and thus therapeutically relevant, or might be shared by endogenous bile salts. Comprehensive reviews on molecular actions of UDCA have been published in the recent past [205, 224, [264] [265] [266] . UDCA is currently considered the mainstay of therapy for PBC [28, 29] . Randomized, double-blinded, placebocontrolled trials have consistently shown that UDCA, administered today in standardized doses of 13 to 15 mg/kg/day, improves serum biochemical markers including bilirubin [215] [216] [217] [218] 220] , an important prognostic marker in PBC [267] . A number of studies demonstrated an improvement of histological features by UDCA. In a combined analysis of four clinical trials including a total of 367 patients [268] and in an independent in 103 patients [221] , UDCA therapy was associated with delayed progression and a marked decrease in progression rate from early-stage disease to late histologic stages. Despite these effects on disease progression and severity, the assessment of the effect of UDCA treatment on long-term survival has been more difficult. A combined analysis of three randomized controlled trials including 548 patients in total treated with UDCA for up to 4 years revealed improved survival free of liver transplantation in patients with moderate or severe disease [222] . Another combined analysis of five studies, all with a follow-up of at least 4 years, quantified the reduction in the risk of death or liver transplantation in patients treated with UDCA to 32% [269] . A long-term follow-up study in a cohort of 225 patients, in part overlapping with the aforementioned analyses, reported 10-year survival without liver transplantation to be significantly higher in UDCA-treated patients compared with survival predicted by the Mayo model [270] . Recent long-term trials from France, Spain, and The Netherlands have shown comparable results [11] [12] [13] 271] . The survival rate of patients in early stages of the disease who biochemically responded to therapy was similar to that in the control populations. "Response" here is defined as a decrease in AP to <40% of pretreatment levels or normalization at 1 year (Barcelona criteria) [12] or serum bilirubin <1 mg/dl, AP≤3 × N and AST≤2 × N after 1 year of UDCA treatment (Paris criteria) [13, 14] . An additional nonrandomized controlled study from Greece confirmed that in most patients with PBC treated with UDCA (particularly those who are in early stage of disease), the 10-year survival is comparable to that in the general population [272] . Despite these intriguing findings, the beneficial effects of UDCA on survival have repeatedly been questioned: A large randomized Swedish trial failed to confirm an effect of UDCA on disease progression and survival at a dose of approximately 8 mg/kg/day [273] . It was concluded from this and other studies that doses of UDCA lower than 10 mg/kg/day are of little benefit. In the follow-up of patients in a US [274] and a Canadian trial [217] , UDCA had no significant influence on incidence of endpoints liver transplantation or death. Also, a meta-analysis of 11 randomized trials could not confirm a significant effect of UDCA on survival and incidence of liver transplantation [275] . In this meta-analysis, however, six studies with only 2 years of follow-up were included, as were two studies administering UDCA at low doses of 10 mg/kg/day or less. Other meta-analyses suffered from similar shortcomings [276, 277] and may therefore have missed a beneficial effect of UDCA. Accordingly, metaanalyses which only included long-term trials with followup of at least 2 years and those using an effective dose of UDCA of more than 10 mg/kg/day verified that treatment with UDCA significantly improves quality of life and transplant-free survival and delays histologic progression in early-stage patients [278, 279] . Current guidelines therefore recommend to treat PBC with UDCA using doses of 13 to 15 mg/kg/day and to start treatment early [28, 29, 280] . About two thirds of patients treated according to these recommendations respond adequately as defined by the Barcelona or the Paris criteria and may have a normal life expectancy. The most recent study from The Netherlands comprising 375 patients with a mean follow-up of 9.7 years again stressed the importance of early treatment showing a clear survival benefit for patients treated in early stages of disease with normal serum bilirubin and albumin levels at the start of therapy [14] . For the remaining one third of patients who fail to achieve biochemical response according to Paris and Barcelona criteria or who are at an advanced histological stage at start of medical treatment, therapeutic options are limited to date and novel approaches are needed. Corticosteroids and other immunosuppressive agents have been evaluated for therapeutic use in PBC. In a 3-year, placebo-controlled trial including 36 patients with PBC, prednisolone significantly improved serum AP levels, IgGs, and AMA and diminished deterioration of liver histology, whereas bone loss was aggravated [281] . Combination of UDCA and prednisolone in comparison to UDCA alone resulted in a significant improvement of histological features [282] . It is yet unclear whether increased expression of AE2 isoforms contributes to the beneficial effect of combined treatment with UDCA and corticosteroids in PBC. Experimentally, enhanced expression of alternative AE2 isoforms and enhanced transport capacity for bicarbonate in human cholangiocytes and in a hepatocyte cell line were demonstrated [283] . AE2 expression and biliary bicarbonate secretion is usually diminished in PBC [241] . Serious side effects of long-term glucocorticoid treatment may outweigh the potential benefit. In this respect, the introduction of budesonide, a nonhalogenated corticosteroid with an extensive first-pass metabolism has been a promising innovation. A 2-year controlled double-blinded trial included 39 patients with early-stage PBC and compared treatment with UDCA plus budesonide against UDCA plus placebo. Combination therapy improved biochemical and histological features and reported only few corticosteroid-related side effects [284] . These promising results, however, could not be confirmed in a subsequent study from the Mayo Clinic. Here, adding budesonide for 1 year in 22 patients with suboptimal response to UDCA alone had only marginal effects on serum bilirubin and AP levels. In contrast, the Mayo risk score increased and there was a significant worsening of osteoporosis [285] . This open trial, however, included latestage patients, which may in part explain the disappointing results. As found in a short-term pharmacokinetic study, administration of budesonide to cirrhotic PBC patients leads to high plasma levels of budesonide associated with serious adverse effects and should therefore be avoided [286] . A Finish study [287] included only patients with stages I to III (n=77) in a 3-year randomized trial and used a lower dose of budesonide. The effects of budesonide plus UDCA were compared to UDCA alone. A significant improvement of histological features was observed in the combination group on top of the beneficial biochemical effects of UDCA alone. Long-term controlled trials are required to define whether combination of UDCA with budesonide provides a significant benefit in patients with early-stage PBC inadequately responding to UDCA monotherapy. Other immunosuppressive agents including azathioprine, cyclosporine, mycophenolate mofetil, or methotrexate and drugs with antifibrotic properties including penicillamine, colchicine, and silymarin have not been shown to markedly improve the natural course of the disease or were associated with significant toxicity during long-term treatment [1, [288] [289] [290] [291] [292] [293] [294] [295] [296] . Novel concepts for medical therapy of PBC, alone or in combination with UDCA, have recently been considered particularly for use in patients with incomplete response to UDCA. Among others, antiretroviral, immunomodulatory, and antioxidant approaches were evaluated. A human betaretrovirus has been controversially debated as a potential pathogen in PBC as outlined above. An antiretroviral strategy has therefore been tested in PBC: Lamivudine in combination with zidovudine (Combivir) normalized AP and reduced bile duct injury in a 1-year pilot trial including 11 patients [150] . This finding still awaits confirmation by a randomized, placebo-controlled study. The peroxisome proliferator-activated receptor α (PPARα) agonist bezafibrate was reported to improve serum liver tests in PBC [297] and should undergo more extensive evaluation in patients with PBC with an incomplete response to UDCA. Future anticholestatic strategies in PBC are targeted towards stimulation of transcription of key hepatocellular and cholangiocellular transporters in order to improve the secretory capacity of the cholestatic liver. First results of pilot studies using the FXR agonist 6-ECDCA are eagerly awaited. Pruritus UDCA is an accepted treatment of cholestatic pruritus in ICP (intrahepatic cholestasis of pregnancy). However, its effect on pruritus is variable in PBC [217] . At present, no convincing data for an antipruritic effect of UDCA in PBC are available [298] . Both peripherally acting pruritogens and central nervous dysfunction have been implicated in the pathogenesis of cholestatic pruritus [9] . According to these two concepts, most therapeutic interventions currently under study for cholestatic pruritus are either directed towards elimination of so far undefined pruritogens or modulation of central neurotransmission. The pharmaceutic options currently recommended include [29] : (1) anion exchange resins, such as cholestyramine and colestipol, which bind anions and amphipathic molecules and reduce their intestinal absorption and systemic accumulation. Despite extensive clinical experience, suggesting a beneficial effect in up to 90% of patients, no large clinical trials have evaluated the efficacy of exchange resins [298] . In patients lacking adequate improvement, treatment with (2) rifampicin may be helpful [298] [299] [300] [301] . Rifampicin is a semisynthetic antibiotic and as a potent pregnane X receptor (PXR) agonist leads to the induction of hepatic microsomal enzymes. Thereby, it may promote metabolism of potential pruritogens. As a thirdline option, (3) opioid antagonists (naloxone, nalmefene, naltrexone) have been found to reduce itch severity in patients with PBC [298, [302] [303] [304] . (4) The selective serotonin reuptake inhibitor sertraline [305, 306] has been reported to improve cholestatic pruritus in small clinical trials and has most recently been considered as a fourth-line treatment option [29, 280] . Other experimental approaches include 5-hydroxytryptamine receptor type 3 antagonists, cannabinoids, subhypnotic doses of propofol, plasmapheresis, albumin dialysis, and nasobiliary drainage in desperate cases, although adequate trials are lacking [29, 63] . Liver transplantation should be considered in serious cases in which all other strategies have failed, even if liver function is still conserved [29] . Specific medical therapies of fatigue associated with chronic cholestasis have not yet been defined. UDCA treatment seems to have no or limited beneficial effects on fatigue in PBC [217, 307] . Oral supplementation with antioxidants showed promising results in a pilot study, but had no beneficial effect in a randomized, placebocontrolled, crossover trial [308] . Independent of the underlying disease, altered serotoninergic neurotransmission has been implicated in the development of fatigue [309] . The 5-HT3 serotonin receptor antagonist ondansetron, however, did not significantly reduce fatigue compared with placebo in a randomized, placebo-controlled trial [310] . The selective serotonin reuptake inhibitors fluvoxamine and fluoxetine also failed to exert beneficial effects on fatigue in this patient group [311, 312] . In a series of 21 PBC patients with excessive daytime sleepiness [313] , the centrally acting agent modafinil was investigated in an open-label study [314] . Significant improvement was seen in Epworth Sleepiness Scale scores and fatigue severity as assessed by PBC-40 fatigue domain score, but the drug had to be stopped due to side effects in a considerable part of the patients before the end of the study. In contrast to pruritus, fatigue may not improve significantly following liver transplantation. Independent of the etiology, bone mineral loss is a recognized complication of cholestatic liver diseases. In PBC, it moderately increases absolute and relative fracture risk compared to the general population [315, 316] . Treatment of the underlying liver disease with UDCA does not prevent bone loss in PBC [317, 318] . Calcium and vitamin D supplementation is usually recommended in osteoporosis, although their role in preventing osteoporosis and fracture in chronic liver disease is unclear [319, 320] . Alternative interventions to prevent and treat osteoporosis in liver disease have only been tested in a few studies. Parenteral calcitonin, tested in one controlled trial for 6 months, was ineffective in halting bone loss in patients with PBC [321] , whereas hormone replacement therapy was demonstrated to improve vertebral bone density in PBC [319, 322, 323] . In large clinical trials, however, hormone replacement therapy was associated with serious side effects, including the risks of breast cancer, coronary artery disease, and thromboembolism, limiting its use for the prevention of fractures in osteoporosis [324] [325] [326] . Use of bisphosphonates is controversial. In a placebocontrolled trial, cyclical etidronate did not significantly improve bone density [327] . Also, a 4-year treatment with clodronate plus calcium/vitamin D did not improve osteopenia in women with PBC in a prospective study [328] . In two controlled trials however, alendronate significantly improved spine and femoral bone mineral density compared with placebo and had greater antiresorptive activity than etidronate [329, 330] . According to current guidelines [29] , supplementation with calcium (1,000-1,200 mg/day) and vitamin D (400-800 IU/day), though not evidence-based, should be considered in all patients with cholestatic liver disease. Alendronate or other bisphosphonates are recommended at a T score < −2.5 (DEXA) or following pathological fracture. Liver transplantation is the treatment of choice in patients with late-stage PBC. Indications are decompensated cirrhosis with treatment-resistant ascites, recurrent spontaneous bacterial peritonitis, encephalopathy, recurrent variceal bleeding, or hepatocellular carcinoma. In highly selected patients (see above), treatment-resistant pruritus in the absence of decompensated cirrhosis may be an indication for transplantation, and so is severe osteoporosis [28, 331] . Survival rates of 80% to 90% at 5 years have been reported. The disease recurs in up to 30% at 10 years after transplantation, but then usually displays a mild course under immunosuppressive treatment [331] . Open Access This article is distributed under the terms of the Creative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.

ORMA: a tool for identification of species-specific variations in 16S rRNA gene and oligonucleotides design

16S rRNA gene is one of the preferred targets for resolving species phylogenesis issues in microbiological-related contexts. However, the identification of single-nucleotide variations capable of distinguishing a sequence among a set of homologous ones can be problematic. Here we present ORMA (Oligonucleotide Retrieving for Molecular Applications), a set of scripts for discriminating positions search and for performing the selection of high-quality oligonucleotide probes to be used in molecular applications. Two assays based on Ligase Detection Reaction (LDR) are presented. First, a new set of probe pairs on cyanobacteria 16S rRNA sequences of 18 different species was compared to that of a previous study. Then, a set of LDR probe pairs for the discrimination of 13 pathogens contaminating bovine milk was evaluated. The software determined more than 100 candidate probe pairs per dataset, from more than 300 16S rRNA sequences, in less than 5 min. Results demonstrated how ORMA improved the performance of the LDR assay on cyanobacteria, correctly identifying 12 out of 14 samples, and allowed the perfect discrimination among the 13 milk pathogenic-related species. ORMA represents a significant improvement from other contexts where enzyme-based techniques have been employed on already known mutations of a single base or on entire subsequences.

During the last decades, different nucleic-acid-based detection techniques have been developed in order to employ identification based on single-nucleotide variations in both genotyping and detection experiments on a multiplicity of targets. These techniques allowed distinguishing alleles and correctly assessing the genotype at the singlebase level. In particular, 16S rRNA gene sequences have been used to resolve bacterial phylogeny and taxonomy issues in different contexts. The DNA sequence coding for the small ribosomal subunit has been by far the most common genetic marker employed by the scientific community, because of: (i) its presence in almost all bacteria, often existing as a multi-gene family, or operons; (ii) the function of the 16S rRNA gene has not changed over time, suggesting that random sequence changes are an accurate measure of time (evolution); and (iii) the 16S rRNA gene (more than 1500 bp) is large enough for informatics purposes (1) with large stretches of conserved regions and few different loci. DNA microarrays represent one of the most popular platforms in molecular technologies, allowing a highthroughput format for the parallel detection of 16S rRNA genes from environmental samples (2) . DNA chips have been developed as a preferred device for the identification of different microorganisms based on 16S gene sequences. The multiplicity of species which can be arrayed on a single-DNA chip allows a high multiplexing capability, with the possibility of identifying many different targets at one time (3) . Single-base variations by microarray analysis can be detected by differential hybridization techniques using allele-specific oligonucleotide probes (4), or by enzyme-mediated detection methods (5) . One of the most critical points of the molecular recognition procedures is the design of the specific probes needed to perform the entire analysis. In genotyping experiments, this is accomplished on the basis of the already-known information about each single-base variation. In detection experiments, on the other hand, in order to explore whether a certain target sequence is present in a DNA sample or not, the main problem is searching for a priori not yet identified specific positions that can discriminate exactly between one target and another. In hybridization-based techniques, mutations are identified on the basis of the higher thermal stability of the perfectly-matched probes as compared to mismatched probes. Although this has been the most frequently applied technique, it is characterized by many hindrances which make hybridization-based strategy function poorly in high-complexity biological samples. Therefore, for analytical and diagnostic purposes, hybridization is generally combined with some other selection or enrichment procedures. Enzyme-mediated ligation methods, on the other hand, rely on interrogation of a mutation by a couple of oligonucleotides annealing immediately adjacent to each other on a target DNA, with one of the probes having its 3 0 -end complementary to the point mutation. In this case, the search is for a single base that characterizes a species against all the others in a group of interest. The presence of a point mutation is assessed by the ligation of the two adjacent oligonucleotides, which occurs only when both are correctly base-paired (6) . The Ligation Detection Reaction (LDR) (7) , for instance, represents a reliable technique for identifying one or more sequences differing by one or more single-base changes, insertions, deletions, or translocations in a plurality of targetnucleotide sequences. This enzymatic in vitro reaction is based on the design of two oligonucleotide probes for each target sequence: a probe specific for the variation (called 'Discriminating Probe', or DS), which is 5 0 -fluorescently labeled, and a 5 0 -phosphorylated 'Common Probe' (or CP), starting one base 3 0 -downstream of the DS. The previously polymerase chain reaction (PCR)-amplified sample, the oligonucleotide probe pairs and a thermostable DNA ligase are blended to form a mixture: the two probes hybridize consecutively along the template and the DNA ligase joins their ends only in the case of a perfect match. This reaction is cycled to increase product yield. The PCR-LDR approach, usually, is associated to the hybridization onto a Universal Array (UA), where a set of artificial sequences, called Zip-codes are arranged (7) . This entire approach was proven to be rapid, flexible and easily adaptable from one target to another, useful, for example, in environmental monitoring (8, 9) , forensics (10) and the food industry (11, 12) . Here we present Oligonucleotide Retrieving for Molecular Applications (ORMA), a series of integrated scripts in Matlab, which performs an accurate search of all the positions able to specifically discriminate one species among homologous ones, based on the 16S rRNA gene sequence. ORMA also performs an accurate selection of high-quality oligonucleotide probes to be used in molecular applications. Automated and computerbased methods can be very useful for performing accurately and rapidly all the requested operations, through the many steps between the original, complete, set of sequences and the final list of application-oriented probes. The problem of designing specific oligonucleotide probes for the identification of target species has already been addressed by a certain number of software (13) (14) (15) (16) . At present, there is no preferential reference strategy for designing microarrays for species identification based on 16S rRNA sequences: many authors rely on academic software (17, 18) , others develop their own scripts (19, 20) . Among the currently available academic software, ARB (21) and PRIMROSE (22) are very diffused, both being tools implemented specifically on 16S rRNA, structured for interacting with and retrieving sequences from specific databases and operating a probe design on the basis of the phylogenesis of the species under analysis. Also, some commercial software, like Oligo 7 (Molecular Biology Insights, Cascade, CO, USA) (23) or AlleleID (Premier Biosoft, Palo Alto, CA, USA) (24) have been applied for probe design in a pathogen characterization experiment (25) . In this article ORMA was used for determining sets of LDR probe pairs in microbiological-related contexts (water safety and food safety applications, respectively). The approach was evaluated and validated using the probe pairs derived from ORMA-determined discriminating positions on a set of cyanobacteria 16S rRNA sequences belonging to 18 different species; the results were compared to those of a previously published study (8) . Secondly, a set of LDR probe pairs for the discrimination of 13 mastitis-or intoxication-related pathogens species in bovine milk was designed and experimentally evaluated. The tool, although here applied on 16S rRNA, can be used on any set of highly correlated sequences. ORMA scripts were developed under Matlab 6.1 (Mathworks, Natick, MA, USA) environment (Release 12.1). No additional toolboxes are required. All statistical analyses and representations were made by the same software. Probe designs and simulations were run onto a hp Workstation xw4100, with a Dual-core 3.2 GHz. Intel Processor and 2.5 GB RAM. ORMA functions and m-code are available for free upon request. Overall structure. ORMA overall structure is tree-like, with a main function that, sequentially, recalls all the side scripts needed to perform each requested operation. The software also comprises a series of scripts for retrieving oligonucleotide sequences, quality-check them and design probes for different applications, such as Ligase Detection Reaction (LDR) or Minisequencing/Primer Extension probes. The overall procedure is accomplished in four main steps (Figure 1, Supplementary Figure 1 ): (i) sequence importing and processing; (ii) discriminating positions finding; (iii) designing of the candidate probes, starting from the positions found and (iv) ranking (i.e. assignment of a quality score to each) and exporting of the candidates (in tabular format). (i) Sequence import and processing. The search for discriminating positions on 16S rRNA starts from the import of a set of already-aligned sequences (which can be optionally used for the creation of consensus sequences, grouping them in homogeneous clusters, before being used for the discriminating position search algorithm). Standard multiple-alignment formats (Clustal-like, Multi Sequence Files, or aligned FASTA format) can be used. A careful check of multiple alignment scores should be made, in order to avoid designs on sequence datasets of distantly related species, which can occur in base misalignments. The scripts also include a procedure for consensus determination from a set of user-defined sequences, according to four different rules: (a) majority rule, in which the consensus base is the most frequently present in the aligned sequences and no degenerated bases are used. In case of equal occurrences, 'N's are used in the consensus; (b) threshold rule 'simple', which assigns a specific base to the corresponding position in the consensus only if its frequency is above a given threshold. Different thresholds can be set for gaps and bases. Degenerated bases are not used and are substituted by 'N's in the consensus; (c) threshold rule 'complex', which comprises also degenerated bases. The algorithm is the same as point (b) option, but requires a threshold for substituting positions with multiple bases above the threshold with the corresponding IUPAC code degenerated base and (d) ARB-like algorithm, with separate thresholds for gaps and bases. All the bases above the given threshold are used to compute eventual degenerated bases. For each of these four options, consensus score accuracy is calculated, as the percentage of original sequences that carried the same base as the consensus in each position. (ii-iii) Design of candidate probes. We have implemented a Single Base Seeker (SBS) algorithm, for the determination of positions able to discriminate one sequence among a set of homologous ones. The discriminating position finding procedure can be summarized as follows in four basic steps: (a) Choice of a user-defined subset of sequences of the dataset (indicated as the 'positive set'). The remaining sequences are used as a group of the discriminating positions must be different from; these are addressed, in the present article, as the 'negative set'. 'Positive' and 'Negative' sets differ for the fact that every consensus in the 'positive set' group will be subjected to probe design, whereas those of the 'negative set' will not; (b) For each sequence, determination of a list of the positions of nondegenerated bases; (c) For each position on point b, calculation of a score as the sum of all the sequences carrying the same base as the considered sequence, in the same position. If the only sequence carrying the base is the tested one, the position is set as discriminating and (d) Re-calculation of the score on point c, substituting to each (eventual) degenerated base its two or three alternatives (an 'N' automatically flags the position as nondiscriminating). ORMA, then, retrieves the sequences flanking each of the putative discriminating positions. Actual oligonucleotide design is dependent on the molecular application chosen. The maximum length and the thermodynamic model for calculation of the parameters of the probes can be specified by the user. For the LDR experiments here described, two oligonucleotide probes are designed, one upstream (Discriminating Probe, DS, comprising the discriminating position) and one downstream (Common Probe, CP) of each position. (iv) Discriminating position related filters and scores. The putative discriminating positions and related candidate probes are subjected to a series of constraints and quality filters. The software keeps track of all the designed candidates, assigning a quality score, depending on how many filters they pass. The current options of the script on the discriminant base are: (a) limiting the range of positions, in order to exclude candidates insisting on positions too close to the 5 0 -or 3 0 -end of the sequences, where, usually, the majority of errors in the alignment or characterization of the sequences occur and (b) testing the presence of other species with probes insisting on the same position, thus excluding eventual interactions between a single CP and multiple DS, with subsequent non-specificity. The candidate probes can also be filtered and ranked according to their thermodynamic properties (length, melting temperature, number of degenerated bases, low complexity regions), evidencing the candidates having a certain length, a melting temperature comprised in a user-specified range, having no more than the inputted number of degenerated bases (which can be a real issue for the oligonucleotide specificity), having short homopolymeric regions and not comprising short tandem repeats. Then, ORMA calculates some specific statistics for the qualitative evaluation of the candidates designed on consensus sequences, compared to the original dataset (i.e. the subset of sequences from which every consensus is built): (a) the intra-group score, as the number of initial sequences having the same discriminating base as the consensus and (b) the inter-group score, as the number of sequences other than those used for that consensus having the same discriminating base as the candidate one. This latter score is calculated only when the consensus were created inside ORMA, starting from a single-global alignment. These scores allow the choice of probes that best discriminate between the target and the non-target sequences (i.e. having the highest intra-group and the lowest inter-group score). The software output can be exported as a comma-separated spreadsheet reporting: (a) the list of all the discriminating bases, grouped per species, with absolute (referring to the global alignment) and relative (referring to the specific consensus) positions of the discriminating base, and the base distributions of all the other consensus sequences in the same position; (b) the thermodynamic parameters of the candidate probe pairs, including the T m , the length of DS and CP probes and the number of degenerated bases in each and (c) the qualitative filtering and the specificity-related scores, including the sequence score, as the average of the consensus scores along all the bases constituting the DS and CP, with penalties for degenerated bases. Cyanobacteria dataset experiment. The complete cyanobacteria 16S rRNA data set comprised a total of 352 sequences, which were organized by phylogenetical similarity and grouped in a total of 18 clusters, as described in (8) . Multiple alignments of all the sequences was performed by ClustalW (26) and the resulting file was imported into ORMA, where 18 consensus, one per cluster, were built. Consensus sequences were determined following the 'ARB-like' algorithm (as described in 'Materials and Methods' section and in Supplementary Methods), setting 50% as the threshold for gaps and 40% as the threshold for other bases. Melting temperature calculations followed the 'salt-adjusted' method, with 50 mM Na + and 0% formamide. Candidate probe pairs were filtered on the basis of their length (minimum 25 nt, maximum 60 nt per probe), melting temperature (63-688C) and number of degenerated bases (maximum 4), on both DS and CP. The best probe pairs for all the species were selected, according to their best intra-and inter-group scores. We required that no less than 80% of the sequences constituting each of the 18 clusters carried the same base as the consensus in the candidate discriminating position (intra-group score). When only one candidate was designed or the intra-group score of the best candidate was below 80%, we still picked that candidate for further evaluations. On the other hand, the inter-group score was set to be below a 2% threshold, with the same exceptions as above. The 'Unicyano' probe, which allowed the identification of any of the species in the study, was the one proposed by Castiglioni et al., with minor refinements for adjusting its melting temperature. At first, the LDR mix made by all probe pairs (250 fmol/ml each probe) was tested on specific synthetic templates (perfectly complementary to each probe pair) to assess the feasibility of the LDR procedure with the ORMAdesigned probe pairs. Then, a total of 14 DNA samples, corresponding to 13 cyanobacteria species (kindly provided by MIDI_CHIP project partners, http://www.cip .ulg.ac.be/midichip/) (Table 1) , were tested in duplicate, independent, LDR experiments, with both ORMA and Castiglioni et al. probe pairs. Milk-pathogen dataset experiment. Milk pathogensrelated 16S sequences were retrieved from RDP-Ribosomal Database Project II (release 9.51, http:// rdp.cme.msu.edu/) (27) for a total of 738 sequences and divided into 13 subgroups, according to their phylogenetic classification. Only sequences of length >1200 bp and flagged as of 'good' quality were retrieved. Each subgroup was aligned independently in ClustalW, since the overall number of 16S sequences was >500 (above the maximum limit of the alignment tool) and imported into ORMA. The consensus sequence for each group was calculated with the same parameters specified for the cyanobacteria data set. Then, a new multiple-alignment step was performed before proceeding to actual probe design. One probe pair for each of the main six subspecies of the Streptococcus group (Streptococcus agalactiae, S. bovis, S. equi, S. canis, S. dysgalactiae S. uberis) was designed; moreover, the Staphylococcus aureus probe pair was designed independently from all the remaining coagulase negative Staphylococci (grouped in the 'Staphylococcus, no aureus' probe), because of its relationship with outbreaks of mastitis in dairy ruminants (28) and with major health issues, like food-related intoxications (29) . In order to have the best homogeneity among the species within each group, the design was actually performed in three rounds: (a) Salmonella spp. was aligned against Escherichia coli and related spp. consensus sequence only; (b) S. canis was aligned against Streptococcus group sequences only; (c) All the remaining positions were selected considering the alignment of all other subspecies. One probe pair per species was designed, except for Campylobacter spp. for which two probe pairs were evaluated in terms of reproducibility and specificity. The thermodynamic parameters were the same described for the cyanobacteria data set, except for the melting temperature, which was required to be in the range 67-698C. The inter-group score of the candidates was required to be above a threshold of 80%, as in the cyanobacteria dataset. Probe pair specificity was checked by both RDP II database and BLAST (Basic Local Alignment Search Tool, http://www.ncbi.nlm.nih.gov/ blast/Blast.cgi) (30) analysis, carefully examining the 3 0 -region of the discriminating probe, in order to exclude any interaction between probe pairs targeting different species. LDR probe pairs were mixed at a final concentration of 1 pmol/ml and tested on 13 DNAs from ATCC reference strains (LGC Promochem, Middlesex, UK) and bacterial collections (Supplementary Table 1 ). Genomic DNA was extracted following the protocol described by (31) , PCR amplified and analyzed in duplicate, by separated LDR reactions. PCR and LDR/Universal Array approach. Complete experimental procedures concerning the amplification of 16S rRNA sequences (including primers and thermal cycling), LDR mixes, Universal Arrays preparation and hybridization are reported in Supplementary Data. Data analysis. All arrays were scanned with ScanArray 5000 scanner (Perkin Elmer Life Sciences, Boston, MA, USA), at 10 mm resolution, with different acquisition parameters on both laser power and photo-multiplier gain, in order to avoid saturation. Intensities of fluorescence (IF) were quantitated by ScanArray Express 3.0 software, using the 'Adaptive circle' option, letting diameters vary from 60 to 300 mm. No normalization procedures on the IFs were performed. To assess whether a probe pair was significantly above the background (i.e. was 'present' or not), we performed a one-sided t-test (a = 0.01). At the same time, also the type II error was calculated and 1-b used as the estimate of the power of the statistical test. The null distribution was set as the population of 'Blank' spots (e.g. with no oligonucleotide spotted, n = 6) IFs. Two times the standard deviation of pixel intensities of the same spots was added to obtain a conservative estimate. For each Zip-code, we considered the population of the IFs of all the replicates (n = 4) and tested it for being significantly above the nulldistribution (H 0 : m test = m null ; H 1 : m test > m null ). Signal-to-noise ratios, SNR p and SNR np were calculated, for each 'present' and 'non-present' probe pairs, respectively, indicating the ratio between the mean IF of each probe pair and the mean 'Blank' IF, divided on the probe-type. Searching, designing and selecting oligonucleotide probes for molecular applications experiments on sets of highly similar sequences, such as the 16S rRNA, is a non-trivial procedure, which involves many complex and timeconsuming steps. In this article, this procedure was accomplished by the use of ORMA, an integrated architecture of Matlab scripts. The 16S rRNA, a gene sequence of more than 1500 bp, is the preferred genomic target for analyses in the microbiological field (17) (18) (19) (20) . It should be noted that 16S region is commonly used in taxonomical classifications involving in silico alignment and procedures for its two basic properties: (i) 16S presents highly conserved regions which can be used to correctly align all the sequences in the database; (ii) on the other side, 16S presents highly polymorphic regions that can be used in clusterization, phylogenetic tree construction and molecular discrimination of microbiological families even very close one to each other (32) . Use of an automated method for discriminating positions determination, probe retrieval and filtering has obvious and evident advantages over the manual design, often used in previously published papers (8, (33) (34) (35) . These advantages become more significant with increasing dimension of the databases and of the sequences length. ORMA can perform all these operations with user-specified parameters in an automated way and calculates a series of Where sequencing has been performed, the result of the classification is also reported. Sample ID refers to the numbers used in Figure 2 . a Clonal DNA from environmental sample. b Specific indicates that only the probe corresponding to the species was present; non-specific means that no probe was present (except for the universal cyanobacteria probe); aspecific means that the species-specific probe was present, but also other probes showed an IF significantly above background signal. The number of replicates is reported within brackets. c According to RDP II database, release 9.60. qualitative parameters which help in the choice of candidate probes that best discriminate between the sequences of the positive and those of the negative set. The general idea of these scores is to distinguish the sequences/groups which are of interest in a given experiment from those who aren't and that can potentially have a cross-contamination with the positive set, because they could be amplified by PCR, contributing to the molecular complexity of the sample. In this article, performances of ORMA were evaluated by considering the experimental evidences coming from the design of LDR probe pairs on two different 16S rRNA datasets. First, a new set of cyano-specific probe pairs was designed and compared to the original one (8), generated on the same database of sequences. Then, the tool was used to setup LDR probe pairs for the identification of pathogenic species present in bovine milk. Species-specific probe pairs were designed in a single round, starting from the whole dataset of 352 ClustalWaligned cyanobacteria 16S rRNA sequences, imported, converted and grouped into 18 group-specific consensus sequences by ORMA. Calculated consensus sequences were highly similar, (ClustalW score = 87.31 AE 2.13, n = 18), had a high consensus score (average score 89.20 AE 4.16, n = 352) and a very low content of degenerated bases (average < 2%, max = 6%). ORMA identified a total of 192 candidate probe pairs for the 18 species, with an overall duration of the whole procedure of less than 5 min (Table 2 ). More tests on speed performances of the SBS algorithm on simulated data available as Supplementary Data and Supplementary Figure 2 . One probe pair per species was chosen, according to its ranking after ORMA filtering steps. The probe pair for Anabaena + Aphanizomenon group was flagged as inadequate by ORMA, having six degenerated bases in the CP, which could negatively influence its thermodynamical properties. However, this probe pair insisted on the only discriminating position found for that cluster. The mix containing all probe pairs was tested on the corresponding synthetic templates and, as expected, all except Anabaena+Aphanizomenon gave positive results. Duplicate LDR experiments on 18 probe pairs (17 species-specific + 1 universal) were carried out on 14 16S rRNA PCR products. We performed side-by-side tests of the same DNA samples by the two probe pairs datasets, ORMA and the one described in Castiglioni et al., comparing their performances and specificity. Probe pairs used in Castiglioni et al. identified correctly (P < 0.005, average beta power of the test: 0.85) 6 out of 14 analyzed DNAs (in both duplicate LDR), whereas other two completely failed. Six other DNAs somehow showed a degree of aspecificity (i.e. the correct probe pair was present, but non-specific probe pairs were also called present) (Table 1, Figure 2 ). Cyanobacteria universal probe pair was called as statistically over the background in all the experiments. Evaluations on ratio of signal intensities suggested that hybridizations went well and were not responsible for the aspecificity. In fact, excluding non-specific signals, SNR np had an average value of 1.18 AE 0.61 and SNR p varied between 10 and 680, with an average of about 149 (data not shown). The Anabaena + Aphanizomenon probe pair of Castiglioni et al. study resulted specific on both synthetic and environmental samples (data not shown). This probe pair, however, was designed with its DS insisting on a position which did not discriminate univocally the Anabaena + Aphanizomenon consensus from the consensuses of the other species. Thus, it would never be identified by ORMA as discriminating (because of the way the algorithm is built). Instead, the presence of some internal mismatches (especially the one on the second base before the 3 0 -end of the DS) is probably the reason for this finding. In fact, the mismatch gives instability to the 3 0 -end of the DS when annealing on the 16S rRNA sequences of species other than those of Anabaena + Aphanizomenon cluster, impeding the ligase to join the two adjacent end of the DS and CP oligonucleotides. ORMA designed probe pairs have been capable of correctly identifying (P < 0.005, average beta power of the test: 0.85) 12 out of 14 analyzed cyanobacteria samples, on both replicates. Also in this experimental set, the cyanobacteria universal probe pair was called as statistically over the background in all the experiments (as expected, since this probe pair and the ones used in Castiglioni et al. coincided). Performances of the LDR procedure, in terms of signal-to-noise ratios were comparable to those obtained with the Castiglioni et al. probe set, having a SNR np of 1.1 AE 0.26 and a SNR p ranging from 7 to 387 (average $131) (data not shown), indicating a certain variability. In this case, we had no signs of aspecificity in the experiments (Figure 2 ), even in those cases which were critical with Castiglioni et al. probe pairs. In fact, probes were chosen in order to maximize the intra-group similarity (i.e. having the maximum number of sequences in the positive set carrying the discriminating base) and minimize the possibility of an inter-group cross-talk (i.e. having a minimum number of sequences in the negative set carrying the discriminating base) (Figure 3 ). The average of intra-group scores of the candidates was 95.1% AE 10.1% (n = 17), varying in the range 60-100%. The minimum value was that of the cluster of Gleotheceae, in which we had only five sequences, whereas 13 out of the 17 clusters were characterized by a score of 100%. Intergroup scores, on the other hand, were always very low, with an average of 0.4% AE 1% (n = 17). Thus, where ORMA probe pairs failed, we had a false negative call (with the cyanobacteria universal probe pair called as 'present'), but not a false positive. Experiments on the two Nostoc DNAs gave no results on the species-specific probe pair; anyway the presence of a cyanobacterial DNA was correctly assessed by the Universal probe pair. Sequencing of the two products revealed that one of them has been correctly classified by microbiological methods, whereas the other DNA was very uncertain and classified as 'cylindrospermum' (58% confidence) by RDP 'Classifier' tool, release 9.60. [On 22 May 2008, RDP II database for cyanobacteria (release 9.61) underwent a major change in hierarchical classification of the species. The taxonomies here presented refer to older versions, which at present, can be found within genus GpI of On the x-axis, the IDs of the tested samples (see Table 1 for full description) are reported. On the y-axis, the probe pair name is reported. The line 'Other' represents the mean of all the remaining Zip-codes in the universal arrays that were not associated to any actual probe. Experiments on Nostoc samples were repeated twice on different DNAs because of the failure of the first test. Halotolerans probe pair in one replicate of sample 8 (classified as Nostoc) has a P-value of 0.02, above the threshold of 0.01 chosen for significance. family Family I, philum cyanobacteria.] Both sequences found very little similarity with our probe pairs, with internal mismatches and a different base in the discriminating position. Anyway, the probe pair itself was successfully tested on the synthetic template. The failure of ORMAdesigned Anabaena + Aphanizomenon probe pair suggests the possibility of making a re-design in the near future, increasing the number of sequences of the database and improving the information content of the dataset. Another strategy could be designing probe pairs on subclusters of Anabaena + Aphanizomenon, building new consensuses from more homogeneous groups; in this way, the presence of such two species would be assessed by multiple probe pairs and not only by one. 16S rRNA sequences of pathogens contaminating bovine milk or related to bovine mastitis were used to design LDR oligonucleotide probes by ORMA, providing a further confirmation of its reliability and specificity. In this study, three round of design were actually performed, in order to have the best homogeneity between the species used in each round. A single round would have caused the loss of discriminating positions due to misalignment of some species (e.g. Salmonella) which are somehow different from all others. ORMA found a total of 392 candidate positions (34, 4 and 354 in the design for Salmonella, S. canis and all remaining species, respectively), which were selected according to the quality ranking scores assigned by ORMA. In this experiment, ORMA calculated only the intra-group score, but not the inter-group score, because of the fact that the sequences for each group were imported separately and the software was unable to recall the position corresponding to discriminating ones in all the sequences constituting each of the consensuses. The candidate probes were all characterized by an optimal specificity of the discriminating base, as suggested by the intra-group scores which were above 90% in 11 out of 14 cases. The scores were, in any case, above the fixed threshold of 80%, having an average of 94.0% AE 6.9% (n = 14). Also in this case, the lowest score (i.e. 80%) was that of the cluster (i.e. S. equi) constituted by the lowest number of sequences (n = 5). The final evaluation on the candidate probe pairs was made by RDP and BLAST checks, because of the multiplicity of species, whose 16S rRNA gene was amplified by means of universal primers, potentially present in milk-derived matrixes and the lack of a complete internal negative set in ORMA. The probes were slightly longer than the ones on cyanobacteria dataset, with an average length of about 40 nt, with very homogeneous melting temperatures (mean T m = 67.6 AE 0.4, n = 28) and a very low number of degenerated bases (only the DS probe for S. equi had 1 degenerate base) ( Table 3 ). The consensus scores for both the DS and CP confirmed the overall quality of the probes (average score of 96.5 AE 4.2, n = 28, with 60% of the probes having a score >99%). The procedure showed optimal specificity, with excellent signal-to-noise ratios, as shown in detail in the article of Cremonesi and co-workers (36) . Results were in complete concordance with sample identification made by ATCC; only probes associated to the supposed species were present (P-values always <0.005), whereas all remaining probes were well below any acceptable P-value for the t-test (Figure 4) . In this dataset, SNR p varied from 4.31 to 238.3, with an average of 34.28; at the same time, SNR np varied between 0.12 and 0.83, with an average of 0.48 AE 0.18. The two probes on Campylobacter species (insisting on two different positions) performed nearly the same in terms of specificity, both giving P-values far below the acceptance threshold of 0.01, whereas performances in terms of signal intensity varied, with one probe having average IFs about 2-fold higher than the other, in both replicates, suggesting a somehow different sensitivity in the two competing probes. Thus, ORMA helped in developing a reliable PCR-LDR-UA assay, which allowed the identification of pathogenic species in milk, based only on 16S rRNA gene, whereas other assays (37) needed multiple genes. The molecular procedure permitted the discrimination between the most frequently isolated or emerging The exact probes sequence from ORMA is reported. For synthesis purpose, any degenerated base was substituted with inosine (I). The description of the reported columns is the same as those in Table 2 . pathogens in mastitis (e.g. S. aureus, S. agalactiae, S. uberis), or potentially dangerous for human health (e.g. E. coli and related species, Salmonella, S. aureus and Bacillus spp). Streptococcus spp. was identified at the species level, even in the cases, like the one of S. uberis and S. parauberis, where the molecular identification on the basis of the 16S rRNA gene required PCRs with species-specific primers. Moreover, the ORMA-based LDR technique represents a significant improvement of the existing detection methods for Mycoplasma spp. strains (38) , known to be contagious causes of intramammary infection in herds, overcoming the long and laborious standard-detection methods based on microbiological procedures (36) . These results confirmed the ability of this tool to determine discriminating positions in complex datasets. The ability to identify 'fingerprint' positions within a set of homologous sequences, like those of 16S rRNA gene, is the main feature of ORMA. To achieve optimal results, the starting set of sequences should be carefully selected, because, if sequences are characterized by many lowsimilarity regions, the determination of terminally discriminating position could be biased by badly aligned subsequences. In that case, a different algorithm (actually not included, but under development) for the determination of detection probes by means of the hybridization strategy, can be more appropriated. On the other side, using sequences nearly identical one to each other can cause the opposite behavior, where no discriminating positions can be determined. A careful grouping of the sequences in clusters (as we did for both of our examples, building 18 consensus out of 352 sequences in cyanobacteria dataset and 13 consensus out of 752 sequences in milk pathogens dataset) is strongly suggested. In this latter application three rounds of design were applied, in order to compensate the non-perfect homogeneity of some species. Experimental results demonstrated the correctness of this approach and the specificity of the probe pairs obtained with this design strategy. Experimental data on the 16S rRNA cyanobacteria and milk-pathogens dataset demonstrated that ORMA specifically addressed discriminating positions within a set of highly similar sequences. Nonetheless, our tool identified a total of 192 and 392 candidate positions, respectively. The intra-and intergroup scores were demonstrated to be very helpful in determining the best probes for discrimination and avoiding cross-talk between species. ORMA is a bioinformatic tool for the search and determination of single-discriminating positions among a set of highly homologous sequences and represents a The scale varies between non-significance (>0.05) to high-significance (<0.005). The line 'Other' represents the mean of all the remaining Zip-codes in the universal arrays that were not associated to any actual probe. Complete association between samples numbers and names is given in Supplementary Table 2. significant improvement from other contexts where enzyme-based techniques have been employed on already known single-nucleotide polymorphisms (SNPs) (39) or on entire subsequences (11) . This unique feature makes ORMA completely different from all the other available software for probe design in detection experiments. During the past years, academic software for species detection have been developed. ProDesign (13) is a tool based on a 'spaced seed algorithm' for the determination of probes capable of discriminating multiple pathogenic species, at different hierarchical levels. Similarly, YODA (14) performs design tasks on complete genomes against non-target species. TOFI-beta (15) implements a suffixtree-based algorithm for isolating suitable candidate probes from a target genome and filters the list according to thermodynamical and specificity requirements. These three software are implemented for the design of probes for hybridization-based detection assays. PathogenMIPer (16) , instead, is based on a different strategy (i.e. molecular inversion assays), which starts from the selection of unique sequences on a reduced dataset and then does a global comparison to all those potentially matching. All these software perform smart designs where the probes have to be selected on the whole genomic DNA; this is the typical pipeline in contexts where no preselection of the target sequences has been made, which is not the case of ORMA. In fact, in both the presented datasets, the molecular complexity of the genomic material has been reduced by PCR on the 16S rRNA. The probe pairs design, then, was performed only on the basis of a specific subset of the whole 16S dataset, limited for the specific environment in which the target species have to be detected: cyanobacteria DNA was selected and amplified by cyano-specific PCR primers, while milk pathogens 16S rRNA sequences, although amplified by universal primers, were compared only to context-specific species. The double check in RDP and BLAST, performed after the complete probe pair design by ORMA, confirmed that our choice to work with such a reduced dataset was indeed correct, because the detected species accounted for the majority of the biological diversity present in the target matrix (i.e. milk). Moreover, many of the aforementioned software perform the specificity checks by extensive BLAST searches, which is a reasonable choice for designing specific probes starting from the whole genomic DNA; in case of datasets with limited complexity (or in which the complexity has been reduced by means of molecular procedures), this approach results too computationally intensive and unnecessary for the scope. ARB (21) and PRIMROSE (22) are tools widely used for the classification and the phylogenesis of bacterial species, structured for interacting with databases specific for the same molecular target (i.e. 16S rRNA) and operate a probe design on the basis of the phylogenesis of the species under analysis. None of these two software, however, is built specifically for the determination of discriminating positions within a set of very similar sequences and they provide probe design functionality only for hybridization assays or PCR primers. When used for probe design in detection application, the strategies are based on internal mismatches or on unique stretches of nucleotides (40) . In this case, the discrimination power resides more in the decreased melting temperature of mismatched duplexes, rather than on a perfectly matched base pair between the probe and the target. Although our tool was applied on the design of probes for a specific technique (LDR) and on a specific target gene (16S rRNA), the software is not limited to this combination. LDR technique approach implied the retrieval of a pair of sequences, one of which (the DS probe) insisted on the discriminating position, whereas the other (the CP probe) is designed to anneal one base 3 0 -downstream of the discriminating position; the design of probes for minisequencing application would have implied only the determination of one probe with its 3 0 -end one base before the variation. At the same time, the design of a reporting probe for a TaqMan Real-time PCR assay would have implied the determination of one oligo with the single-base variation in the middle of the sequence. Due to its modular structure and to the straightforwardness of other applications from the already implemented one, probe sets retrieval and filtering methods could be easily added, starting from the discriminating positions found by the SBS algorithm. A further extension to hybridization probes/standard PCR primer design will be evaluated, changing the strategy for determining the positions to be tested. Provided that the initial database of sequences is accurate, updated and complete as much as possible, ORMA can retrieve discriminating positions and design specific probes on every set of sequences. Its implementation, in fact, is not based on an internal database of sequences (which are, instead, retrieved and loaded from external resources) and can be extended to any gene. In any case, the database should be critically built by only context-specific sequences. Standard procedures, like PCR with specific primers, can help in isolating only the subsets of sequences which constitute the actual database from those completely unrelated to the biological context under investigation, avoiding any interference with actual probes, as exemplified by the cyanobacteria dataset experiment. Sequences of off-target or distantlyrelated species could negatively act in the process of multiple-alignment, leading to poorly aligned datasets and biased designs. Since the databases, cyanobacterias in particular, are constantly and frequently upgraded, ORMA capability of determining discriminating positions can be refined, depending on the completeness of the initial datasets (both positive and negative set). Moreover, the continuous changing of classification and the addition of new sequences make an exhaustive and definitive design of the best cyanobacteria probes absolutely not trivial. ORMA represents a good alternative solution to the troublesome problem of searching specific positions within a large set of homologous 16S rRNA sequences and provides tools for performing a series of probe-related operations, such as sequence retrieval, filtering and scoring, allowing the user to have a set of candidates on which the actual and definitive selection can be done. The calculation on intra-and inter-group scores allows the selection of highly specific probes for molecular applications, covering the highest number of species of the positive set and having the lowest interaction with the negative set. In silico checks versus public databases (e.g. RDP or BLAST) are necessary only in case of lack of a reference among the sequences imported in ORMA or when the species eventually present in the biological context under study are too many for being comprised into a reasonably small negative set (e.g. all the microorganisms potentially present in bovine milk). Appropriate experimental designs, comprising context-specific PCRs for reducing the molecular complexity of the target can also be helpful. A complete set of major thermodynamic parameters are reported in the output, helping the researcher to carefully select the best probes. Our data assessed and demonstrated the performances of ORMA in designing probes for molecular applications on 16S rRNA gene and their feasibility for experimental use, with improved specificity and sensitivity.

Initial psychological responses to Influenza A, H1N1 ("Swine flu")

BACKGROUND: The outbreak of the pandemic flu, Influenza A H1N1 (Swine Flu) in early 2009, provided a major challenge to health services around the world. Previous pandemics have led to stockpiling of goods, the victimisation of particular population groups, and the cancellation of travel and the boycotting of particular foods (e.g. pork). We examined initial behavioural and attitudinal responses towards Influenza A, H1N1 ("Swine flu") in the six days following the WHO pandemic alert level 5, and regional differences in these responses. METHODS: 328 respondents completed a cross-sectional Internet or paper-based questionnaire study in Malaysia (N = 180) or Europe (N = 148). Measures assessed changes in transport usage, purchase of preparatory goods for a pandemic, perceived risk groups, indicators of anxiety, assessed estimated mortality rates for seasonal flu, effectiveness of seasonal flu vaccination, and changes in pork consumption RESULTS: 26% of the respondents were 'very concerned' about being a flu victim (42% Malaysians, 5% Europeans, p < .001). 36% reported reduced public transport use (48% Malaysia, 22% Europe, p < .001), 39% flight cancellations (56% Malaysia, 17% Europe, p < .001). 8% had purchased preparatory materials (e.g. face masks: 8% Malaysia, 7% Europe), 41% Malaysia (15% Europe) intended to do so (p < .001). 63% of Europeans, 19% of Malaysians had discussed the pandemic with friends (p < .001). Groups seen as at 'high risk' of infection included the immune compromised (mentioned by 87% respondents), pig farmers (70%), elderly (57%), prostitutes/highly sexually active (53%), and the homeless (53%). In data collected only in Europe, 64% greatly underestimated the mortality rates of seasonal flu, 26% believed seasonal flu vaccination gave protection against swine flu. 7% had reduced/stopped eating pork. 3% had purchased anti-viral drugs for use at home, while 32% intended to do so if the pandemic worsened. CONCLUSION: Initial responses to Influenza A show large regional differences in anxiety, with Malaysians more anxious and more likely to reduce travel and to buy masks and food. Discussions with family and friends may reinforce existing anxiety levels. Particular groups (homosexuals, prostitutes, the homeless) are perceived as at greater risk, potentially leading to increased prejudice during a pandemic. Europeans underestimated mortality of seasonal flu, and require more information about the protection given by seasonal flu inoculation.

Medical interest in Influenza A, H1N1 has been considerable [1] . However, despite dramatic warnings in the media, little is known about behavioural responses to pandemic threats. 'Common sense', lay beliefs about those most likely to be at risk, and the appropriate behaviours to adopt to avoid infection, are often not taken into account by medical practitioners. Such beliefs have been shown to influence adherence and self-care behaviours [2] . During the SARS and Ebola outbreaks, association of the viruses with Chinese or African 'others' permitted Europeans to feel relatively safe from infection [3] , and contributed to the victimisation of some Chinese in Toronto [4] . Stockpiling by the worried well can rapidly lead to shortages; cancellations of travel and closure of businesses can rapidly have profound economic consequences [5, 6] . Faced with concerns about their mortality, individuals may turn to others for reassurance, but these social networks, by sharing their uncertainties, may sometimes contribute to greater stress [7] . The international threat posed by H1N1 calls for a necessary pooling of international data, both by medical teams and by social scientists. Particular concern has been expressed about the pandemic spreading to Asia, and the potential for mixing with other variants, such as avian influenza [8] . Our team in the UK, Portugal and Malaysia sought to explore initial responses to the pandemic influenza threat. Responses to a new pandemic can be very time specific, with public reactions liable to change almost daily with media coverage [9] . Particularly important may be the gathering of data during the escalating responses that accompany a WHO pandemic alert phase 4 and 5 [10] . The WHO raised their flu alert to level 5 on 29 th April [11] , a mass media information campaign began in the UK on 5 th May 2009. We collected data from 30th April 2009 (at which time there had been 9 deaths and 212 confirmed cases worldwide) until 6th May 2009 (31 confirmed deaths, 1569 cases). By 6 th May there had been 27 confirmed cases in Europe, but none in Asia. Our study sought to gather a snapshot of the attitudinal and behavioural responses during the early stages of a pandemic, knowledge about the differences between seasonal and pandemic flu, and those groups seen as most 'at risk' from infection. Understanding such attitudes and levels of knowledge may have important public health implications for information campaigns aimed at encouraging appropriate precautions against infection, while comprehending risk perceptions can help identify those groups most likely to be at risk of stereotyping and prejudice during a pandemic. Following ethical approval by the relevant University ethics boards in London and Malaysia, data was collected from a total of 328 respondents (mean age 31.2, SD 13.37, 62% female). A paper version of the questionnaire was distributed in Malaysia, with students recruiting 180 respondents from their own classes, and community members from residential areas and local offices in Kuala Lumpur (age range 18-70, mean age 29.0 (SD 13.36), 59% female)). Response rate was generally good, with 180 out of 200 approached to participate (90%) completing the questionnaire. In Europe, data was collected between 30 th April and 6 th May 2009 from 158 respondents (age range 18-69, mean age 33.9, SD 12.8, 68% female) via an online questionnaire in English or Portuguese, linked to the website http://www.swinefluques tionnaire.com. The website link was pasted onto a variety of general, non-health networking websites (e.g. 'I love London'). Respondents were primarily from the UK and Portugal but also included 30 respondents living outside these countries and resident in Finland (19 respondents), Poland (6 respondents), Malta (3 respondents) and France (2 respondents). Ten non-European based residents were then removed from the online survey before analysis. Respondents were asked to complete a questionnaire 'about perceptions of swine flu' which comprised a number of closed and open-ended format questions. Beliefs about comparative risk were analysed for a number of group members associated with risk in previous pandemics. We included these 'high risk' groups on the basis of previous research on representations and reactions to HIV/AIDS, the Ebola virus and SARS [3, 9, [12] [13] [14] . 'Outgroup' members, who may be marginalized by society (prostitutes, homosexuals) have been linked to higher risk in previous epidemics [12, 13] , while the association of poverty with disease spread in previous infections [3] led us to include homeless people. Concerns about the risk of respiratory diseases from proximity to animals [3] meant we included pig farmers and farmers in general in our list of risk groups. We also included the elderly and immune compromised, two groups at higher risk from seasonal influenzas. Respondents completed closed-format 3point scales, indicating the extent to which they believed these groups were more at risk than me, the same risk as me or less at risk than me. Anxiety indicators were assessed through two closed-format questions assessing personal worries about catching the virus (measured on a 4-point scale, from very concerned to not at all concerned), as well as questions assessing friends and families' estimates of the risk (4 point scale, from very high to very low). We used closed format questions for two questions assessing the purchase, or intention to purchase, specific items, such as face masks ("Have you already bought (or are you intending to buy) anything in preparation for a swine flu epidemic (e.g. face masks, food, tissues, cleaning materials)?" (yes/no responses). We accompanied this with an open-ended question for those who indicated they had/intended to purchase an item ("what have you bought?") with responses categorised for frequency in Malaysia and Europe, with the most common categories reported below. We also assessed whether respondents expected to modify their public transport use as a result of the threat (3 point scale: increase transport use, decrease transport use, or remain the same, collapsed into the categories 'less' and 'much more or the same' for analysis), and whether respondents intended to cancel or delay travel plans (yes / no). An open-ended question asked 'how can you protect yourself from infection?' with the most frequent responses coded into categories by our research team in Malaysia and Europe. In Europe only, we asked additional closed-ended questions about stopping or reducing eating pork as a result of the pandemic (yes/no), and the precautionary desire for anti-viral drugs at home ('would you like to have your own anti-viral drugs at home just in case'? 'Have you tried to obtain any anti-viral medicines to keep just in case?' (both yes/no answers)). We asked about the mortality rate in ordinary, seasonal flu (five point scale, under 50,000, 50-000-99,000, 100,000-249,000, 250,000-500,000 and over 500,000), as well as whether the symptoms of swine flu differed from seasonal flu, and whether seasonal flu vaccination provided immunity against swine flu (both yes/no answers). Throughout, statistical analysis for structured questions was through chi-square analyses and Pearson product moment correlation. As shown in table 1 approximately a third of respondents reported they would use public transport less (116/320) or had contemplated cancelling or delaying flights (124/ 312), with this response more pronounced in Malaysia (respective x 2 (1) = 21.91, 49.20, both p < .001). Few had already bought products (25/328), but 41% (74/180) of our Malaysians were preparing to do so (x 2 (1) = p < .001 for differences between Malaysia and Europe). Most likely to be purchased were masks (mentioned 46 times in Malaysia, 14 in Europe) and food (14 times Malaysia, 5 times in Europe). Asked in the free response question how to avoid infection, 37% (121/328) of our respondents cited washing hands and good hygiene, 28% (92/328) wearing a mask, 13% avoiding infected others and 12% (39/328) shunning crowded places. Five groups were seen as particularly at risk by more than half of our respondents: those with weakened immunity, pig farmers, the elderly, the homeless and prostitutes/ highly sexually active. Malaysians were more likely to see pig farmers, general farmers, homosexuals and prostitutes as at greater risk (respective x 2 (2) = 68.03, 11.44, 31.82, and 12.10, p < .001 for each), Europeans were more likely to see the elderly and those with weakened immunity at risk (respective x 2 (2) = 8.27, 3.49, p < .05). Whilst around half (165/325) of our respondents reported they were at least 'somewhat concerned' about being a victim of the pandemic, this anxiety was stronger in Malaysia, where 71% (127/178) indicated they were at least 'somewhat concerned' (x 2 (3) = 91.67, p < .001). Nearly three quarters of our overall respondents (241/325) felt they had at least some control over whether they were infected. Europeans were more likely to have discussed their fears with their friends (90/142) (x 2 (1) = 66.56, p < .001). Across the sample, personal perceptions of risk about the pandemic were related to those of families and friends (respective rs .57, .58, p < .001). Those most anxious about personally being a victim of the outbreak were the most likely to reduce their use of public transport (r (320) = .48, p < .001) and cancel/delay air travel (r (322) = .37, p < .001). In our additional (Europe only) data, respondents underrated the dangers of ordinary season flu, with 64% (95/ 148) claiming that this killed under 100,000 worldwide (actual numbers are between 250,000 and 500,000) [15] . 26% (38/148) of our European respondents wrongly believed that a vaccination for seasonal flu gave immunity against swine flu. The same percentage believed swine flu symptoms differ from those of seasonal flu. While only 3% (4/148) had already obtained anti-viral drugs for use against swine flu, 32% (47/148) claimed they would like to have these at home in case of infection. Few (7%, or 10/ 148) claimed they had stopped or reduced their eating of pork as a result of the pandemic. At present, it is unclear as to whether the outbreak of Influenza A, H1N1 will prove to be a "false alarm", or whether the virus will mutate and spread in a new, more dangerous form, perhaps this Autumn [16] . Our data collection in the early stages of the pandemic/pandemic of Spring 2009 suggests that respondents felt they had some control over potential infection. Respondents identified 'washing hands', avoidance of infected people, avoidance of crowded areas and mask wearing as strategies for avoiding infection, reflecting generally approved public health measures [17] . Malaysians were particularly anxious about a pandemic, despite the lack of cases of this influenza in Malaysia during our research period, probably reflecting the recent avian influenza alert in this country [8] . As with previous health alerts, personal anxieties can feed behavioural changes [18] , with many Malaysians contemplating significant changes in their use of transport, and anticipating the purchasing of goods, particularly masks, in preparation. European respondents were particularly likely to have discussed the pandemic with their friends, while a quarter of respondents overall had discussed the pandemic thereat with their family. Our correlational data suggests that such conversations may reinforce existing levels of anxiety. Practitioners need to be aware that rumours spread fast during times of pandemic threat, with significant risks of emotional as well as physical 'contagion' between populations [19] . Any increase in anxiety can lead to rapid behavioural changes that can soon lead to shortages, and enhance the desire for medication available at home. An unrealistically optimistic belief that others are at greater risk than ourselves can reduce our willingness to enact healthy behaviours [20] . During pandemics, particular 'out-groups' may be vulnerable to discrimination [21] . Although respondents correctly identified groups such as the immunocompromised as at greater risk [22] , half our respondents saw the sexually active as at greater risk, almost a third of Malaysians suggested homosexuals were at more risk of infection. This may reflect a popular belief in Malaysia that homosexuals are likely to be already immunocompromised through infection with HIV/AIDS. The homeless were also perceived as at greater risk in both Malaysia and Europe. Political and health authorities need to be wary of increased stereotyping and prejudice towards particular societal groups during an influenza pandemic. Our Europe-only data suggested that individuals underestimate the threat of regular seasonal flu, while a quarter of our respondents incorrectly believed seasonal flu and swine flu symptoms were different, and that seasonal flu vaccination could help immunise against swine flu. Despite major media and governmental campaigns across Europe, there is obviously still a need for greater information with respect to symptom logy and immunisation against infection. Our study was a rapid, cross-sectional analysis in response to a particular outbreak, and as such suffers from a number of limitations. Although this research was unique in tracing initial behavioural responses to this pandemic, our respondents were not true random samples in either continent, and we assessed only a small range of potential behaviours. Our Malaysian sample was drawn from one large city -Kuala Lumpur -and may therefore not be representative of other, more rural populations in that country. Similarly, our European data was drawn primarily from the UK and therefore cannot be seen as representative of the whole continent. Self-report biases in questionnaire completion may mean that our respondents were unwilling to provide openly prejudicial responses, whilst our study in Europe was further limited by including only those who had access to the Internet. To fully model likely behavioural changes, and their consequences for public health services, larger, more representative longitudinal studies are now needed to track public anxieties and health behaviours in the continuing battle against pandemic influenzas. Numerous studies have identified behavioural interventions valuable in prevention of epidemic/pandemic influenza, but we know little about individuals' own perceptions of risk at the beginning of a pandemic, which groups in society they believe most at risk of infection and how they have changed and intend to change their behaviours as a pandemic develops. Our findings suggest culture and individual anxiety are important predictors of behavioural responses to pandemic influenza, with higher levels of anxiety about swine flu in Malaysia compared to Europe, and with greater levels of behavioural change in Malaysia. Particular 'out-groups' (e.g. prostitutes, homosexuals) were judged to be at relatively high risk of infection, with Malaysian respondents particularly likely to emphasise the infection risk in these groups. Such judgements of risk may have important implications for the equitable treatment of socially marginalised group, particularly as the pandemic continues to accelerate worldwide. Note Asterisks indicate significant regional differences (Europe vs. Asia) using Pearson chi-square statistic * p < .05; ** p < .01. We reran these analyses using logistic regressions controlling for age and sex, with similar results to the chi-square analyses. Further details of these findings are available from the first author. Percentages are rounded so may not all always add to 100. Ns range from 312-328 due to some missing data from Malaysia.

On managing complex adaptive systems motivated by biosystems application to infections

Many attempts to control Complex adaptive systems (CAS) have failed. Here we try to learn from biosystems to derive some principles for CAS management. An application to managing infections is given.

Complex adaptive systems are special cases of complex systems. They are complex in that they are diverse and made up of multiple interconnected elements and adaptive in that they have the capacity to change and learn from experience. i.1) CAS consists of adaptive (capable of learning) entities. Some properties of such systems cannot be understood just by studying the constituents. They are called emergent properties [3, 4] . i.2) Examples of CAS include the brain where the emergent property is cognition which is a property of the brain not of individual cells. We propose that IS is a CAS with tolerance as an emergent property since tolerance is a property of IS as a whole not individual B or T or macrophage etc.... cells. In fact almost every biological, social and financial system is a CAS. i.3) Emergent properties imply that reductions approach is not suitable for CAS. Hence these systems have to be studied as a whole in addition to the reductionest approach. It is crucial to see that both approaches complement each other and are not contradictory [5] . i.4) CAS have some generic properties e.g. they are open (connected to other systems). Interactions in CAS are nonlinear hence long range predictions in CAS are highly unlikely. Typically they have a network (e.g. evolving network) structure. Optimization in such systems is multiobjective. Control of most CAS is difficult, it should be targeted and integrated i.e. using diverse approaches not just one. Most CAS show delay and memory behavior. Spatial effects should not be neglected in many CAS. Most, but not all, CAS are decentralized. i.5) CAS should be studied as a whole in addition to studying its constituents. Hence in vivo experiments are crucial to understand them. i.6) CAS are difficult to control for the following reasons: They are nonlinear. They are open. Delay is a generic property of CAS. Some failure cases in controlling CAS are the Measles-Mumps Rubella (MMR) vaccination and the change of fish type in Lake Victoria [2] . Therefore in the next sections we try to learn from some CAS systems to derive some principles for "managing" CAS. By management we mean taking decisions with as least bad side effects as possible. And if bad side effects occurred then decisions are taken to remove or contain them as best as possible. Some of the basic concepts for optimization in CAS are: ii.1) Multi-objective optimization [6] especially metaheuristics i.e. methods which contain random element and which give a good approximate solution to the optimization problem within a reasonable time. ii.2) Game theory specially evolutionary game theory [7] . In some cases it is important to include the evolution of strategies or players c.f. the immune system deals with almost continuously evolving antigen e.g. bacteria, viruses or macroparasites. ii.3) Fractals and networks which play a crucial role in nonlocal spread of effects e.g. infections. ii.4) Collective animal behavior. Some of its principles include [8] : Non-selfish key individuals, positive and negative feed back, diversity and threshold responses. From the above results we conclude the following: iii.1) CAS are extremely difficult to control hence managing CAS may be a more feasible goal than controlling them. iii.2) Diverse and integrated approaches to the problem are crucial. In may cases single approaches to the problem may lead to counter-productive results. Also this increases the robustness of the management. iii. 3) The existence of non-selfish key individuals is necessary to solve many problems. iii.4) The goals should be multi-objective and only guide lines (metaheuristic) approaches should be adopted. One cannot attempt exact determination or control in nonlinear and open systems. Hence no detailed plans should be laid out, instead guide lines may be more realistic. iii.5) Small world network structure should be followed hence a kind of hierarchical network with short cuts can significantly improve the efficiency of the system [9] . iii.6) Feedback hence adaptive management is expected to be more efficient than decisions taken once and for all. Also considering the system as game theory with evolving strategies or opponents may help. iii.7) Emergent problems "messes" happen and will happen. Now the above conclusions are applied to an overview of infection management. We propose to call such problems "emergent problems" [10] . An emergent problem is characterized by: 1) It has several reasons not just one. 2) It cannot be solved locally (e.g. by one country) hence "collective efforts" are needed. 3) It needs a long time to be solved hence evolution in solution strategies has to be taken into consideration. Collective efforts needed to solve such emergent problems require large scale cooperation. Humans are known not to be good in cooperation. In fact a core problem in most of the above problems is the tragedy of commons [11] which simply states that in the case of finite resources e.g. drugs (which is always the case) a conflict exists between individual's interest and the group's (common) interest e.g. rich and poorer countries. However it is known that for such cooperation to exist a feeling of imminent danger has to exist [12]. Therefore early preparations to deal with such problems are not expected to be as efficient as should be. It is clear that the absence of non-selfish key individuals is a key reason for the inability to solve this problem. However some clusters of cooperators do exist. This should be accompanied by retaliation against non-cooperators to lower their benefits from defection (c.f. prisoner's dilemma game [7] ). Some benefits from our discussion are: i) Diverse approaches to the problem is highly desirable e.g. using multi-drugs to reduce the danger of drug resistance (c.f. the sole dependence on Tami flu for avian influenza). ii) The network (long range) effects are crucial in many infections e.g. foot and mouth disease, SARS, Avian flu and H1N1 flu. iii) We should take into account the evolution of antigens (c.f. the new strains of Tuberculosis).

Prevalence of human herpesvirus 6 antibodies and DNA in allogeneic stem cell transplant patients: two-year single centre experience

INTRODUCTION: Human herpesvirus 6 (HHV-6) has been recognized as a potentially significant pathogen in hemopoietic stem cell transplant (HSCT) recipients. Different clinical manifestations have been described, including fever, skin rash, bone marrow suppression, and encephalitis. MATERIALS AND METHODS: A retrospective review of a group of 26 adult recipients of allogeneic HSCTs was conducted. Serum samples taken before transplant were examined for the presence of specific anti-HHV-6 IgM and IgG antibodies. After transplantation, quantitative real-time PCR was used to determine viral load in plasma samples from days 0–80 post-transplant. RESULTS: HHV-6 DNA was detected in plasma samples in 8 (30%) of the 26 recipients between days 18 and 40 after transplantation. All of them developed fever of unknown origin and over 50% had graft-versus-host disease features. Three individuals from this group died during detectable HHV-6 viremia. Another two recipients showed a single positive PCR result at a later time. Infection with HHV-6 was thus confirmed in 10 (38.5%) of the 26 graft recipients. CONCLUSIONS: There is a high frequency of detectable HHV-6 viral load in stem cell transplant recipients in Poland. Further investigation to monitor HHV-6 reactivation in graft recipients will be important to improve outcome for these patients.

Human herpesvirus type 6 (HHV-6) belongs to the subfamily β-herpesvirinae and is closely related to HHV-7 and HHV-5 (formerly known as cytomegalovirus -CMV) [5] . Most people have had a primary HHV-6 infection in early childhood [3] and the virus is widespread in the human population, as shown by the presence of specific antibodies in 90% of healthy adults [17] . As a result of the primary infection, HHV-6 is assumed to establish a latent infection and viral DNA can be detected in salivary glands [7] , monocytes [10] , and early bone marrow progenitor cells [13] . HHV-6 is reported to be a causative agent of exanthema subitum [20] and has been associated with a broad spectrum of diseases, including febrile convulsions [11] , encephalopathy [9, 18] , hepatitis [6] , and lymphoproliferative disorders [15] . An active HHV-6 infection can cause fatal disease in an immunocompromised host, including patients after stem cell transplantations (SCTs), but fatal outcome in immunocompetent persons due to HHV-6 infection is extremely rare [14, 16] . HHV-6 presence in plasma or serum was documented in 33-48% of hemopoietic stem cell transplant (HSCT) recipients using molecular techniques [22] . The peak viral load occurs early after transplantation and usually within the first four weeks after transplantation [4, 21] . Allogeneic SCT, advanced hematological disease, young age, gender mismatch between the donor and recipient, and treatment with corticosteroids are commonly reported risk factors associated with HHV-6 infection after transplantation [4, 21] . In a recent study we summarized retrospective results of the determination of HHV infection status in allogeneic stem cell transplant recipients. As infections with HHV-6 are rarely accompanied by clinical symptoms, our first aim was to compare HHV-6 infection status, measured as viral DNA presence in the blood in the post-transplantation period and patients' anti-HHV-6 serological status by detection of IgG and IgM antibodies. This allowed us to determine both the presence of active infection and whether a recent infection was a result of primary contact with the virus or the reactivation of latent virus, which should be valuable information in terms of comparing HHV-6 infection status and clinical status of an infected patient. Another virus belonging to the β-herpesvirinae subfamily, HHV-5 (CMV), is one from the most important viral pathogens causing clinical infections in patients receiving immunosuppressive treatment. As there is information in published data about a connection between HHV-5 and HHV-6 reactivation in graft recipients, our aim was to determine the appearance of simultaneous reactivation of both β-herpesviruses. We believe that this is probably the first report from Poland involving this kind of examination in HSCT recipients. This retrospective study involved patients who received an allogeneic HSCT and were hospitalized at the Hematology, Oncology, and Internal Medicine Clinics, Medical University of Warsaw. In the period under consideration (December 2004 to October 2006) there were 38 patients receiving allogeneic stem cell transplants. The criteria for patient selection for the study included the appearance of at least one syndrome from those listed below within a period of 100 days after HSCT: pneumonia, appearance or intensification of graft-versus-host disease (GvHD), skin rash, or pyrexia of unknown origin. Introduction of these criteria resulted in a group to 26 adult receivers of HSCTs ( Table 1) . Monitoring of clinical status of the patients and viral load in blood samples comprised a period of 180 days after HSCT. IgG and IgM antibodies against HHV-6 were measured from the panel of 26 serum specimens, collected once before HSCT, from the patients with different hematological disorders using a commercial enzyme immunoassay (PanBio). The results obtained with both HHV-6-specific IgG and IgM tests were expressed in PanBio units (PBU). The PBU is calculated as the ratio of the sample absorbance to the cut-off absorbance and multiplied by 10 (absorbance of sample/mean absorbance of cut-off ×10). The collection of plasma samples from all patients for HHV-5 PCR investigations, according EBMT guidelines, began at a median of 3 days after transplantation (range: 1-7 days) and lasted until a median of 105 days (range: 30-180 days). Routine collection of plasma samples was performed once a week until the 100th day after allogeneic HSCT, thereafter once every two weeks, up to 180th day after HSCT. The median number of blood samples per patient was 12 (range: 3-21). A total of 294 samples from the 26 patients were collected. For the purpose of recent study, viral DNA was extracted from 200 µl of each plasma sample using a QIAmp ® DNA Blood Mini Kit (Qiagen) in accordance with the manufacturer's instructions and retrospectively used for detection of HHV-6 DNA. For the detection of HHV-6, a real-time PCR assay with fluorescent probes complementary to the sequence lying within the amplified product was used. The tests were run on a LightCycler 2.0 instrument (Roche) with the commercial quantitative MutaREAL ® HHV-6 kit (ALPCO). All samples were examined according to the manufacturer's instructions. Every tested sample was amplified with an internal control (positive control of the DNA extraction and amplification process). Each amplification reaction also included, besides the tested samples, positive HHV-6-specific controls and a negative control of the DNA extraction and amplification process. HHV-5 DNA was detected using the commercial real-time test CMV Quant Kit ® (Roche) developed for the LightCycler 2.0 instrument. The test uses internal SCORPIONS™ fluorescent probes for the PCR amplification product. Analogously to HHV-6 detection, for every sample an internal control was added and the amplification was performed in the presence of amplification-specific controls (positive, negative, and extraction process control). Specific IgM antibodies were present in the serum of 2 (8%) patients, while the others were negative (92%). Twenty of the 26 persons (77%) had IgG antibodies against HHV-6 before HSCT and the other 6 (23%) were negative. Of the group of 16 patients who had no detectable HHV-6 during the entire post-transplantation period, 10 (62.5% of the HHV-6-negative patients) had a positive result only for the anti-HHV-6 IgG test, but no anti-HHV-6 IgM, in serum sample taken directly before transplantation, 4 (31.3%) were both IgM and IgG negative, and the remaining patient had anti-HHV-6 antibodies of both classes ( Table 1) . The serological status of the 10 patients who had HHV-6 DNA detected in one or more blood samples was very similar: 7 patients (70%) had only IgG, 2 patients (20%) had no IgG or IgM antibodies, and one patient had HHV-6-specific IgG and IgM. Regarding the amplification results of DNA isolated from plasma using HHV-6-specific PCR, expressed as the presence of exponential increases in fluorescence, products were detected in 29 samples (10%) taken from 10 patients (38%). Two of them (8%) had HHV-6 DNA in only a single positive sample and another 8 (30%) had positive results in two or more subsequent tests. In the patients with two or more plasma samples containing HHV-6 DNA, viremia was observed between days 18 and 40 after transplantation (Table 1 ). In the two patients who had a single HHV-6-positive blood sample, viral DNA was detected at a later time. The quantitative results obtained by the MutaREAL ® HHV-6 test in all positive cases were at low levels, between 700-1600 copies/ml. HHV-5 DNA was detected in the plasma samples collected from four HHV-6-negative patients. In all four cases, HHV-5 DNAemia was observed in the typical period of 40-65 days after transplantation and, moreover, a second HHV-5 viremia course occurred in two patients starting at days 103 and 123 after transplanta- Abbreviations: ALL -acute lymphoblastic leukemia, AML -acute myeloid leukemia, CML -chronic myeloid leukemia, MDSmyelodysplastic syndrome, rPBSCT -related peripheral blood stem cell transplantation, uPBSCT -unrelated peripheral blood stem cell transplantation, uCBSCT -unrelated cord blood stem cell transplantation, rBMT -related bone marrow transplantation, uBMT -unrelated bone marrow transplantation, -: negative, +: positive in one plasma/serum sample, ++: positive in two or more plasma samples. Abbreviations: ARDS -acute respiratory distress syndrome, CNS -central nervous system, GvHD -graft-versus-host disease, N/A -not applicable, -: negative, +: positive in one plasma/serum sample, ++: positive in two or more plasma samples. tion. None of the HHV-6-positive patients had HHV-5 DNA in their plasma samples during the 180-day period. Overall mortality in the entire group of HSCT recipients during the first 180 days after transplantation was 34.6% (9 of the 26 patients) and the most frequent direct cause of death was acute respiratory distress syndrome (6 patients), in some patients accompanied by sepsis (2 patients), pneumonia (1 patient), or central nervous system (CNS) infection (1 patients). Other death causes included pneumonia (1 patient), bleeding within the CNS during infectious meningitis (1 patient), and shock during subsequent HSCT (1 patient). Seven of the 9 patients died during the first 34 days after HSCT (Table 2 ). In the HHV-6-positive patients, mortality was 50% (5 out of 10 patients). Four of those patients died during the HHV-6 viremia period, which occurred within the first 34 days after HSCT. Table 2 summarizes the clinical features and the times and causes of death of the HSCT recipients. Herpesviruses persist throughout life after primary infection. Viral proliferation occurs either spontaneously or under conditions of immunosuppression. Reactivation can lead to illnesses that typically differ in their clinical presentation from the disease associated with the primary infection. After SCT, reactivating members of the β-herpesvirinae subfamily (among them HHV-6) frequently cause serious, sometimes life--threatening disease [8, 12] . HHV-6 infection or reactivation in these individuals has been associated with a delay or suppression of marrow engraftment [12] , pneumonia [2] , skin rash, and fever [16] . Although it is difficult to prove an etiologic association of the virus with these disease events, their propinquity with HHV-6 activity in the absence of other possible causes suggests that at the very least a subset of these events is due to HHV-6 activity. In the present study we found that 38% (10 individuals) of the graft recipients developed HHV-6 infection ( Table 1) . Eight patients (30%) had detectable viral DNA levels in two or more samples in subsequent tests during the six weeks following SCT and another two (8%) had a single sample positive at a later time. It is likely that the HHV-6 infection that occurred after transplantation was due to activation of the virus in the bodies of the recipients, since most of the recipients were immune to HHV-6 and no virus was found before SCT. However, we do not know whether HHV-6 strains detected in the present study were derived from the donor or were transfered from other sources to the seronegative recipients. The virus probably remains latent in the body after primary infection, as do other human herpesviruses. If the HHV-6 was derived from the donor, it must have been latently infecting the donor's hemopoietic stem cells and was activated to replicate after transfer to the recipient. In other cases, the virus was probably reactivated from the recipient's own body by factors such as a profound immune dysfunction or an allogeneic reaction after transplantation. We did not find any correlation between viremia and the applied conditioning regimen or anti-GvHD prophylaxis. All of the HHV-6-positive graft recipients had fever of unknown etiology during the six weeks after SCT and half of them (4 persons) had acute GvHD features. Sixty-three percent of the described patients had pneumonia and 38% skin rash. No Epstein-Barr virus or CMV DNA was found in all the plasma samples during HHV-6 onset ( Table 1) . The exact association between HHV-6 reactivation and mortality found in this study is not clear. Three of the patients (38%) with detectable HHV-6 DNA levels died during viremia shortly after transplantation, all due to coexisting pneumonia of unconfirmed etiology. Acyclovir in typical doses was used as an antiviral prophylaxis during this period, but without any visible clinical success. Studies in vitro have shown that HHV-6 DNA replication is inhibited by ganciclovir, foscarnet, and cidofovir, but not by acyclovir [1] . There is a high frequency of detectable HHV-6 viral load in SCT recipients and it may lead to an increased risk of fatal symptomatic disease [19] . The availability of quantitative real-time PCR means that results are available in a clinically helpful time-frame, which should assist with implementing timely therapeutic intervention and assessing response to treatment. Further investigation to monitor HHV-6 reactivation on a larger group of SCT recipients will be important in improving outcome for these patients.

The infection of primary avian tracheal epithelial cells with infectious bronchitis virus

Here we introduce a culture system for the isolation, passaging and amplification of avian tracheal epithelial (ATE) cells. The ATE medium, which contains chicken embryo extract and fetal bovine serum, supports the growth of ciliated cells, goblet cells and basal cells from chicken tracheas on fibronectin- or matrigel-coated dishes. Non-epithelial cells make up less than 10% of the total population. We further show that ATE cells support the replication and spread of infectious bronchitis virus (IBV). Interestingly, immunocytostaining revealed that basal cells are resistant to IBV infection. We also demonstrate that glycosaminoglycan had no effect on infection of the cells by IBV. Taken together, these findings suggest that primary ATE cells provide a novel cell culture system for the amplification of IBV and the in vitro characterization of viral cytopathogenesis.

Embryonated egg inoculation is commonly used to amplify avian RNA viruses. Successful viral replication mainly depends upon viral adaption to and replication in the cells of the embryonic and extraembryonic tissues. Growth retardation and malformation of the embryo are the main indicators of viral reproduction in this system and signal the proper period for harvesting the amplified virions from the amniotic or chorioallantoic fluids. The low efficiency of viral amplification may prevent etiology determinations, intensive surveillance of epidemiology and the tracing of potential natural domestic hosts for viral pathogens during an outbreak. For instance, avian metapneumovirus (MPV), which primarily attacks the respiratory tract, replicates inefficiently in embryonic eggs. Diagnosis of MPV infection is mainly dependent on the detection of viral RNA or viral antigens, but not viral isolation [5, 6] . In addition, to investigate the epidemiology of avian influenza viruses (AIV) requires efficient and sensitive viral detection and amplification system for comprehensive collection of the viruses from waterfowl and migratory birds [17, 31] . However, the wildbird isolated AIV are usually mildly or nonpathogenic to chicken embryos [20, 26] and the efficiency of viral isolation is often unsatisfactory in embryonated eggs [11, 19] . It has been shown that serial passage using chicken embryos enforces the selection of adapted viruses and inevitably alters the genetic codes of the primary viral isolates, especially for genes involved in cell adsorption and viral replication [8, 9] . The environment of the viral replication may also selectively modify the original cell tropism and pathogenesis of the isolated virus [15, 25] . It has been shown that a serially embryo-passaged of infectious bronchitis viruses (IBV) M41 strain exhibited significantly attenuated pathogenesis within the oviduct compared to the parental strain [7] . In addition, egg-mediated epitope alterations in amplified isolates might attenuate the protective efficiency of these viruses when used to vaccinate against the challenge of a wild-type virulent strain [16, 24] . For instance, after egg adaption, human influenza A virus, originally showing a2-6 sialic acid tropism only, exhibits increased hemagglutinin (HA) binding affinity to a2-3 sialic acid-containing gangliosides due to amino acid substitutions in the vicinity of the receptor binding site of the HA protein [8, 15] . For primary viral isolates, providing natural host cells for replication might eliminate the selection pressure of an exotic growth environment and avert the alteration of cell tropism and consequent genomic mutations. Recently, a study reported a primary culture system for tracheal epithelial cells from an embryonic day (E) 17 chick embryo, in which dissociated culture cells exhibited ciliary movement and were positive for the expression of pan-cytokeratin [33] . In addition, the global profile of mRNA expression in the cultured cells showed that cytokeratin 14 (K14), a basal cell marker, was highly expressed. However, the protein expression of cell-type specific markers and the existence of mucin-secreting goblet cell are not well-characterized. In this study, we developed a novel culture system to isolate, amplify and passage chicken tracheal epithelial cells in an efficient manner. The avian tracheal epithelial (ATE) cell types were identified using immunocytostaining, and the ratio of cell types is statistically illustrated. We further demonstrate that primary ATE cells support IBV replication. The susceptible cell types and the effect of glycosaminoglycan (GAG) on IBV attachment to ATE cells was also investigated. Two IBV, 2575/98 and 2296/95, were isolated in Taiwan and maintained using serial passage through specific pathogen-free (SPF) eggs [29] . IBV were inoculated into the embryonic eggs at the E9 stage and harvested at the E11 stage from chorioallantoic fluid. The original viral titers of 2575/98 and 2296/ 95 are 10 7.4 EID 50 /0.1 mL and that of 2296 reached 10 9.6 /0.1 mL, respectively [14] . The applied viral titer of both IBV was 1 · 10 5 EID 50 /mL. Plaque-purified Taiwan-isolated Japanese encephalitis virus (JEV) strain RP-9 [28] was used as a positive control for the study of GAG effect on the virus-cell attachment. The adsorption of JEV was performed with serum-free RPMI medium at 37°C for 1 h. Virus-infected cells were incubated in RPMI containing 2% fetal bovine serum (FBS). Virus titers were determined by plaque assay in BHK-21 (baby hamster kidney) cells, which were maintained in RPMI 1640 with 5% FBS, 2 mM L-glutamine, 100 mg/mL of streptomycin, and 100 IU/mL of penicillin. The medium and culture reagents are obtained from Invitrogen (Carlsbad, CA, USA). Tracheas were obtained from one-day-old SPF chicks and rinsed in DMEM medium (Invitrogen) under sterile conditions. After the removal of surrounding adipose and muscular tissues, tracheas were digested with dispase I solution (2.5 U/mL dispase I) (Roche, Nutley, NJ, USA) for 2 h at 37°C to disrupt the basal membrane. Forceps were used to force the outflow of detached cell sheets of tracheal epithelium from the tracheal lumen. The epithelial cells were harvested and further digested with collagenase I (1 mg/mL, Roche) for 5 min at 37°C. The cells were gently pipetted, and the tissues were homogenized into small cell clumps. FBS was added to the digesting solution to stop the reaction. The cells were centrifuged at 1000 rpm for 5 min to remove residual digesting enzymes. The cell pellets were resuspended in ATE medium, containing 10% FBS (Invitrogen), 10% chicken embryo extract (CEE) (US Biological, Swampscott, MA, USA; or self-made [27] ), 1% glutamine (Invitrogen), 0.1 mM b-mercaptoethanol (Sigma-Aldrich, St. Louis, MO, USA), 1% nonessential amino acids (Invitrogen) and 1% penicillin/streptomycin (Invitrogen) in DMEM-F12 medium (Invitrogen). Cells were seeded on 2% matrigel-or 20 lg/mL fibronectin-coated 24-well or 96-well plates (Corning Inc., Corning, NY, USA). When the cells reached confluence, the cultured cells were passaged by digestion with both 0.01% (w/vol) protease (type XIV, Sigma-Aldrich) and 2.5 U/mL dispase I for 5 min at 37°C. GAG compounds, including the heparin (H), heparan sulfate (HS), dextran sulfate (DS) and chondroitin sulfate (CS), were obtained from Sigma-Aldrich. The MTT (3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) assay kit was obtained from Sigma-Aldrich. FITC-conjugated BSI-B4 isolectin (Sigma-Aldrich) was used as a marker for basal cells from the tracheal epithelium [12] . Commercial culture media for human and rat airway epithelial cells were obtained from CELLnTEC advanced cell system (Postfach, Bern, Switzerland) and CELL Applications Inc. (San Diego, CA, USA), respectively. Cells or cryosections of infected trachea were fixed in 4% cold paraformaldehyde and permeabilized with 0.3% Triton-X 100. Immunocytochemistry was performed with the following primary antibodies: E-cadherin (1:500, BD Biosciences, Franklin Lakes, NJ, USA), b-tubulin IV and pan b-tubulin (1:100, Sigma), ZO-1 (1:100, Zymed-Invitrogen), mucin 5AC (1:50, Abcam, Cambridge, MA, USA), cytokeratin 14 (1:100, Convance, Princeton, NJ, USA), smooth muscle actin (SMA, 1:50, DakoCytomation, Glostrup, Denmark) and vimentin (1:500, Convance). Immune-serum of IBV-infected chicken, screened by ELISA, was used for the detection of IBV antigens in infected cells (1:200) [30] . Fixed cells were washed twice with 0.1% Tween-20 in phosphate buffer saline (PBS). Appropriate fluorescence-tagged secondary antibodies (all from Jackson ImmunoResearch, West Grove, PA, USA) were used for visualization. Blue 4 0 ,6-Diamidino-2-phenylindole (DAPI) was used for nuclear counter-staining. Images of immunostaining were captured using a fluorescent microscope (Nikon ECLIPSE 80I) or confocal microscope (LSM510 Meta, Zeiss). Total RNA was extracted from TW2575/98infected ATE cells using TRIzol TM C&T (Protech, Taiwan) according to the manufacturer's protocol. Positive-sense and negative-sense viral RNA (0.5 lg) were reverse transcribed into cDNA using SuperScript TM III reverse transcriptase (RT) (Invitrogen) with oligo dT and sense primer 5 0 -ACT GAA AAT GAT AGT GTT ATG-3 0 , respectively. PCR was performed with a proofreading DNA polymerase (KOD-Plus, Toyobo, Osaka, Japan). The PCR cycling conditions consisted of 28 cycles at 94°C for 30 s, 62°C for 30 s, and 68°C for 1 min on a PCR machine (ASTEC 818). The primer sequences for detection of the nucleocapsid gene of IBV were 5 0 -AAT GCA TCT TGG TTT CAA GC-3 0 and 5 0 -TCC TCA TCT GAG GTC AAT GC-3 0 ; for detection of GAPDH, the sequences were 5 0 -GTG AAG GTC GGT GTG AAC G3 0 and 5 0 -GGT GAA GAC ACC AGT AGA CAC TC-3 0 . The effect of GAG on JEV and IBV infections was measured in BHK-21 cells and ATE cells, respectively. The cells were seeded at a density of 5 · 10 3 cells/well in 96-well plates. The GAG were serially diluted and applied at 0, 7.5, 15, and 30 lg/mL for BHK-21 cells and at 0, 0.5, 1.0, 2.0 mg/mL for ATE cells. The BHK-21 and ATE cells were infected with 500 plaque-forming units of JEV and 10 3 EID 50 of IBV at 37°C for 1 h. Cells were washed three times with serum-free medium, and then normal growth medium was added. The number of infected cells per 96-well, revealed by IBV immunocytostaining, was manually counted at 8 h post-infection (h.p.i.) from five individual fields. No obvious cell damage was observed after GAG treatment or viral infection at 8 h.p.i. [18, 28] . For the isolation of tracheal epithelial cells, the intact cell sheet of the tracheal epithelium was first isolated from dispase-treated tracheas (Fig. 1A) . The epithelial membrane sheet was further mechanically disrupted into small pieces by pipetting. Floating unattached ATE cells rapidly underwent cell death or growth arrest after a one-day culture. The optimal cell matrix for ATE cells was evaluated; both 2% matrigel or 20 lg/mL fibronectin efficiently promoted cell attachment, but gelatin, collagen I, collagen IV and laminin were less effective (data not shown). The primary culture system for mammalian tracheal epithelial cells is well-established. To culture the ATE cells, specialized commercial media for the tracheal epithelia of humans and rats were first tested (described in the Materials and methods section). None of them supported cell growth or long-term viability of ATE cells. Adjustment of the concentration of retinoic acid, a differentiating factor for tracheal epithelial cells [10] , did not affect attachment or cell growth of ATE cells. Increasing FBS from 10% to 20% in the culture medium modestly enhanced the viability of the cells (data not shown). We speculated that unidentified chicken nutrients essential for the promotion of ATE cell growth were absent in the culture medium. CEE, an essential supplement for the culture of mouse neural crest stem cells [27] , was added to DMEM-F12 medium with 10% FBS. We found that CEE significantly improved cell growth and reduced cell death of the primary ATE cells (Figs. 1B and 1C) . The cell number of initial harvest from one trachea on day 1 is about 2 · 10 5 cells. With this optimized culture medium, isolated cells displayed increased MTT activity, and the celldoubling time was 46 h during the first 6 days of culture (Fig. 1D) . The ATE cells could sustain proliferative activity for 3-5 passages. These results show that the established culture system provides sufficient nutrients and proper Tracheal epithelial cells are connected to each other by E-cadherin molecules [23] . Among the isolated ATE cells, 88.3 ± 13.2% expressed E-cadherin and a gap-junction molecule, ZO-1 [23] . E-cadherin-negative cells, including SMA-positive muscle cells and vimentin-positive fibroblast cells, accounted for 5.0 ± 1.4% and 3.5 ± 1.2% of the cultured cells, respectively, indicating the high purity of this primary culture system. In tracheal epithelial cells, b-tubulin IV/b-tubulin, mucin 5AC and cytokeratin 14 (K14) are specific markers of ciliated cells, goblet cells and basal cells, respectively [12, 13, 23] . Basal cells also show high affinity for GSI-B4, a plant lectin that specifically binds the glycoconjugates of basal cells of tracheal epithelia [12, 13] . Figure 2 shows that these three major cell types -ciliated cells, goblet cells and basal cells -were all present in the ATE population, and accounted for 22.8 ± 9.1%, 44.8 ± 9.7% and 41.6 ± 8.4% of the ATE cells, respectively. To determine whether ATE cells support the replication of IBV, two egg-adapted Taiwan IBV strains, 2575/98 and 2296/95, were used in this study. These two IBV viruses cannot infect primary chicken embryonic fibroblast cells or immortalized DF1 chicken fibroblast cells. The ATE cells (5 · 10 4 ) were infected with 50 lL 2575/98 (EID 50 = 10 5 /mL) for 1 h at 37°C. At 24 h.p.i., IBV proteins were detected in 2 500-3 000 cells by staining with chicken anti-serum against IBV (Fig. 3B) . We found that the IBV only infected N-cadherin + tracheal epithelial cells, revealing that neither smooth muscle cells nor fibroblast cells are the major target cells for IBV infection (Figs. 3D and 3E) . The infection rate gradually increased to 55.3 ± 12.8% of the total cells at 72 h.p.i. (Fig. 3C) , as estimated by immunocytostaining for IBV antigens in all adherent cells. Similar results were observed in 2296/95-infected cells (data not shown). In addition, both positive-and negative-sense viral RNA present were detected in the cytosol, and positive-sense viral RNA could be detected in the supernatant of ATE cells at 24 h.p.i. (Fig. 3F) , demonstrating that mature IBV virion could be produced and released from the cells. These results indicate that cultured ATE cells support viral infection, replication and spreading of IBV. It has been reported that IBV can infect both ciliated cells and goblet cells in the tracheal epithelium [1, 22] . However, whether basal cells are also susceptible to IBV infection remains obscure. We infected isolated intact tracheas with 50 lL 2575/98 (EID 50 = 10 5 /mL) for 48 h and found that IBV protein expression only colocalized with b-tubulin IV-or mucin 5AC-positive cells (Figs. 4A and 4B) . No viral protein was detected in the K14-positive basal cells of the tracheal epithelium (Figs. 4C and 4D). To further examine whether basal cells are resistant to IBV infection in vitro, 5 · 10 4 ATE cells were also infected with 50 lL of IBV 2575/98 strain (10 5 EID 50 /mL) for 48 h. In vitro results consistently showed that viral proteins of IBV were mainly detected in ciliated cells and goblet cells, but not in basal cells (Figs. 4E-4H). Similar cell tropism results were obtained when IBV 2296/95 or a higher dosage of viral loading was used (data not shown). These in vivo and in vitro experiments clearly delineate the cell tropism of IBV in the avian respiratory tract. Virus binding to host cell surface receptors is critical for determining tissue tropism and viral pathogenesis. It has been shown that H or HS may serve as a coreceptor in fibroblast cells for infection by the IBV Beaudette strain, but not the M41 strain [18] . In this study, highly sulfated GAG, such as H and DS, and common GAG, such as HS and CS, were tested to determine whether they possess different neutralizing effects on binding of Taiwan-isolated IBV to ATE cells. JEV was used as a positive control for the evaluation of the GAG effect [28] . Figure 5A shows that both H and DS at 7.5 lg/mL effectively blocked JEV infection in BHK-21 cells, as previously reported [28] . Complete inhibition of JEV infection was achieved when the concentrations of H and DS were elevated to 30 lg/mL. Although low concentrations of HS and CS did not affect JEV infection, a modest neutralizing effect was observed when the concentrations of HS and CS were elevated to 30 lg/mL. In contrast to the JEV results, even when all of the tested GAG were applied at 2.0 mg/mL, no significant inhibitory effects were observed on the infection of primary ATE cells by TW2575/98 (Fig. 5B ). In addition, the GAG did not alter the cell tropism of IBV on cilia cells or goblet cells (data not shown). These results strongly suggest that the binding affinity of GAG for IBV is too low to interfere with viral entry into tracheal epithelial cells. In this study, we showed that primary ATE cells exhibit the same cell composition as tracheal epithelia and can be passaged and amplified in a convenient and efficient way. These cells support IBV viral replication and viral release, providing an ideal system to amplify respiratory viruses and characterize their pathogenesis. In studies of tracheal infection, ciliated cells and goblet cells have been shown to be the major target cells of IBV [1, 22] . Our results with primary ATE cells also showed that IBV viral protein can be detected in both ciliated cells and goblet cells. This cell tropism may account for the pathogenesis of IBV, such as the ciliostasis in observed IBV-infected TOC and the reduction of sialic acid secretion [3, 22] . In addition, we are the first to demonstrate that IBV does not appear to infect K14-positive basal cells. In an uncomplicated IBV-infected chick, clinical signs such as gasping, coughing, tracheal rales and nasal discharge persist for only 5-to-7 days and disappear within 2 weeks [3, 22] . In the recovery stage, unaffected basal cells may be responsible for epithelial hyperplasia after desquamation of the ciliated and goblet cells and reconstruction of the epithelium of the injured respiratory tract to reestablish normal physiological functioning [12, 13, 22] . The insensitivity of basal cells to IBV infection also suggests that coinfection with other viral or bacterial pathogens is required to disrupt basal membrane integrity and cause hemorrhage in IBV-infected tracheas. At present, quantification of the virulence of IBV is still a problem [3, 22] . Measuring ciliostasis in TOC is not a sensitive approach for determining the severity of respiratory tract injury caused by IBV infection [21] . In addition, different species of chicken may also vary in the vulnerability of their respiratory tissue to IBV infection [2, 22] . In our ATE system, IBV shows high affinity for ciliated cells and goblet cells. Cell tropism, viral replication and viral spread can be determined and quantified by immunocytostaining, suggesting that ATE cells could serve as a quantitative platform to determine the pathogenesis of IBV. It has been shown that sialic acid and HS may help IBV bind to its receptor on the cell membrane of target cells [18, 32] . Receptor binding triggers endocytosis and delivers the viral genome into the cytosol through pHdependent membrane fusion [4] . In this study, we demonstrate that none of the tested GAG interfered with IBV binding to ATE cells. Further investigation is required to test whether this ineffectiveness is due to a lack of the XBBXBX H consensus sequence in the S protein sequences of TW IBV. In this study, we established a primary ATE cell culture system for the maintenance and amplification of epithelial cells of the chicken respiratory tract. Using these culture conditions, immortalized cell lines could potentially be established by overexpression of oncogenes, telomerase or anti-apoptotic genes. Both primary cells and respiratory cell lines will provide a new window to reveal viral life cycles, viral persistence, virus-cell interaction and the pathogenesis of avian respiratory viruses.

Using Dynamic Stochastic Modelling to Estimate Population Risk Factors in Infectious Disease: The Example of FIV in 15 Cat Populations

BACKGROUND: In natural cat populations, Feline Immunodeficiency Virus (FIV) is transmitted through bites between individuals. Factors such as the density of cats within the population or the sex-ratio can have potentially strong effects on the frequency of fight between individuals and hence appear as important population risk factors for FIV. METHODOLOGY/PRINCIPAL FINDINGS: To study such population risk factors, we present data on FIV prevalence in 15 cat populations in northeastern France. We investigate five key social factors of cat populations; the density of cats, the sex-ratio, the number of males and the mean age of males and females within the population. We overcome the problem of dependence in the infective status data using sexually-structured dynamic stochastic models. Only the age of males and females had an effect (p = 0.043 and p = 0.02, respectively) on the male-to-female transmission rate. Due to multiple tests, it is even likely that these effects are, in reality, not significant. Finally we show that, in our study area, the data can be explained by a very simple model that does not invoke any risk factor. CONCLUSION: Our conclusion is that, in host-parasite systems in general, fluctuations due to stochasticity in the transmission process are naturally very large and may alone explain a larger part of the variability in observed disease prevalence between populations than previously expected. Finally, we determined confidence intervals for the simple model parameters that can be used to further aid in management of the disease.

Feline Immunodeficiency Virus (FIV) infects numerous feline species worldwide [1] . This Lentivirus from Retroviridae family is closely related to Human Immunodeficiency Virus (HIV) and Simian Immunodeficiency Virus (SIV) [2] . This is a virus of major importance because it is lethal to the domestic cat (Felis silvestris catus) and can affect several other cat species, most of which are threatened or endangered e.g., the European wildcat F. s. silvestris in Europe [3] [4] [5] . There is thus a need to better understand the risk factors affecting the spread and patterns of persistence of FIV in natural populations of domestic cats. In natural domestic cat populations, FIV is mainly transmitted through bites arising from aggressive or sexual contacts [3, [6] [7] [8] [9] [10] . As a consequence, the spread of FIV in domestic cat populations is highly influenced by the mating system; a higher FIV prevalence is observed in aggressive and polygynous cat populations that involve more fights and bites than in much less aggressive and promiscuous urban ones [8, 9] , where FIV can be absent [11] . Basically, factors affecting cats'aggressiveness can be divided into two categories. At the individual level, some cats are more aggressive than others. Typically, this is the case for dominant males [8, 9] or orange cats [12] . In the field, they are generally more often infected than subordinates, females or other colour morphs [8, 12, 13] . At the population level, the overall aggressiveness of cats largely depends on the population social structure. A male-biased sex-ratio may make the entire population more aggressive, making virus transmission more efficient and, thus, lead to higher disease prevalence. Until now, to our knowledge, all studies of FIV risk factors have focused on individual risk factors. Factors that may increase the overall virus transmission rate are at least as important for controlling the disease spread but, paradoxically, have been largely overlooked until now. Here, we investigate how some characteristics of cat populations, such as cat density or sex-ratio, e.g., as indicators for aggressiveness in contacts within the population, may act as population risk factors that increase or decrease the virus prevalence within populations. Understanding the factors that may increase the FIV transmission rate within populations requires the sampling of a set of neighboring cat populations (which, until now, has rarely been done), and then examination of how FIV prevalence correlates with the suspected risk factors. For that purpose, we sampled 15 cat populations in North-Eastern France and measured, within each population, FIV prevalence in males and females. We found significant variability in disease prevalence between populations, especially in males. We also measured five social indicators in order to measure how they correlated with FIV prevalence. Commonly, risk factors are analyzed with logistic regression models. However, these models are built on the assumption that individuals become infected independently of each other; a hypothesis that contradicts the fundamental communicable nature of infectious diseases [1, 14, 15] . Moreover, as described below, assumptions of independence lead to underestimate the variability in disease prevalence between populations that would be observed in the absence of risk factors. Our method is inspired by previous works based on the comparison of stochastic dynamic models of the disease spread within host populations to the data [14, [16] [17] [18] [19] [20] . The idea is that each combination of population risk factors leads to a different model. Our objective is to determine the model (i.e. the combination of risk factors) that best fits the data. Beyond the simple analysis of the risk factors associated with FIV, this work aims to understand why the variability observed in our disease prevalence data is so large -data on disease prevalence in males exhibited significant extra-Binomial variations. Can we isolate population risk factors that would explain particularly high disease prevalence in some populations? Does the spatial aggregation of populations with high virus prevalence help to explain the variability in disease prevalence? Or, in contrast, is the large variability observed in disease prevalence a natural consequence of the transmissible nature of the virus? The work presented here supports this last hypothesis: random fluctuations in the transmission process lead to much greater variation in disease prevalence than with a simple Binomial distribution, underlying classical risk factor analyses. So the simplest model describes well the data and explains the large variability observed in disease prevalence between cat populations without invoking any risk factor. Finally, we determine confidence intervals for the model parameters. The model is very simple, explains the data well, and hence constitutes an interesting tool for further understanding and control of the spread of FIV in these cat populations. The approach developed here can easily extend to many host-parasite interactions. The field work has been made by qualified people according to current French legislation. Accreditation has been granted to the UMR-CNRS 5558 (accreditation number 692660703) for the program. Fifteen spatially separated rural cat populations were sampled during 2007 in North-Eastern France near the city of Nancy (Fig. 1, black rectangles) . The distance separating neighboring cat Figure 1 . The study area. We identified three metapopulations (grey areas). Studied cat populations are represented with black rectangles and solid arrows represent the suspected interactions between the studied populations. Some unstudied populations may interact with the studied ones (dashed arrows) and are represented by white rectangles. doi:10.1371/journal.pone.0007377.g001 populations varied from 1.2 to 4 km. The study zone covers a territory of approximately 250 km 2 . In order to delimit the study area, we considered the geographical characteristics that might limit movements between the studied cat populations and those outside of the studied area. The spatial organisation of geographical barriers suggests that the populations may be organized into three distinct metapopulations, with rare contacts between cats of different metapopulations (Fig. 1, grey areas) . At a finer scale, behavioral observations reveal that males can disperse between populations along roads. By adopting a basic assumption that populations are considered connected when they are not too distant from each other (i.e. less than 2 km) and are connected by roads, we propose a connection network between the different populations (see Fig. 1 , solid arrows). Unfortunately, it was not possible to establish a fully isolated perimeter (at least in relation to spread of FIV), so some populations of the study area are in fact connected to unstudied populations ( Fig. 1 , white rectangles -the connections to the study populations are represented by dashed arrows). In particular, Saulxure-Les-Vannes is connected to several study populations. However, it was not considered for sampling because of potential bias due to a cat culling programme there. The village of Sepvigny, which is connected to Champougny, could not be sampled for technical reasons. Most of the cats were captured using baited traps; others being caught directly in their owner's houses. Upon capture, cats were anaesthetized with an intramuscular injection of ketamin chlorhydrate (Immalgène 1000 15 mg/kg, Rhône-Mérieux) and acepromazin (Vétranquil 5.5% 0.5 mg/kg, Sanofi). They were marked permanently using an electronic passive integrative transponder (pig-tag) to allow all individuals to be identified in case of recapture. For each cat, we have recorded, among other data, information on sex, age and serological status in relation to FIV. Blood samples were taken from the jugular vein and the cats were then released. The ELISA method (SNAP Combo +, Idexx) was used to detect the presence of FIV-specific antibodies, which generally identifies virus carriers [6] . All positive sera for FIV were confirmed by Western blot analysis [21] . FIV was scored as present or absent for each sampled cat. 2.1 General approach. The approach we use here is very similar to the classical approach based on multifactorial logistic regression, which consists of: -Step 1: Choice of a model H 0 against which the data is compared. In the case of the classical logistic regression approach, it is assumed that all individuals have a same probability p to be infected, independently of the other individuals' status. As a result, under model H 0 the distribution of the number of infected cases in the population follows a binomial distribution of parameter p and N, where N is the number of individuals of the population. -Step 2: Some of the model parameters are expected to depend on risk factors. Choosing p as a function of risk factors means that each individual has its own probability of being infected (depending on its characteristics in terms of risk factors). It is classically assumed that the logit of p is a linear function of the different risk factors: logit(p) = a 0 +ga i X i , where X i denotes the i-th risk factor value for the individual. -Step 3: Model selection process. Different models are defined by setting some coefficients (a i ) to 0. Hence, the probability of being infected only depends on the risk factors which associated coefficients are non-zero. The different models are compared (usually using an Akaïke Information Criterium, AIC) to determine which model best describes the data. Note that there are two equivalent ways of presenting the classical approach. Firstly, the expected proportion of infected captured individuals is taken as a function of risk factors plus a random term based on a centered binomial distribution. Secondly, the probability that each captured individual is infected is taken as a function of risk factors, with random fluctuations in expected proportions naturally arising from these probabilities. Here, we present the second format because it allows us to easily illustrate how our approach is, in fact, a natural extension of the classical one. The main difference between our approach and the classical one comes from the model used to describe the data. It is quite obvious that for transmissible diseases the probability of one individual being infected is not independent of the infection status of the other individuals [1, 14, 15] . Here we consider the probability of individuals becoming infected as the result of a dynamic process of between-host virus transmission (described in the next section). These types of models are widely recognized as common tools for representing infectious disease data. We also make some minor changes to steps 2 and 3. In step 2, the logit function is chosen in the classical approach mainly because the model parameter p is bounded by 0 and 1. Since, as described below, our model parameters are not bounded by 1, we have no reason to consider their logit value. Lastly, in step 3 for model comparison we use likelihood ratio tests (LRT) rather than AIC. LRTs are chosen to test one particular assumption, which is here whether the simplest model, i.e. where no model parameter depends on risk factors, is sufficient to describe the data. 2.2. The dynamic epidemiological model -Model H 0 . The aim of this framework is to study population risk factors, i.e. factors that affect the rate at which the virus is transmitted within the population. Individual risk factors, i.e. factors that make some individuals more prone to infection than others in the same population, are not studied here. Our mathematical model extends the classical Susceptible-Infected (SI) model (Fig. 2) . We assume that all individuals of each population are equivalent, apart from their sex, the effect of sex on FIV transmission being too significant to be ignored. Indeed, of the 250 males captured in the study, 58 were seropositive (23.2%) compared to 22 of 249 (8.8%) females, which is highly significant (x 2 = 13.80, 1 df, p,10 24 ). Moreover, males and females play different roles in the transmission of FIV [8, 12] . Since females rarely bite, they can be considered as non-transmitting of the virus. The sexual structure of the model is simply represented by splitting classes S and I into two sub-classes, one for each sex. The age of individuals is not considered in our model, even though it may affect their behavior and, thus, their risk of becoming infected [3, 8, 13] . Moreover, due to long FIV infection duration, an accumulation of infected cases develops in older age cohorts. To represent the effects of age in a simplified way, we assume that the mean age of cats in the population may act as a risk factor for FIV transmission. This is justified since, here, we mainly focus on the global prevalence of FIV within populations without reference to the age-distribution of infections. We assume a proportionate mixing law for the incidence function of FIV between males, which is more appropriate in social species [22, 23] . Transmission between males of the same population occurs at a rate b M /M, where M is the total number of males in the population, and susceptible females are infected by infected males from their population at a rate b F /M. The constants b M and b F are proportional to the rate at which males are involved in fights and to the rate at which females mate, respectively. We assume constant numbers of males (M) and females (F) within each population, whereby dead cats are instantaneously replaced by newborn cats. Since vertical transmission is very unlikely in the field [7, [24] [25] [26] , all newborns are classified as susceptible to infection. Infected cats die at a rate a. Susceptible cats also die, but since they are instantaneously replaced by susceptible (and thus equivalent) newborn cats, their death is not explicitly modeled. For the sake of simplicity, we assume that populations are not explicitly connected, such that the numbers of infected cats in the different populations are independent random variables. To avoid the definitive extinction of the virus from the populations we assume regular infections from an external source, e.g. another population. Males and females are infected from external sources at a rate e M /M and e F /F, respectively. The rates e M and e F are termed the external transmission rates. The model is based on a continuous-time Markov process. Since we consider independent populations and constant numbers of males and females, the following set of (M+1)(F+1) ordinary differential equations describes the model (see [27] for an example of demonstration of differential equations representing continuoustime Markovian processes). For 0#m#M and 0#f#F we have: where p m, f (t) is the probability of having exactly m infected males and f infected females in the population at a time t (I M = m and In this model, the spread of FIV in males is independent of the number of infected females. As a result, the probability of finding exactly m infected males in the population (given by p m~P F f~0 p m, f ) is independent of the female transmission rates (b F and e F ) and hence of the proportion of infected females in the population. The distribution of the number of infected males given by the model can also be compared with male infection prevalence data, independently of female infection prevalence. We define this model as the ''male transmission model''. It is equivalent to a classical SI model [28] . Models H 1 . As discussed earlier, our purpose here is to measure the influence of some factors on the rate at which the virus spreads within or between populations. Two types of risk factors are tested here. The first ones concern the impact of demographic parameters (such as the number of cats within the population) on the virus transmission rate between cats of the same population. The second ones are not really risk factors. Behavioral observations suggest networks of connectivity between the different populations. The objective is to estimate whether introducing this information on the probability of disease reintroductions within populations produces significant predictive improvements, compared to models where external reintroduction rates are simply constants. Firstly, we try to improve the goodness-of-fit of the observed data by assuming that both within-population transmission rates b M and b F depend on the demographic characteristics of the cat population: where SR obs , N obs , M obs , AF obs and AM obs are the observed values for the sex-ratio, the population size, the number of males in the population (M obs = SR obs N obs ) and the mean age of captured males and females, respectively; considering these characteristics is intuitive since all of them may affect the social structure of the population and, hence, the transmission rates of are the (linear) parameters that quantify the effects of these five demographic characteristics on the transmission rates b M and b F . Note that, a priori, the coefficients can have negative values and hence predict negative transmission rates. We fix a minimum value (10 24 ) below which b M and b F cannot fall since negative transmission rates are not allowed in the model. For the sake of simplicity, we assume that the external transmission rates e M and e F are not affected by the risk factors presented above. We define as H 0 the model where b M~b 0 M and b F~b 0 F , the four model parameters (b 0 M , b 0 F , e M and e F ) being positive. As a general definition, models involving other parameters are called H(l), where l denotes the set of free (non-zero) parameters in the model that are notb 0 M , b 0 F , e M and e F . Then we investigate the possibility that, all other parameters being equal, the external transmission rates (e M and e F ) may differ between cat populations due to their spatial organization. Indeed, behavioral observations suggest a network of contacts between the different populations (see Fig. 1 , solid arrows), which can be simplified by dividing the study area into three distinct metapopulations (see Fig. 1 , grey areas). Since we do not model spatial structure explicitly, we assume that connectivity between populations affects external transmission rate. We define the resulting ''neighboring'' models and ''metapopulation'' models as follows. A potential neighboring network has been suggested by behavioral observations (see Fig. 1 ). Intuitively, when there is a high FIV prevalence in males in neighboring populations, the external transmission rate of FIV should be higher. For this reason, we propose that the external transmission rate of FIV within a population could be considered as an affine function of the number of infected males in the neighboring populations (I neigh obs ): We refer to this model as the ''neighboring model'' H neigh (l), where l denotes the set of free parameters in the model (in addition to b 0 M , b 0 F , e 0 M and e 0 F that are always freely variable). The metapopulation model considers that viral exchange is more intense between populations from the same metapopulation than between populations from different metapopulations. A simple way to test this hypothesis is to assume that populations belonging to the same metapopulation have the same external transmission rate, and that this external transmission rate differs between populations from different metapopulations. We define the ''male metapopulation model'' H M meta l ð Þ, where e i M represents the value of e M in metapopulation i, and l denotes the set of free parameters in the model (in addition to to b 0 M , b 0 F , e 1 M , e 2 M and e 3 M that are always freely variable). Note that in this model the only parameter that differs between cat populations is e M , which varies among metapopulations (e 1 M , e 2 M , e 3 M depending on the metapopulation). Finally, we also define the ''female metapopulation model'' H F meta l ð Þ, which is strictly equivalent to H M meta l ð Þ, except that it pertains to female external transmission rates. 2.4. Comparing models to data. Models cannot be directly compared with data because they predict distributions for the total number of infected and susceptible males and females in the population, whereas data are just samples of the real total numbers, i.e. the probability of capture is strictly below 1. To simplify, we assume that the total number of males and females in the populations are proportional to their observed values, i.e. M = M capt (1+p NC ) and F = F capt (1+p NC ), where p NC is a constant (1/(1+p NC ) is the proportion of captured cats) and M and F are the real numbers of males and females within the population, respectively. Based on the ratio between the number of cats captured through baited traps and the number of cats observed through intense monitoring in each population, we estimate that p NC is equal to 0.3 in average. We assume that FIV is present in this area for a long period of time, corresponding to the stationary state of the distribution. So data are compared to this state. Note that the fact that the distribution is stationary does not mean that the population is at the equilibrium (i.e. endemic state), but only that epidemic, endemic and extinction events may succeed, and this being considered a population has a time-independent probability of being in each of its possible states. Stationary distributions of the model, i.e. probabilities of finding exactly m infected males (for all 0ƒmƒM) and f infected females (for all 0ƒf ƒF ) in the population, generate a distribution of possible outcomes d 0 for the total number of cats. To incorporate the fact that data are missing for non-captured individuals, we add a hypergeometric sampling element to the distribution d 0 (in other words data are the result of a random sampling of the entire population). This leads to the distribution d to which data can be compared [29] : where H x,y,z is the hyper-geometric law of integer parameters x, y and z, which is defined when max(y,z) #x for all integers t satisfying t #min(x,y) and z2t #x2y by: The distribution d is then equal to the distribution d 0 after sampling a proportion 1/(1+p NC ) of the population. In other words, d 0 is the asymptotic distribution of the number of infected males and females, after sampling a proportion 1/(1+p NC ) of the population. 2.5. Model selection. Each of the models presented above can be summarized by the set of parameters that may vary freelyother parameters being fixed. Let us consider a model H. For each value h of the free parameters in the freely variable parameter space H (h is a vector of the values of all the free parameters), we can calculate for each population k the probability of generating the number of infected males and females actually observed. We call it L k (h|D k ), where D k represents the data restricted to population k; D k is defined by the number of infected males and females in population k. Since we assumed that populations are independent, we can easily calculate the likelihood of the data D with the model H: Now, if we consider two models H 1 and H 2 , the two models are compared using the maximum likelihood ratio statistics defined by: We use the classical approximation that, under regularity conditions, the likelihood ratio follows a x-square distribution with r degrees of freedom, where r is the difference in the number of free parameters between models H 2 and H 1 . 2.6. Determining confidence intervals for the parameters. Another important objective of mathematical modeling is to calibrate the selected model, i.e. the model selected from the previously described process (see Section 2.4 above) by determining confidence intervals for its parameters. We consider a model H with a given set of freely varying parameters that defines a vector (h); i.e. each component of h is a parameter of the model. Within each population, the model predicts a distribution for the number of infected cats (male or female). Each possible model outcome (defined as a vector of 30 integers representing the number of infected males and females within each of the 15 populations) has a probability of occurrence. What we want to determine is the values h of the free parameters for which the observed data is a plausible outcome of the model. We accept that the data is a plausible outcome of the model when its likelihood is within the range of likelihood values of typical model outcomes, as described below. For each vector h, we determine the threshold L 0.05 (h) such that 95% of the model outcomes have a likelihood value larger than L 0.05 (h). We now look at the likelihood of the observed data under the model parameters, defined above as L(h|D). Again we explore the parameter space. The confidence region H C can be defined as H C = {hMH / L(h|D).L 0.05 (h)}. Thus, for H C the observed data is a likely model outcome. Since the model often has several free parameters, then the 95% confidence interval is, in fact, a region of the multi-dimensional parameter space of the free parameters (H). For that reason we use the term ''confidence region'' rather than ''confidence interval''. Finally, note that in the models the only parameter value we fix a priori is the mortality rate of FIV infected individuals (a). Since the model is analyzed at equilibrium, changing the mortality rate of infected individuals only results in a change in time scale. To remain consistent with cat-FIV interaction characteristics, we fix a = 0.0208 month 21 , so that infected cats have a 4-year life expectancy [8] . The model time unit is the month. 2.7. Computational procedure. Computationnal procedures are performed with Matlab. Stationary distributions of FIV prevalence in males and females are obtained by resolving the linear system corresponding to dp m,f /dt = 0. Maximum of the likelihood function are computed using a conjugate gradient method. The cat number, sex-ratio, mean age of the males and females and percentage of FIV positive males and females captured in each population are given in Table 1 . A total of 499 cats were sampled and tested for FIV in the 15 populations. There was large variability in the number of cats sampled due to large differences in population sizes, ranging from 13 cats in Clerey-la-Côte to 71 in Sauvigny. The overall sex-ratio is close to 50% but with differences between populations, although it does not differ statistically from a 50:50 binomial distribution (x 2 = 17.21, 15 df, p = 0.31). However, in Ruppes the sex-ratio is rather high (0.79) and this value significantly differs from 0.5 when applying a Bonferroni correction for multiple tests (p,0.05). For each captured cat, we estimated its age following Pascal and Castanet [30] , and then the mean age of males and females in each population. For the entire study area the mean age is 3.08 years for males and 3.55 years for females; ranging from 1.54 years in Champougny to 5.25 years in Jubainville for males and from 2.32 years in Ruppes to 5.60 years in Barisey-la-Côte for females. It is also interesting to note that there is a strong correlation between the mean age of males and females in the studied populations (r = 0.85). Finally, as previously documented, the global prevalence of FIV differs greatly between sexes (23% in males compared to 9% in females), with an average FIV prevalence in the entire study area of approximately 16%. There is significant variability in FIV prevalence between populations, especially in males, where data show significant extra-Binomial dispersion (Fisher's exact test with Table 1 . Total number of sampled cats, adult sex-ratio, number of FIV seropositive individuals (FIV+) and mean age of captured males and females in each population. Here we perform a rapid analysis of the mathematical model, this type of model having been studied in more detail elsewhere [28, 31] . For the sake of simplicity, we focus on the real distribution of FIV prevalence in males (the results are thus independent of b F , F and e F ); we assume p NC = 0, i.e. all individuals of the population have been sampled. First, we look at the distribution of FIV prevalence in males for arbitrarily fixed values of the parameters: b M = 0.025, M = 50 and e M = 0.01 (Fig. 3a, solid line) . For clarity, we plot the distribution of FIV prevalence as a continuous line, although the distribution is discrete. The probability of finding no infected cats in the population is high (17%). The mean prevalence of FIV is 12.66% and in 95% of the model outcomes the FIV prevalence ranges between 0 and 32%. This predicted distribution of FIV prevalence in males differs from a binomial one (a distribution frequently used in risk factor analysis, [8, 9] ) having the same mean (Fig. 3a , dashed line). For a binomial distribution, the probability of finding no infected individuals in the population is much smaller (0.1%) and the confidence interval for FIV prevalence is [0.01; 0.20]. In Fig. 3b we analyze the effect of the external transmission rate on the mean and standard deviation of FIV prevalence in males. We focus on the distribution conditioned to non-extinction and, in parallel, we plot the probability of FIV extinction from the population (dashed line, right axis). Unsurprisingly, the probability of FIV extinction decreases with increasing external transmission rate (e M ). More interestingly, below a given threshold (here e M = 10 23 ) the distribution of FIV prevalence is not affected by e M , meaning that infrequent infections of FIV from external sources have almost no effect on FIV transmission within already infected populations. Under these circumstances, external infections only affect the frequency of extinction of the virus. Above the threshold, the mean prevalence of FIV increases with e M . Thus, external infections are an important component of FIV transmission, even within already infected populations. This result may have important implications. For example, in our data only two of the 15 populations have no infected males. This indicates that the external transmission rates of FIV within our populations must be large enough such that there are infected males in at least 13 out of the 15 populations. Under such external transmission rates, is the spread of FIV within already infected populations affected by external infections or is external infection only important for the long-term persistence of the virus? This question will be addressed later when we provide estimates for the parameters. 3. Analysis of the observed data using the dynamic model 3.1. Effect of demographic risk factors. Now we consider the full model, including both males and females, and compare how integrating the different risk factors increases the goodness-offit to our observations using likelihood ratio tests. We performed the data analysis with each of the following values for the proportion of non-captured cats (p NC ): 0.15, 0.30 and 0.45. Since the results obtained from these three values are very similar, we only show those obtained for p NC = 0.30. It is important to note that from now on, likelihoods are calculated with the distributions of FIV prevalence in males and females, without removing the cases of extinction (i.e. we use the distribution d). The probability of observing zero infected individuals is an important characteristic of the models, and removing extinction cases would lead to lose very important information, especially relating to the external virus transmission rate. To start with, we look at the impact of the population characteristics (the sex-ratio in captured cats SR obs ; the estimated population size, i.e. the number of captured cats N obs ; the number of captured males in the cat population M obs = SR obs N obs ; and the mean age of captured males and females, AM obs and AF obs , respectively). Results are summarized in Table 2 . The only significant effect we found is associated with the effect of mean age of males (x 2 = 4.09, 1 df, p = 0.043) and females (x 2 = 5.335, 1 df, p = 0.02) on the male-tofemale transmission rate (b AM F and b AF F , respectively). Interestingly, there is a negative correlation between the mean ages of males and females and FIV prevalence (b AF ). This means that the effect of age on FIV prevalence observed here is not due to the accumulation of FIV cases with age. The p values are rather large (p = 0.02 and p = 0.043), especially considering the large number of tests performed. Unfortunately, we cannot apply a simple Bonferroni correction for multiple tests because of dependence among the different tests performed. Due to the strong correlation between the mean age of males and females, it is not surprising that both variables have the same significant effect on the male-to-female transmission rate. It seems more likely that only one of the two variables has a real biological effect, the effect of the other one being due to correlation. Due to the strong correlation between the two variables, we cannot rule out a role for mean age of males on the male-to-female transmission rate. H neigh e neigh F , respectively). Under these circumstances, we find no significant improvement in the models compared to H 0 (see Table 2 ). In summary, we found two potential risk factors for FIV: the mean ages of males and females that influence the FIV prevalence in females. These two factors are certainly linked because a large correlation exists between the two variables. Yet, considering the number of tests we performed and the relatively high p values we obtained, we cannot exclude the possibility that the simplest model We remove female prevalence data from the analysis because there is no female to male transmission of the virus. We determine the confidence region of the transmission rates b M and e M, of the ''male transmission model'' parameter space for three different values of the proportion of non-captured cats: p NC = 0.30 (Fig. 4a, b) , p NC = 0.15 and p NC = 0.45 (Fig. 4b) . In Fig. 4b we superimpose these three confidence regions; only showing their boundaries. We conclude that p NC has a slight impact on the edge of the confidence region. If we project the region onto the b M 0 axis we obtain a 95% confidence interval for the male-to-male transmission rate b M ([1. Finally we look at the impact of the external transmission rate e M on FIV spread in already infected populations. We estimate the average size of a population as the mean number of observed males per population multiplied by 1+p NC , which is equal to 21 for p NC = 0.30. We divide the mean number of infected hosts calculated with the model for a population of average size conditioned to FIV non-extinction by the value obtained with the same parameters, but with an external transmission rate a hundred times lower. We denote R as this value minus one , where I is the mean number of infected individuals in the population. R is a proxy of the impact of external infections on FIV transmission in already infected populations. If external infections have a small effect compared to the withinpopulation transmissions, then R will be close to 0 (see Fig. 3c ). In contrast, if external infections have an important effect compared to the within-population transmissions, then R will be quite larger than 0. For p NC = 0.30 we calculate R in a square region of the male transmission rate (b M and e M ) parameter space (Fig. 4) and we superimpose on the same graph the edge of the confidence region. We observe that in the upper left corner of the parameter space (Fig. 4c) R is around 0.02, which means that at low external transmission rates, external infections only increase by 2% the prevalence of FIV and, so, have a very limited impact on the spread of FIV in already infected populations. In contrast, in the lower right corner of the confidence region R is around 2.5, which means that frequent external infections greatly increase FIV prevalence even in already infected populations. prevalence in females. Now we investigate for which parameters values in the model including both males and females data are a plausible outcome of the model. We focus on the simplest model. First, it is interesting to know if, and for which set of parameters, the simplest model can fit the data. Second, since the effect of the mean age of populations is not highly significant, we do not believe it makes biological sense to take this factor into account here. Here, the parameter space is four-dimensional, so we cannot plot the confidence region. Since we are interested in determining the parameters directly influencing FIV prevalence in females, we simply plot the projection of the confidence region in the female transmission rate (b F , e F ) parameter space (Fig. 5) . 5 shows that there is an important dependency between b F and e F . Increasing the value of e F increases the mean prevalence in females and so the parameter b F must be decreased in order to explain the observed data. As a first approximation, the confidence region can be characterized by the relationship 0:015ƒb F z3:46e F ƒ0:11. Interestingly, the confidence region crosses the X and Y axis (see Fig. 5 ). This means that even if one of the two rates (b F or e F ) equals zero, then the model can still explain the data. In other words, the data may be explained by considering only infection of females by males of the same population, without external infections or, conversely, by only considering infections by males from other populations without within-population male-to-female infections. Overall, we cannot determine which source of infection for females (internal or external) is the most important in our study area. In the previous section we have seen that data are a plausible outcome of the simple model for a large region of the parameters. In the present section we show how the simple male-transmission model (where the transmission rate is independent of risk factors) (Fig. 6a) . For comparison we show the same graph using a binomial model (assuming independence between individuals regarding FIV, Fig. 6b) . As seen previously, the dynamic model predicts a very large variability of FIV prevalence in males within population (see Fig. 6a ), which is larger than with the binomial model (see Fig. 6b ). As a result observed FIV prevalence in males is always in the 95% confident region for the dynamic model (see Fig. 6a ), but not for the binomial model (see Fig. 6b ). In Fig. 6c we show the variance predicted by the different models (with maximum likelihood estimations of their parameters) and we compare it with that estimated from the data (usinĝ 14, where p i is the prevalence of FIV in males in population i and p p is the mean FIV prevalence). Again results show that the binomial model predicts a smaller variance than what is observed in the field (FIV prevalence data in males show around 73% more variance than what is expected by the binomial model), whereas the dynamic model shows an overestimated variance compared to what is estimated from data (but only around 42% larger). To investigate whether the variance estimated from data is consistent with predictions of the dynamic model, we study the distribution of FIV prevalence in males expected by the male transmission model with maximum likelihood estimation of the parameter. In each population we simulate a male FIV prevalence according to the distribution d and then we estimate the variance in FIV prevalence in males between the 15 populations. We run 10,000 replicates and obtain a theoretical distribution of the estimated variance of FIV in males (ŝ s 2 , Fig. 6d ). We find that the observed value ofŝ s 2 (black bar) is in fact a plausible outcome of with the dynamic model. The spread of a transmissible disease in a host population is a dynamic process where the probability of individuals becoming infected depends on the number of infected individuals in their neighborhood. Nowadays, dynamic models of epidemics are widely accepted as efficient tools to help understand the spread and management of infectious diseases (see e.g. [32] [33] [34] [35] ). So it is not surprising that stochastic versions of these models have emerged during the past decade as the best way to analyze infectious diseases data (see e.g. [1, [14] [15] [16] [17] [18] [19] 20, 36, 37] ). Methods based on the comparison of stochastic epidemic models to data hence constitute natural tools to estimate how different factors may affect the spread and impact of infectious diseases. Our dataset exhibits large variability in FIV prevalence in both males and females among populations. However, a rapid study of the dynamic model shows that, in such a model, great variability in FIV prevalence may be expected. The rate at which susceptible individuals become infected depends on the proportion of infected individuals in a population. If, by chance, the proportion of infected individuals becomes large then the number of new infections will increase, maintaining high infection prevalence for the next generation. By contrast, a low proportion of infected individuals decreases the number of infections in subsequent generations. To investigate the possibility that the cats density, the sex-ratio, the number of males or the mean age of cats within the population may act as risk factors influencing the disease transmission rate, we performed a statistical analysis of the data using the sexuallystructured SI model. Which population characteristics correlate with large FIV prevalence and so explain, in part, the variability in FIV prevalence? We found no such factors, except for mean ages. Interestingly, these ages have a negative effect on FIV prevalence in females, despite the accumulation of cases that occurs with age. One possible explanation is that the presence of older territorial males in some populations ensures greater social stability, which decreases the rate of at-risk (mating) contacts. Reversely, a negative correlation between FIV prevalence and age of cats could be due to the additional mortality induced by the virus. However, considering the weak impact of the infection on the lifeexpectancy of individuals, this explanation seems rather implausible to us. In fact, these effects are not highly significant (p = 0.02 for the mean age of females and p = 0.043 for the mean age of males). Determining whether or not age affects the probability of becoming infected by FIV would require i) correction for the multiple tests performed and ii) correction for the effect of the accumulation of cases with age. Since this is beyond the scope of the work presented here, we cannot make definite conclusions on the effect of age. The cat populations observed in this study are of small size, and certainly are not large enough to retain the virus over long periods of time. Since we detected infected cats in 14 out of the 15 populations (and infected males in 13 of them), we can assume regular viral exchange between populations. Previous theoretical studies have shown the importance of the spatial dispersal of the FIV virus between populations [38] . Due to the topographic isolation of our study area, it seems reasonable to assert that viral exchange between the studied populations is primarily responsible for the reintroduction of the virus into populations where it has become extinct. We proposed two different virus dispersal networks between the populations, but neither significantly improved the goodness-of-fit to the observed data. Although our observations are most likely insufficient to capture the exact dispersal network between populations, the networks we analyzed should be quite realistic, because they are consistent with the natural barriers in the study area. Lastly, it is important to note that a spatial correlation in FIV prevalence between connected populations can be observed only if external infections have a substantial impact on FIV prevalence within the population. An analysis of the confidence region of the male transmission parameters shows that the impact of external infections on FIV prevalence within populations is very limited for the smallest values of the external transmission rate (see Fig. 4c ). In this case, the connectivity between populations cannot be revealed by a corresponding correlation in FIV prevalence. In contrast, for the highest values of external transmission rate in the confidence region, we can expect a correlation in FIV prevalence between connected populations. To sum up, the fact that no spatial correlation in FIV prevalence is observed may simply be due to the fact that external infections are relatively rare and thus play almost no role for disease prevalence in already infected populations. Logistic regression models are still widely used for the analysis of risk factors associated with infectious diseases, even though their over-simplified independence hypothesis is largely recognised as a limitation to their use [1, 14, 15] . The main difference between the two approaches, based on binomial and dynamic models, comes from the variability expected by their respective H 0 models, as illustrated in Fig. 3a . Binomial models predict much narrower distributions than dynamic models. The consequence is illustrated in Fig. 4 , where we can see that the simple SI model accounts for the observed variability in FIV prevalence in males for a wide range of parameters. In contrast, the binomial test on the distribution of the infected cats among the 15 populations rejects the global binomial distribution hypothesis (p<0.006). To explain the data with a logistic regression model that assumes binomial distributions, additional risk factors need to be invoked. With dynamic models, risk factors are not required to explain the variability in the male disease prevalence observed here. The implication of that is that bringing evidence for population risk factors in infectious disease requires large sample sizes. In our present case n = 499 is not large enough and further sampling is required to bring evidence of population risk factors for FIV transmission. The model developed here is quite simple. In particular, it does not account for a potential difference in individuals' infectivity between the acute and chronic phase of the infection. Such levels of complexity could be added to the method. This would make the model more realistic but also more complex, which was not our purpose here. The most important conclusion of the paper, i.e. that dynamic models predict much more variability than models where individuals are independent and hence are sufficient to explain highly variable prevalence data, would remain true for more complex model. Another model assumption is that we neglected the contacts with populations outside the study area (white rectangles in Fig. 1 ). Since we did not find an important effect of the number of infected neighbors on the disease prevalence in populations, we are confident that adding the neglected populations would not deeply affect our results. The approach developed here is general and can easily extend to a wide variety of cat populations, but also to other host-parasite systems. It facilitates selection of the best model to describe data, which can be calibrated by determining confidence regions for the model parameters. The model can be used, for example, to test virtual management plans and to look at the expected results in the entire confidence region. This should assist in predicting the success one might expect with different management strategies. In the case of FIV, this study could help to rationalize the use of potential future vaccines or castration campaigns to limit the spread of the virus between males. In the case of FIV, the approach gives us a 95% confidence interval for the model parameters, in particular for the basic reproductive number R 0 ([0.626, 1.942]), with a maximum likelihood estimate of 1.285. This value appears rather low, meaning that virus transmission is rather rare at the level of the population. This is not surprising, since experimental results indicate that most of the virus present in the saliva is not infectious [39] , suggesting a weak efficiency in disease transmission [7] . Given the high frequency of fights between males in such populations and the low rate at which males acquire the infection (around once every four years in a hypothetical scenario where all males are infected), our results are consistent with the concept of a low probability of virus transmission from bites [9] .

Multiplex primer prediction software for divergent targets

We describe a Multiplex Primer Prediction (MPP) algorithm to build multiplex compatible primer sets to amplify all members of large, diverse and unalignable sets of target sequences. The MPP algorithm is scalable to larger target sets than other available software, and it does not require a multiple sequence alignment. We applied it to questions in viral detection, and demonstrated that there are no universally conserved priming sequences among viruses and that it could require an unfeasibly large number of primers (∼3700 18-mers or ∼2000 10-mers) to generate amplicons from all sequenced viruses. We then designed primer sets separately for each viral family, and for several diverse species such as foot-and-mouth disease virus (FMDV), hemagglutinin (HA) and neuraminidase (NA) segments of influenza A virus, Norwalk virus, and HIV-1. We empirically demonstrated the application of the software with a multiplex set of 16 short (10 nt) primers designed to amplify the Poxviridae family to produce a specific amplicon from vaccinia virus.

Researchers employ numerous approaches for viral detection and discovery, including metagenomic sequencing (1) (2) (3) , microarrays (4) (5) (6) (7) (8) (9) or multiplex PCR followed by other methods of characterization such as mass spectrometry (10) (11) (12) (13) (14) , suspension arrays (15, 16) or amplicon sequencing (17) . Multiplex PCR followed by analysis of amplified products fills a niche for viral identification when singleplex PCR has failed or there are a few dozen likely candidates but the expense of metagenomic sequencing or high-density microarrays is unwarranted (18) . However, multiplex primer design for many highly divergent targets is challenging since no universally conserved primers may exist, and finding sets of primers likely to function well in multiplex (e.g. isothermal T m 's, no primer dimers) adds to the complexity of finding conserved primer candidates. Primer design software that requires a multiple sequence alignment (MSA) as input can be problematic for diverse target sets, as MSAs can be difficult to construct, exhausting memory or available time before an alignment is completed. Even if an alignment does complete for divergent target sets such as all members of a family of RNA viruses or gene homologues across species, alignments may show little nucleotide sequence conservation, and multiple primers are required to amplify all targets. Considering the challenges of primer design for targets showing sequence variation, it is not surprising that many of the PCR-based assays in the literature are predicted to fail to detect desired targets when compared against available sequence data, and this problem is worst at higher taxonomic levels like family (19) . Most currently available multiplex primer prediction tools require an MSA (20) (21) (22) (23) (24) . None of those tools build multiplex sets in which no primers in the set are predicted to form primer-dimers, although some avoid homodimers. We attempted to run a number of alternative software tools for multiplex or degenerate primer prediction for several species level target sets ranging in size from 41 to $6000 sequences: Primaclade (24) , FastPCR (www .biocenter.helsinki.fi/bi/programs/fastpcr.htm), GeneUp (25) , PDA-MS/UniQ software (26) , Greene SCPrimer (20) and HYDEN (22) . Only Greene SCPrimer and HYDEN could handle the two smallest, species-level target sets (Norwalk virus and FMDV), and none of the tools completed for the larger target sets that we attempted to run. Therefore, we designed the MPP software to avoid the requirement of an MSA and to scale better for large and diverse target sets. MPP builds a multiplex-compatible set of primers capable of amplifying all target sequences, attempting to minimize the number primers in the set. We set out to determine a set of highly conserved 'universal' primers for viruses, akin to the highly conserved 16S rRNA universal primers for bacteria. Throughout, we use the term 'detected' to mean that there should be at least one PCR product. This is a loose definition of 'detected' adopted for convenience in the following discussions, and we recognize that a PCR product may prove insufficient for viral characterization. We predicted a set of 'universal viral primers' for all available complete genomes of all viruses, and found that the number of universal viral primers would be impractical to implement, even if short, highly conserved priming sequences were used. Then we predicted family-level primer sets for every viral family, as well as for several highly diverse species of RNA viruses, for primers of a traditional length as well as nontraditional, shorter, more highly conserved primers, which are more likely to amplify novel, unsequenced viruses and which could be an alternative to degenerate primers. While the software uses a greedy algorithm that may settle at a local minimum which is above the true minimal set, it generated primer sets for each viral family with fewer than half the number of primers that would be expected without optimization. Although family-level primer sets for some families are too large to be practical, 64% of the families had primer sets of no more than 20 primer pairs, quite feasible for multiplex PCR. Finally, we empirically multiplexed a set of 16 short (10-mer) primers designed for the Poxviridae family. This demonstrated specific amplification of the expected viral fragment from vaccinia Lister strain. This preliminary demonstration suggests that specific amplification with family-level multiplexed sets of short primers might be feasible for viral detection and discovery, particularly as a means for selective enrichment of viral target for downstream amplicon characterization (e.g. sequencing or probe hybridization). Here, we outline a greedy algorithm used to calculate conserved sets of multiplexed primers to amplify fragments from each member of a target set of sequences, and provide a more detailed description in the Supplementary Methods. First, we enumerate all candidate oligos fitting user-specified requirements for length, T m , and lack of hairpin formation. We rank pairs of these by the number of targets in which that pair occurs within a distance range d 1 bases to d 2 bases of one another, where these might be specified so as to bound a reasonable range for a downstream characterization method such electrophoretic discrimination of bands, probe hybridization or sequencing. The most frequent pair is selected as primers. The process is repeated for the remaining targets that would not have an amplicon from the first pair, with the added consideration that new primers selected be predicted not to form dimers with other primers already selected, as well as can be predicted based on nearest neighbor thermodynamic predictions (27) , although free energy calculations cannot predict with certainty that primer dimers will be excluded in practice. There is an option to bin primers into reaction subsets, if desired. With binning, primers are added to a bin until that bin contains b primers, at which point a new bin is begun, following the same process. Binning primers in smaller groups avoids exclusion of the most highly conserved oligos because of primer dimer free energy constraints. The universal set of primers is the set of selected primers to amplify all genomes in the target set. The primers within a bin should be multiplexed into a single PCR reaction, but each bin should be run separately. This is a simple binning strategy, and alternative strategies could be employed such as starting a new bin with any primer pair that dimerizes with other previously selected primers regardless of the number of primers in the bin, but this could have the possible disadvantage of bin explosion to numerous singleplex or small multiplex reactions. The graph-based algorithm of MuPlex (28, 29) is another binning strategy that could be incorporated. Output on the number of primers rejected due to the various filters (T m , hairpin free energy, etc.) is printed to standard output after each round of primer pairs are selected and also cumulatively, so a user can monitor if/which parameters might be too stringent. If an alternative set of primers is desired that does not overlap with the set selected, one can replace with N's the subsequences matching the selected primers and their reverse complements in the target sequences and rerun the software with the modified input sequences. The script find_amplicons.pl predicts all the amplicons that should be generated by a list of (multiplexed) primers mixed with a set of sequences. This PCR-simulating script lists amplicon sequences, their length and position, and the forward+reverse primer combination to yield each product, as well as a summary file of the fragment length distributions enabling a quick assessment of how well each target sequence can be discriminated based purely on amplicon length patterns. We used this script to check whether some viral primer sets are predicted to generate amplicons from the human genome. For all T m and free energy predictions, we used the following Unafold settings: [Na+] = 0.2 M, [Mg+] = 0.0015, annealing temperature = 30 C, and strand concentration of each strand = 10 -7 M, making the total strand concentration of both strands = 2 Â 10 -7 . The MPP algorithm described here focuses on finding conserved primers, and does not require that the primers be family-or species-specific. For the runs predicting family-specific primers, we designed viral family primers that were unique relative to nontarget viral families by replacing any substring of 17 nt or longer that occurs in any non-target family with a substring of 'N''s, using the suffix array software vmatch (http://www.vmatch.de/). MPP eliminates oligos with N's from consideration as primers. This approach is simple, although it risks being overly strict by eliminating some potentially successful candidate primers, since it disallows those cases where a single nonunique primer pairs with a unique primer, a pair of nonunique primers are too far apart to actually generate an amplicon in a nontarget sequence, or candidate oligos partially overlap a nonunique stretch of N's. In general, searching for primers from sequence that is specific to one set of targets and excluding candidate substrings present in nontarget sequence is a useful strategy to design signatures for pathogen detection. This can be done by replacing nonunique or nonspecific substrings with N's in the sequences input to MPP using software such as vmatch. For runs predicting universal viral primers including all viruses at once in the target set, all viral complete genomes and segments downloaded from publicly available sequence databases (Genbank, Baylor, TIGR) as of 25 April 2007 were used. Draft sequences with multiple contigs were merged into a single sequence entry, with contigs separated by 1000 N's, a stretch sufficiently long (>d 2 ) so that primer pairs would not be designed to fall on different contigs, although there were very few draft sequences in contigs where this was necessary. Because of the large numbers of sequences in two families, only the MP segment sequences from Orthomyxoviridae and the L segment from Bunyaviridae were included, as these are the more conserved segments, reducing the number of targets by 23 017 sequences. The total number of target sequences for these all virus runs were 11 477. We predicted a multiplex set of universal 10-mers without binning the primers into separate reactions, but it was necessary to use a very low x dimer and x homodimer of À11 kcal/mol to be possible to predict multiplexed primers to amplify every target. For the next calculations, we subdivided the primers into sub-reactions with 20 primers per reaction bin, to avoid excluding primers that were predicted to form primer dimers, and raised x dimer and x homodimer to -7 kcal/mol. Primer sets of length 5-18 were predicted ( Figure 1 ). To assess the effect of removing T m constraints, we generated universal primer sets with length but no T m requirements, and compared the primer counts to those with T m constraints ( Figure 2 ). Primer sets are available by contacting the authors. We evaluated how the growth of sequence availability could affect the size of the universal set of viral primers, as well as our ability to detect unknown/unsequenced viruses. We predicted a universal viral primer set for all viral genomes and segments available as of 1 January 2004, totaling 9965 sequences, for primers of length 7-15 nt ( Figure 2 ). It was not necessary to exclude any Orthomyxoviridae or Bunyaviridae segments, because these segments were not so deeply sequenced at that time. We then determined how much of the 2007 sequence data would have been detectable using the 2004 primer sets (Figure 3 ). To predict how contamination with human DNA might affect the ability to detect specific amplification of viruses, the average number of amplicons for the human genome was also predicted based on these primer sets with 20 primers per bin ( Figure 4 ). The effect of reducing the multiplex size to 10 primers per reaction on human genome amplification was calculated by dividing priming sets in half, with 10 primers per reaction. Subdividing viral targets into families, we used an updated set of sequences, downloaded 5 February 2009. For familylevel primers, we computed primer sets using three alternative parameter settings (Table 1) : (i) 17-21-mer primers with T m = 55-60 C, primers not checked for uniqueness; (ii) same length and T m specifications as in (i) but eliminating from consideration as primers any oligos of at least 17 nt that occur in nontarget viral families, as described earlier; and (iii) 10-15-mer primers with T m = 40-45 C. These primers are available as Supplementary Data. We also generated primer sets for several species with high sequence diversity: HIV-1, FMDV, Norwalk virus and influenza A segments HA and NA (summary in Tables 2 and 3, primer sequences available as Supplementary Data). MPP parameters were the same as in Table S1 , with T m = 35-50 C for 10-mer primers and T m = 55-70 C for 17-18-mer primers, and sequences were downloaded in 2007 or 2008 (influenza only). To illustrate the challenges of designing primers from an alignment, we aligned these organisms using MUSCLE (30) when possible. For the HIV-1 and influenza A segments HA and NA, MUSCLE ran out of memory before completing. An alternative alignment tool, ClustalW (31), had completed only a small fraction of the alignment after running for days. Therefore, for these large data sets, a random selection of $35 sequences for each target was aligned with MUSCLE, and this alignment was used to build a profile Hidden Markov Model (HMM) (hmmbuild) using HMMer (http://hmmer.wustl.edu/). (32) The full sequence set was then aligned to the HMM using hmmalign. For Norwalk virus and FMDV, we designed multiplex-degenerate primer sets using GreeneSCPrimer (20) and HYDEN (22) , but the other three targets sets were too large for those programs. None of the other primer prediction programs we tried (as indicated in the 'Results' section) would scale to handle even these two smallest target sets. For an empirical demonstration of our algorithm, we tested a multiplex set of 16 short, 10-nt primers designed for the Poxviridae viral family against commercially purified vaccinia virus extracts. The primer sequences are provided in Table 4 along with the predicted single amplicon for vaccinia Lister strain. We chose the Poxviridae family multiplex for these first empirical demonstrations because extracted viral nucleic acid was readily available for experiments. All primers were purchased from Integrated DNA Technologies (Coralville, IA, USA) and resuspended to 100 mM stock solutions in TE buffer (pH 8.0, Teknova, Hollister, CA, USA). Working solutions containing equimolar concentrations of each of the 16 primers were used in all experiments. Purified, quantitated vaccinia Lister strain DNA was purchased at a concentration of 1.3 Â 10 4 copies/ml in nuclease-free water from Advanced Biotechnologies, Inc. (Columbia, MD, USA). All PCR experiments were prepared using the Superscript III RT-PCR kit from Invitrogen (Carlsbad, CA, USA). We selected the RT-PCR kit in order to establish a protocol that could later be readily applied to additional multiplex viral family reactions with viral DNA and/or RNA. Each 25 ml reaction contained 1 Â SSIII buffer, 1 U of SSIII RT/ Taq enzyme, 4.8 mM MgSO 4 , 0.1 mM each primer, and a viral template mass of 2.7 pg ($10 4 copies). Tests were performed in triplicate and corresponding negative controls were run under identical conditions except that viral template was replaced with nuclease-free water (Ambion, Austin, TX, USA). All reactions were thermocycled on the Bio-Rad DNA Engine (Hercules, CA, USA) as follows: one cycle of 2 min at 94 C; 40 cycles of 15 s at 94 C, 30 s at 43.9 C, 1 min at 68 C; one cycle of 5 min at 68 C. In line with our goal to create a protocol applicable to both DNA and/or RNA templates, we verified that inclusion of an RT step (one cycle of 30 min at 45 C) does not alter the outcome of the subsequent PCR when working with DNA template (data not shown). As such, for all Poxviridae 16-plex results reported here, an RT step was not used. Another goal was to establish reaction conditions that can be applied to any multiplex viral primer set without the need for re-optimization, assuming the primer set is designed with the same general parameter contraints (T m s, etc.) used to compute the 16-plex presented here. As such, we first optimized master mix conditions, where it was determined that a 4.8 mM MgSO 4 concentration resulted in most optimal amplification. Similarly, we determined the optimal annealing temperature based on results from annealing temperature gradient experiments. A more detailed discussion on the optimization of reaction conditions for amplification with our short primer viral multiplexes is the subject of a separate paper in preparation (Hiddessen and co-workers). This multiplex was compared against the human genome, predicting 233 amplicons between 50 and 1000 bp from the Poxviridae 16-plex. For empirical tests against background human nucleic acids, we followed the same reaction conditions as above. In these tests, vaccinia DNA was held at a constant concentration of 2.7 pg (1.3 Â 10 4 copies), and serial dilutions of human genomic DNA (Novagen, Madison, WI, USA), were added to the vaccinia PCR reaction mix, starting at 2.7 pg and titrating down over 4 orders of magnitude to 2.7 Â 10 -4 pg. These experiments were performed using mass ratios of vaccinia:human DNA at 1 : 1, 10 : 1, 100 : 1, 1000 : 1 and 10000 : 1. These correspond to copy number ratios of vaccinia:human genomes of 1. All PCR experiments were analyzed on 3% agarose TBE gels containing ethidum bromide that were purchased from Bio-Rad (Hercules, CA, USA). Blue juice TM 10Â loading dye was purchased from Invitrogen (Carlsbad, CA, USA) and diluted to 2Â before use. A 50-bp DNA ladder was purchased from Novagen (Madison, WI, USA). For analysis, 15 ml from each separate 25 ml PCR reaction were combined with 2 ml of of loading dye and 15 ml of the loading-dye/product mixture was loaded per well and electrophoresed for 1 h 40 min at 85 V. For confirmation of amplicon sequence, 9 ml of amplified PCR product was mixed with 4 ml of ExoSap-it (USB, Cleveland, OH, USA) and incubated at 37 C for 15 min followed by a denaturation step at 80 C for 15 min. Sanger sequencing was then performed with the BigDye V3.1 Terminator Kit (Applied Biosystems, Inc., Foster City, CA, USA). Single reactions contained 4 ml of Ready Reaction Mix, 0.2-mM primer, 2 ml 5Â Sequencing Buffer, 11.5 ml of nuclease free water (Applied Biosystems, Inc.) and 2 ml of post-Exosap-it PCR product. The sequencing reaction used the following thermocycling profile: 1 cycle of 94 C for 1 min; 25 cycles of 94 C for 15 s, 38 C for 30 s and 60 C for 4 min. Two microliters of sequencing reaction product was combined with 18 ml of Hi-Di Tm Formamide (Applied Biosystems, Inc.) and run on the ABI 3130 (Applied Biosystems, Inc.). The results were analyzed using Sequencing Analysis v5.2 (Applied Biosystems, Inc.). Predicting a set of universal viral primers that are all mixed in a single large reaction, we predict that 1008 primers of length 10 nt would be required to ensure amplification of at least one fragment of length 80-620 bp from every viral genome. These are predicted to generate a mean and maximum number of amplicons per viral genome of 13.2 and 948, respectively. Binning the primers into smaller sets of 20 primers/bin doubled the number of primers required, as Figure 1 shows the fraction of viral genomes amplified versus the number of primers, for primers ranging from 6 to 18 nt in length. About 2000 binned 10-mers are required to amplify 100% of sequenced viruses, generating a mean of 1 and maximum of 2.9 amplicons per genome, on average. Using traditional-length 18-mer primers nearly doubles the number to $3700 primers. The incomplete curve for 7-mers was from inappropriate settings for this oligo length, since some genomes do not contain pairs of 7-mers with the required T m , as given in the Supplementary Tables S1 and S2, within the desired amplicon length range. The concave curves showing diminishing returns reflect biased sequence availability rather than any particularly highly conserved primers: the primers that amplify the largest number of sequences are all from influenza, as far more sequences are available for this species than for any other. Increasing availability of sequence data on universal primer sets The increase in sequence data requires $700 more 10-mer primers to amplify all sequenced viruses in 2007 compared to 2004 (Figure 2 ). While the increase in the number of sequences used between the two dates was only $15%, the number of primers required increased by 48%, illustrating the substantial increase in diversity represented by the additional sequence data. Figure 2 also shows that removing all T m constraints allows fewer primers to be used since no conserved primers are eliminated due to T m , as some AT rich subsequences tend to be fairly conserved. Figure 3 shows that a universal primer set predicted using the 2004 sequence data would amplify only 35% of the 2007 sequences using primers of at least 10 nt in length. Shorter primers increase this fraction to over 60%, due to the higher likelihood of occurrence and conservation of shorter oligos, but even so, a multiplex of 7-mers is not guaranteed to amplify a fragment from every virus. The minor differences in the number of genomes detected between 12, 13, 14 and 15-mers can be attributed to the facts that a greedy but not necessarily optimal algorithm is used to select one solution from among many, that the primers in a particular set depend on the T m ranges we used as well as length differences, and the unpredictable nature of novel viral sequences accumulated between 2004 and 2007, rather than any real difference in the ability to detect genomes among those primer sets. If all human DNA cannot be removed from the sample, simulations indicate that on average, multiplexes of 10-mer primers are expected to produce hundreds of amplicons from the human genome, which would appear as a smear on a gel ( Figure 4 ). With primers of length 11-14, there is a large variation among bins. While most short sequences do occur in the human genome (33) , an amplicon requires that two occur in proximity, and primers !11 bases make this a sporadic event for bins with only 10 primers. Nonetheless, imperfect sample purification to eliminate eukaryotic nucleic acids could be problematic for universal viral priming using primers shorter than 15 bases, particularly for multiplexes of 10 or more primers. Family identification by sequencing products from universal 10-mer primer set versus randomly amplified fragments If universal viral primers amplified fragments from a newly emerged virus, would the product sequences show similarity to others in the same family? We predicted (2). The number of family-level primers for each family, and the number of genomes available for generating those primer sets, is given in Table 1 , for three alternative settings: short primers of 10-15 nt with T m 40-45 C, standard length primers of 17-21 nt and T m 55-60 C, and the same but requiring that each primer subsequence of at least 17 nt be unique to the target viral family relative to other viral families. At least one amplicon of 200-800 bp was required from every genome. Hypothetically, the worst-case scenario to amplify a target set of N sequences would require 2N primers. MPP requires on average only 37% or 45% of this number, for primers of length 10-15 nt or 17-21 nt without the requirement for family specific primers, respectively (averaged across families, omitting those families for which the computations were not run to completion). The most diverse families, in particular Bunyaviridae, Geminiviridae, Polydnaviridae, Reoviridae, Retroviridae, Siphoviridae and Orthomyxoviridae require so many primers that actually applying family-level amplification is probably infeasible. For these, more restricted target sets may be necessary, such as limiting to a single segment for the segmented families or to subclades, and possibly the incorporation of primers with degenerate or inosine bases. Some families with many genomes can be amplified with relatively few primers, such as Coronaviridae, Hepadnaviridae, Poxviridae, Togaviridae, Microviridae and Polyomaviridae. Using primers of length 10-15-mers or 17-21-mers, 66% or 63%, respectively, of the viral families have primer sets of 40 or fewer primers (20 primer pairs), which is feasible for typical multiplexes. The sizes of the family-level primer sets show a trend for an increase with the number of available sequences in the family ( Figure 5 , P = 0.14). There is no clear relationship between the number of primers in the set and whether the genomes are single or double stranded RNA or DNA. All family primer sequences are available as Supplementary Data. Primer design with the MPP software indicates that relatively few primers are required to amplify all sequenced genomes of HIV-1, FMDV and Norwalk virus (Table 2) , and these can be calculated in minutes. Influenza A HA and NA segments demand large numbers of 10-mer or 17-18-mer primers and hours to calculate, so one could break these into subgroups, possibly by serotype, as shown for several HA serotypes in Table 3 . The percentage of genomes amplified versus the number of primers used, for primers of either 10-mers or 17-18-mers, is shown in Figure 6 . This plot shows that a large fraction of targets are amplified with only 2 primers, and the addition of subsequent primers shows diminishing returns in amplifying fewer, more divergent targets not detected by the initial, more conserved, primer pair, although the true diminishing returns depend on the extent to which available sequence data is an unbiased representation of diversity. The more traditional method of attempting to find primers from a MSA would be problematic, probably requiring manually designed primer multiplexes or highly degenerate primers. GAAGAAGCG starting at 9071. These regions are too far apart to be used as primers for most polymerases used in diagnostic PCR protocols, where amplicons must typically be less than 300 bases long for efficient amplification. A recently published study (35) selected primers from the 5 0 LTR U5 end to the Gag-Pol start (5 0 -TAGC AGTGGCGCCCGA-3 0 and 5 0 -TCTCTCTCCTTCTAGC CTCCGC-3 0 ), but a comparison against available genomic data indicates that 487 of the 1175 genomes (41%) do not contain a sequence match for this primer pair, so may fail to be amplified. For influenza A segment HA, the size of the longest conserved region from the 95% consensus is only 5 bases, and for segment NA, only 6 bases, insufficient for even a single primer. For FMDV and Norwalk virus, the longest 100% conserved regions are 9 and 6 bases, respectively. The MPP software makes it straightforward for a nonexpert to predict a multiplex-compatible set of primers to amplify all targets, even for enormous and heterogeneous target sets that cannot be aligned. For comparison, we considered other software options for designing primers for these heterogeneous viruses. FastPCR (www.biocenter.helsinki.fi/bi/programs/fastpcr .htm) and GeneUp (25) were the only programs that did not require an MSA as input. The FastPCR algorithm for group-specific PCR (i.e. universal amplification) designs PCR primer pairs individually for each target sequence without regard for primer conservation among targets, and then compares each primer pair to the other targets. This is a brute force strategy that is only suitable for small target sets and short target sequences (appropriate for gene lengths, but not for viral genome-length sequences). We ran the FastPCR software on our internal servers, but it did not complete 'group-specific PCR' for the smallest data set, Norwalk virus, after running for 18 h, and for 'multiplex PCR' gave the error message 'No compatible combination of pair primers for multiplex PCR found'. GeneUp simulates PCR with pairwise combinations of candidate primers which pass length, T m , GC%, and palindrome filters against all target sequences, and uses a greedy algorithm to build a primer set to amplify all of the targets. Since testing all possible pairwise oligonucleotide combinations against each target explodes in time and memory for large target sets, a cap on the maximum number of candidate primers to be tested must be imposed, presumably using the most common oligos, although this is not explicit in the paper. Then a Grey shaded entries indicate where calculations were not run to completion. In other cases (not shaded) where fewer than 100% of targets were predicted to be amplified, the algorithm failed to find primer pairs that met all the required specifications for the remaining targets, that is, primers in the right length, T m , and amplicon length range, with hairpin and dimer avoidance with other primers already selected to be in the set, could not be found. None of the calculations that completed required more than the 16 GB of RAM that was available. PCR simulation (performing a text search for the primer in the target sequence) of each pair versus each target is performed. Unfortunately, the most common oligos may not occur in the correct orientation or distance to serve as primer pairs, so that multiple iterations of the entire process must be performed before a set covering all targets is obtained. MPP, in contrast, uses an efficient ranking algorithm to favor pairs of primer candidates that will produce amplicons of the right size in the most targets. Because of the MPP hashing algorithm and data structures, those targets that are amplified is easily determined without simulating PCR. A copy of the GeneUp software for testing could not be obtained from the authors. The PDA-MS/UniQ approach (26) uses a hash index of 4-mers and scoring heuristic to identify common regions in the target sequences with the most shared tetramers. Simulating combinations of candidate primers selected from these common regions is then performed using a genetic algorithm. Their genetic algorithm is a more scalable approach than those above, and the quality of the solution may be improved with more compute power, although a potential disadvantage of genetic algorithms is that they can be slow, particularly if there are very many possible combinations. For computational tractability, they limited primer size to 12 nt, and avoidance of hairpins and primer dimers was not modeled. This software was not available for download or on a public web server. The other software all required MSA as input. While these tools could not be tested on larger (e.g. family level) primer prediction problems where alignments would not be possible or appropriate, we nevertheless attempted to run these tools on the alignable species-level target sets. The Primaclade (24) webserver timed out for the smallest alignments we tested, Norwalk and FMDV. CODEHOP (23) requires protein alignment as input so is not appropriate for whole-genome (nucleotide) alignments. GreeneSCPrimer (20) did generate a number of degenerate primer candidates from the MSAs for Norwalk and FMDV, requiring the user to manually select a combination of forward and reverse groups from a set of options. We ran GreeneSCPrimer using length, T m , etc. settings mirroring or more lenient than those we used for Table 3 (T m = 55-65 C, GC% = 20-80%, length 17-25 bp, 100% coverage, product size 80-620 bp, allowed T m difference 10 C, others left as defaults). HYDEN (22) also generated degenerate candidates from a MSA, although it does not check T m and the length is limited to a single value rather than a range. For our tests using HYDEN, we used a length of 18 rather than 17 because of the lack of T m control, and allowed 0 mismatches. The GreeneSCPrimer option requiring the fewest total primers for the Norwalk set required 18 primers, four of which had either 2-fold or 4-fold degeneracy so the actual number of priming sequences would be 26, compared to a total of 20 non-degenerate primers predicted by MPP (Supplementary Data). HYDEN generated four degenerate primers covering only 34 of 41 sequences, each with 3-or 4-fold degeneracy for Norwalk, which translates to 15 priming sequences in the reaction. One would need to find primers to amplify the remaining seven sequences. Small degenerate priming sets (e.g. four primers in this case) are less expensive to purchase, but because of dilution effects from the many sequence combinations actually present (15 priming sequences in the PCR), sensitivity may be reduced compared to nondegenerate priming. However, using a smaller set of degenerate signatures such as those from HYDEN may be preferable, and is a capability that could improve MPP in a future version. For FMDV, MPP predicted six nondegenerate multiplex compatible primers to amplify all targets. GreeneSCPrimer generated a number of candidates, and manual inspection identified that the best of those primer combinations would require 6 primers, one of which had 2-fold and another had 3-fold degeneracy, totaling 9 actual priming sequences in a reaction (Supplementary Data). This compares with two primers each with 4-fold degeneracy using HYDEN (eight priming sequences in a reaction), to amplify 98% (183 of 187) targets. Again, the small number of signatures predicted by HYDEN is desirable for some applications, although the degeneracy is high and these must be supplemented to pick up the few outlying sequences The aim of the MuPlex software (28, 29) is to partition primer pairs into multiplex-compatible bins for SNP genotyping, and it does not employ any algorithm during primer selection to minimize the number of primers to amplify all targets, yielding a one-to-one correspondence between number of primer pairs versus number of targets. However, the MuPlex graph-based algorithm to partition primers into bins is more sophisticated than the simple binning scheme of MPP. In future work, MPP could be used to identify a universal set of primers and the backend of MuPlex used to optimize how they are binned into separate reactions. In summary, no software except MPP was capable of predicting primer sets for the larger target sets we examined or generating a multiplexed primer set without a MSA for any of the target sets we examined, even the smallest ones containing only a single species. Only HYDEN and GreeneSCPrimer, both requiring MSA inputs, completed for the two smallest target sets, selecting smaller sets of degenerate primers than the nondegenerate MPP primers. If target sets are sufficiently conserved so that a reasonable MSA can be built, and degenerate primers are acceptable, these tools may be preferable over MPP. However, for larger and more diverse target sets, only MPP completed. Using the experimental conditions described in the 'Methods' section, we tested the Poxviridae 16-plex (Table 4 ) against vaccinia virus, Lister strain DNA. As assessed by agarose gel analysis, we achieved specific amplification of the predicted 617-bp amplicon for the target vaccinia genome (lane 3, Figure 7 ). The band does not appear in the no template control shown in lane 4 ( Figure 7 ). To confirm that the amplicon observed in gel analysis was a specifically amplified product from vaccinia virus, Lister strain, we sequenced the product according to the procedures described above. An analysis of the high-quality sequence read data taken from the electropherogram yielded a 95% identity (maximum) to vaccinia virus, Lister strain (AY678276.1) with a query The forward and reverse primers (FP, RP) in bold are predicted to amplify vaccinia Lister with the indicated product length (617-bp amplicon size). coverage of 99% and an e-score of 0 (no data shown). These values indicate that the sequenced product was vaccinia DNA and not amplification from an exogenous nucleic acid source, e.g. host cell. Notably, our multiplex reaction did not produce a smear (lane 3, Figure 7 ) that would be indicative of nonspecific priming of either the target or exogenous host cellular nucleic acids. While the exact manufacturer's extraction protocol is proprietary, it is generally known that a standard sucrose gradient ultracentrifugation step is used to enrich for the viral capsids prior to viral nucleic acid extraction. However, to the best of our knowledge, no nuclease digestions are performed prior to viral capsid lysis. Furthermore, while the viral 'extract' contains a mixture of both viral DNA and cellular nucleic acids, the exact proportions of host cell and viral nucleic acids cannot be determined. Thus, our results indicate that a multiplex of 16 primers of 10 nt each amplifies only the specific predicted band from vaccinia. In some applications, such as clinical or biodetection applications, the exogenous nucleic acids from other eukaryotic sources, notably human sources, may be present in varying and unknown concentrations. To test the effects of background human genomic DNA on the performance (amplification specificity) of the Poxviridae 16-plex primer set, we conducted PCR reactions across a series of mass ratios of vaccinia Lister DNA:human genomic DNA at 1 : 1, 10 : 1, 100 : 1, 1000 : 1 and 10000 : 1. These correspond to approximate copy number ratios of vaccinia : human genomes of 1.3 Â 10 4 : 0.82, 1.3 Â 10 4 : 0.082, 1.3 Â 10 4 : 0.0082, 1.3 Â 10 4 : 0.00082 and 1.3 Â 10 4 : 0.000082, using genome sizes of 1.9 Â 10 5 bp and 3 Â 10 9 bp for vaccinia and human DNA, respectively. For context, there are 6.58 pg or two copies of the genome in one human cell. Our algorithm predicted 233 amplicons between 50 and 1000 bp from the human genome. In experiments, this would appear as a 'smear' on the agarose gel, which indeed was observed for the mass ratios of 1 : 1 and 10 : 1 (Figure 7 , lanes 6 and 7, respectively). However, even at a ratio of 10 : 1 vaccinia:human DNA (lane 7), the 617-bp vaccinia amplicon is clearly visible on the gel, despite the numerous nonspecific amplicons. At ratios of 100 : 1, 1000 : 1 and 10000 : 1 (lanes 8, 9 and 10, respectively, Figure 7) , the smear is drastically reduced to nonexistent (at the resolution of the agarose gel), and the vaccinia amplicon is readily visible. For comparison, we tested the 16-plex primers against the same mass of human DNA in the absence of vaccinia DNA at 2.7 pg (the 1 : 1 ratio mass), 0.027 pg (100 : 1 mass) and 2.7 Â 10 À4 pg (10000 : 1 mass) (lanes 12, 13, and 14, respectively, Figure 7) . The data show a similar smear at 2.7 pg as was observed when vaccinia DNA was present (lane 6). However, the low-intensity 617-bp amplicon from vaccinia is visible in lane 6 (with vaccinia) while a similar amplicon is not present in lane 12 (without vaccina). These results provide evidence that amplification with these family level primer sets will depend on viral titers in the actual sample, and that there are cases where amplification will either not be possible or require additional sample purification steps. However, as discussed further in the next section, specific amplification with a short-primer multiplex could be used for selective enrichment of viral targets by generating amplicons with known sequence, which may be feasible for viral detection if combined with a probebased amplicon detection method such as TaqMan Õ or Luminex bead based suspension arrays (http://www .luminexcorp.com/). We compared recently published conserved Orthopoxvirus primers (36) to the 148 Poxviridae genomes, and computational predictions suggest that 67 of the available genomes might not be amplified by the two primers they designed for the Orthopox genus. These included a number of monkeypox, ectromelia, several vaccinia and a couple of variola strains. However, in permissive hybridization conditions it is possible that primers would anneal despite mismatches to target, allowing more of the targets to be amplified. Four conserved Orthopoxvirus primers from an earlier publication (37) , before many of the Poxviridae genomes were available, do not match 53 genomes, including a number of monkeypox, ectromelia, camelpox and one variola minor genome. The primers in (37) included inosine bases, which we replaced with each possible A, T, G and C base in all possible combinations for our primer-target comparison. Thus, published Orthopoxvirus primers may fail to amplify many desired targets. Primer design for amplification and detection of divergent target sequences can be challenging, and this problem will only grow as sequencing technologies improve. Some methods are limited in scalability, particularly those requiring a MSA as input. Developing a PCR multiplex is often a tedious mix-and-match process from among primers originally designed to work in singleplex. We describe the MPP algorithm based on hashing of conserved k-mer subsequences that requires no MSA and where multiplex-compatible primer sets are built de novo to avoid primer dimer and hairpin formation to the extent that can be predicted based on free energy calculations. The algorithm seeks to minimize the number of primers to amplify all targets, although because the algorithm is heuristic and not an exhaustive search of all combinations of primers, the smallest primer sets may not always be selected. This is an NP complete problem (38) , so an exhaustive search for the global optimum is only practical for very small target sets. It is also beyond the scope of the current software to predict priming with mismatches between target and primer, which may occur under nonstringent hybridization conditions or for mismatches at the 5 0 end of a primer. While allowing for such priming would reduce the number of primers needed to amplify all targets, it would substantially slow the algorithm and increase memory requirements. Nor do we claim to have taken all steps to optimize (e.g. through parallelization, see Supplementary Methods) for speed or memory, but instead present this software as a simple embodiment of one alternative to MSA for multiplex primer design. The software handles a large number of input sequences, although for very diverse targets the predicted primer multiplex may be too large to be empirically feasible. For many target sets, such as those including all the genomes of a species, predicting a universal primer set requires only minutes up to a few hours, although for inputs with thousands of sequences a run may take days. We used MPP to design a universal primer set for a target set of all virus genomes, and showed that even if short primers are used, thousands of priming sequences would be required to amplify all sequenced viruses. We then applied MPP to design multiplex family level primers for every viral family, as well as for some diverse species target sets too large for other available primer prediction software. In addition to finding primers for viral detection, another application of MPP could be to design multiplex primers for homologs in a gene family. The user can specify an appropriate product length range to ensure amplification of an adequate span across the gene. MPP could also be used to design multiplex primer sets for unrelated target sequences, for example, multiple bacterial and viral species or gene families, in a single reaction. Since no sequence alignment is required, there is no need for any sequence conservation among targets. Recent work showed that a large-scale multiplex of 800 primer pairs specifically designed to detect a diverse set of genes from nine pathogens improved sensitivity on a microarray by up to 1000-fold (39) . The MPP software could be used to design multiplex primer sets for similar work. We demonstrated the application of the algorithm experimentally using purified vaccinia DNA, amplifying the expected band with a Poxviridae short primer multiplex PCR, and confirmed the band's expected sequence. We provided an example of the effect of background nucleic acids by spiking human DNA into the PCR reaction over a series of mass ratios. While a band for vaccinia was visualized by slab gel electrophoresis for all mass ratios, background DNA smears were also clear at the higher ratios. At a mass ratio of 1 : 1 vaccinia:human DNA, there are 2.7 pg or 0.82 copies of the human genome present in the reaction. To place this into context, there are $6.58 pg or two copies of genomic DNA in one human cell. Thus, the results from the 1 : 1 mass ratio reaction represents the impact that $41% of the total DNA content from one human cell could have on the results, when analyzed by slab gel electrophoresis. In a clinical sample, there could be a much greater number of human cells and DNA. Amplification with these family level primer sets will depend on viral titers in the actual sample, and these experiments show that there are cases where amplification will either not be possible or require additional sample purification steps. While short primer PCR multiplexes may enable amplification of diverse target sets, they will not match the specificity of longer primers. Combination of these short primer multiplexes with digitized, microfluidic-based picoliter reactions (40) and/or next-generation microfluidic-based sample purification technologies that have the ability to isolate target from contaminating nucleic acids may help to overcome this limitation. However, even in the absence of such technologies, a highly conserved short primer multiplex could enrich amplification products for members of a particular viral family compared to randomly amplified or unamplified sample. Then it would need to be followed by product sequencing or array hybridization, such as TaqMan Õ or Luminex bead-based suspension arrays (http://www.luminexcorp.com/), to provide more specific information about the organism(s) present, since electrophoretic banding patterns may contain unexpected bands, particularly if the sample contains a significant amount of nonviral nucleic acids. Previous studies have shown the utility of short primer singleplex PCR using 9-mers or 10-mers followed by gel electrophoresis for genetic fingerprinting of eukaryotes and bacteria (41, 42) . For viruses, with much smaller and more diverse genomes, the large numbers of 9-mer or 10-mer primers required to generate at least one band from every virus as predicted by our analyses implies that primer size would need to be as short as 5-mers to rely on a gel banding pattern using only one priming sequence for fingerprinting viruses (unpublished data). However, the analyses here predict that imperfect sample purification to eliminate eukaryotic nucleic acids could be problematic for universal viral priming using primers shorter than 15 bases, particularly for multiplexes of 10 or more primers. Nanda et al. (43, 44) were able to achieve sufficient viral isolation from cell culture samples to allow viral identification using viral PCR with priming sequences as short as pentamers, so the problem of contaminating host nucleic acids for specific, short primer PCR of viruses is not insurmountable. They found specific pentamer PCR to be several logs more sensitive than nonspecific amplification, provided that they purified encapsidated viral nucleic acids prior to PCR. Another method that has been used for virus discovery is VIDISCA (Virus discovery cDNA-AFLP) using restriction enzyme digestion, adaptor ligation, and PCR by priming with the adaptor sequence (45, 46) . This method, like the pentamer priming used by (43, 44) , requires prior separation of encapsidated viral nucleic acids, as it generates fragments from any DNA present, viral, host or otherwise. Multiplex PCR with primers 10-15 nt in length may be yet another alternative strategy lying between these nonspecific methods and PCR with standard primers of at least 18 nt, as we have shown that it can add some measure of specificity for a viral family. In summary, we applied the MPP software to generate multiplex-compatible primer sets for every viral family and several divergent viral species and experimentally demonstrated application of one multiplex set to show nonrandom amplification with a set of short primers. Primer sets are available as supplemental material or by request from the authors. The MPP software is freely available for academic and nonprofit use at http://mpp .llnl.gov.

Short-hairpin RNAs delivered by lentiviral vector transduction trigger RIG-I-mediated IFN activation

Activation of the type I interferon (IFN) pathway by small interfering RNA (siRNA) is a major contributor to the off-target effects of RNA interference in mammalian cells. While IFN induction complicates gene function studies, immunostimulation by siRNAs may be beneficial in certain therapeutic settings. Various forms of siRNA, meeting different compositional and structural requirements, have been reported to trigger IFN activation. The consensus is that intracellularly expressed short-hairpin RNAs (shRNAs) are less prone to IFN activation because they are not detected by the cell-surface receptors. In particular, lentiviral vector-mediated transduction of shRNAs has been reported to avoid IFN response. Here we identify a shRNA that potently activates the IFN pathway in human cells in a sequence- and 5′-triphosphate-dependent manner. In addition to suppressing its intended mRNA target, expression of the shRNA results in dimerization of interferon regulatory factor-3, activation of IFN promoters and secretion of biologically active IFNs into the extracellular medium. Delivery by lentiviral vector transduction did not avoid IFN activation by this and another, unrelated shRNA. We also demonstrated that retinoic-acid-inducible gene I, and not melanoma differentiation associated gene 5 or toll-like receptor 3, is the cytoplasmic sensor for intracellularly expressed shRNAs that trigger IFN activation.

A specific double-stranded RNA (dsRNA) structure, $21-22 bp dsRNA with 3 0 overhangs, plays a critical role in initiating both microRNA (miRNA)-and small interfering RNA (siRNA)-mediated gene silencing, as it is the structure recognized by the RNA interference (RNAi) machinery, the RNA-induced silencing complex (RISC) (1) (2) (3) . Except for preformed siRNA duplexes of $21 bp, the RISC-loaded small RNAs are generated by a ribonuclease (RNase) III-like enzyme that is found in virtually all eukaryotic organisms. This enzyme, aptly named Dicer for its ability to cleave a variety of larger (>30 bp) dsRNA molecules into the $21 bp dsRNA with a characteristic 3 0 overhang of 2 nt, is a multidomain RNA-binding protein and itself a component of RISC. The primary sequence of the RNAs is not important in RISC formation, and RNAi can suppress virtually any target as long as rules of sequence complementarities between the small RNA and the target RNA are satisfied. dsRNAs are also a type of pathogen-associated molecular pattern (PAMP) that are detected by cellular innate immunity sensors named Pattern Recognition Receptors (PRRs) (4) . The interaction between a PAMP and a PRR triggers activation of the interferon (IFN) pathway in mammalian cells, which significantly changes the gene-expression profile in the cells and contributes to the well-documented off-target effect of RNAi. IFN induction is especially problematic in antiviral studies employing RNAi, where the antiviral effect of IFN must be distinguished from that of RNAi. Typical IFN-inducing structure patterns include dsRNA of certain length, single-stranded RNA (ssRNA) containing 5 0 -triphosphates (5 0 -ppp), the dsRNA analogue polyinosinic-polycytidylic acid (poly I:C), and certain dsDNA molecules. These RNA patterns are generally believed to possess 'non-self' properties to allow the cell to recognize foreign (often viral) RNAs specifically. Various forms of siRNA duplexes have been reported to trigger IFN induction both in vitro and in vivo (5) (6) (7) (8) (9) , probably through the cell surface-and/or endosomeexpressed Toll-like receptors (TLRs), including TLR3 and TLR7 (6, 8, 9) . Short-hairpin RNAs (shRNAs) expressed from a DNA plasmid have also been shown to activate IFN (10) . The double-stranded form of these RNAs is below the size limit of the stem-loop RNAs that can be detected by the RNA-activated protein kinase (PKR) (11) and is probably detected by other cytoplasmic PRRs. Two cytoplasmic RNA helicases, retinoic-acid-inducible gene I (RIG-I) and melanoma differentiation associated gene 5 (MDA5), signal to the IFN-b promoter when activated by specific RNA structures (12) (13) (14) . Although both PRRs signal through the mitochondrial antiviral signaling protein MAVS/Cardif/VISA/IPS-1 (15) (16) (17) (18) , studies of ligand specificity suggest that RIG-I and MDA5 are parallel sensors with overlapping substrates. For example, although both PRRs are activated by poly I:C in cell culture systems (12, (19) (20) (21) (22) (23) , MDA5 appears to be more important in mediating the poly I:C response in vivo (13, 14) . In addition, RIG-I can bind and respond to ssRNAs bearing 5 0 -ppp, whereas MDA5 is not activated by 5 0 -ppp-containing RNA (24, 25) . Finally, several cytosolic sensors for dsDNA has been recently reported (26) (27) (28) (29) (30) (31) . Nevertheless, current data on what constitutes effective substrates for either PRR are incomplete and sometimes controversial. Here we report for the first time that shRNAs delivered by lentiviral transduction triggered IFN activation and that RIG-I and MAVS, but not MDA5 or TLR3, mediated the IFN activation triggered by intracellularly expressed shRNA, which could activate both IFN-a and IFN-b promoters. IFN activation depended on sequence, a 5 0 -ppp and correct processing of the RNA hairpin by Dicer; it was independent of promoter choice, presence of blunt ends, route of delivery and RNAi potency. GS5 and LH86 cells have been described earlier (32, 33) . Huh-7 and 293FT cells were maintained in DMEM supplemented with 10% FBS. We used the following antibodies: anti-CyPA (Biomol, Plymouth Meeting, PA, USA); anti-CyPB (AfEnity BioReagents, Rockford, IL, USA); anti-Ku80, anti-Flag and anti-actin (Sigma-Aldrich, St Louis, MO, USA); anti-IFN stimulate gene (ISG)15 (Rockland Immunochemicals, Gilbertsville, PA, USA); anti-NS5A (Virogen, Watertown, MA, USA) and anti-NS3 (in-house). GSB1 and H801 cells have been described earlier (34) . Poly I : C was purchased from Sigma-Aldrich, and synthetic hairpin RNA was purchased from Integrated DNA Technologies (Coralville, IA, USA). Synthetic siRNA was purchased from Ambion (Austin, TX, USA). Protein contents of cell lysate were quantified with the Bio-Rad DC protein assay (Bio-Rad, Hercules, CA, USA), and an equal amount of total protein was loaded in each lane. Samples for IRF-3 dimerization assay were run on a polyacrylamide gel under non-denaturing conditions (35) . Other samples were denatured and separated by sodium dodecyl sulfate polyacrylamide gelelectrophoresis (SDS-PAGE). Proteins were then transferred onto a nitrocellulose membrane and stained with the appropriate antibodies with the SNAP i.d. TM system (Millipore, Worcester, MA, USA) according to the manufacturer's instructions. For luciferase assays, cells were seeded to a confluency of 50%, and for all other assays, cells were seeded to a confluency of 30%. The next day, transfections of DNA plasmids and synthetic RNAs were performed with Lipofectamine TM 2000 (Invitrogen, Carlsbad, CA, USA) according to the manufacturer's instructions. Plasmids pGL3-IFNA1, pGL3-IFNB, pRL-TK, pCMV-Flag-IRF-3 and pCR3.1-IRF-7A have been described earlier (36) . shRNAs were expressed from a human immunodeEciency virus (HIV)-based lentiviral vector (32, 37) , and sh-PCAF was constructed on the basis of a previously reported sequence (38) . Plasmid sh-B971/H1 was constructed by cloning of the DNA fragment encoding the sh-B971 RNA into pSilencer 3.0-H1 (Ambion, Austin, TX, USA) according to the manufacturer's instructions. The RIG-I and TLR3 constructs have been described (39, 40) . The RIG-I C construct encodes Flag-tagged, C-terminal 707 aa of human RIG-I cloned into a bicistronic expression vector modified from pBICEP-CMV-1 (Sigma-Aldrich, St Louis, MO, USA), in which the CMV promoter was replaced with the elongation-factor-1 promoter. The MDA5, MDA5-C constructs were kindly provided by Fujita (12) . HCV genotype 2a NS3-4A protease was expressed from the pCMV-3Tag-1a plasmid (Stratagene, La Jolla, CA, USA). 293FT cells were seeded in 24-well plates and were transfected 16 h later with 400 ng of a shRNA expression vector, 40 ng of pGL3-IFNA1 or pGL3-IFNB, 20 ng of pRL-TK and 50 ng of pCR3.1-IRF-7A. Cells were collected 48 h after transfection. Luciferase assays were performed with the Dual-Glo Õ Luciferase Assay system reagents (Promega, Madison, WI) and luminescence quantified with a Modulus Microplate reader (Turner BioSystems, Sunnyvale, CA, USA). Ratios of firefly luciferase (from the pGL3 vectors) to Renilla luciferase (from the pRL-TK vector) were calculated, and that of the sh-B971 sample was normalized to 100%. Sequences of shRNA are shown in Table 1 . Lentiviral vector production and transduction were performed as described earlier (37) . Viral vectors were pelleted by ultracentrifugation at 50 000g at 4 C for 3 h and resuspended in a volume of PBS that was 1% of the original medium volume. The titers of the concentrated vectors were then measured with a p24 ELISA kit (ZeptoMetrix, Buffalo, NY, USA). Real-time reverse transcription PCR (RT-PCR) was performed as described earlier (32) . The primers used were OAS1 forward, 5 0 -AGG TGG TAA AGG GTG GCT CC-3 0 and OAS1 reverse 5 0 -ACA ACC AGG TCA GCG TCA GAT-3 0 ; RIG-I forward 5 0 -GAG GCA GAG GAA GAG CAA GAG G-3 0 and RIG-I reverse 5 0 -CGC CTT CAG ACA TGG GAC GAA G-3 0 ; GAPDH forward 5 0 -TCA CTG CCA CCC AGA AGA CTG-3 0 and GAPDH reverse 5 0 -GGA TGA CCT TGC CCA CAG C-3 0 . The primers for HCV detection were 5 0 -CGC TCA ATG CCT GGA GAT TTG-3 0 and 5 0 -GCA CTC GCA AGC ACC CTA TC-3 0 . For flow cytometry, GS5 cells were fixed 48 h after treatment in a solution of 2% paraformaldehyde and analyzed with a FACSCanto flow cytometer (BD Biosciences, San Jose, CA, USA). Mean GFP intensity was plotted, and that of the sh-NTC sample was normalized to 100%. Total RNA from transiently transfected 293FT cells was extracted with RNA STAT-60 (Tel-Test, Friendswood, TX, USA) and separated on a 7.5% urea polyacrylamide gel. The transfer of RNA onto nitrocellulose membrane and hybridization were performed according to standard molecular biology protocols. The probe for detecting the expression of sh-B971 and its variants was a synthetic DNA oligomer corresponding to the bottom strand of sh-B971. Radioactive labeling of the probe was performed with an end-labeling protocol with T7 polynucleotide kinase (Ambion, Austin, TX, USA). The exposure and detection of the radioactive signal was performed with a Typhoon Imager (GE Healthcare, Piscataway, NJ, USA) with Quantity One software (Bio-Rad, Hercules, CA, USA). A short-hairpin RNA directed at CyPB induces IFN production in human embryonic kidney cells To investigate the potential role of the cyclophilins (CyPs) in HCV replication (41), we delivered several shRNAs directed at mRNAs of three CyPs into HCV replicon cells by means of a lentiviral vector, using a murine U6 promoter to drive the expression of the shRNA ( Figure 1A ) (37) . We observed a discrepancy between two anti-CyPB shRNAs (B971 and B710) in their relative efficiency in knocking down CyPB expression and in suppressing HCV. Lentiviral vector sh-B971 was less efficient in knocking down CyPB expression but potently inhibited HCV NS5A expression in a human hepatoma cell line containing replicating HCV RNA ( Figure 1B , left). Viral inhibition was independent of CyPB knockdown, as control medium from transfected 293FT cells that did not contain any lentiviral vector particles, generated by omission of the packaging plasmids during transfection, also inhibited HCV replication ( Figure 1B , right) without affecting CyPB expression. The fast kinetics of viral inhibition (complete inhibition with 48 h, data not shown) was also more consistent with IFN than with RNAi-based inhibition. The presence of IFN in the lentiviral vector preparation of sh-B971 was confirmed by strong induction of 2 0 -5 0 -oligoadenylate synthetase 1 (OAS1), a classic IFN-induced gene, in both naı¨ve Huh-7 and the HCV replicon cell line (GS5) treated with the medium ( Figure 1C ). In addition, HCV replication in an IFN-resistant HCV replicon cell line (H801), in contrast to that in a wildtype replicon cell line (GSB1) (34), was not inhibited by the sh-B971 medium ( Figure 1D ), suggesting the lack of additional viral inhibiting agents in the sh-B971 medium. Expression of sh-B971 in 293FT cells also induced dimerization of IRF-3, confirming the activation of the IFN production pathway in these transfected cells ( Figure 1E ). Finally, sh-B971 was able to activate both IFN-a and IFN-b promoters, although the activation of the IFN-a promoter required coexpression of IRF-7, which is normally expressed at very low levels in 293-based cells ( Figure 1F ). These results demonstrate that sh-B971 is a potent activator of IRF-3 and IRF-7, master regulators of IFN expression in human cells. We next investigated the role of the different viral/ exogenous RNA sensors, RIG-I, MDA5 and TLR3, in sh-B971-triggered IFN production. Mammalian expression plasmids encoding each of these proteins, as well as the dominant negative (DN) mutants of RIG-I and MDA5, were transfected into 293FT cells with shRNAs and an IFN-b promoter reporter construct. The signaling to IFN-b promoter and the expression of the PRR proteins were then examined 48 h after transfection. In the absence of sensor proteins, the sh-B971 increased activation of the IFN-b promoter by 2.6-fold ( Figure 2A ). Coexpression of MDA5 or TLR3 did not increase or decrease sh-B971's ability to activate IFN-b promoter relatively to the negative control shRNA (sh-NTC), but in the presence of RIG-I coexpression, the induction of IFN-b promoter by sh-B971 was increased to $30-fold. Moreover, ectopic expression of a DN mutant of RIG-I (RIG-I C), but not that of MDA5 (MDA5-C), completely abrogated IFN promoter activation by sh-B971. With the exception of TLR3, which required prolonged exposure of the western blot to be detected, the cytoplasmic sensors and their mutants were expressed at comparable levels ( Figure 2B ). Moreover, activation of IRF-3 ( Figure 1E ) and IFN promoters ( Figure 1F ) in 293FT cells, which do not contain a functional TLR3 signaling pathway (42) , indicates that TLR3 plays a negligible role, if any, in IFN induction by sh-B971. The combination of sh-B971 and RIG-I produced the highest level of IFN-b promoter activity, which were confirmed by western blotting showing that endogenous ISG15 induction was only detectable in cells cotransfected with sh-B971 and wild-type RIG-I ( Figure 2B ). To confirm further that biologically active IFN was released from these cells, we applied the culture medium of the transfected 293FT cells to an HCV replicon cell line (GS5) in which NS5A-GFP expression is used for monitoring viral RNA replication (43) . HCV replication in this cell line is extremely sensitive to IFN, and the effect of the cytokine can be readily measured as the change in the mean GFP intensity of the treated cells. As shown in Figure 2C , culture medium from sh-B971 efficiently suppressed HCV replication, resulting in a decrease in Figure 1 . A small-hairpin RNA directed at CyPB induces IFN production in human embryonic kidney cells. (A) Sequence of sh-B971, which was expressed from a self-inactivating human immunodeficiency virus (HIV) vector with a murine U6 promoter (59) . (B) Inhibition of HCV expression by culture media of sh-B971-transfected 293FT cells. GS5 cells were treated with culture supernatant taken from 293FT cells transfected with various shRNA plasmids with (left) or without (right) the packaging plasmids overnight. Cells were then cultured in fresh media for an additional 6 days before being lysed for western blotting. (C) OAS1 induction by culture supernatant from 293FT cells transfected with sh-B971. Huh 7 and GS5 cells were treated with culture supernatant from 293FT cells transfected with either sh-Luc or sh-B971 for 24 h before RNA extraction and real-time RT-PCR analysis. OAS1 RNA level was normalized to that of GAPDH RNA. (D) Transfected culture media failed to suppress HCV replication in an IFN-resistant cell line. HCV replicon cells were cultured as described earlier (34) and then treated with the indicated culture medium from transfected 293FT cells. HCV RNA was analyzed with real-time RT-PCR. (E) IRF-3 dimerization in response to sh-B971 expression. Flag-IRF-3 was cotransfected with a shRNA into 293FT cells. Cells were lysed 24 h after transfection, and total cell lysate was separated on a polyacrylamide gel under non-denaturing conditions, transferred and stained with an anti-flag antibody. (F) IFN-a and IFN-b promoter activation by sh-B971 expression. Sh-NTC, sh-C454 (an shRNA directed at CyPC), or sh-B971 was cotransfected along with luciferase reporter plasmids with or without IRF-7. The ratios of firefly luciferase readings to Renilla luciferase readings were plotted. the NS5A-GFP intensity within 48 h of treatment. Cotransfecting wild-type RIG-I produced a medium with stronger inhibition, whereas the RIG-C drastically suppressed the antiviral effect of the medium. Finally, real-time RT-PCR analysis revealed that sh-B971, but not the negative control shRNA, strongly activated expression of endogenous RIG-I, a well-characterized ISG whose induction requires paracrine/autocrine action of IFN (44, 45) . As expected, poly I : C activated RIG-I expression in the same assay ( Figure 2D ). These results, taken together, show that RIG-I is the cellular sensor that mediates the IFN induction by sh-B971. The majority of the shRNAs that we use in the lab do not activate RIG-I expression and IFN signaling despite having essentially the same structure as sh-B971, so we wanted to determine whether the sequence of sh-B971 is distinctive enough to trigger the production of IFN. We first tested a synthetic siRNA duplex with the same target sequence as sh-B971. This siRNA (si-B971-syn) should resemble the final Dicer product of sh-B971 except for the 5 0 -ends. The synthetic siRNA contains 5 0 -OH groups, whereas the Dicer products probably Figure 3A ) while failing to activate IFN production, as measured by the GFP-HCV assay ( Figure 3B ). To determine whether the sequence of the intact hairpin RNA before Dicer cleavage is sufficient to trigger IFN, we tested a synthetic shRNA (sh-B971-syn) that had exactly the same sequence as the predicted intracellular sh-B971 transcript generated by the U6 promoter. Again, the 5 0 -end of the synthetic sh-B971 had a 5 0 -OH group instead of any phosphate. Sh-B971-syn behaved similarly to si-B971-syn in that it knocked down CyPB expression without activating IFN response (Figure 3) . These results suggest that the 5 0 -end status of sh-B971 is important for IFN activation, consistent with the previously finding that a 5 0 -triphosphate is required for RIG-I activation (24, 25) . To determine the contribution of the individual residues of the sh-B971 sequence, we introduced a series of point mutations into the shRNA and tested them for IFN induction. We changed the first nucleotide from A to G, C, or T while maintaining base-pairing between nucleotides +1 and +47. These mutant shRNAs lacked the ability to activate IFN production (Table 1) . Changing the +1 nucleotide to G while leaving the +47 nucleotide intact also abolished IFN activation by the shRNA (A1/G), as did the reciprocal mutation U47/C. The importance of the first nucleotide was further confirmed by the inability of sh-B971+1 to activate IFN. The target of sh-B971+1 was shifted 1 nt downstream on the CyPB mRNA, producing an shRNA starting with a G at the +1 position. The presence of an A at the +1 position was not, however, sufficient to render a shRNA competent for IFN activation, as replacing the first nucleotide of the sh-NTC with an A did not generate an IFN-inducing shRNA (NTC-A and NTC+1). These results indicate that a protruding/unpaired A at the end of the hairpin or the RNA duplex, a potential result of 'breathing' at the end of the dsRNA, is not sufficient to trigger IFN induction as previously suggested (38) . Two point mutations located farther into the stem structure of the shRNA (9G9 and B18A1) also reduced its ability to induce IFN even though the base-pairing was perfectly maintained in these mutants. Finally, replacing the 9-nt hairpin loop with a 7-nt loop that had been previously shown to abolish shRNA-mediated RNAi (loop A mutant) (46) eliminated sh-B971's ability to induce IFN, suggesting the importance of RNA processing in the induction. To determine whether the inability of the mutant shRNAs to induce IFN was due to lower expression levels, we performed northern blotting analysis of the shRNA expression on the wild-type and two mutants. The mutants A1/G and Loop A were chosen because their final siRNA products have exactly the same sequence as that of the wild-type sh-B971 and can thus be detected with the same efficiency by the same probe. Although sh-A/G and sh-Loop A were clearly unable to activate IFN-b promoter ( Figure 4A ), they were both expressed at levels comparable to those of the wild-type sh-B971 product ( Figure 4B) . Interestingly, the final siRNA product of sh-Loop A was slightly smaller than those of sh-B971 and sh-A1/G, suggesting that cleavage did occur and perhaps occurred one or 2 nt into the stem to compensate for the shorter loop. Blunt-ended siRNA has been previously reported to be stronger inducers of IFN than the siRNAs with overhangs (47) . Indeed, a previously reported IFN-inducing shRNA, sh-PCAF (p300/CREB-binding protein-associated factor), contains a blunt end (38) and was more potent in activating IFN than sh-B971 ( Figure 5A ), which is predicted to form an overhang of 2-3 Ts at each end of the final siRNA. We therefore constructed a version of the sh-B971 that would be blunt at the end that is not processed by Dicer by adding two extra As to the 5 0 -end of the shRNA. This modification (Blunt sh-B971) did not increase the ability of sh-B971 to activate IFN-b promoter ( Figure 5A ). We confirmed, in two independent experiments, that IFN induction by sh-PCAF was also mediated by RIG-I. First, cotransfection of DN RIG-I resulted a 50-to 100-fold inhibition of IFN induction by sh-PCAF ( Figure 5B ), whereas wild-type RIG-I increased IFN induction by several fold in the same assay. Second, when HCV NS3-4A protease, which cleaves MAVS, thereby blocking the RIG-I pathway, was coexpressed with either sh-B971 or sh-PCAF, IFN induction by these shRNAs were severely compromised ( Figure 5C ), further substantiating a role of the RIG-I and MAVS pathway in mediating IFN induction by both the blunt-ended sh-PCAF and the sh-B971 with overhang. The proper expression of NS3-4A protease was confirmed by western blotting ( Figure 5D ). To assess the contribution of the promoter choice in IFN activation by intracellular expressed shRNA, we expressed sh-B971 from another commonly used pol III promoter, the human H1 promoter. Both the original, mU6-driven sh-B971 and the H1-driven sh-B971 activated IFN-b promoter ( Figure 6A ) and resulted in secretion of IFN into the transfected cell-culture media, which in turn suppressed HCV replication ( Figure 6B ). Proper expression of the siRNA ( Figure 6C ) and the subsequent knockdown of CyPB expression ( Figure 6D ) all appeared normal for sh-B971 expressed from the H1 promoter plasmid, which has a backbone different from that of our lentiviral vector carrying the mU6 promoter. These data suggest that IFN induction by sh-B971 is not restricted to a particular promoter or expression construct. Further supporting this conclusion was the observation that the expression cassette by itself, removed and isolated from the lentiviral plasmid by restriction digestion, could also activate IFN production in transfected 293FT cells (data not shown). To this point, all the IFN induction experiments were done with transient transfection of DNA vectors and it was possible that certain features of the double-stranded plasmid DNA are responsible for IFN induction. We first tried to address this point by transfecting just the shRNAexpressing cassette, generated either by PCR or restriction enzyme digestion, into 293FT cells and confirming that these fragments of $200 bp were sufficient to trigger IFN induction (Supplementary Figure S1) . To definitively rule out any contribution by dsDNA, we used a lentiviral transduction system which has been suggested to express shRNAs that can escape detection by PRRs and IFN activation (48) . We produced lentiviral particles containing shRNAs from 293FT cells using standard . Sh-B971 expressed from an H1 promoter triggers IFN activation. Sh-B971 expressed from an H1 promoter was capable of (A) activating IFN-b promoter and (B) triggering IFN production to inhibit HCV replication in GS5 cells. (C) Intracellular levels of U6-and H1-driven sh-B971 products. RNA extraction and northern blotting were performed as described in Figure 4B . (D) Knockdown of CyPB expression by sh-B971 expressed from an H1 promoter. methods, centrifuged them to separate the vectors from the IFN-containing media, and then used them to infect naı¨ve 293FT cells ( Figure 7A ). Both sh-B971 and sh-PCAF vectors induced IFN production when delivered as concentrated lentiviral particles, measured both by HCV suppression ( Figure 7B ) and by OAS induction ( Figure 7C ) in Huh-7 cells. To rule out the possibility that residual IFN in the concentrated viral particles was responsible for these results, we added 100 U/ml IFN to the negative control vector sample before the concentration step. This preparation, designated sh-NTC*, was not able to trigger IFN production in naı¨ve 293FT cells, suggesting that the concentration step effectively removed the soluble IFN from the viral particle pellet. Proper knockdown of the siRNA target of sh-B971 was confirmed by this route of shRNA delivery ( Figure 7D ). To prove definitively that IFN induction by the shRNAs was mediated by the lentiviral infection route, we tested the effect of an inhibitor of HIV reverse transcriptase, Nevirapine, on IFN induction by sh-B971 and sh-PCAF. As shown in Figure 7E , inclusion of Nevirapine at the time of transduction effectively blocked the ability of both shRNAs to induce IFN in the transduced cells, suggesting the importance of the reverse transcription step in the expression of the shRNAs delivered by the lentiviruses. To determine whether lentiviral vector-delivered shRNA can trigger IFN induction in cells other than 293FT cells, we transduced a human hepatoma cell line, LH86, which has been reported to produce IFN upon viral infection (33) , and examined IFN induction in these cells. Culture medium from LH86 cells transduced with sh-PCAF contained biologically active IFN, which suppressed HCV replication in GS5 cells ( Figure 7F ), indicating that the ability of shRNAs delivered by lentivirus to induce IFN response was not limited to 293FT cells. It has been reported that certain chemically synthesized and phage polymerase in vitro transcribed siRNAs can non-specifically induce IFN responses and produce offtarget effect via various PRRs, including TLRs. However, the induction of IFN response by shRNAs and its underlying mechanisms have not been as well studied. The actual number of shRNAs that are capable of triggering IFN response will certainly be larger than the few that have been reported in the literature, yet very little is known about the unique characteristics of the select shRNAs and the pathway that they use to activate IFN production. The present study identifies RIG-I, but not MDA5 or TLR3, as the mediator for activation of IFN responses by two shRNAs that are distinct in sequence and structure but both capable of IFN induction in human cells. This was demonstrated by induction of IRF-3 dimerization, activation of IFN promoters, induction of endogenous ISGs (ISG15, OAS and RIG-I), and secretion of IFN, all of which depended on RIG-I and its downstream adaptor, MAVS. In addition, we show that delivery of these shRNAs via lentiviral transduction does not reduce their IFN-inducing capacity, indicating that the ability of lentiviral vector transduction to avoid IFN induction by shRNAs, as reported previously (48), may not be universally applicable to all the shRNAs. Specific recognition of dsRNAs or ssRNAs bearing 5 0 -triphosphates by RIG-I is presumably determined mostly by structural features other than the nucleotide sequence of the RNA. Yet IFN activation by sh-B971 exhibited a stringent dependence on specific nucleotides at multiple positions of the shRNA. An AA dinucleotide at the beginning of the U6 transcript has previously been suggested to result in aberrant transcription, and preserving a C/G sequence at positions À1/+1 suggested to avert IFN induction (38) . We indeed observed a strict requirement for an adenylate at the +1 position of sh-B971 for RIG-I recognition and IFN activation, but we observed no difference in expression levels or the apparent sizes of the sh-B971 RNAs bearing either an A or a G at the +1 position. Furthermore, mutations introduced elsewhere in the shRNA also abolished or diminished sh-B971's ability to activate IFN, suggesting additional sequence requirement for efficient RIG-I recognition and IFN triggering. Despite these results, because we were not successfully in cloning and sequencing the vectorexpressed siRNA, we cannot exclude the possibility that the adenylate at the +1 position interferes with transcription and that the resultant abnormal transcript contributes to IFN induction. Interestingly, the loop A mutant, which contains a predicted loop of 7 nt, generated a siRNA duplex inside the cells that is slightly smaller than that of the shRNAs with a wild-type hairpin loop, suggesting the processing by Dicer into the stem, perhaps fulfilling the requirement of a length of 9 nt for the hairpin loop (46) . This mutant form of sh-B971 was not, however, able to trigger IFN activation. Despite the abilities of both sh-B971 and sh-PCAF to activate the RIG-I pathway, the two shRNAs are unrelated in sequence. Two short stretches of siRNA sequences, GUCCUUCCAA and UGUGU, that have been previously defined as IFN-or cytokine-activating motifs (8, 9) are not found in either sh-B971 or sh-PCAF. Any common sequence motifs of IFN-activating shRNAs, if any, remain to be defined. The two shRNAs also differ in that one is predicted to contain one blunt end and the other two ends with overhangs. These results suggest that, although blunt ends may increase siRNA's ability to be recognized by RIG-I (47), they are not required for IFN activation by an endogenously expressed shRNA. The best-characterized RNA structure motif recognized by RIG-I is the 5 0 -ppp, which is absent from virtually all the cellular RNAs as a result of either 5 0 -capping or internal cleavage before their appearance in the cytoplasm. A synthetic shRNA that has the same sequence as sh-B971 but lacks the 5 0 -ppp failed to induce IFN, suggesting the 5 0 -end status of the intracellularly expressed sh-B971 contributes to IFN activation. Whether or not the 5 0 -end of an shRNA is capped has not been investigated. Murine U6 RNA does not contain the trimethylguanosine cap that is present on mRNAs and other U small nuclear RNAs; instead it contains a g-monomethyl phosphate cap at its 5 0 -end (49) . Capping of heterologous transcripts produced from the mU6 promoter, however, requires a stem loop at the 5 0 -end of the transcript and an AUAUAC sequence immediately after (50) . Most shRNAs, including sh-B971 and sh-PCAF, would not meet these requirements and thus should contain unmodified 5 0 -ppp. Similarly, no evidence of a cap structure for H1 transcripts could be found in the literature. We attempted to express sh-B971 using a miRNA expression cassette and the pol II promoter (51) . The primary transcript generated with this construct would be capped at 5 0 -end by a trimethylguanosine cap and the final siRNA duplex would bear a monophosphate at the 5 0 -ends of both strands because of Drosha and Dicer cleavage. This version of the sh-B971 vector was much weaker in its ability to trigger IFN activation. Unfortunately the intracellular expression of the RNA duplex was also much weaker and barely detectable by northern blotting. In addition, no knockdown of the target CyPB mRNA was seen with this miRNA-based sh-B971 (data not shown). As a result, whether sh-B971, if expressed at higher level from this construct, could effectively activate IFN remains unclear. So far as we know, ours is the first report of IFN activation in the target cells by shRNAs delivered by lentiviral transduction. A previous report of IFN induction by lentiviral vector-expressed shRNA only examined the IFN generated in the vector-producing cells, which then up-regulated IFN-stimulated genes in the transduced cells (10) . The distinction is important as lentiviral vectors used in a gene-therapy setting will likely be purified and free of any IFN that has been generated during the vector preparation step, but IFN activation in the target cells would pose a more serious concern. Our data suggest the importance of screening shRNAs for IFN induction in the transduced cells in vitro before largescale studies. An HIV reverse transcriptase inhibitor efficiently blocked IFN production by both sh-B971 and sh-PCAF when delivered by transduction, indicating the virion-encapsulated RNA was not able to trigger IFN activation. In this respect, it is interesting to note that positive-stranded RNA viruses, which produce dsRNA intermediates in the cytoplasm during replication (52) (53) (54) (55) , often replicate in membrane enclosed vesicles (56) , This sequestration of viral dsRNA in membranous structures may shield the RNA from the cytoplasmic PRRs and contribute to a successful infection. IFN-induction and RNAi by shRNAs appear to be independent functions of the same RNA (57). Our results also showed that IFN-induction by sh-B971 is independent of its ability to suppress target mRNA expression through RNAi. On the other hand, it might be possible to screen for duel functional siRNAs that confer therapeutic benefits by both RNAi and immunostimulation (58) . For example, siRNAs that target either viral genomes or cellular cofactors of the viruses can be screened for their ability to trigger IFN activation in hopes of find 'super siRNAs' with increased efficacy against IFN-sensitive viruses.

Peptide-Mediated Cellular Delivery of Oligonucleotide-Based Therapeutics In Vitro: Quantitative Evaluation of Overall Efficacy Employing Easy to Handle Reporter Systems

Cellular uptake of therapeutic oligonucleotides and subsequent intracellular trafficking to their target sites represents the major technical hurdle for the biological effectiveness of these potential drugs. Accordingly, laboratories worldwide focus on the development of suitable delivery systems. Among the different available non-viral systems like cationic polymers, cationic liposomes and polymeric nanoparticles, cell-penetrating peptides (CPPs) represent an attractive concept to bypass the problem of poor membrane permeability of these charged macromolecules. While uptake per se in most cases does not represent the main obstacle of nucleic acid delivery in vitro, it becomes increasingly apparent that intracellular trafficking is the bottleneck. As a consequence, in order to optimize a given delivery system, a side-by-side analysis of nucleic acid cargo internalized and the corresponding biological effect is required to determine the overall efficacy. In this review, we will concentrate on peptide-mediated delivery of siRNAs and steric block oligonucleotides and discuss different methods for quantitative assessment of the amount of cargo taken up and how to correlate those numbers with biological effects by applying easy to handle reporter systems. To illustrate current limitations of non-viral nucleic acid delivery systems, we present own data as an example and discuss options of how to enhance trafficking of molecules entrapped in cellular compartments.

Oligonucleotide-based strategies which can be used to modulate a vast variety of cellular functions represent a promising alternative to conventional therapies (for a review see: [1, 2] ). Among the different oligonucleotides with therapeutic potential are aptamers, transcription factor-binding decoy oligonucleotides, ribozymes, triplex-forming oligonucleotides (TFO), immunostimulatory CpG motifs, antisense oligonucleotides, small interfering RNAs (siRNAs) and antagomirs. Nowadays, these potential macromolecular drugs are generally either relatively easily derived by rational design (e.g. antisense or siRNA) or straightforward selection processes (e.g. aptamers). One of their main advantages over protein-or peptide-based approaches comprises the high specificity for their target while being non-immunogenic. However, despite these advances, a major impediment to the development of nucleic acid-based strategies for treatment and prevention of diseases is the relatively inefficient means to effectively deliver these macromolecules into the desired target cells. Although viral vectors have been widely used to transfer genetic material into cells [3, 4] , they bear an inherent risk for the patient to encounter severe immunological responses or even develop cancer [5] [6] [7] [8] . As a result of these problems much attention has been paid in recent years to the development of non-viral delivery systems. This conception *Address correspondence to this author at the Institut für Molekulare Medizin, Universität zu Lübeck, Ratzeburger Allee 160, 23538 Lübeck, Germany; Tel: +49-451-500-2745; Fax: +49-451-500-2729; E-mail: restle@imm.uni-luebeck.de includes an assortment of fairly unrelated approaches yielding various degrees of enhanced cellular uptake of nucleic acids. Currently, liposomes and cationic polymers are used as a standard tool to transfect cells in vitro. However, these procedures are characterized by a significant lack of efficiency accompanied by a high level of toxicity rendering them mostly inadequate for in vivo applications. In this context cell-penetrating peptides (see below) represent an interesting alternative as they generally are less toxic than liposomes or cationic polymers. Moreover, they are commonly better suited to transfer cargo into different cell types like non-adherent cells and primary cells, which are hard to transfect using commercially available standard protocols. The most advanced approaches in the field, which are not subject of the present article, are complex carrier systems combining vantages of assorted strategies to generate nanoparticles with better defined properties aimed towards enhanced uptake as well as intracellular trafficking in combination with cell-specific functionalities. For example, there are attempts to combine peptides with cationic liposomes [9] [10] [11] [12] [13] [14] [15] [16] or polyethyleneimine (PEI) [17] . Other strategies are aimed towards the synthesis of high or low molecular weight branched polymers and/or peptides [18] [19] [20] [21] [22] or dendrimers [23, 24] . Even more complex systems are particularly promising with respect to in vivo delivery [25] [26] [27] [28] [29] . In this review we will report about particular aspects of non-viral oligonucleotide delivery in vitro, pinpointing the current limitations, and provide quantitative means for determining where the bottlenecks of such strategies at present are. The focus of this article is recent progress in the field of peptide-mediated cellular delivery of siRNA and steric block oligonucleotides in cell tissue culture as a starting point for further developments illustrated by own experimental data. Our intention is not to provide the reader with easy solutions on how to solve the existing problems encountered with such approaches but give some hints where to start optimizing a particular approach. The idea of using peptides as carriers goes back some twenty years when it was discovered that the HIV-1 transactivating protein Tat is taken up by mammalian cells [30, 31] . A few years later, the Antennapedia homeodomain of Drosophila melanogaster was shown to act similarly [32] . Later on, it could be shown that peptides derived from Tat and Antennapedia as well as other proteins are capable of transporting macromolecular cargo molecules into cells [33] [34] [35] . Based on such promising results, a rapidly expanding field focusing on the so-called cell-penetrating peptides (CPPs), also referred to as protein transduction domains (PTD), began to develop. Since the first reports about Tat, a large number of naturally occurring as well as engineered CPPs have been discovered [36] [37] [38] [39] [40] [41] [42] . Table 1 gives an overview of selected "classical" CPPs. Generally, CPPs are short polycationic sequences of less than 30 amino acids that are able to translocate different cargoes (e.g. nucleic acids, peptides and even entire proteins) into cells. The only common characteristic of these peptides appears to be that they are amphipathic and net positively charged at physiological pH. Frequently the cargo is covalently attached to the CPP which can be achieved by expression as a fusion construct or by chemical coupling (for a review see: [43] ). In particular cases, cargo and carrier bind each other non-covalently through mainly ionic interactions [40, 44, 45] . Despite the widespread interest in peptide carriers, the mechanisms underlying the cellular translocation of CPPs are poorly understood. Early work relied upon fluorescence imaging or flow cytometry analysis of chemically fixed cells to examine intracellular localization of fluorescently labeled peptides in the absence or presence of cargo. According to these experiments peptides appeared to be internalized very rapidly within minutes even at 4 °C. From such observations it was concluded that CPPs penetrate cell membranes by an energy-independent mechanism [46] [47] [48] [49] [50] . Although it had been reported quite early on that certain fixation procedures may cause artefacts leading to an overestimation of the cellular uptake rates [51] [52] [53] the dimension of this problem was not commonly recognized until a side by side comparison of fixed and living cells was published [54] . Based on these findings, many groups re-examined their data. However, despite considerable technical improvements, there are still puzzling controversial results concerning the exact mechanism of CPP uptake. Though in most cases endocytosis has been suggested to be the main route of internalization (Fig. (1A) ), substantial difficulties are encountered identifying the exact pathway ( [42, 55] and references therein). Prior to endocytosis CPPs interact electrostatically with the extracellular matrix of the cell surface mostly through binding to negatively charged glycosaminoglycans, i.e. heparan sulfate proteoglycans [56] [57] [58] [59] . Recent studies indicate that the uptake mechanism of CPPs can be influenced by the attachment of cargos. For example, Richard et al. [54, 60] reported a colocalization of Tat 48-59 with markers of clathrin-mediated endocytosis, whereas Fittipaldi et al. [61] found a caveolae/lipid raft-dependent process for a Tat-GFP fusion protein and Wadia et al. [62] described a macropinocytotic uptake pathway for a fusion construct of Tat peptide with Cre recombinase. In summary, the precise mechanism of internalization remains elusive and strongly depends on the properties of both CPP and cargo as well as on the transfection conditions and the cell lines used [63] [64] [65] [66] [67] [68] . As opposed to the majority of CPP applications reported, which rely on covalent linkage of carrier and cargo, limiting their general use considerably as a new construct has to be generated as well as tested for any given nucleic acid cargo, we will focus in this article on a peptide termed MPG which forms highly stable non-covalent complexes with nucleic acids (Fig. (1) ). The peptide is a derivative of the original MPG peptide described by Morris and coworkers [47] and differs by five amino acids in the hydrophobic part. These changes result in an alteration of the overall structure of the peptide towards a higher tendency of adopting a helical conformation [69] . Accordingly, the two peptides behave penetratin (Antp 43-58 ) RQIKIWFQNRRMKWKK [203] transportan GWTLNSAGYLLGKINLKALAALAKKIL [204] TP10 AGYLLGKINLKALAALAKKIL [205] Oligoarginine (R8) RRRRRRRR [50] MAP KLALKLALKALKAALKLA [163] MPG GALFLGFLGAAGSTMGAWSQPKKKRKV [47] MPG GALFLAFLAAALSLMGLWSQPKKKRKV [69] differently with respect to their interaction with artificial lipids as well as Xenopus oocytes [70, 71] and most probably, their exact mechanism of uptake is not the same. Besides ionic interactions responsible for the initial peptide/nucleic acid complex formation, hydrophobic peptide/peptide interactions drive the maturation of large nanoparticles in a sandwich-like assembly reaction (Fig. (1B) and Fig. (2) ). In recent years, RNA interference (RNAi) has gained a lot of interest as a tool for functional genomics studies and probably equally important as a promising therapeutic approach for the treatment of various diseases [72, 73] . RNAi is a highly evolutionally conserved and specific process of post-transcriptional gene silencing (PTGS) by which double stranded RNA (dsRNA), when introduced into a cell, causes sequence-specific degradation of homologous mRNA sequences [74, 75] . Mechanistically the process can be divided into two steps. An initiator step where dsRNA is cleaved by dicer, a member of the RNase III family, into 21-25 nt long small interfering RNA (siRNA) fragments [76] . In a consecutive step, these fragments are transferred to RISC (RNAinduced silencing complex) where one of the strands, the so called guide strand, serves as a molecular template to recognize homologous mRNA that is cleaved by Argonaute [77, 78] , a protein component of RISC. Once the guide strand is bound to RISC this complex can undergo many rounds of mRNA binding and cleavage ( [79] , Fig. (3A) ). To circumvent application of long double stranded RNAs, which inevitably trigger an interferon response, it is sufficient to extracellularly supply 21 nt long dsRNAs [80, 81] . Alternatively, siRNAs can be expressed endogenously using DNA vectors which code for short hairpin (sh) RNAs [82] [83] [84] . These shRNAs are than cleaved by dicer to siRNAs. Short hairpin RNA constructs have advantages over siRNA because the effects of these constructs can lead to a more stable and long-term result. On the other hand, besides the fact that shRNAs might interfere with the microRNA pathway [85, 86] , this strategy requires a gene therapy approach in the long run [87] . For this reason we will not cover shRNAs. As described above, siRNAs represent a valuable tool to inhibit the expression of a target gene in a sequence-specific manner. In the following section, selected examples of CPPmediated siRNA delivery will be presented which are summarized in Table 2 . Only a few studies describe the covalent attachment of nucleic acid cargo and peptide carrier (confer Table 2 ). In one approach, simple mixing of siRNA targeted against GFP or CDK9 and Tat peptide did not generate any measurable RNAi effect whereas cross-linked siRNA-Tat 47-57 led to a significant down-regulation of the target proteins. However, high concentrations of siRNA (about 300 nM) had to be used [88] . Both LF-and Tat 47-57 -mediated transfections resulted in a perinuclear localization of siRNA. In contrast, fluorescently labeled Tat 47-57 without cargo was mainly found in the nucleolus, suggesting that interactions with RISC influence subcellular localization. In another approach, significant uptake of siRNAs targeted against luciferase or GFP could be observed after disulfide coupling the 5'-end of the sense strand to penetratin or transportan [89] . Compared to LF2000, slightly higher levels of transfection were achieved. Interestingly, after LF2000-mediated transfection, basal luciferase activity returned to normal levels one day earlier than after CPP-mediated transfection although the same concentration of siRNA was applied. A remarkably strong RNAi effect in hard to transfect primary neuronal cells was reported by Davidson et al. [90] . Here, siRNAs directed against several endogenous proteins were coupled to penetratin via a disulfide bond. The observed down regulation of the target proteins after peptide-mediated siRNA delivery was found to be far more effective compared to LF2000. This was in part attributed to the toxicity of the lipids. As one of the first groups to report on Tat 48-60 -or penetratin-mediated siRNA delivery in vivo, Moschos et al. showed, that intratracheal administration of the conjugates did not lead to any intensification of the knockdown of the target gene p38 mitogen-activated protein kinase in mouse lungs in comparison to unmodified non-formulated siRNA [91] . Strikingly, it was found that the peptides alone triggered a detectable decrease in target gene expression and that the penetratin-conjugate induced elevated levels of the immune markers IFN-, TNF-, and IL-12p40 in lung tissue. Besides technical difficulties arising from the syntheses of conjugates consisting of short cationic or hydrophobic peptides and highly negatively charged siRNAs, Dowdy and his group [92] present a rather critical point of view referring to previous studies with CPP-siRNA-conjugates. They claim that the successful delivery described therein is solely the result of excess free peptide, which leads to additional complexation, and thereby cellular import of the siRNA. This is in accordance with Turner et al. [93] , who were the first to observe that careful purification of CPP-antisense-conjugates abrogates their biological effect. Among other things, this might be the reason why most of the studies reporting on successful peptide-mediated delivery of siRNAs use a noncovalent complexation approach (confer Table 2 ). In 2003, Simeoni et al. [94] were the first who noncovalently complexed siRNA with the peptide MPG. At a 1:10 ratio of negative nucleic acid to positive peptide charges a decrease in luciferase activity of about 80 % was detectable in HeLa or Cos-7 cells. This effect was further enhanced to about 90 % down-regulation by a mutation in the NLS sequence of the carrier peptide (MPG NLS ), presumably due to an increased delivery to the cytoplasm, where RISC is localized. Recently, Veldhoen et al. [55] used a derivative of the MPG peptide for the delivery of siRNA, which will be described in the chapter "MPG -mediated delivery of siRNA and steric block oligonucleotides". Leng et al. [21] presented promising results with a prospect for cell-specific siRNA delivery. Different versions of a branched histidine/ lysine-polymer (H3K8b) yielded up to 80 % knockdown of the target gene in several cell types. Structure-function studies revealed an important role of the composition of the histidine-rich domain as well as its position within the peptide and the branches for siRNA delivery, whereas size and surface charge did not have any effect. Furthermore, the toxicity was much lower than for the commercial cationic lipids Oligofectamine and LF2000. Finally, the attachment of the tripeptide RGD, an integrin-ligand, slightly enhanced siRNA delivery and turned this carrier into a cell-specific system. A Fig. (1) . Simplistic scheme of peptide-based nucleic acid delivery systems (A). Interaction of CPP and cargo is either achieved by covalent attachment or by non-covalent complexation through mainly ionic interactions. In case of non-covalent complex formation, a further assembly of cargo/carrier complexes occurs, leading to the formation of large nanoparticles (confer Fig. (2) ). In case of covalently joined molecules a similar scenario is less likely, yet cannot be excluded. Prior to the translocation process the particles attach to the cell surface by ionic interactions of positively charged CPP residues with negatively charged membrane components. Subsequently, complexes are taken up via an endocytotic pathway. Although less likely, direct penetration cannot be excluded and may occur simultaneously. Once inside the cell, the cargo has to escape from vesicular compartments, otherwise it eventually gets degraded in the lysosome. Red: negative charges, blue: positive charges, green: hydrophobic domains. Three-dimensional model of MPG /siRNA interactions (B). The model was generated by iterative rigid body docking cycles of siRNA (PDB 1R9F) and peptide using the program Hex 4.2 [201] . The PDB file of MPG was generated with the program ICM (Molsoft LLC) taking into consideration different secondary structure predictions and energy minimization protocols. Out of many docking solutions particular ones were picked for illustration purposes using the program Chimera [202] . The phosphate backbone of the siRNA is shown in red, the nucleobases in light gray. Aliphatic, aromatic and hydrophobic residues of the peptide are shown in green, positive charged residues in blue and the remaining amino acids in gray. It is assumed that formation of larger particles is driven by hydrophobic peptide/peptide interactions generating free positive charges where other siRNA molecules can interact. This eventually drives complex formation in a sandwich or mesh like assembly reaction. In principle such a scenario holds true for any given nucleic acid cargo. similar concept has very recently been used by Kumar et al. [95] for a specific delivery approach into the brain. A peptide derived from rabies virus glycoprotein (RVG) interacts specifically with the nicotinic acetylcholine receptor (AchR) on neuronal cells to enable viral entry. The authors could show that the biotinylated form of the 29-amino-acid peptide (YTIWMPENPRPGTPCDIFTNSRGKRASNG) was taken up by neuronal cells. In order to transport nucleic acids with this vehicle, R 9 was conjugated to RVG peptide. Systemic treatment of mice with siRNA in a non-covalent complex with this modified peptide promoted a highly specific cellular import of siRNA only into cells expressing AchR. Even more important, an antiviral siRNA treatment resulted in successful protection of mice against encephalitis caused by Japanese encephalitis virus (JEV). This is the first study to report on a non-toxic method to deliver siRNA across the blood brain barrier which could help to circumvent dangerous and ineffective injections into the brain. To date it presents one of the most promising tissue-specific delivery approaches which might be expandable to other in vivo applications. Along these lines, most studies today are performed with the aim of CPP-mediated siRNA delivery in vivo. Although many of them are already showing promising results, e.g. concerning tumor-targeting and ocular delivery [96] [97] [98] , this is beyond the scope of this review and will be discussed elsewhere in this issue [99, 100] . With the aim to increase the endosomal escape of siR-NAs after peptide-mediated delivery, Lundberg et al. [101] rationally modified penetratin to form a CPP (termed EB1) with improved endosomolytic properties. They achieved a pH-dependent conformational change of the peptide to a higher degree of helicity by the replacement of two basic amino acids with histidines and the N-terminal addition of six amino acids. In this study, several CPPs were compared in a non-covalent approach by measuring the overall cellular stearyl-R8 n-c EGFP, MAP2B primary rat hippocampal neurons [206] R8-MEND (siRNA/stearyl-R8 core) n-c luciferase HeLa [207] uptake via fluorescence and biological effect of siRNA targeted to luciferase mRNA. Penetratin-as well as TP10mediated transfection did not lead to any silencing of luciferase gene expression, despite high amounts of intracellular siRNA [101] and in contrast to previous reports using siRNA-penetratin-conjugates [90] or TP10/DNA-complexes [102] . EB1-mediated delivery of 100 nM siRNA led to approximately 50 % reduction of luciferase activity. This silencing effect was slightly better than for bPrPp and in the same range as for MPG NLS , but still not as pronounced as for LF2000-mediated transfection of 100 nM siRNA. As it was described earlier, that addition of a pH-sensitive peptide derived from hemagglutinin (HA2) can promote endosomal escape [62] , the authors linked HA2 to penetratin [101] . It turned out that although HA2-penetratin improved the silencing effect when coincubated with penetratin, EB1 was more potent than this combination of peptides. Together with confocal microscopy studies the authors concluded that the lack of biological effect after penetratin-mediated siRNA delivery is due to a lack of endosomal escape and that EB1 has a superior endosomolytic activity in comparison to HA2penetratin. Endoh et al. [103, 104] very recently presented an innovative strategy, called CLIP-RNAi (i.e. CPP-linked RBPmediated RNA internalization and photo-induced RNAi) combining delivery of a specific RNA sequence with enhanced photoinduced release of RNA from endosomes. This goal was accomplished by fusing the U1A RNA-binding domain (RBD) to the Tat peptide and extending the siRNA with a short stretch of nucleotides specifically recognized by this RBD. These complexes were efficiently internalized but exhibited a punctuate cytoplasmic localization pattern, indicative of endosomal entrapment. However, photostimulation of a fluorophore attached to the peptide led to a redistribution of complex into the cytosol followed by efficient RNAi-mediated gene silencing. Human pre-mRNAs contain on average eight expressed sequences (exons) with an average length of 150 nt and up to 60 intervening sequences (introns) which can vary in length between 35 and 10,000 nt, therefore comprising up to 90 % of each transcriptional unit. In the nucleus, ribonucleoprotein complexes called spliceosomes recognize exon-intron boundaries and catalyze the precise removal of introns and subsequent joining of exons in a process called RNA splicing [105] [106] [107] . Additionally, each primary transcript can yield different mature RNAs through alternative splicing, thereby expanding the information content and versatility of the transcriptome, e.g. through the production of protein isoforms. A recent study of 10,000 human genes revealed, that at least 70 % of all multi-exon genes are alternatively spliced [108] . There are several different types of alternative splicing, amongst others affecting transcription start sites, splice sites, polyadenylation sites or even whole introns and exons. Disruptions of these intricate splicing patterns are tightly coupled with human pathophysiology, either as a determinant or a direct cause of disease or as a modifier of disease susceptibility and severity [109] . Among these diseases are -thalassemia, cystic fibrosis, muscular dystrophies, Frasier syndrome, certain kinds of dementia and cancer. A more de-tailed description of the underlying mechanisms is beyond the focus of this article and can be found in a number of reviews [110] [111] [112] [113] . López-Bigas et al. [114] proposed that 60 % of mutations that cause disease lead to splicing defects rather than changes in the amino acid sequence. Two common forms of mutations are depicted in Fig. (3B) . On the one hand, a mutation in the splice donor can favor recognition of a cryptic splice donor and result in a mutant mRNA containing additional intronic sequences (part I). On the other hand, a mutation in the splice acceptor can lead to skipping of a whole exon and result in a shortened mRNA (part II). Both scenarios have been used in the context of antisense oligonucleotide-mediated approaches targeting alternative splicing. The use of antisense oligonucleotides interacting with mRNA to affect protein production goes back some 20 years [115, 116] . Since then, three principle mechanisms have been exploited for this purpose (for a review see: [117, 118] ): (I) the oligonucleotide/RNA duplex forms a substrate for endogenous RNase H, leading to mRNA cleavage; (II) the oligonucleotide/RNA duplex prevents the productive assembly of the ribosomal complex or arrests a ribosomal complex already engaged in translation, in both cases affecting protein biosynthesis; (III) the oligonucleotide/RNA duplex alters pre-mRNA splicing in the nucleus. The following section will focus on the last approach with the aim to treat splicing disorders and give examples of possible applications for CPPs in this context. Different forms of human -thalassemia are caused by mutations within in the -globin intron 2, which activate cryptic splice sites and thus lead to the formation of nonfunctional transcripts (Fig. (3B part I) ). Those aberrantly used sites can be blocked by antisense steric block oligonucleotides, which leads to the synthesis of functional protein [119, 120] . Kole and his group adopted this principle for the development of a splice correction assay [121] . In this model system, a firefly luciferase construct leads to the synthesis of inactive enzyme because the reporter gene pre-mRNA is interrupted by the human -globin intron 2 containing an aberrant splice site. Upon binding of a steric block oligonucleotide, correct splicing is restored which in turn yields a functional luciferase protein. To achieve this, the oligonucleotide has to be delivered to the nucleus. Furthermore, only oligonucleotides that don't activate RNase H are applicable [2] , e.g. phosphorodiamidate morpholino oligomers (PMO, [122] ), locked nucleic acids (LNA, [123] ), peptide nucleic acids (PNA, [124] ) or 2'-O-methyl-modified oligonucleotides (OMe). In addition to their inability to activate RNase H, most of these modifications confer higher affinity to the target RNA and increased resistance against enzymatic degradation than unmodified versions. Compared to RNAibased model systems, this assay is less susceptible to side effects like cytotoxicity or off-target effects because the reporter gene activity is turned up rather than turned down. The splice correction assay has been successfully applied for the analysis of several carrier systems [24, 93, [125] [126] [127] [128] [129] [130] [131] [132] [133] [134] [135] . In the following section, selected examples will be presented, which are summarized in Table 3 . In contrast to many noncovalent CPP-mediated siRNA delivery approaches, efficient splice correction was only achieved with conjugates of peptide and steric block oligonucleotide. Astriab-Fisher et al. [128] described delivery of OMe RNA phosphorothioate oligonucleotides linked via a disulfide bridge to Tat peptide and penetratin. A few hours after transfection, the CPP-oligonucleotide conjugates were detected both in cytoplasmic vesicles and in the nucleus and caused a dose-dependent increase in luciferase activity. These findings are in contrast to results of Turner et al. [93] , who could not find a biological effect for several CPP-oligonucleotide conjugates in a HeLa cell assay for Tat-mediated transactivation of the HIV-1 long terminal repeat. The authors observed vesicular uptake but no nuclear import for their highly pure conjugates. Interestingly, the rate of uptake could be enhanced by addition of free CPP to the conjugates, though still no biological activity was detected. Based on these findings, Turner et al. [93] concluded that these free CPPs form complexes with CPP-cargo conjugates, which play a significant role in the uptake process. This is in accordance with observations by Meade et al. [92] for the uptake of CPP-siRNA conjugates described above. Moulton et al. [136] achieved correction of missplicing at low micromolar concentrations of a R 9 F 2 -PMO conjugate but not with complexes of peptide and PMO. The steric block activity of the R 9 F 2 -PMO conjugates could be further increased with longer spacers whereas variations in the conjugation chemistry did not result in any differences. Furthermore, transfection rates were higher than for conjugates with Tat peptide, penetratin or a Tat peptide analogue. Using the HIV-1 transactivation assay mentioned above, Turner et al. [137] could show that most CPP-oligonucleotide conjugates attained biologic activity only through co-administration of the endosomolytic substance chloroquine. Fluorescence microscopy analyses revealed that this treatment released fluorescently labeled conjugates from endosomal compartments into the nucleus. Besides the addition of chloroquine, different endosome disrupting strategies have been evaluated using the splice correction assay, for example co-treatment with endosome-disruptive peptides [129] or photochemical internalization [138] (see chapter "Strategies to enhance endosomal escape"). However, the most promising results have been achieved with two newly developed derivatives of classical CPPs (reviewed in [130] ). The modification of oligoarginines with non-natural, uncharged amino acids [139] led, amongst others, to the peptide (R-Ahx-R) 4 , in which Ahx represents a six-atom aminohexanoic acid spacer. Abes et al. demonstrated that in contrast to Tat or oligoargine, PMO-conjugates of this peptide led to dose-dependent splice correction at low micromolar concentrations in the absence of endosomolytic agents. The underlying mechanism for this superior activity is not clear yet, as the uptake of (R-Ahx-R) 4 constructs was less efficient than the uptake of Tat or oligoarginine constructs and also involved endocytotic routes [126] . The second peptide is a derivative of penetratin, to which six arginine residues were added at the N-terminus (R 6 Pen). R 6 Pen-PNA conjugates were shown to promote efficient splice correction at low concentrations and in the absence of endosomolytic agents [127] . Again, uptake of R 6 Pen-conjugates seemed to involve endocytosis and there was hardly any difference in splice correcting activity regardless of the nature of the linker used for conjugation, e.g. a stable thioether versus a reducible disulfide linker [130] . Part II of Fig. (3B) illustrates a phenomenon that represents a strategy for the treatment of Duchenne muscular dystrophy (DMD). DMD is a severe progressive neuromuscular disorder caused by several different mutations in the dystrophin gene that abolish the production of functional protein [140] . Depending on the location of the mutation, the corresponding exon is skipped by covering the responsible splice sites with steric block oligonucleotides. This allows the transcription of internally deleted, but largely functional, dystro-phin proteins and converts a severe DMD into a milder Becker muscular dystrophy phenotype. A more detailed description of this approach and its application in a number of animal models can be found in several excellent recent reviews [141, 145] . Successful systemic delivery of splice switching oligonucleotides with or without chemical modifications (PMO, LNA, OMe) has been accomplished via injection of naked nucleic acids [146, 147] , with the help of viral vectors [148] , through re-implantation of ex vivo manipulated stem cells [149] or in combination with CPPs [150] [151] [152] [153] [154] . For the latter purpose, several studies were carried out with novel derivatives of arginine-rich peptides containing different numbers of non-amino acids, e.g. aminohexanoic acid and/or -alanine. These CPP-PMO conjugates showed higher serum stability, less endosomal trapping and led to efficient exon skipping in myoblasts and mice at lower dosages than the splice switching PMO alone [152, 153] . Yin et al. [154] used a PNA-modified splice-switching oligonucleotide conjugated to Tat, a muscle specific peptide (MSP) or different functional domains of the adenovirus capsid protein VP1 (AAV6, AAV8) and examined exon skipping efficiency in vitro and in vivo. Surprisingly, both after transfection and intramuscular injection, the activity of these PNA-peptide conjugates was not significantly better than that achieved by naked neutral PNA, presumably due to endosomal trapping. Intracellular trafficking represents one of the major limitations of current non-viral nucleic acid delivery approaches [155] . In other words a large percentage of intracellular cargo molecules are entrapped in vesicular compartments and thus will not trigger the desired effect. Moreover, degradation or retrograde transport might further reduce the number of active molecules. So in order to determine the overall efficacy of a given delivery approach it is essential to know the numbers of intact cargo molecules inside the cell along with the minimal numbers of molecules required to cause a particular effect. Based on such information one can easily calculate the percentage of bioactive molecules. In the following chapter we will briefly describe selected examples of variable suitability for a quantitative determination of nucleic acids in a cellular context. In principle, either the peptide or the cargo can be labeled by a reporter group, e.g. a radioisotope [156] or a fluorophore [157] . Fluorescent peptides or cargos have been quantitatively evaluated by FACS [54] or fluorescence correlation microscopy (FCS) [158] or FRET [159] . In all cases, it is crucial to distinguish between internalized and membraneassociated signals. For this purpose, a simple wash step with just buffer is not sufficient to completely remove membranebound peptide/cargo-complexes [54, 55] . Extracellularly bound complexes can be efficiently removed for example by enzymatic digestion with trypsin [160] , acid wash [161] or heparin treatment [55, 162] . Alternatively, discrimination between intra-and extracellular material is possible through chemical modification of extracellular components [163] or fluorescence quenching [164] . Having established that only intracellular signals are taken into account, it is still challenging to distinguish between intact and degraded forms of peptide or cargo. In two studies, a fluorescence-based quantification method was combined with either HPLC analysis [163] or "cell activity by capillary electrophoresis" [165] to verify the integrity of cargo and carrier. Recently, a technique to measure cellular uptake of CPPs by matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) was reported by the group of Burlina [166] [167] [168] . This quantification is based on the addition of an internal standard, i.e. a peptide with a stable isotope label. The method has been used to determine the amount and stability of intact internalized peptides, e.g. Penetratin, R9 and several novel CPPs, and can also be used for the quantification of peptidic cargoes, e.g. an inhibitor of protein kinase C [166] . Varga et al. [169] [170] [171] developed an "integrative systems" approach, combining quantitative experiments and computational modeling studies of vector uptake and trafficking kinetics, with the aim to take multiple potentially rate-limiting cellular and molecular processes into account. By applying their mathematical model to plasmid delivery with either Lipofectamine, several PEI-based vector formulations or an adenoviral vector, they could successfully predict experimentally observed effects and identify endosomal escape as the most important rate-limiting intracellular barrier for non-viral vectors. Recently, Zhou et al. [172] applied a similar strategy for the characterization of a novel lipopolymer (WLSP). This carrier shows an increased rate of endosomal escape compared to conventional PEI-based carriers. With the aim to quantify rhodamine-labeled plasmid DNA in cellular compartments while avoiding problems arising from subcellular fractionation, like recovery and leakage, Akita et al. [173] developed a novel quantitative strategy, the confocal image-assisted three-dimensionally integrated quantification (CIDIQ) method. To distinguish endosomes/lysosomes and the nucleus from the cytosol, they were stained with LysoSensor DND-189 and Hoechst 33258, respectively, and sequential Z-series images were captured by CLSM. By applying this quantification method, the authors could show that due to a rapid endosomal escape, Lipofectamine Plus delivered more plasmid into the nucleus than R8 or stearylated R8. Hama et al. [174] used the same method in combination with TaqMan PCR to evaluate the uptake and intracellular distribution of plasmid DNA after delivery with viral as well as non-viral vectors. Due to superior cell surface binding, the efficiency of cellular uptake was significantly higher for Lipofectamine Plus than for adenovirus whereas intracellular trafficking, i.e. endosomal escape and nuclear transfer, were essentially the same. However, to achieve comparable transgene expression, 8000-fold higher intranuclear plasmid numbers were required in case of Lipofectamine Plus. This finding suggests a difference in nuclear transcription efficiency after non-viral delivery. In another approach, Jiang et al. [175] extended the 3' end of the sense strand of a siRNA with a nuclease-resistant DNA hairpin to obtain a so-called "crook" siRNA. This modification had no effect on RNAi-mediated reporter gene inhibition and served as a primer for a filling-in reaction followed by PCR. Parameters were chosen so that the initial rate of template amplification correlates with the initial con-centration of the "crook" siRNA. Under these conditions, quantification of attomolar siRNA levels per cell was possible after liposomal transfections with Oligofectamine. A highly sensitive method was developed by Overhoff et al. [176] for the detection of siRNA after phosphorothioatestimulated uptake [177] and adapted for the quantification of siRNA or steric block oligonucleotides after non-covalent peptide-mediated delivery ( [55] and Laufer et al., manuscript in preparation, see chapter "MPG -mediated delivery of siRNA and steric block oligonucleotides"). This so-called liquid hybridization assay is based on the extraction of total cellular RNA and the subsequent hybridization in solution of a radioactively labeled probe which is complementary to the oligonucleotide to be detected. Finally, following PAGE analysis, absolute amounts of internalized oligonucleotide can be quantified with high accuracy down to ~10 molecules per cell using internal standards [55] . In addition to this outstanding sensitivity, no amplification step is needed and only intact oligonucleotides are taken into account. Considering the multitude of available CPPs, nucleic acid cargos and cellular as well as animal model systems, a comparison of different delivery strategies seems nearly impossible. In this context, we have for the first time undertaken a detailed side by side comparison of two different model systems using the peptide MPG as delivery agent. In the following paragraph we present own experimental data to exemplarily illustrate particular aspects regarding current limitations of peptide-based delivery systems. In contrast to many other CPP procedures, which rely on covalent linkage of carrier and cargo, the peptide MPG forms highly stable non-covalent complexes with nucleic acids, displaying binding constants in the low nanomolar range ( [55] , A. Trampe, unpublished data). The high flexibility of this non-covalent approach can be exploited to easily transport a wide variety of nucleic acid cargos without having to synthesize a new construct for each oligonucleotide. We have used MPG for the delivery of siRNAs in the context of an RNAi-based reporter system and for the delivery of steric block oligonucleotides in the context of the splice correction assay described above [121] . After MPG -mediated transfection of a luciferase-targeted siRNA, we observed strong inhibition of reporter gene readout with an IC 50 in the subnanomolar range [55] . After MPG -mediated transfection of a luciferasetargeted steric block oligonucleotide (further on also referred to as ON-705), we observed a moderate up-regulation of reporter gene readout, representative of low splice correcting activity (Laufer et al., manuscript in preparation). One possible explanation for the different degree of reporter gene regulation could be the different intracellular target sites, i.e. the cytoplasm for siRNAs and the nucleus for splice correction oligonucleotides. To attain more information about the subcellular localization of MPG -oligonucleotide complexes, we performed confocal laser scanning as well as conventional fluorescence microscopy studies with fluorescently labeled siRNAs or steric block oligonucleotides. In both cases, a punctuate non-homogenous distribution of the nucleic acids inside the cells was observed. This pattern is indicative of an accumulation of nucleic acids in endocytotic vesicles, which was verified by coincubation with LysoSensor (Fig. (4A) ). A quantitative computational analysis of fluorescence microscopy data yielded an average of approximately 50 % colocalization between endosomes and siRNA (Fig. (4B) ). In contrast to earlier assumptions that CPPs directly traverse the lipid bilayer, it has commonly become accepted that for most peptide-cargo combinations endocytosis plays a major role in cellular uptake. As described above and in the chapter "Strategies to enhance endosomal escape" below, administration of endosome disruptive substances like chloroquine can greatly increase endosomal release of trapped nucleic acids. Chloroquine is a weak base, non-charged at neutral pH but charged at pH 5.5 [178] . It is able to pass easily through membranes in its uncharged form, but becomes protonated and accumulates within acidic vesicles in its positively charged, membraneimpermeable form. Although its exact mode of action has not yet been resolved, it is generally accepted that chloroquine works via prevention of endosome acidification which in turn increases the residence time of cargo within the endosomes eventually resulting in a higher probability of transfer to the cytoplasm. Fluorescence microscopy analyses in the presence of 100 M chloroquine yielded two quite contrary outcomes. While the localization of siRNA did not change after addition of chloroquine (Fig. (4C) ), for the steric block oligonucleotide the picture changed completely (Fig. (4D) ). In the latter case, a diffuse fluorescence all over the cytoplasm with an accumulation of ON-705 in the nucleus could be observed. Nonetheless, considerable amounts of nucleic acid molecules were still visible as a punctual pattern, which indicates that the chloroquine treatment liberates only a certain fraction. On the whole, these qualitative observations are in full agreement with the observed biologic effects in the absence or presence of chloroquine (Fig. (5A) ). For MPG -mediated transfection of siRNA, even under conditions where the amount of bio-available siRNA was severely limited, only a minor increase in RNAi (ca. 30 %) was measurable. For MPG -mediated transfection of steric block oligonucleotide, on the other hand, a dramatic increase of reporter gene up-regulation by a factor of 50 -100 was observed. However, in both cases, the overall uptake did not change upon incubation with chloroquine ( Fig. (5B) ), which proves that the endosomolytic substance does not interfere with uptake but leads to a re-distribution of internalized nucleic acids. The underlying mechanism for the different effects triggered by chloroquine in case of peptide/siRNA and peptide/steric block oligonucleotide complexes remains unclear. Though, this is a good example that the cargo can substantially affect the properties and thereby intracellular trafficking of a particular carrier system. Ultimately, to assess the overall efficacy of this carrier system, we were interested to elucidate which percentage of molecules taken up after MPG -mediated delivery is biologically active. To derive such information, the exact intracellular amount of intact oligonucleotide, the corresponding reporter signal and the minimal number of molecules necessary to trigger a specific degree of reporter gene modulation have to be known. For the quantification of internalized cargo, we adapted a highly sensitive method first described by Overhoff et al. [176] , enabling us to detect intracellular oligonucleotide amounts down to 10 copies per cell [55] . The method is based on the liquid hybridization of a radioactively labeled probe with the corresponding oligonucleotide in cellular lysates. In this context it should be noted that a stringent heparin wash following the transfection procedure is crucial to avoid an overestimation of intracellular nucleic acid molecules due to complexes attached to the outside of the cell membrane [54, 55] . In order to correlate the numbers derived from the quantification experiments with the minimal number of molecules essential to trigger the observed effect, an independent assay had to be established. The gold standard in this case is microinjection as this technique enables one to deliver definite amounts of nucleic acids with a high degree of bioavailability into the cytoplasm or the nucleus of a mammalian cell along with a low toxicity profile and great accuracy. Considering that after cytoplasmic microinjection only 12 siRNA molecules are sufficient for halfmaximal inhibition of reporter gene expression, one can estimate that of the 10,000 molecules measured after MPGmediated transfection, only ca. 0.1 % are biologically active ( Table 4) . For the splice correction assay, the numbers are different in terms of absolute numbers but the outcome remains the same. Compared to the 300,000 molecules sufficient for maximal splice correction after nuclear microinjection, of the 70,000,000 molecules required following peptide-mediated delivery only ca. 0.5 % are biologically active ( Table 5 ). In both cases >> 99 % of internalized oligonucleotides are most likely retained in endosomes and subsequently degraded in lysosomes after peptide-mediated delivery. Frankly, this is a sobering result and puts in numbers how much room for improvement there actually is. Though it certainly is not legitimate to generalize these findings, there are countless reports in the literature suggesting similar limitations for the majority of non-viral strategies. According to actual conceptions, siRNA enters a multiple-turnover pathway with one siRNA molecule capable of RISC-mediated cleavage of 50 or more mRNA molecules [79] . Even though the fate of steric block oligonucleotides is not really clear, it can be assumed that per splicing event one molecule is used up and translated into a functional mRNA molecule (e.g. single-turnover pathway). As a result the number of molecules needed to trigger an apparent effect is much higher in the splice correction assay compared to the RNAi-based reporter system (confer Table 4 and Table 5 ). On the other hand, being a single-turnover pathway, the splice correction assay should be much more sensitive to even minor changes in intracellular steric block oligonucleotide concentrations whereas the catalytic nature of the multiple-turnover RNAi mechanism might mask such small variations. This is in accordance with the data described above. Taken together, based on the results presented as well as unpublished data (Laufer et al., manuscript in preparation) , the splice correction assay appears to be the superior tool for a quantitative assessment of nucleic acid delivery strategies. As outlined above, endosomal release is one of the major rate-limiting steps for cellular delivery of macromolecules via cationic lipids, polyplexes and especially CPPs. In the following chapter we will present some examples of how to increase endosomal release. Transfections were performed with 2.1 M MPG and 1 nM siRNA or 10 g/ml LF2000 and 0.02 nM siRNA, i.e. in the range of the IC50 value [55] . Quantification was performed after 24 h according to the liquid hybridization protocol [55] . Molecules per cell were calculated based on the cell number seeded for transfection. For microinjection experiments, molecules per cell were calculated on the basis of the injection volume. Endosome-disrupting substances, like chloroquine, calcium or sucrose, were used to significantly enhance the activity of antisense PNA oligonucleotides conjugated to Tat, oligoarginines or oligolysines [125, 179, 180] . This effect did not result from increased uptake, but rather improved bioavailability in the cytoplasm or nucleus after endosomal escape. Takeuchi et al. [181] showed that by incubation of the target cells with pyrenebutyrate, delivery of arginine-rich peptides could be shifted from endocytic uptake to direct membrane translocation, yielding a rapid distribution of the peptide throughout the cytoplasm, even at 4 °C. Pyrenebutyrate acts as a counteranion and, by interacting with the positively charged peptide, increases the overall hydrophobicity, thereby facilitating a direct translocation through the lipid bilayer, as earlier shown with artificial membranes [182] . This method, which works only in the absence of a medium or serum, was successfully applied for administration of a fluorescent protein and an apoptosis-inducing peptide into dividing as well as non-dividing cells [181] . However, in general the strategies described above are not feasible for in vivo applications, due to high cytotoxicity or other undesirable secondary effects. The imidazole group of histidine (His) can absorb protons in the acidic environment of the endosome, leading to osmotic swelling, membrane disruption and eventually nucleic acid escape. Accordingly, Lo et al. [183] modified Tat, which can bind and condense DNA through ionic interactions but has no acidic residues that can promote endosomal release, with different numbers of His residues. Highest reporter gene expression could be achieved after plasmid delivery with a Tat peptide covalently fused to 10 His residues (Tat-10H). Insertion of two additional cysteine residues into Tat-10H further enhanced stability of peptide/DNA complexes and transgene expression through formation of interpeptide disulfide bonds. Youngblood et al. [184] evaluated the influence of the stability of arginine-rich peptide PMO conjugates on cellular uptake and antisense activity. They could show that the stability is affected by the amino acid composition and the type of linkage to the cargo. Moreover, they found that degraded fragments could not escape anymore from endosomal or lysosomal compartments. Another concept makes use of photosensitive substances, which induce the release of macromolecules from vesicles by light exposure. This so-called photochemical internalization (PCI) has, in the past, been used for intracellular delivery of a large variety of macromolecules (reviewed in [185] ) and, more recently, for the endosomal release of nucleic acids after delivery mediated by liposomes [186, 187] , polyplexes [188] or CPPs [138, 189] . Shiraishi et al. [138] investigated the biological activity of PNAs conjugated either to Tat, R7 or KLA-peptide in combination with a PCI treatment. Depending on the peptide, nuclear as well as cytosolic antisense effects could be enhanced by up to two orders of magnitude. Similar results were presented by Folini et al. [189] for a PNA targeting human telomerase reverse transcriptase conjugated to Tat. In both studies, lower nucleic acid doses were sufficient, thereby reducing the probability of off-target effects. In light of encouraging data from ongoing anticancer clinical trials employing photodynamic therapy [190, 191] , an in vivo application of PCI seems feasible and will be discussed in more detail by Oliveira et al. [192] in this issue. Furthermore, target specificity could be increased by local illumination of cells or tissue. Viral fusion proteins drive the fusion process between the viral membrane and the endosomal host cell membrane in a pH-dependent manner, which is required to translocate the viral genome into the cytoplasm after receptor-mediated endocytosis. Fusogenic peptides, usually hydrophobic, rich in glycine residues and found at the amino terminus of these proteins, were shown to have membrane perturbing and lipid mixing activities [193] . Many well studied representatives of this group are derived from the fusion sequence of influenza virus HA or HIV-1 gp41 [194] and have been used to improve the transfection efficiency of non-viral delivery systems [195] . Addition of the influenza-derived dimeric peptide diINF-7 to LF2000/siRNA complexes had no effect on the particle size of ca. 120 nm, but significantly improved gene silencing activity of siRNAs targeting the epidermal growth factor receptor or the K-ras oncogene [196] . Similar results were obtained for plasmid delivery through addition of a fusogenic peptide derived from herpes simplex virus glycoprotein H to Lipofectamine/DNA complexes [197] . To the same end, Futaki et al. [198] used the peptide GALA, which was specially designed to mimic the function of viral fusion sequences, together with various commercially available cationic liposomes. Although they could not detect significant differences in cellular localization, plasmid transfection efficiency was increased and liposomal dosage could be reduced. PEI covalently modified with the HIV-1 gp41-derived peptide HGP led to a 38-fold increase of gene expression after plasmid DNA delivery and also enhanced siRNAmediated knockdown of GAPDH by approximately 2-fold [199] . 30 % of cells incubated with PEI-HGP polyplexes showed not only the punctuate plasmid DNA staining observed with PEI polyplexes alone, but also a diffuse fluorescence throughout the cell, indicative of endosomal release of vectors. This would explain the observed increase in transfection efficiency, since the overall uptake was unaffected. Wadia et al. [62] were the first to use the influenza hemagglutinin-derived fusogenic peptide HA2 in combination with a CPP. Delivery of a Tat-HA2 conjugate with increasing concentrations of a Tat-Cre conjugate enhanced reporter protein activity, presumably through an increased release from macropinosomes. The same fusogenic peptide, linked to a polyarginine-p53 conjugate, promoted release of p53 from macropinosomes and subsequent translocation to the nucleus, accompanied by an enhanced anti-cancer effect [200] . The development of delivery systems for therapeutic oligonucleotides is a fast growing field. Owing to the enormous potential of short nucleic acids as alternative drugs such a growth is not unexpected. Besides viral vectors there is a highly diverse and constantly increasing number of non-viral systems evolving. However, despite considerable progress achieved in recent years, even the most advanced systems either lack the efficiencies required for downstream drug development or do show a substantial degree of toxicity or both. Of the many factors which limit their use, cellular uptake of the cargo/carrier complexes and subsequent intracellular trafficking to reach the target site are the most important. In addition to such essential considerations there are various additional parameters to be taken into account like serum stability, pharmacokinetic features and tissue barriers as well as target cell specificity. Now the question arises where to start optimizing a given delivery system. Currently, there is a clear trend towards in vivo testing. In principle such a development is a step in the right direction since many of the experimental data derived from artificial cell tissue culture systems with established cell lines are not applicable to the in vivo situation. On the other hand, it is questionable to what extent such animal experiments will eventually pay off as long as important fundamental problems remain largely unsolved. As outlined above, our quantitative studies along with microscopic analyses of siRNAs and steric block oligonucleotides using either a peptide or a commercially availably cationic lipid as carrier clearly show that less than 0.1% -5% of molecules taken up are involved in a biological response, i.e. RNAimediated down regulation or splice correction-mediated up regulation of reporter gene activity. Evidently, the vast majority of internalized cargo never reaches the target. This implies that uptake per se is not the limiting factor here. Although there are no such detailed quantitative numbers available in the literature for other systems, there are countless reports showing that to various degrees this applies to almost any non-viral delivery approach currently available. Taken together, if one would succeed to optimize intracellular trafficking this holds the potential to boost overall efficacy by up to 3 orders of magnitude. Moreover, it is reasonable to assume that such fundamental cellular restrictions can be adequately investigated in tissue culture without the use of animal models. So it might be worthwhile to reconsider the concept of maybe premature in vivo testing by moving backwards one step and first optimizing the systems with regard to intracellular limitations before dealing with the next level of complexity. Accordingly, in vitro model systems like the ones described above for siRNAs or steric block oligonucleotides are valuable tools to study particular aspects of nucleic acid delivery. However, currently available data are based on studies using a variety of different cell lines and techniques, which renders a direct comparison of different delivery approaches impossible. In this context it would be highly desirable to introduce standardized protocols for in vitro testing (e.g. the splice correction system developed by Kole and coworkers [121] ) together with methods for detailed quantitative analyses (e.g. the liquid hybridization protocol [55, 176] ). This would facilitate a direct quantitative comparison of exceedingly diverse approaches on at least the cellular level. One reason for the problem with intracellular trafficking of oligonucleotides arises from the mode carrier/cargo complexes are taken up by cells. Today it is well established that the majority of these complexes are taken up via endosomal pathways and therefore end up in vesicular compartments from which they have to escape in order to reach their target. Although there are many attempts reported to trigger endosomal escape by various strategies, they either proved to be toxic or did not achieve sustained success. Additionally, there might be a further reason for the encountered difficulties to overcome these intracellular barriers. It is not unreasonable to speculate that during evolution cells might have evolved mechanisms to avoid large amounts of foreign nucleic acids freely floating in the cytoplasm and thus safely contain them in vesicular compartments where they are eventually degraded. Alternatively they might be exported by retrograde transport. Co-evoluting viruses evidently have developed strategies to circumvent such defense mechanisms. So it might be somewhat naive to expect that a rather simple man-made carrier system is capable to efficiently overcome such an intrinsic barrier. Current developments towards more complex and elaborate carrier systems take into account such considerations. In any case it appears there is no simple solution for this problem. In conclusion, despite significant progress in the field of nucleic acid delivery in vitro as well as in vivo, there still is a long way to go before this will become a standard procedure in the clinic. Certain problems like endosomal escape are known for more than twenty years and still far from being resolved. In order to develop new strategies, more information about intracellular processes involved in nucleic acid trafficking is needed. Moreover, it would be desirable if the field would move from a qualitative description towards a quantitative evaluation preferentially using standardized model systems. This would allow for comparison of different approaches with one another. While animal studies are inevitable in the long run, there still is a lot of room for improvement on the cellular level. So it might be worthwhile to fathom how far we can push the different systems on this level.

Human pregnancy-associated malaria-specific B cells target polymorphic, conformational epitopes in VAR2CSA

Pregnancy-associated malaria (PAM) is caused by Plasmodium falciparum-infected erythrocytes (IEs) that bind to chondroitin sulphate A (CSA) in the placenta by PAM-associated clonally variant surface antigens (VSA). Pregnancy-specific VSA (VSA(PAM)), which include the PfEMP1 variant VAR2CSA, are targets of IgG-mediated protective immunity to PAM. Here, we report an investigation of the specificity of naturally acquired immunity to PAM, using eight human monoclonal IgG1 antibodies that react exclusively with intact CSA-adhering IEs expressing VSA(PAM). Four reacted in Western blotting with high-molecular-weight (> 200 kDa) proteins, while seven reacted with either the DBL3-X or the DBL5-ε domains of VAR2CSA expressed either as Baculovirus constructs or on the surface of transfected Jurkat cells. We used a panel of recombinant antigens representing DBL3-X domains from P. falciparum field isolates to evaluate B-cell epitope diversity among parasite isolates, and identified the binding site of one monoclonal antibody using a chimeric DBL3-X construct. Our findings show that there is a high-frequency memory response to VSA(PAM), indicating that VAR2CSA is a primary target of naturally acquired PAM-specific protective immunity, and demonstrate the value of human monoclonal antibodies and conformationally intact recombinant antigens in VSA characterization.

People living in areas of intense transmission of Plasmodium falciparum parasites acquire protective immunity to malaria during childhood, and the bulk of mortality and severe morbidity from P. falciparum malaria is therefore concentrated among young children. Protective immunity acquired in response to P. falciparum exposure appears to be mediated mainly by IgG antibodies specific for variant surface antigens (VSA) that mediate sequestration of infected erythrocytes (IEs) in various tissues (reviewed by Hviid, 2005) . Despite pre-existing protective immunity, women become highly susceptible to P. falciparum infection when they become pregnant, and pregnancyassociated malaria (PAM) is a major cause of mother/ offspring morbidity (Guyatt and Snow, 2001; 2004) . However, in areas of stable P. falciparum transmission, susceptibility to PAM rapidly declines with increasing parity, consistent with acquisition of PAM-specific protective immunity (reviewed by Hviid, 2004) . PAM is caused by P. falciparum-IEs selectively accumulating in the placental intervillous space through VSA PAM-mediated adhesion to chondroitin sulphate A (CSA). VSAPAM differ in several ways from VSA expressed on IEs obtained from males and non-pregnant females. Thus, only VSAPAM mediate binding to CSA in vitro (Fried and Duffy, 1996) and only VSAPAM-expressing IEs are consistently not recognized by IgG in the plasma of P. falciparum-exposed women who have never been pregnant or by IgG in plasma from similarly exposed men (Beeson et al., 1999; Ricke et al., 2000) . These observations, and the fact that plasma levels of VSAPAM-specific IgG increase with increasing parity (Fried et al., 1998; Ricke et al., 2000) , are consistent with evidence that these antibodies are the mediators of protective immunity to PAM (Duffy and Fried, 2003; Staalsoe et al., 2004) . The molecular identity of VSAPAM remains controversial, although current evidence points to VAR2CSA, an interclonally conserved member of the PfEMP1 molecules encoded by the multigene var family. Thus, transcription of the gene encoding VAR2CSA is increased among CSAadhering and placental isolates, VAR2CSA is exposed on the surface of CSA-adhering IEs (Salanti et al., 2003; 2004; Tuikue Ndam et al., 2005) , and plasma levels of VAR2CSA-specific IgG increase with increasing parity and correlate with protective immunity to PAM (Salanti et al., 2004) . However, the importance of VAR2CSAspecific antibodies relative to antibodies specific for other putative VSA PAM in acquired protective immunity to PAM remains to be established. The clonal analysis of memory B cells represents a powerful tool to dissect the immune response to complex pathogens such as P. falciparum (Lanzavecchia et al., 2006) . In this study, we used an improved Epstein-Barr virus (EBV) immortalization method (Traggiai et al., 2004) to analyse memory B cells from multiparous PAM-exposed women. Frequency analysis and isolation of specific monoclonal antibodies identified polymorphic, linear and conformationdependent epitopes in VAR2CSA as dominant targets of the human memory B-cell response to PAM. We first used flow cytometry to screen plasma from 27 PAM-exposed and recently pregnant multigravidae for IgG antibodies capable of staining P. falciparum-IEs expressing VSAPAM Ricke et al., 2000) . We selected three donors (one parity 2 and two parity 3 women) with high VSAPAM-specific plasma antibody levels and used frozen peripheral blood mononuclear cells (PBMC) obtained 1 month post-partum. Memory B cells were immortalized with EBV in the presence of CpG oligonucleotides and allogeneic, irradiated PBMC as described (Traggiai et al., 2004) . A total of 5760 replicate cultures of 100 immortalized B cells per well were set up, and after 3 weeks the culture supernatants were screened for their capacity to stain erythrocytes infected with each of three P. falciparum lines. Two of the lines (FCR3-CSA and NF54-VAR2CSA) had been previously selected in vitro to express VSA PAM, characterized by reactivity with IgG from multiparous women and lack of reactivity with IgG from P. falciparum-exposed men ( Fig. 1) (Fried et al., 1998; Beeson et al., 1999; Ricke et al., 2000) . The third line (3D7-SM) was selected to express non-PAM-type VSA equally recognized by IgG from P. falciparum-exposed men and women ( Fig. 1) Jensen et al., 2004) . Supernatants from 105 of the polyclonal B-cell lines stained one or both of the VSAPAM-expressing lines. The frequency of VSAPAM-reactive polyclonal supernatants varied from 6/1920 [0.3% (95% confidence interval: 0.1-0.7%)] to 33/1344 [2.5% (1.8-3.4%)] in the three donors. These results suggest that the frequency of VSAPAM-specific B cells can be high (at least up to 1 in 4000 memory B cells) in recently pregnant multigravidae. The higher memory B-cell frequencies in the present study compared with earlier reports for PfEMP1 (Dorfman et al., 2005) and total P. falciparum antigens (Fievet et al., 1993; Migot et al., 1995) probably reflect the efficient method of B-cell immortalization employed here. Cloning of EBV-immortalized IgG + B cells from 28 of the VSAPAM-specific lines by limiting dilution resulted in eight clones producing VSAPAM-specific IgG1. Lines were selected for cloning on the basis of their IgG synthesis Flow cytometry analysis of human VSA-specific plasma IgG reactivity with the surface of P. falciparum-IEs. Labelling of FCR3-CSA, NF54-VAR2CSA and 3D7-SM by IgG in individual plasma samples from P. falciparum-exposed pregnant women ( ), from sympatric men (᭢) and from non-exposed adult control donors (᭹) are shown. and growth characteristics. Six of the clones (PAM1.4, PAM2.8, PAM3.10, PAM5.2, PAM6.1, PAM7.5) produced antibodies recognizing antigens on the surface of erythrocytes infected by both the VSAPAM-expressing lines used to screen for antibody specificity (Table 1) . Antibodies from the two remaining clones (PAM4.7 and PAM8.1) only recognized FCR3-CSA. In contrast, none of the monoclonal antibodies recognized the 3D7-SM control line not expressing VSAPAM ( Fig. 2A-C) . Testing of monoclonal antibody reactivity with erythrocytes infected by a panel of additional parasite lines provided further evidence that all were indeed specific for PAM-type VSA expressed on the surface of CSA-adhering IEs (Table 1) . However, the monoclonal antibodies did not all recognize all VSAPAM-expressing lines, probably because the epitopes they recognize are polymorphic. IgG antibodies produced by a control B-cell clone (D7) did not recognize any of the tested parasite lines. Monoclonal antibody recognition patterns for individual parasite lines were tested in parallel, and repeated assessments of recognition patterns yielded consistent results. The flow cytometry evidence of antibody reactivity with antigens on the surface of IEs expressing VSAPAM and the absence of reactivity with non-PAM-type VSA (Table 1) was confirmed by immunofluorescence microscopy of live IEs ( Fig. 2D and E). Denaturing Western blots of the VSAPAM-expressing sublines yielded single, distinct bands (of similar size for each antibody) when probed with PAM3.10, PAM5.2, PAM6.1 and PAM7.5 monoclonal antibodies ( Fig. 2F , and data not shown). Proteins were not detected when blots were probed with the monoclonal antibodies PAM1.4, PAM2.8 or PAM4.7 (data not shown) despite their reactivity with the surface of intact VSAPAMexpressing IEs (Table 1) , pointing to reactivity with conformation-dependent epitopes. No bands were observed when the monoclonal antibodies were used to probe Western blots of the non-PAM-type VSAexpressing parental lines (Fig. 2F , and data not shown). PAM8.1 was not tested by Western blotting with IEs expressing VSAPAM, but was tested with VAR2CSAspecific constructs (see below). The high molecular weight of the proteins detected by Western blotting (Fig. 2F ) suggested that the monoclonal antibodies were specific for members of the so far bestcharacterized family of VSA, PfEMP1 (Leech et al., 1984) . This family includes VAR2CSA (predicted molecular weight: 355 kDa), which is the only PfEMP1 described so far that has the characteristics expected of VSAPAM (Salanti et al., 2003; 2004) . We therefore used a panel of recombinant proteins spanning the entire extracellular b. 3D7 (Walliker et al., 1987) was originally cloned from, and appears genetically identical to, NF54 ( Delemarre and Van der Kaay, 1979) . c. Line used in screening of B-cell supernatants for production of VSAPAM-specific IgG. n.d., not determined. VAR2CSA-specific IgG in pregnancy-associated malaria 337 part of VAR2CSA from 3D7 ( Fig. 3A ) and FCR3 (data not shown) to examine the antigen specificity of the monoclonal VSAPAM-specific IgG antibodies further. Antibodies PAM2.8, PAM3.10, PAM5.2, PAM6.1 and PAM7.5 tested positive in 3D7-VAR2CSA domain-specific ELISA ( Fig. 3A and Table 2 ), while antibodies PAM2.8, PAM3.10, PAM4.7, PAM5.2 and PAM8.1 tested positive in the FCR3-VAR2CSA ELISA (Table 2) . Control ELISA employing scrambled constructs and constructs from other PfEMP1 not implicated in the pathogenesis of PAM were consistently completely negative (data not shown). VAR2CSA constructs produced in Escherichia coli cells that should promote disulphide bond formation in secreted proteins (Barfod et al., 2006) were also consistently negative in ELISA (data not shown). Each of the VAR2CSA-reactive monoclonal antibodies had absolute specificity for either DBL3-X (PAM2.8, PAM6.1 and PAM8.1; originating from two donors) or DBL5-e VAR2CSA-specific IgG in pregnancy-associated malaria 339 (PAM3.10, PAM4.7, PAM5.2 and PAM7.5; also originating from two donors) ( Fig. 3A and Table 2 ). This pattern of reactivity was confirmed when the monoclonal antibodies were used to detect surface-expressed 3D7-VAR2CSA domains on transfected Jurkat cells in a flow cytometry assay (Fig. 3A) . Competition ELISA to examine the epitopes recognized by the DBL3-X and DBL5-e-reactive monoclonal antibodies showed that the DBL3-X-reactive antibodies PAM2.8 and PAM6.1 targeted antigenically distinct epitopes (Fig. 3B ), while two (PAM3.10 and PAM7.5) of the DBL5-e-reactive antibodies appeared to target neighbouring or overlapping epitopes (Fig. 3C) . Human VAR2CSA DBL3-X-specific monoclonal IgG antibodies recognize epitopes that vary between parasite isolates The pattern of monoclonal antibody recognition of IEs varied between parasite isolates. This is consistent with the finding that the var2csa sequence is composed of conserved stretches separated by stretches with substantial interclonal diversity (Duffy et al., 2006a; Trimnell et al., 2006) , and the prediction that B-cell epitopes in VAR2CSA DBL3-X locate mainly to polymorphic, surface-exposed parts of VAR2CSA . We therefore cloned and sequenced 43 VAR2CSA DBL3-X domains from placental parasite isolates. A subset of 29 of these domains selected to represent the overall VAR2CSA DBL3-X diversity was expressed as Baculovirus recombinant proteins and used in ELISA to test the specificity of the three VAR2CSA DBL3-X-specific monoclonals. The PAM2.8 antibody reacted with 25, PAM6.1 with eight and PAM8.1 with 20 of the domain variants (Fig. 4A) . A multiple sequence alignment of all the proteins indicated that the main difference between the PAM8.1-negative and -positive proteins was a C-terminal 16-amino-acid stretch that maps to a polymorphic region of 3D7-VAR2CSA DBL3-X, which is predicted to be a surfaceexposed loop (Fig. 4B) . Residues in this region either were deleted in the PAM8.1-negative proteins or had a different amino acid composition compared with the PAM8.1-positive variants (Fig. 4A) . To substantiate this possibility we constructed a chimeric protein where the 16-amino-acid stretch from a PAM8.1-positive domain variant (FCR3) was transferred to the corresponding site in a PAM8.1-negative variant lacking this sequence (3D7) (Fig. 4A, bottom) . The recombinant proteins corresponding to the unmodified FCR3 sequence and the chimeric construct both tested positive in Western blots probed with PAM8.1, in contrast to the recombinant protein representing the authentic 3D7 sequence (Fig. 4C) , thus confirming the predicted position of the PAM8.1 epitope. It was not possible to predict the exact targets of PAM2.8 and PAM6.1 by multiple alignments of the primary sequences. Human monoclonal antibody PAM1.4 effectively selects for expression of VSA PAM and increased transcription of VAR2CSA PAM1.4 stained VSAPAM-expressing IEs, but did not yield any bands in Western blots, and did not react with any of the VAR2CSA constructs when tested in ELISA or by flow cytometry (Tables 1 and 2 ). These observations are compatible with recognition by this antibody of a conformational epitope in VAR2CSA, but also with recognition of an unidentified non-VAR2CSA PAM-specific IE surface antigen. To address this question, we tested the ability of PAM1.4 to enrich VSAPAM-expressing IEs in two parasite lines (EJ24 and EJ27) initially expressing non-PAM-type VSA and only marginally recognized by PAM1.4 ( Fig. 5A and B, and data not shown). Although both isolates were originally obtained from the peripheral blood of pregnant women, and thus expected to express VSAPAM, isolates expressing non-PAM VSA -such as EJ24 and EJ27 -are occasionally found (Ofori et al., 2003, and our unpublished data) . Remarkably, a single round of PAM1.4 antibody selection of EJ27 ( Fig. 5C and D) and EJ24 (data not shown) resulted in rapid emergence of IEs uniformly recognized by PAM1.4 and expressing VSAPAM ( Fig. 5C and D). Quantitative real-time polymerase chain reaction (PCR) analysis of the isolates showed increases in var2csa transcription in response to the selection for PAM1.4 reactivity (EJ24: twofold and EJ27: 30-fold). In addition, EJ24 acquired reactivity with the VAR2CSAspecific antibodies PAM2.8, PAM3.10, PAM6.1 and PAM7.5 following selection for PAM1.4 reactivity ( Table 1) . EJ27 did not acquire additional reactivity following PAM1.4 selection, probably because of interclonal C. Western blots of recombinant 3D7-and FCR3-specific VAR2CSA DBL3-X constructs, and of the above-mentioned chimeric construct, probed with loading control antibody V5 (left) and PAM8.1 (right). MW, molecular weight. differences in the VAR2CSA epitopes recognized by the other monoclonal antibodies. Taken together, these findings are consistent with VAR2CSA being the antigenic target of PAM1.4. We have shown that it is possible to interrogate the memory B-cell repertoire of malaria-immune donors to estimate frequencies of P. falciparum-specific B cells, and to isolate specific monoclonal antibodies with specificity for the VSA repeatedly implicated as the main targets of acquired protective immunity to malaria. We have used this approach to demonstrate that PAM can result in acquisition of high frequencies of B cells producing IgG with specificity for VSA PAM, and in particular VAR2CSA, strengthening previous evidence that these antigen specificities are critically important in acquired protective immunity to PAM. We furthermore show that VSAPAM-specific memory B cells acquired in response to PAM primarily target polymorphic, conformationdependent epitopes that are reproduced by Baculovirusproduced recombinant antigen constructs. Our data thus underscore the importance of VAR2CSA in acquired B. Pre-selection non-PAM VSA-type recognition pattern of EJ27 by IgG in plasma from P. falciparum-exposed men and women and in plasma from non-exposed adults. C. Reactivity of PAM1.4 antibody (heavy line) and negative control antibody (thin line) with the surface of erythrocytes infected by the EJ27 after a single round of selection for reactivity with PAM1.4. D. Post-selection VSAPAM-type recognition pattern of EJ27 by IgG in plasma from P. falciparum-exposed men and women and in plasma from non-exposed adults. immunity to PAM. However, the findings reported here and elsewhere also suggest that var2csa diversity (Duffy et al., 2006a; Trimnell et al., 2006) is driven by protective immunity to PAM, a situation that may complicate development of VAR2CSAbased vaccines against PAM (Beeson et al., 2006) . IE adhesion to CSA, which is thought to be a critical element in the pathogenesis of PAM (Fried and Duffy, 1996) , is mediated by VAR2CSA as documented by recent knockout studies (Viebig et al., 2005; Duffy et al., 2006b) , and several CSA-adhesive domains have been identified in the antigen (Gamain et al., 2005) . Recent studies in mice suggest that vaccination can elicit broadly reactive antibodies that can block VAR2CSAdependent IE adhesion to CSA (Gamain et al., 2004; Bir et al., 2006) , but whether such antibodies are ever produced in humans in response to PAM, how clinically relevant they are, and whether they can be induced by vaccination in humans remain unanswered questions. Understandably, present research is highly focused on the identification of functionally constrained epitopes in these domains that are critical for IE adhesion to CSA and of intergenomically conserved epitopes that may serve as targets of antibodies interfering with it. Human monoclonal antibodies appear to be a powerful tool in this research. All P. falciparum parasites used in this study were grown in 0 + erythrocytes (Cranmer et al., 1997) . 3D7, FCR3 and NF54 are long-term in vitro cultured lines. All expressed non-PAM-type VSA, meaning that intact IEs were recognized to a similar extent by IgG in the plasma of P. falciparum-exposed men and sympatric, multigravid women in a flow cytometry assay of VSA expression (Fig. 1 , 3D7-SM) . The 3D7 subline 3D7-SM was derived by human plasma antibody selection of 3D7 for expression of non-PAM-type PfEMP1 associated with severe malaria in children as described Jensen et al., 2004) . The VSAPAM-expressing subline 3D7-BeWo was selected by repeated panning of IEs on the choriocarcinoma line BeWo as described elsewhere (Haase et al., 2006) . Parasites were considered as expressing VSAPAM if the level of labelling of intact IEs by IgG in a panel of plasma samples from P. falciparum-exposed multigravid women was significantly higher than the level in plasma from sympatric men (Fig. 1, FCR3 -CSA and NF54-VAR2CSA). The characteristics of the plasma IgG recognition pattern of VSAPAM and non-PAM-type VSA have been documented in detail elsewhere (Ricke et al., 2000; Staalsoe et al., 2001) . Sublines of FCR3 and NF54 (FCR3-CSA and NF54-CSA respectively) were selected for expression of VSAPAM by repeated panning of IEs on CSA in vitro (Fried and Duffy, 1996; Ricke et al., 2000) . NF54-CSA was further selected for IE reactivity with rabbit antiserum specific for VAR2CSA DBL5-e, resulting in subline NF54-VAR2CSA (Salanti et al., 2004) . Additional sublines of FCR3 (FCR3-A745 and FCR3-CD36) expressing non-PAM VSA were selected by repeated panning on CSA-negative CHO cells (CHO-A745) and recombinant CD36, respectively, essentially as described for BeWo and CSA selection. Isolates EJ24 and EJ27 were obtained from the peripheral blood of pregnant, P. falciparum-exposed women and adapted to in vitro culture . Both isolates were selected for expression of VSA reacting with the human VSAPAMspecific monoclonal antibody PAM1.4 (see below), essentially as described , but using Protein A-coated magnetic microbeads, as VSAPAM-expressing IEs are prone to non-specific labelling by second-step antisera (Creasey et al., 2003; Rasti et al., 2006) . Peripheral blood mononuclear cells (PBMC) from P. falciparum-exposed, recently pregnant multiparous women were isolated and cryopreserved as described (Hviid et al., 1993) . At the day of use, PBMC were thawed and IgG + memory B cells were isolated using CD22 microbeads (Miltenyi) followed by cell sorting as described (Traggiai et al., 2004) . Cells were immortalized at 100 cells per well in multiple 96-well plates using EBV in the presence of CpG ODN2006 (Microsynth, Switzerland) (Hartmann and Krieg, 2000) and irradiated PBMC as described (Traggiai et al., 2004) . Polyclonal B-cell culture supernatants were screened by flow cytometry for IgG reactivity with the surface of intact, unfixed erythrocytes infected by FCR3-CSA, NF54-VAR2CSA and 3D7-SM. VSAPAM-reactive B-cell lines, selected on the basis of their rate of IgG synthesis and growth rates, were cloned by limiting dilution as described (Traggiai et al., 2004) and the selectivity of the human monoclonal antibodies produced by the clones for IEs expressing VSAPAM was confirmed as above. The reactivity of the antibodies with the surface of wet-mounted antibody-labelled IEs was further verified by immunofluorescence microscopy, using an LSM5 scanning microscope (Carl Zeiss MicroImaging) (Salanti et al., 2004) . The IgG subclass of all the human monoclonal antibodies was determined by ELISA and verified by flow cytometry using isotype-specific antibodies (Megnekou et al., 2005) . Parasite cultures were enriched for erythrocytes infected by late trophozoite/schizont-stage parasites by exposure to a strong magnetic field (Paul et al., 1981; Staalsoe et al., 1999) . Protein extracts of purified IEs were prepared with 2% SDS in PBS containing complete protease inhibitor (Roche, VAR2CSA-specific IgG in pregnancy-associated malaria 343 Basel, Switzerland). The extracts were boiled in denaturing loading buffer and separated in pre-cast tris-acetate 5-8% SDS gradient gels (Invitrogen, Tåstrup, Denmark) with trisacetate running buffer (Invitrogen), employing pre-stained broad-range molecular weight markers (Bio-Rad, Herlev, Denmark) . The separated proteins were transferred to PVDF membranes by wet blotting in transfer buffer containing 20% isopropanol, 20 mM tris-acetate and 0.1% SDS, followed by blocking with 5% skimmed-milk powder in TBS-T buffer. Membranes were incubated with the human monoclonal antibodies or a monoclonal mouse anti-exon 2 antibody, followed by incubation with a secondary anti-human or anti-mouse AP-conjugated antibody (Sigma, MO, USA) and developed using a 5-bromo-4-chloro-3-indoyl phosphate and nitroblue tetrazolium solution (Sigma). Baculovirus-produced proteins and a pre-stained ProSieve protein marker (Cambrex) were run on pre-cast 4-12% SDS gradient gels (Invitrogen, Tåstrup, Denmark) with NuPage MOPS SDS running buffer (Invitrogen). Proteins were transferred to a nitrocellulose membrane by wet blotting using a buffer of 20% methanol, 25 mM Tris and 192 mM glycine. Following blocking in 5% skimmed-milk powder in TBS-T buffer, membranes were incubated for 1 h with either a 1:5000 dilution of horseradish peroxidase-conjugated loading control antibody anti-V5 (R960-25, Invitrogen) or a 1:1000 dilution of PAM8.1. The PAM8.1-probed membrane was further incubated with a 1:1000 dilution of a secondary anti-human IgG antibody (P0214, Dako Cytomation). Membranes were developed using 3-amino-9-ethyl-carbazole tablets dissolved in acetone, 50 mM sodium acetate and 30% H2O2. Regions of re-codonized 3D7-var2csa (PFL0030c) and FCR3-var2csa covering the entire exon 1 were subcloned into the pBAD-TOPO vector, transferred with the V5 and HIS tag to the pAcGP67-A transfer vector (BD Biosciences), produced as recombinant proteins in Baculovirus-infected insect cells, and purified as described (Salanti et al., 2004) . We have previously shown that Baculovirus-produced VAR2CSA constructs are conformationally intact, as they induce production of rabbit antisera reactive with native VAR2CSA on the surface of IEs (Barfod et al., 2006) . The following regions (indicated by encoded amino acids numbers) were produced: -1967, N: 1380-1579, O: 1559-1992, Q: 1776-2131, R: 1965-2666, S: 1981-2336, U: 1981-2524, V: 2147-2666, X: 2210-2666 and Y: 2317-2666 . Different variants of DBL3-X were cloned and produced as described . For the cloning of the chimeric construct composed of 5′ 3D7-VAR2CSA DBL3-X and 3′ FCR3-VAR2CSA DBL3-X, we used the primers 5′: cggaattcGATACAAATGGTGCCTGT and 3′: CATTTCTTT TCATTCTTACCATTATTAGTGCA to generate the 3D7specific, and 5′: AAGAATGAAAAGAAATGTATTAATTC and 3′: atttgcggccgcATATACTGCTATAATCTCC to generate the FCR3-specific part of the chimera. These primers amplify a slightly smaller PCR product than the original primers used for making the FCR3 and 3D7 DBL3-X constructs, and this is reflected in the smaller molecular size of the chimeric construct. The two PCR products were gel-purified and used in a second PCR using the two outer primers to generate a PCR product consisting of 5′ 3D7 and 3′ FCR3, with an EcoRI site and a NotI site. The PCR product was cloned into a modified pAcGP67-A vector (BD Biosciences) and expressed in insect cells as described (Salanti et al., 2004) . VSAPAM-reactive monoclonal IgG-containing supernatants were tested in ELISA (Dodoo et al., 2000) for reactivity with the recombinant VAR2CSA proteins produced in Baculovirusinfected insect cells. In addition, the epitope specificities of monoclonal antibodies targeting DBL3-X (PAM2.8 and PAM6.1) and DBL5-e (PAM3.10, PAM5.2 and PAM7.5) were analysed by competition ELISA. PAM3.10 and PAM6.1 were purified on ÄktaXpress (GE Healthcare, Brøndby, Denmark) using a HiTrap Protein G HP 1 ml column with subsequent desalting on a HiPrep 26/10 desalting column (GE Healthcare). Purified IgG was biotinylated using EZ-link maleimide-PEO solid phase as described by the manufacturer (Pierce, Bonn, Germany). Microtitre plates (Nunc, Roskilde, Denmark) were coated with recombinant DBL3-X (8.3 mg ml -1 ) or DBL5-e (10.4 mg ml -1 ) in PBS (1 h, 37°C). After blocking of the plates, biotinylated PAM3.1-specific IgG (0.72 mg ml -1 ) or biotinylated PAM6.1-specific IgG (2.5 mg ml -1 ) and increasing concentrations of the competitor monoclonal culture supernatants were added to triplicate wells. PAM1.4 (unknown VSAPAM-specificity) and PAM2.8 (DBL3-X-specific) were added as negative controls to DBL3-X-coated and DBL5-e-coated plates respectively. Bound biotinylated IgG was detected by incubation (1 h, room temperature) of wells with horseradish peroxidase-conjugated streptavidin (1 mg ml -1 , 100 ml well -1 ; Pierce, Bonn, Germany). The human T-cell line Jurkat (Gillis and Watson, 1980) was cultured in RPMI 1640, supplemented with 25 mM HEPES and L-glutamine (Gibco, Tåstrup, Denmark), 10% FCS, 100 IU ml -1 penicillin and 100 mg ml -1 streptomycin. Two million cells were seeded into each well of a six-well plate and transfected with 3-4 mg of plasmid DNA (see above) and 4 ml of DMRIE-C transfection reagent (Invitrogen) according to the manufacturer's instructions. Within 48 h of transfection, the cells were washed and re-suspended at 1 ¥ 10 6 ml -1 in PBS supplemented with 2% FCS. Cells (1 ¥ 10 5 ) were incubated with human monoclonal antibodies or with a haemagglutinin mouse antibody for 30 min followed by two washes and labelling by secondary FITC-conjugated anti-human IgG or anti-mouse IgG antibody. Flow cytometry analysis was essentially as above. Recombinant VAR2CSA DBL3-X constructs from 29 genotypically distinct P. falciparum isolates were produced in Baculovirus-infected insect cells and tested in ELISA essentially as described above. The three-dimensional structure of the 3D7-VAR2CSA DBL3-X sequence (PFL0030c, amino acids 1217-1559) was modelled in silico as described elsewhere . Briefly, the crystal structure of EBA-175 F1 (PDB code 1ZRO chain A) (Tolia et al., 2005) , which has 28% sequence identity to the 3D7-VAR2CSA DBL3-X domain, was used as a template. The model was evaluated with respect to locations of conserved cysteine bridges and buried hydrophobic residues in the structures of DBL domains from EBA-175 F1 and F2 (Tolia et al., 2005) and Pk-a DBL (Singh et al., 2005) . Quantitative real-time PCR was performed on cDNA from unselected and PAM1.4 antibody-selected isolates EJ24 and EJ27 using a Rotorgene thermal cycler system (Corbett Research, Cambridge, UK) and a primer set specific for a highly conserved part of the var2csa DBL-4e domain, targeting all var2csa genes without bias (Salanti et al., 2003; Tuikue Ndam et al., 2005) . Selection-induced changes in var2csa transcription were quantified as described (Salanti et al., 2003) .

Early Assessment of Anxiety and Behavioral Response to Novel Swine-Origin Influenza A(H1N1)

BACKGROUND: Since late April, 2009, a novel influenza virus A (H1N1), generally referred to as the “swine flu,” has spread around the globe and infected hundreds of thousands of people. During the first few days after the initial outbreak in Mexico, extensive media coverage together with a high degree of uncertainty about the transmissibility and mortality rate associated with the virus caused widespread concern in the population. The spread of an infectious disease can be strongly influenced by behavioral changes (e.g., social distancing) during the early phase of an epidemic, but data on risk perception and behavioral response to a novel virus is usually collected with a substantial delay or after an epidemic has run its course. METHODOLOGY/PRINCIPAL FINDINGS: Here, we report the results from an online survey that gathered data (n = 6,249) about risk perception of the Influenza A(H1N1) outbreak during the first few days of widespread media coverage (April 28 - May 5, 2009). We find that after an initially high level of concern, levels of anxiety waned along with the perception of the virus as an immediate threat. Overall, our data provide evidence that emotional status mediates behavioral response. Intriguingly, principal component analysis revealed strong clustering of anxiety about swine flu, bird flu and terrorism. All three of these threats receive a great deal of media attention and their fundamental uncertainty is likely to generate an inordinate amount of fear vis-a-vis their actual threat. CONCLUSIONS/SIGNIFICANCE: Our results suggest that respondents' behavior varies in predictable ways. Of particular interest, we find that affective variables, such as self-reported anxiety over the epidemic, mediate the likelihood that respondents will engage in protective behavior. Understanding how protective behavior such as social distancing varies and the specific factors that mediate it may help with the design of epidemic control strategies.

An ongoing outbreak of novel influenza A(H1N1), colloquially referred to as ''swine flu,'' has caused over 200,000 confirmed cases (as of 28 August 2009 [1] ). Because of under-reporting, the actual number of people infected is substantially larger, particularly considering that many countries have ceased testing for A(H1N1) [1] . As human-to-human transmission became widespread in at least one region of the world, WHO rapidly declared the outbreak an imminent pandemic [2] and with widespread human infection, WHO declared a phase 6 pandemic on 11 June 2009, where it remains at the time of submission [3] . The virus appears to have a higher reproduction number and somewhat higher case fatality ratio than recent seasonal influenza viruses [4, 5] , and has certainly caused great concern in the population, fueled by both extensive media coverage and an initially high level of uncertainty about mortality rates and transmissibility of the virus. Mathematical and computational models are useful for predicting the fate of an epidemic, and while such models have become increasingly complex and realistic in recent times, a key ingredient is often ignored: human behavioral responses to the threat and/or presence of a disease [6] . How people assess risk of infection and how such risk assessment drives behavioral change is of great interest as individual social distancing can greatly affect the spread of an epidemic [7, 8, 9] . While the effect of behavioral change in response to perceived health threats on the spread of infectious diseases has been investigated theoretically for some time, particularly in the context of sexually transmitted diseases [8] , recent work has started addressing the problem in a broader context that is also applicable to the spread of influenza [6, 7] . This work has a strong, though as yet under-explored relationship to work on risk perception and health threats [10, 11, 12] . Data on risk perception and behavioral response in the general population have rarely been collected right from the outset of an epidemic. Instead, they are usually gathered with a substantial delay in the case of influenza A(H1N1) [13] , after the epidemic has run its course, as in the case of SARS [9] , or before sustained human to human transmission is established, as in the case of highly pathogenic avian influenza A(H5N1) [14] . However, the feasibility of halting or mitigating the spread of an infectious disease is highest during the very early phases of an outbreak, and thus data on behavioral response during this time would provide valuable information for public health policy and research. Here, we report the results from an online survey that gathered data (n = 6,249) about risk perception of the outbreak during the first few days of widespread media coverage (April 29 -May 5, 2009) of the emergence of novel swine-origin Influenza A(H1N1). The research presented here was approved by the Stanford University Non-Medical Human Subjects Institutional Review Board on 28 April 2009 (Protocol #16782). We fielded in internet-based survey starting on 29 April 2009 using Opinio survey software [15] . The URL for the survey is (https://opinio.stanford.edu/opinio/s?s = 1403). The sampling universe for this study was adults 18 and older with access to a networked computer. The initial seed for the sample was generated using social networking software, and a request sent to a standing subject pool comprised of Stanford alumni and social science students at a nearby community college maintained by the Institute for Research in the Social Sciences at Stanford University. The survey was picked up by a variety of internet media sources including several science general media blogs. Directly following publicity in these blogs, we received the most responses. Table 1 summarizes the sample. The survey was designed to get a rapid assessment of respondents' affective state, sources of information on the emerging pandemic, and the behaviors undertaken for protection while minimizing respondent burden. As such, it included only 17 questions. Questions probing subjective assessment of risk perception, level of anxiety, and ability to avoid flu infection were asked on a 9 point ordinal scale with anchors at the extrema (''very high'', ''very low'') and the center (''intermediate''). Subjective emotional status (i.e., respondents' affective state regarding the epidemic) was anchored by the terms ''very calm'' through ''intermediate'' to ''very anxious.'' Comparative questions of subjective risk percep-tion for eight health threats were asked using a five-point ordinal scale with anchors at all points: ''very low,'' ''low,'' ''intermediate,'' ''high,'' and ''very high.'' Questions regarding media (both respondents' frequency of getting information from a particular source and their judgment of each source's accuracy) were asked on a five point ordinal scale with anchors at all points (''very often/ accurate'' to ''never/almost certainly inaccurate''). Respondents' knowledge of swine flu was assessed with a series of six True/False questions. Respondents gave free-text responses to questions about their current age, the number of people currently living in their household (including themselves) and their zip code if they currently live in the United States. Respondents who reported not currently living in the United States were asked to report their current country of residence in a free-text field. A screen-shot of the full survey instrument is included in the Supplementary Material. For our analysis of participants' response to the threat of swine flu, we use a variable we call ''survey day.'' The survey went online in the evening of 28 April, Pacific time, so we combined responses from 28 and 29 April into a single day. This combined day of 28-29 April represents survey day 1. Subjects were asked to state the number of contacts in the past 24 hours. Contacts were defined by close physical contact as operationalized by a face to face conversation of more than two words in the presence of another individual or physical exposure involving skin contact such as a handshake, hug, or contact during sporting activities. Respondents were provided five ordered categories: less than 5, 5-10, 11-20, 21-50, 51-100, more than 100. Handcock and Jones [16] discuss the phenomenon of heaping and related problems for statistical inference in answering epidemiological questions regarding contact number. Structuring responses within broad ordinal categories avoids many of the pitfalls of contact-heaping encountered in epidemiological investigations. To measure the response in epidemiologically relevant behavior to information on the potential pandemic, we asked a series of questions about protective actions taken by the respondents. In the survey, we asked: ''Given the current status of the epidemic, which of the following precautionary actions will you take?'' Avoid people who cough/sneeze Avoid large gatherings of people Wash hands more often Avoid people who are in contact with infected people Avoid public transportation Avoid school/work Avoid travel to infected areas Use disinfectant Wear a mask Not all of these behaviors are necessarily effective or recommended protective measures (e.g., wearing a mask), but our aim was to gauge people's attempt at self protection so even non-efficacious behavioral change is interesting in that it indicates willingness to act on the part of the respondent. We constructed an index of protective behavior by summing the answers to the questions regarding actions taken to avoid influenza infection. The index ranged from 0-9, with larger values indicating more protective measures taken. Using a binomial GLM with canonical logit-link, we modeled the protection index as a function of covariates. Our primary interest was the possible mediating effect of affective variables on action taken to protect against swine flu infection. To evaluate the hypothesis that respondents' affective state (subjective anxiety, fatalism about infection) predicts protective measures, we include in the model demographic (age, gender), epidemiological (household size, number of contacts, survey day), and media (source of information on the outbreak) conditioning variables. For the media variables, we constructed dummies with a value of 1 corresponding to answers of ''very often'' or ''often'' and a value of 0 for all other responses to the question of ''How often do you use the following sources to get information about swine flu?'' With such a large number of conditioning variables, in addition to the affective variables of greatest interest, there is a distinct danger of overfitting the GLM. To address this problem, we used likelihoodbased model selection [17] to search the model space set up by our conditioning variables [18] . Of the nine protective behaviors, increased hand-washing is both the simplest and probably most effective at curbing transmission. As such, it is strongly advocated in infection control educational material [19] . In addition to our tests for predictors of the protection index, we therefore also tested the effect of measured covariates on the odds of increased hand washing using a binomial GLM again with canonical logit-link. A concern regarding the relationship of people's self-reported anxiety and their protective behavior is that some people might generally be more anxious regarding health. We probed general anxiety by asking about respondents' anxiety with regard to a number of infectious, chronic, and violent threats to health. We asked a series of questions probing respondents' perceived subjective risk on a 5-point scale for a variety of health threats, including other infectious diseases (A(H5N1) ''bird flu'', seasonal flu, HIV/AIDS), chronic diseases (heart disease, diabetes, cancer), and violence (unintentional accidents, terrorism). We calculated the correlation matrix for answers to these threat questions and used Principal Components Analysis (PCA) to explore potential structure in the responses to different categories of threat [20] . We begin by presenting descriptive results of the survey and follow with our primary analytical questions from the survey, namely, testing the hypothesis that respondents' affective state mediates their protective action. We gathered 6,249 responses from 28 April to 5 May 2009. Table 1 presents descriptive statistics of the sample. Figure 1 presents the distributions of respondents' contacts within the 24 hours prior to taking the survey. Figure 2 presents the means of the subjective threats. Swine flu had a mean second only to injury, and the highest among the infectious sources of threat. The mean of perceived threat from swine flu fell above the Bonferroni-corrected 95% confidence interval for all other threats but unintentional injury. Figure 3 presents the frequency distribution of perceived personal risk. There is a notable bimodality to this plot. This apparent bimodality is not simply attributable to sampling error since the difference between the responses = 4 vs. those = 5 vs. those = 6 is in excess of 300. Further analysis using finite mixture models [21] provides strong statistical support for the reality of the bimodal pattern (results not shown). While the majority of respondents felt that their personal risk was low, there is a second mode rating their risk as intermediate ( = 5) . This same bimodal pattern can be seen in the frequency distribution of personal empowerment (i.e., ability to avoid infection) shown in figure 4 . While most respondents indicate that they are confident they can avoid infection, a substantial second mode appears at the intermediate value. Figure 5 shows the frequency distribution of protective behaviors. We can see that nearly 80% of respondents report washing hands more frequently, while very few avoid work or school or wear protective masks. Figure 6 shows the means for respondents' information sources. Not surprisingly, the most common source of information reported was the Internet. Again, mean values are plotted with their 95% Bonferroni-corrected confidence intervals. With the exception of social-networking tools (e.g., Facebook, Twitter), all other media sources are statistically indistinguishable from each other, with the social-networking tools being used significantly less. The results of the model for the protection index show a number of robust trends (table 2). In particular, we find that age increases and male gender decreases the protection index. Receiving a large amount of information from the internet, television, and health officials all increase the protection index while receiving large amounts of information from print media, friends, or social networking media has no effect. The number of household members has no discernible effect, though the number of contacts outside the home does. For the ordered factor ''contacts,'' the first category (,5 contacts in the past 24 hours) is the reference category. Interestingly, relative to respondents with the fewest number of contacts, all other contact categories have reduced protection indices, indicating that people with fewer contacts take more protective actions. Not surprisingly, residence in Mexico has a large positive effect, while residence in Canada or Europe decreased the index. The day that the survey was taken (29 April = 1) had a negative effect on the index, indicating that respondents took less protective action as the epidemic proceeded. Respondents' reported subjective anxiety has a substantial impact on the index with high anxiety increasing protection, supporting our hypothesis that affective state mediates protective behavior. Increased hand-washing showed similar trends to the model for the protection index (table 3) . Male gender decreases while age and survey day increase the odds of increasing hand-washing. Receiving a large amount of information from the internet, radio, television, and health officials increase, while living in Europe or Australia/New Zealand decrease the odds. As with the overall protection index, perception of risk and subjective anxiety significantly increase the odds of increased hand-washing modestly. A key epidemiological question is how people's affective status and protective behaviors undertaken change as the epidemic proceeds. To develop a measure for this, we cross-tabulated individual values of the protection index and affective status by survey day. Pearson's chi-square test for independence of both tables was strongly significant (affective: x 2 = 135.6, df = 48, p,0.001; protection: x 2 = 113.1, df = 54, p,0.001), indicating substantial departure of cells from the expected values. To visualize the pattern of departure from the expected values, we calculated an expected tables taken as the cross-product of the marginals of the observed table normalized by the grand sum. We combined rows of these tables to simplify the presentation, plotting the difference between observed and expected tables for a high, medium and low emotional status/protection index respectively. For example, a value of 251 on the calm affective status on day one means that there were 51 fewer responses in the calm categories than would be expected by the overall marginal distribution of responses across all days. In figures 7 and 8, we plot the change in respondent's protective behavior and emotional status over the first week of the survey. The lines represent the differences between observed and expected frequencies of responses for the 9-point scale simplified to three levels each. We see that by day three of the survey (May 1st), the relative number of people reporting a calm emotional state was very high, while the number of people reporting high values of the protective index declined dramatically. We interpret these results to indicate that people's response to a potential pandemic is quite sensitive to media reports. In general, individuals' survey responses to perceived risk for the eight health threats were only moderately correlated, with pairwise correlations typically well under 0.5. PCA did not reveal that a substantially reduced number of dimensions explained these correlated data -six principal components were required to explain 85% of the variance. Nonetheless, some intriguing PC loadings revealed themselves. In particular, the second PC, which explained 15.6% of the variance in the data, showed strong positive correlations with swine flu (r = 0.516), bird flu (r = 0.530), and terrorism (r = 0.467). All three of these threats receive a great deal of media attention and their fundamental uncertainty are likely to generate an inordinate amount of fear vis-a-vis their actual threat [22] . Our results indicate that respondents' behavior varies systematically with covariates from demographic, epidemiological, media, and affective domains. People's anxiety about swine flu and the preventative actions they took to avoid infection declined as the perceived gravity of the novel outbreak waned. Overall, subjective risk perception was low and people's belief in their ability to avoid infection was high. Both of these distributions nonetheless showed a marked bimodality, with a large proportion of respondents indicating a higher subjective risk and more protective actions taken than the majority (Figures 3-4) . The results of our statistical modeling suggest that respondents' deployment of protective behavior is mediated by their subjective anxiety level, controlling for demographic, epidemiological, and geographic variables. There is good and bad news in this result. The literature on risk perception and public health shows that there is generally a very weak correlation between people's anxiety over a particular risk and the probability of death or disability arising from that risk [11, 12] . Overall, it is unclear whether anxiety over perceived risk will lead to efficacious protective behaviors [10] . This said, by far the most common protective behavior reported in our survey was increased hand-washing, which has been shown to be effective at removing Influenza A(H1N1) virus from subjects' hands [23] and is promoted by CDC and other health organizations as an effective infection control intervention [19] . One curious result from the model for the protection index is that people with the fewest contacts have marginally higher protection indices. There are two potential explanations for this finding: (1) individuals with small social support networks (and consequently, few contacts outside the home) are more anxious, making it more likely that they will take greater protective actions or (2) people concerned about infection and taking relatively many protective actions also reduce the number of contacts they have and therefore had a small number of contacts in the past 24 hours. The first explanation is consistent with work in social epidemiology on the role of social networks in mediating infection risk [24, 25] Because of the nature of the sample, we are unable to evaluate the direction of causality that leads to this result. Nonetheless, it remains an intriguing hypothesis. Many questions about this Novel Swine-Origin influenza A(H1N1) virus remain. Of particular concern is the possibility that the virus could mutate during the flu season in the southern hemisphere and selection could drive it to become more virulent as it returns to the northern hemisphere in Autumn. Worryingly, such a pattern of multiple waves with an increased proportion of the total influenza-associated mortality burden has been reported for all three past influenza pandemics [26, 27] . Finding a means to simultaneously communicate to the public the structural uncertainty inherent in projecting pandemics and the seriousness of a pandemic after the media frenzy about its emergence has died down remains a major challenge to public health. Pharmaceutical interventions such as vaccines and antiviral drugs may form a strong line of defense, but the efficacy of such measures remains unclear and depends on the particular biology of a given pathogen. This is exacerbated by people's reluctance to be vaccinated [28] . With more than 300 infectious diseases emerging within less than a century [29] , the threat of pandemic influenza is only the latest out of many public health threats posed by infectious diseases in a globalized world. Unlike pharmaceutical interventions, non-pharmaceutical interventions like social distancing may be effective in halting or at least mitigating an epidemic independent of the specific biology of a pathogen, and thus provide a reliable set of strategies to combat infectious diseases that warrant further attention [30] . Our results that people do not rely on social networking tools to the extent that they rely on other media may have implications for CDC strategies for spreading public health information via social media [19] . In particular, public health messages spread via social media will need to backed up by information spread via more traditional channels, which respondents list as being common sources of trusted information on the outbreak. Our study is subject to a number of weaknesses. The advantage of our internet-based sampling strategy is the ability to quickly deploy a survey and thereby track responses in near real-time. The clear disadvantage of this strategy is a sacrifice of population representativeness. Despite its general availability on the internet, our sample shows a pronounced bias for highlyeducated respondents living in the Western United States. These biases clearly limit the generality of our results, but the large number of respondents filing out the survey as information on the potential pandemic changed nonetheless provides a uniquely valuable data source. Within one week (the cutoff point for the current analysis), we had a sample of 6,249 responses. In contrast, the telephone-based study of Rubin et al. [13] employed a random-digit-dial sampling design, allowing a more representative sample of the general UK population, but their sample was only 997 respondents and the survey was undertaken after media attention had abated, beginning 8 May 2009. Nonetheless, the results reported in this paper are largely congruent with our own results and we see the studies as strongly complementary. Our respondents began filling out the survey on the day that WHO upgraded the pandemic threat category of the H1N1 outbreak from 4 to 5 and spans the week in which the number of WHO-confirmed cases increased tenfold. While our sampling design is subject to many of the usual criticisms of internet-based surveys and is not necessarily representative of the general population, the unparalleled immediacy, longitudinal nature, and the large number of respondents it contains make our data set unique and scientifically important for the study of the spread of information and distribution of risk perception and behavioral change during the most uncertain time (i.e. the initial phase) of an epidemic of a virus novel to the human population.

Accurate noise projection for reduced stochastic epidemic models

We consider a stochastic susceptible-exposed-infected-recovered (SEIR) epidemiological model. Through the use of a normal form coordinate transform, we are able to analytically derive the stochastic center manifold along with the associated, reduced set of stochastic evolution equations. The transformation correctly projects both the dynamics and the noise onto the center manifold. Therefore, the solution of this reduced stochastic dynamical system yields excellent agreement, both in amplitude and phase, with the solution of the original stochastic system for a temporal scale that is orders of magnitude longer than the typical relaxation time. This new method allows for improved time series prediction of the number of infectious cases when modeling the spread of disease in a population. Numerical solutions of the fluctuations of the SEIR model are considered in the infinite population limit using a Langevin equation approach, as well as in a finite population simulated as a Markov process.

The interaction between deterministic and stochastic effects in population dynamics has played, and continues to play, an important role in the modeling of infectious diseases. The mechanistic modeling side of population dynamics is well known and established. 1, 2 These models typically are assumed to be useful for infinitely large, homogeneous populations, and arise from the mean field analysis of probabilistic models. On the other hand, when one considers finite populations, random interactions give rise to internal noise effects, which may introduce new dynamics. Stochastic effects are quite prominent in finite populations, which can range from ecological dynamics 3 to childhood epidemics in cities. 4, 5 For homogeneous populations with seasonal forcing, noise also comes into play in the prediction of large outbreaks. [6] [7] [8] Specifically, external random perturbations change the probabilistic prediction of epidemic outbreaks as well as its control. 9 When geometric structure is applied to the population, the interactions are modeled as a network. 10, 11 Many types of static networks which support epidemics have been considered. Some examples include small world networks, 12 hierarchical networks, 13 and transportation networks of patch models. 14 In addition, the fluctuation of epidemics on adaptive networks, where the wiring between nodes changes in response to the node information, has been examined. 15 In adaptive network models, even the mean field can be high dimensional, since nodes and links evolve in time and must be approximated as an additional set of ordinary differential equations. Another aspect of epidemic models, which is often of interest, involves the inclusion of a time delay. The delay term makes the analysis significantly more complicated. However, it is possible to approximate the delay by creating a cascade consisting of a large number of compartments. 16 For example, one could simulate the delay associated with a disease exposure time with several hundred "exposed" compartments. These model examples are just a few of the types of very high-dimensional models that are currently of interest. As a result of the high dimensionality, there is much computation involved, and the analysis is quite difficult. In particular, real-time computation is not currently possible. However, there are usually many time scales that are well separated ͑due typically to a large range in order of magnitude of the parameters͒ when considering such high-dimensional problems. In the presence of well-separated time scales, a model reduction method is needed to examine the dynamics on a lower-dimensional space. It is known that deterministic model reduction methods may not work well in the stochastic realm, which includes epidemic models. 17 The purpose of this article is to examine a method of nonlinear, stochastic projection so that the deterministic and stochastic dynamics interact correctly on the lower-dimensional manifold and predict correctly the dynamics when compared with the full system. Because the noise affects the timing of outbreaks, it is essential to produce a low-dimensional system that captures the correct timing of the outbreaks as well as the amplitude and phase of any recurrent behavior. We will demonstrate that our stochastic model reduction method properly captures the initial disease outbreak and continues to accurately predict the outbreaks for time scales, which are orders of magnitude longer than the typical relaxation time. Furthermore, in practice, real disease data include only the number of infectious individuals. Our method allows us to predict the number of unobserved exposed individuals based on the observed number of infectious individuals. For stochastic model reduction, there exist several potential methods for general problems. For a system with certain spectral requirements, the existence of a stochastic center manifold was proven in Ref. 18 . Nonrigorous stochastic normal form analysis ͑which leads to the stochastic center mani-fold͒ was performed in Refs. [19] [20] [21] [22] . Rigorous theoretical analysis of normal form coordinate transformations for stochastic center manifold reduction was developed in Refs. 23 and 24. Later, another method of stochastic normal form reduction was developed, 25 in which any anticipatory convolutions ͑integrals into the future of the noise processes͒ that appeared in the slow modes were removed. Since this latter analysis makes the construction of the stochastic normal form coordinate transform more transparent, we use this method to derive the reduced stochastic center manifold equation. Figure 1 shows a schematic demonstrating our approach to the problem. We consider a high-dimensional system along with its corresponding reduced low-dimensional system. In this article, two types of low-dimensional system are discussed: a reduced system based on deterministic center manifold analysis and a reduced system based on a stochastic normal form coordinate transform. Regardless of the type of low-dimensional system being considered, a common noise is injected into both the high-dimensional and lowdimensional models, and an analysis of the solutions found using the high and low-dimensional systems is performed. In this article, as a first study of a high-dimensional system, we consider the susceptible-exposed-infected-recovered ͑SEIR͒ epidemiological model with stochastic forcing. As previously mentioned, we could easily consider a SEIR-type model where the exposed class was modeled using hundreds of compartments. Since the analysis is similar, we consider the simpler standard SEIR model to demonstrate the power of our method. Section II provides a complete description of this model. Section III describes how to transform the deterministic SEIR system to a new system that satisfies the spectral requirements needed to apply the center manifold theory. After the theory is used to find the evolution equations that describe the dynamics on the center manifold, we show in Sec. IV how the reduced model that is found using the deterministic result incorrectly projects the noise onto the center manifold. Section V demonstrates the use of a stochastic normal form coordinate transform to correctly project the noise onto the stochastic center manifold. A discussion and the conclusions are contained, respectively, in Secs. VI and VII. We begin by describing the stochastic version of the SEIR model found in Ref. 26 . We assume that a given population may be divided into the following four classes that evolve in time: ͑1͒ Susceptible class s͑t͒ consists of those individuals who may contract the disease. ͑2͒ Exposed class e͑t͒ consists of those individuals who have been infected by the disease but are not yet infectious. ͑3͒ Infectious class i͑t͒ consists of those individuals who are capable of transmitting the disease to susceptible individuals. ͑4͒ Recovered class r͑t͒ consists of those individuals who are immune to the disease. Furthermore, we assume that the total population size, denoted as N, is constant and can be normalized to S͑t͒ + E͑t͒ + I͑t͒ + R͑t͒ = 1, where S͑t͒ = s͑t͒ / N, E͑t͒ = e͑t͒ / N, I͑t͒ = i͑t͒ / N, and R͑t͒ = r͑t͒ / N. Therefore, the population class variables S, E, I, and R represent fractions of the total population. The governing equations for the stochastic SEIR model are Schematic demonstrating the injection of a common noise into both the high-dimensional system and its associated low-dimensional system. where i is the standard deviation of the noise intensity D i = i 2 / 2. Each of the noise terms i describes a stochastic, Gaussian white force that is characterized by the correlation functions Additionally, represents a constant birth and death rate, ␤ is the contact rate, ␣ is the rate of infection, so that 1 / ␣ is the mean latency period, and ␥ is the rate of recovery, so that 1 / ␥ is the mean infectious period. Although the contact rate ␤ could be given by a time-dependent function ͑e.g., due to seasonal fluctuations͒, for simplicity, we assume ␤ to be constant. Throughout this article, we use the following parameter values: = 0.02͑yr͒ −1 , ␤ = 1575.0͑yr͒ −1 , ␣ =1/ 0.0279͑yr͒ −1 , and ␥ =1/ 0.01͑yr͒ −1 . Disease parameters correspond to typical measles values. 26, 27 Note that any other biologically meaningful parameters may be used as long as the basic reproductive rate R 0 = ␣␤ / ͓͑␣ + ͒͑␥ + ͔͒ Ͼ 1. The interpretation of R 0 is the number of secondary cases produced by a single infectious individual in a population of susceptibles in one infectious period. As a first approximation of stochastic effects, we have considered additive noise. This type of noise may result from migration into and away from the population being considered. 28 Since it is difficult to estimate fluctuating migration rates, 29 it is appropriate to treat migration as an arbitrary external noise source. Also, fluctuations in the birth rate manifest itself as additive noise. Furthermore, as we are not interested in extinction events in this article, it is not necessary to use multiplicative noise. In general, for the problem considered here, it is possible that a rare event in the tail of the noise distribution may cause one or more of the S, E, and I components of the solution to become negative. In this article, we will always assume that the noise is sufficiently small so that a solution remains positive for a long enough time to gather sufficient statistics. Even though it is difficult to accurately estimate the appropriate noise level from real data, our choices of noise intensity lie within the huge confidence intervals computed in Ref. 29 . The case for multiplicative noise will be considered in a separate paper. Although S + E + I + R = 1 in the deterministic system, one should note that the dynamics of the stochastic SEIR system will not necessarily have all of the components sum to unity. However, since the noise has zero mean, the total population will remain close to unity on average. Therefore, we assume that the dynamics is sufficiently described by Eqs. ͑1a͒-͑1c͒. It should be noted that even if E͑t͒ + I͑t͒ = 0 for some t, the noise allows for the reemergence of the epidemic. One way to reduce the dimension of a system of equations is through the use of deterministic center manifold theory. In general, a nonlinear vector field can be transformed so that the linear part ͑Jacobian͒ of the vector field has a block diagonal form where the first matrix block has eigenvalues with positive real part, the second matrix block has eigenvalues with negative real part, and the third matrix block has eigenvalues with zero real part. 30, 31 These blocks are associated, respectively, with the unstable eigenspace, the stable eigenspace, and the center eigenspace. If we suppose that there are no eigenvalues with positive real part, then orbits will rapidly decay to the center eigenspace. In order to make use of the center manifold theory, we must transform Eqs. ͑1a͒-͑1c͒ to a new system of equations that has the necessary spectral structure. The theory will allow us to find an invariant center manifold passing through the fixed point to which we can restrict the transformed system. Details regarding the transformation can be found in Sec. III A, and the computation of the center manifold can be found in Sec. III B. Our analysis begins by considering the governing equations for the stochastic SEIR model given by Eqs. ͑1a͒-͑1c͒. We neglect the i i ͑t͒ terms in Eqs. ͑1a͒-͑1c͒ so that we are considering the deterministic SEIR system. This deterministic system has two fixed points. The first fixed point is and corresponds to a disease free or extinct equilibrium state. The second fixed point corresponds to an endemic state and is given by To ease the analysis, we define a new set of variables, S, Ē , and Ī, as S͑t͒ = S͑t͒ − S 0 , Ē ͑t͒ = E͑t͒ − E 0 , and Ī͑t͒ = I͑t͒ − I 0 . These new variables are substituted into Eqs. ͑1a͒-͑1c͒. Then, treating as a small parameter, we rescale time by letting t = . We may then introduce the following rescaled parameters: ␣ = ␣ 0 / and ␥ = ␥ 0 / , where ␣ 0 and ␥ 0 are O͑1͒. The inclusion of the parameter as a new state variable means that the terms in our rescaled system which contain are now nonlinear terms. Furthermore, the system is augmented with the auxiliary equation d / d = 0. The addition of this auxiliary equation contributes an extra simple zero eigenvalue to the system and adds one new center direction that has trivial dynamics. The shifted and rescaled augmented system of equations is where the endemic fixed point is now located at the origin. The Jacobian of Eqs. ͑5a͒-͑5d͒ is computed to zeroth order in and is evaluated at the origin. Ignoring the components, the Jacobian has only two linearly independent eigenvectors. Therefore, the Jacobian is not diagonalizable. However, it is possible to transform Eqs. ͑5a͒-͑5c͒ to a block diagonal form with the eigenvalue structure that is needed to use the center manifold theory. We use a transformation matrix P consisting of the two linearly independent eigenvectors of the Jacobian along with a third vector chosen to be linearly independent. There are many choices for this third vector; our choice is predicated on keeping the vector as simple as possible. This transformation matrix is given as Using the fact that ͑S , Ē , Ī͒ T = P · ͑U , V , W͒ T , then the transformation matrix leads to the following definition of new variables: U, V, and W, The application of the transformation matrix to Eqs. ͑5a͒-͑5c͒ leads to the transformed evolution equations given by The Jacobian of Eqs. ͑8a͒-͑8d͒ to zeroth order in and evaluated at the origin is which shows that Eqs. ͑8a͒-͑8d͒ may be rewritten in the form where x = ͑U͒, y = ͑V , W͒, A is a constant matrix with eigenvalues that have negative real parts, B is a constant matrix with eigenvalues that have zero real parts, and f and g are nonlinear functions in x, y, and . In particular, Therefore, the system will rapidly collapse onto a lowerdimensional manifold given by center manifold theory. 32 Furthermore, we know that the center manifold is given by where h is an unknown function. Substitution of Eq. ͑14͒ into Eq. ͑8a͒ leads to the following center manifold condition: In general, it is not possible to solve the center manifold condition for the unknown function h͑V , W , ͒. Therefore, we perform the following Taylor series expansion of h͑V , W , ͒ in V, W, and : where h 0 , h 2 , h 3 , h ,¯are unknown coefficients that are found by substituting the Taylor series expansion into the center manifold condition and equating terms of the same order. By carrying out this procedure using a second-order Taylor series expansion of h, the center manifold equation is where ⑀ = ͉͑V , W , ͉͒ so that ⑀ provides a count of the number of V, W, and factors in any one term. Substitution of Eq. ͑17͒ into Eqs. ͑8b͒ and ͑8c͒ leads to the following reduced system of evolution equations, which describe the dynamics on the center manifold We now return to the stochastic SEIR system given by Eqs. ͑1a͒-͑1c͒. The shift in the fixed point to the origin will not have any effect on the noise terms, so that the stochastic version of the shifted equations is As Eqs. ͑19a͒-͑19c͒ are transformed using Eqs. ͑7a͒-͑7c͒, the ␣ = ␣ 0 / scaling, the ␥ = ␥ 0 / scaling, and the t = time scaling, the noise also is scaled so that the stochastic transformed equations are given by where 043110-5 Noise projection for epidemic models Chaos 19, 043110 ͑2009͒ The stochastic terms 4 , 5 , and 6 in Eqs. ͑20a͒-͑20c͒ are still additive Gaussian noise processes. However, Eqs. ͑21a͒-͑21c͒ show how the transformation has acted on the original stochastic terms 1 , 2 , and 3 to create new noise processes, which have a variance different from that of the original noise processes. Also note that we have suppressed the argument of 4 , 5 , and 6 in Eqs. ͑20a͒-͑20c͒. The time scaling means that these noise terms should be evaluated at . The system of equations given by Eqs. ͑20a͒ and ͑21c͒ is an exact transformation of the system of equations given by Eqs. ͑1a͒-͑1c͒. We numerically integrate the original stochastic system of the SEIR model ͓Eqs. ͑1a͒-͑1c͔͒ along with the transformed stochastic system ͓Eqs. ͑20a͒-͑20c͔͒ using a stochastic fourth-order Runge-Kutta scheme with a constant time step size. The original system is solved for S, E, and I, while the transformed system is solved for U, V, and W. In the latter case, once the values of U, V, and W are known, we compute the values of S, Ē , and Ī using the transformation given by Eqs. ͑7a͒-͑7c͒. We shift S, Ē , and Ī, respectively, by S 0 , E 0 , and I 0 to find the values of S, E, and I. Figure 2 compares the time series of the fraction of the population that is infected with a disease I, computed using the original stochastic system of equations of the SEIR model with the time series of I computed using the transformed stochastic system of equations of the SEIR model. Although the two time series shown in Fig. 2 generally agree very well, there is some discrepancy. This discrepancy is due to the fact that the noise processes 4 4 , 5 5 , and It is tempting to consider the reduced stochastic model found by substitution of Eq. ͑17͒ into Eqs. ͑20b͒ and ͑20c͒, so that one has the following stochastic evolution equations ͑that hopefully describe the dynamics on the stochastic center manifold͒: ͑22b͒ One should note that Eqs. ͑22a͒ and ͑22b͒ also can be found by naïvely adding the stochastic terms to the reduced system of evolution equations for the deterministic problem ͓Eqs. ͑18a͒ and ͑18b͔͒. This type of stochastic center manifold reduction has been done for the case of discrete noise. 27 Additionally, in many other fields ͑e.g., oceanography, solid mechanics, fluid mechanics͒, researchers have performed stochastic model reduction using a Karhunen-Loève expansion ͑principal component analysis, proper orthogonal decomposition͒. 33, 34 However, this linear projection does not properly capture the nonlinear effects. Furthermore, one must subjectively choose the number of modes needed for the expansion. Therefore, even though the solution to the reduced model found using this technique may have the correct statistics, individual solution realizations will not agree with the original, complete solution. We will show that Eqs. ͑22a͒ and ͑22b͒ do not contain the correct projection of the noise onto the center manifold. Therefore, when solving the reduced system, one does not obtain the correct solution. Such errors in stochastic epidemic modeling impact the prediction of disease outbreak when modeling the spread of a disease in a population. Using the same numerical scheme previously described, we numerically integrate the complete stochastic system of transformed equations of the SEIR model ͓Eqs. ͑20a͒-͑20c͔͒ along with the reduced system of equations that is based on the deterministic center manifold with a replacement of the noise terms ͓Eqs. ͑22a͒ and ͑22b͔͒. The complete system is solved for U, V, and W, while the reduced system is solved for V and W. In the latter case, U is computed using the center manifold equation given by Eq. ͑17͒. Once the values of U, V, and W are known, we compute the values of S, Ē , and Ī using the transformation given by Eqs. ͑7a͒-͑7c͒. We shift S, Ē , and Ī, respectively, by S 0 , E 0 , and I 0 to find the values of S, E, and I. Figures 3͑a͒ and 3͑b͒ compare the time series of the fraction of the population that is infected with a disease I, computed using the complete stochastic system of transformed equations of the SEIR model ͓Eqs. ͑20a͒-͑20c͔͒ with the time series of I computed using the reduced system of equations of the SEIR model that is based on the deterministic center manifold with a replacement of the noise terms ͓Eqs. ͑22a͒ and ͑22b͔͒. Figure 3͑a͒ shows the initial transients, while Fig. 3͑b͒ shows the time series after the transients have decayed. One can see that the solution computed using the reduced system quickly becomes out of phase with the solution of the complete system. Use of this reduced system would lead to an incorrect prediction of the initial disease outbreak. Additionally, the predicted amplitude of the initial outbreak would be incorrect. The poor agreement, both in phase and amplitude, between the two solutions con-tinues for long periods of time, as seen in Fig. 3͑b͒ . We also have computed the cross correlation of the two time series shown in Figs. 3͑a͒ and 3͑b͒ to be approximately 0.34. Since the cross correlation measures the similarity between the two time series, this low value quantitatively suggests a poor agreement between the two solutions. Using the same systems of transformed equations, we compute 140 years worth of time series for 500 realizations. Ignoring the first 40 years of transient solution, the data are used to create a histogram representing the probability density p SI of the S and I values. Figure 4͑a͒ shows the histogram associated with the complete stochastic system of transformed equations, while Fig. 4͑b͒ shows the histogram associated with the reduced system of equations with a replacement of the noise terms. The color-bar values in Figs. 4͑a͒ and 4͑b͒ have been normalized by 10 −3 . One can see by comparing Fig. 4͑a͒ with Fig. 4͑b͒ that the two probability distributions qualitatively look the same. It is also possible to compare the two distributions using a quantitative measure. The Kullback-Leibler divergence or relative entropy measures the difference between the two probability distributions as where P i,j refers to the ͑i , j͒th component of the probability density found using the complete stochastic system of transformed equations ͓Fig. 4͑a͔͒ and Q i,j refers to the ͑i , j͒th component of the probability density found using the reduced system of equations ͓Fig. 4͑b͔͒. In our numerical computation of the relative entropy, we have added 10 −10 to each P ij and Q ij . This eliminates the possibility of having a Q ij =0 in the denominator of Eq. ͑23͒ and does not have much of an effect on the relative entropy. If the two histograms were identical, then the relative entropy given by Eq. ͑23͒ would be d KL = 0. The two histograms shown in Figs. 4͑a͒ and 4͑b͒ have a relative entropy of d KL = 0.0391, which means that the two histograms, while not identical, are quantitatively very similar. However, one cannot rely entirely on the histograms alone to say that the solutions of the complete system and the reduced system agree. As we have seen in Figs. 3͑a͒ and 3͑b͒, the two solutions have differing amplitudes and are out of phase with one another. It is important to note that these features are not picked up by the histograms of Fig. 4 . To project the noise correctly onto the center manifold, we will derive a normal form coordinate transform for the complete stochastic system of transformed equations of the SEIR model given by Eqs. ͑20a͒-͑20c͒. The particular method we use to construct the normal form coordinate transform not only reduces the dimension of the dynamics, but also separates all of the fast processes from all of the slow processes. 25 This technique has been modified and applied to the large fluctuations of multiscale problems. 17 Many publications [19] [20] [21] [22] discuss the simplification of a stochastic dynamical system using a stochastic normal form transformation. In some of this work, 19, 22 anticipative noise processes appeared in the normal form transformations, but these integrals of the noise process into the future were not dealt with rigorously. Later, the rigorous theoretical analysis needed to support normal form coordinate transforms was developed in Refs. 23 and 24. The technical problem of the anticipative noise integrals also was dealt with rigorously in this work. Even later, another stochastic normal form transformation was developed. 25 This new method allows for the "͓removal of͔ anticipation… from the slow modes with the result that no anticipation is required after the fast transients decay" ͑Ref. 25, p. 13͒. The removal of anticipation leads to a simplification of the normal form. Nonetheless, this simpler normal form retains its accuracy with the original stochastic system. 25 We shall use the method of Ref. 25 to simplify our stochastic dynamical system to one that emulates the long-term dynamics of the original system. The method involves five principles, which we recapitulate here for completeness: ͑1͒ Avoid unbounded secular terms in both the transformation and the evolution equations to ensure a uniform asymptotic approximation. ͑2͒ Decouple all of the slow processes from the fast processes to ensure a valid long-term model. ͑3͒ Insist that the stochastic slow manifold is precisely the transformed fast processes coordinate being equal to zero. ͑4͒ To simplify matters, eliminate as many as possible of the terms in the evolution equations. ͑5͒ Try to remove all fast processes from the slow processes by avoiding as much as possible the fast time memory integrals in the evolution equations. In practice, the original stochastic system of equations ͑which satisfies the necessary spectral requirements͒ in ͑U , V , W͒ T coordinates is transformed to a new ͑Y , X 1 , X 2 ͒ T coordinate system using a near-identity stochastic coordinate transform given as where the specific form of ͑Y , X 1 , X 2 , ͒, ͑Y , X 1 , X 2 , ͒, and ͑Y , X 1 , X 2 , ͒ is chosen to simplify the original system according to the five principles listed previously, and is found using an iterative procedure. To outline the procedure, we provide details for a simple example in Appendix A. Several iterations lead to coordinate transforms for U, V, and W along with evolution equations describing the Y-dynamics, X 1 -dynamics, and X 2 -dynamics in the new coordinate system. The Y-dynamics have exponential decay to the Y = 0 slow manifold. Substitution of Y = 0 leads to the coordinate transforms All of the stochastic terms in Eqs. ͑25a͒-͑25c͒ consist of integrals of the noise process into the past ͑convolutions͒, as given by Eqs. ͑26͒ and ͑27͒. These memory integrals are fast-time processes. Since we are interested in the long term slow processes and since the expectation of G equals e −Ꭽ ‫ء‬ E͓͔, where E͓͔ = 0, we neglect the memory integrals and the higher-order multiplicative terms found in Eqs. ͑25a͒-͑25c͒ so that Note that Eq. ͑28a͒ is the deterministic center manifold equation, and at first order, matches the center manifold equation that was found previously ͓Eq. ͑17͔͒. Substitution of Y = 0 and neglecting all multiplicative noise terms and memory integrals using the argument from above ͑so that we consider only first-order noise terms͒ leads to the following reduced system of evolution equations on the center manifold: The specific form of F and G in Eqs. ͑29a͒ and ͑29b͒ is complicated and is therefore presented in Appendix B. We numerically integrate the complete stochastic system of transformed equations of the SEIR model ͓Eqs. ͑20a͒-͑20c͔͒ along with the reduced system of equations that is found using the stochastic normal form coordinate trans form ͓Eqs. ͑29a͒, ͑29b͒, ͑B1a͒, and ͑B1b͔͒. The complete system is solved for U, V, and W, while the reduced system is solved for X 1 = V and X 2 = W. In the latter case, U is computed using the center manifold equation given by Eq. ͑28a͒. As before, once the values of U, V, and W are known, we compute the values of S, Ē , and Ī using the transformation given by Eqs. ͑7a͒-͑7c͒. We shift S, Ē , and Ī, respectively, by S 0 , E 0 , and I 0 to find the values of S, E, and I. 5. ͑Color online͒ Time series of the fraction of the population that is infected with a disease I, computed using the complete stochastic system of transformed equations of the SEIR model ͓Eqs. ͑20a͒-͑20c͔͒ ͑red solid line͒, and computed using the reduced system of equations of the SEIR model that is found using the stochastic normal form coordinate transform ͓Eqs. ͑29a͒, ͑29b͒, ͑B1a͒, and ͑B1b͔͒ ͑blue dashed line͒. The standard deviation of the noise intensity used in the simulation is i = 0.0005, i =4,5,6. The time series is shown for ͑a͒ t =0 to t = 40 and for ͑b͒ t =40 to t = 100. computed using the complete stochastic system of transformed equations of the SEIR model ͓Eqs. ͑20a͒-͑20c͔͒ with the time series of I computed using the reduced system of equations of the SEIR model that is found using the stochastic normal form coordinate transform ͓Eqs. ͑29a͒, ͑29b͒, ͑B1a͒, and ͑B1b͒ ͔. Figure 5͑a͒ shows the initial transients, while Fig. 5͑b͒ shows the time series after the transients have decayed. One can see that there is an excellent agreement between the two solutions. The initial outbreak is successfully captured by the reduced system. Furthermore, Fig. 5͑b͒ shows that the reduced system accurately predicts recurrent outbreaks for a time scale that is orders of magnitude longer than the relaxation time. This is not surprising since the solution decays exponentially throughout the transient and then remains close to the lower-dimensional center manifold. Since we are not looking at periodic orbits, there are no secular terms in the asymptotic expansion, and the result is valid for all time. Additionally, any noise drift on the center manifold results in bounded solutions due to sufficient dissipation transverse to the manifold. The cross correlation of the two time series shown in Fig. 5 is approximately 0.98, which quantitatively suggests there is an excellent agreement between the two solutions. Using the same systems of transformed equations, we compute 140 years worth of time series for 500 realizations. As before, we ignore the first 40 years worth of transient solution, and the data are used to create a histogram representing the probability density p SI of the S and I values. Figure 6͑a͒ shows the histogram associated with the complete stochastic system of transformed equations, while Fig. 6͑b͒ shows the histogram associated with the reduced system of equations found using the normal form coordinate transform. The color-bar values in Figs. 6͑a͒ and 6͑b͒ have been normalized by 10 −3 . As we saw with Figs. 4͑a͒ and 4͑b͒, the probability distribution shown in Fig. 6͑a͒ looks qualitatively the same as the probability distribution shown in Fig. 6͑b͒ . Using the Kullback-Leibler divergence given by Eq. ͑23͒, we have found that the two histograms shown in Figs. 6͑a͒ and 6͑b͒ have a relative entropy of d KL = 0.0953. Since this value is close to zero, the two histograms are quantitatively very similar. In addition to computing the cross correlation between the solution of the original system and the solutions of the two reduced systems for i = 0.0005, we have computed the cross correlation for the case of zero noise as well as for noise intensities ranging from = 5.0ϫ 10 −10 to = 5.0ϫ 10 −5 . These cross correlations were computed using time series from t = 800 to t = 1000. For the deterministic case ͑zero noise͒, the cross correlation between the time series which were computed using the original system and the reduced system based on the deterministic center manifold is 1.0, since the agreement is perfect. The cross correlation between the original system and the reduced system found using the stochastic normal form is also 1.0. Figure 7 shows the cross correlation between the original system and the two reduced systems for various values of . One can see in Fig. 7 that the solutions found using the reduced system based on the deterministic center manifold compare poorly with the original system at very low noise values. Furthermore, as the noise increases, the agreement between the two solutions gets worse. On the other hand, Fig. 7 shows that the solutions computed using the reduced system found using the normal form coordinate transform compare very well with the solutions to the original system across a wide range of small noise intensities. We have demonstrated that the normal form coordinate transform method reduces the Langevin system so that both the noise and dynamics are accurately projected onto the lower-dimensional manifold. It is natural to consider ͑a͒ the replacement of the stochastic term by a deterministic period drive of small amplitude and ͑b͒ the extension to finite populations. These cases are discussed, respectively, in Secs. VI A and VI B. . 6 . ͑Color online͒ Histogram of probability density p SI of the S and I values found using ͑a͒ the complete stochastic system of transformed equations for the SEIR model with mortality ͓Eqs. ͑20a͒-͑20c͔͒ and ͑b͒ the reduced system of equations of the SEIR model with mortality that is found using the stochastic normal form coordinate transform ͓Eqs. ͑29a͒, ͑29b͒, ͑B1a͒, and ͑B1b͔͒. The histograms are created using 100 years worth of time series ͑starting with year 40͒ for 500 realizations, and the color-bar values have been normalized by 10 −3 . Cross correlation between time series found using the original stochastic system of transformed equations and the reduced system of equations based on the deterministic center manifold ͑"circle" markers͒ and cross correlation between time series found using the original stochastic system of transformed equations and the reduced system of equations based on the stochastic normal form coordinate transform ͑"square" markers͒. The cross correlation is computed using time series from t = 800 to t = 1000. A single time series realization of the noise might be thought of as a deterministic function of small amplitude driving the system. One could rederive the normal form for such a deterministic function. However, since our derived normal form holds specifically for the case of white noise, we show that a simple replacement of the stochastic realization with a deterministic realization does not work. As an example, one could consider the following sinusoidal functions: 1 1 = cos͑10t͒/8000, ͑30a͒ 2 2 = sin͑4t͒/8000, ͑30b͒ where 4 4 , 5 5 , and 6 6 are given by Eqs. ͑21a͒-͑21c͒. Using Eqs. ͑30a͒-͑30c͒ or some other similar deterministic drive, the solution computed using the reduced system based on the deterministic center manifold analysis will agree perfectly with the solution computed using the complete system of equations. On the other hand, since the reduced system based on the normal form analysis was derived specifically for white noise, the transient solution found using this reduced system will not agree with the solution found using the complete system. It is possible to find a normal form coordinate transform for periodic forcing, but the normal form will be different than the one derived in this article for white noise. Figures 8͑a͒ and 8͑b͒ compare the time series of the fraction of the population that is infected with a disease I, computed using the complete system of transformed equations of the SEIR model ͓Eqs. ͑20a͒-͑20c͔͒ with the time series of I computed using the reduced system of equations of the SEIR model that is found using the stochastic normal form coordinate transform ͓Eqs. ͑29a͒, ͑29b͒, ͑B1a͒, and ͑B1b͔͒, but where the stochastic terms of both systems have been replaced by the deterministic terms given by Eqs. ͑30a͒-͑30c͒. Figure 8͑a͒ shows the initial transients, while Fig. 8͑b͒ shows a piece of the time series after the transients have decayed. One can see in Figs. 8͑a͒ and 8͑b͒ that although the two solutions eventually become relatively synchronized with one another, there is a poor agreement, both in phase and amplitude, throughout the transient. The solutions to the original system and both reduced systems are continuous solutions based on an infinite population assumption and are found using Langevin equations having Gaussian noise. It is interesting to examine the effects of general noise by using a Markov simulation to compare solutions of the original and reduced systems. The complete system in the original variables ͑see p. 2͒ will evolve in time t in the following way: transition rate ͑s − 1,e + 1,i͒ ␤si/N ͑s,e − 1,i + 1͒ ␣e ͑s,e,i − 1͒ ␥i ͑s + 1,e,i͒ N ͑s − 1,e,i͒ s ͑s,e − 1,i͒ e ͑s,e,i − 1͒ i. Using a total population size of N =10ϫ 10 6 , we have performed a Markov simulation of the system. After completing the Markov simulation, we divided s, e, and i by N to find S, E, and I. Figure 9͑a͒ shows a time series, after the transients have decayed, of the fraction of the population that is infected with a disease I. The results reflect both the mean and the frequency of the deterministic system. Performing the simulation for 500 realizations allows us to create a histogram representing the probability density p SI of the S and I values. This histogram is shown in Fig. 9͑b͒ , and one can see that the probability density reflects the amplitude, which var- ies with the population size of S and I. The color-bar values in Fig. 9͑b͒ have been normalized by 10 −4 The complete system in the transformed variables has the stable endemic equilibrium at the origin. To bound the dynamics to the first octant, we use the fact that s Ն 0, e Ն 0, and i Ն 0 to derive the appropriate inequalities for the transformed discrete variables u, v, and w. These inequalities can be found in Appendix C as Eq. ͑C1͒. These inequalities enable us to define new discrete variables Y 1 , Y 2 , and Y 3 given by Eqs. ͑C2a͒-͑C2c͒ in Appendix C. In the Y i variables, we define evolution relationships similar to those found in Eq. ͑31͒. The complete transformed system will evolve in time according to the transition and rates given by Eq. ͑C3͒ in Appendix C. After performing a Markov simulation of Eq. ͑C3͒ with a population size of N =10ϫ 10 6 , we can compare the dynamics of the transformed system with the dynamics of the original system by transforming the Y i variables in the time series back to the original s, e, and i variables. Dividing by N yields S, E, and I. Figure 10͑a͒ shows a time series, after the transients have decayed, of the fraction of the population that is infected with a disease I. The mean and the frequency agree with those found from the Markov simulation of the original system. We have performed the simulation for 500 realizations, and a histogram representing the probability density p SI is shown in Fig. 10͑b͒ . The color-bar values in Fig. 10͑b͒ have been normalized by 10 −4 . One can see in Fig. 10͑a͒ that the relative fluctuations of the I component have nearly doubled. While the fluctuation size was 0.152 for the original system, it is 0.310 for the transformed system. Additionally, the two histograms shown in Figs. 9͑b͒ and 10͑b͒ have a relative entropy of d KL = 0.9519, which means they are not in agreement. Because the simulation of the stochastic dynamics in the complete system of transformed variables does not qualitatively ͑or quantitatively͒ resemble the original stochastic system, we cannot expect that the reduced system will agree with either the original or the transformed systems. Therefore, much care should be exercised when extending the model reduction results ͑which show outstanding agreement͒ derived for a specific type of noise in the limit of infinite population to finite populations with a more general type of noise. We have considered the dynamics of a SEIR epidemiological model with stochastic forcing in the form of additive Gaussian noise. We have presented two methods of model reduction, whereby the goal is to project both the noise and the dynamics onto the stochastic center manifold. The first method uses the deterministic center manifold found by neglecting the stochastic terms in the governing equations, while the second method uses a stochastic normal form coordinate transform. Since the original system of governing equations does not have the necessary spectral structure to employ either deterministic or stochastic center manifold theory, the system of equations has been transformed using an appropriate linear transformation coupled with appropriate parameter scaling. At this stage, the first method of model reduction can be performed by computing the deterministic center manifold equation. Substitution of this equation into the complete stochastic system of transformed equations leads to a reduced system of stochastic evolution equations. The solutions of the complete stochastic system of transformed equations as well as the reduced system of equations were computed numerically. We have shown that the individual time series does not agree because the noise has not been correctly projected onto the stochastic center manifold. However, by comparing histograms of the probability density p SI of the S and I values, we saw that there was a very good agreement. This is caused by the fact that although the two solutions are out of phase with one another, their range of amplitude values is similar. The phase difference is not represented in the two histograms. This is a real drawback when trying to predict the timing of outbreaks and leads to potential problems when considering epidemic control, such as the enhancement of disease extinction through random vaccine control. 35 To accurately project the noise onto the manifold, we derived a stochastic normal form coordinate transform for the complete stochastic system of transformed equations. The numerical solution to this reduced system was compared with the solution to the original system, and we showed that there was an excellent agreement both qualitatively and quantitatively. As with the first method, the histograms of the probability density p SI of the S and I values agree very well. It should be noted that the use of these two reduction methods is not constrained to problems in epidemiology, but rather may be used for many types of physical problems. For some generic systems, such as the singularly perturbed, damped Duffing oscillator, either reduction method can be used since the terms in the normal form coordinate transform which lead to the average stochastic center manifold being different from the deterministic center manifold occur at very high order. 17 In other words, the average stochastic center manifold and deterministic center manifold are virtually identical. For the SEIR model considered in this article, there are terms at low order in the normal form transform, which cause a significant difference between the average stochastic center manifold and the deterministic manifold. Therefore, as we have demonstrated, when working with the SEIR model, one must use the normal form coordinate transform method to correctly project the noise onto the center manifold. In summary, we have presented a new method of stochastic model reduction that allows for impressive improvement in time series prediction. The reduced model captures both the amplitude and phase accurately for a temporal scale that is many orders of magnitude longer than the typical relaxation time. Since sufficient statistics of disease data are limited due to short time series collection, the results presented here provide a potential method to properly model real, stochastic disease data in the time domain. Such longterm accuracy of the reduced model will allow for the application of effective control of a disease where phase differences between outbreak times and vaccine controls are important. Additionally, since our method is general, it may be applied to very high-dimensional epidemic models, such as those involving adaptive networks. From a dynamical systems viewpoint, the reduction method has the potential to accurately capture new, emergent dynamics as we increase the size of the random fluctuations. This could be a means to identify new noise-induced phenomena in generic stochastic systems.

Comparison of distance measures in spatial analytical modeling for health service planning

BACKGROUND: Several methodological approaches have been used to estimate distance in health service research. In this study, focusing on cardiac catheterization services, Euclidean, Manhattan, and the less widely known Minkowski distance metrics are used to estimate distances from patient residence to hospital. Distance metrics typically produce less accurate estimates than actual measurements, but each metric provides a single model of travel over a given network. Therefore, distance metrics, unlike actual measurements, can be directly used in spatial analytical modeling. Euclidean distance is most often used, but unlikely the most appropriate metric. Minkowski distance is a more promising method. Distances estimated with each metric are contrasted with road distance and travel time measurements, and an optimized Minkowski distance is implemented in spatial analytical modeling. METHODS: Road distance and travel time are calculated from the postal code of residence of each patient undergoing cardiac catheterization to the pertinent hospital. The Minkowski metric is optimized, to approximate travel time and road distance, respectively. Distance estimates and distance measurements are then compared using descriptive statistics and visual mapping methods. The optimized Minkowski metric is implemented, via the spatial weight matrix, in a spatial regression model identifying socio-economic factors significantly associated with cardiac catheterization. RESULTS: The Minkowski coefficient that best approximates road distance is 1.54; 1.31 best approximates travel time. The latter is also a good predictor of road distance, thus providing the best single model of travel from patient's residence to hospital. The Euclidean metric and the optimal Minkowski metric are alternatively implemented in the regression model, and the results compared. The Minkowski method produces more reliable results than the traditional Euclidean metric. CONCLUSION: Road distance and travel time measurements are the most accurate estimates, but cannot be directly implemented in spatial analytical modeling. Euclidean distance tends to underestimate road distance and travel time; Manhattan distance tends to overestimate both. The optimized Minkowski distance partially overcomes their shortcomings; it provides a single model of travel over the network. The method is flexible, suitable for analytical modeling, and more accurate than the traditional metrics; its use ultimately increases the reliability of spatial analytical models.

Health service research is concerned with the investigation of how social, financial, organizational, technological, and behavioral factors affect access to health care, the quality and cost of health care, and ultimately health and well-being [1] . Distance plays a vital role in studies assessing spatial disease patterns as well as access to hospital services. In a highly complex health care environment, even micro-geographic differences in the availability of tertiary services can affect access to care [2, 3] . The study of distances from patient homes to the nearest hospital is an example where distance is often studied as a crude but objective indicator of geographic accessibility to hospital services [4] . In such situations, the measurement of actual travel distance (or travel time) on a road network is clearly the most appropriate method [5] . Health service research, however, encompasses a much broader investigation area, where spatial analytical models are employed to assist in the provision of effective accessibility to health care services. Distance is often used indirectly in these types of analysis as one of the parameters defining the model's thrust and its results. In a rapidly changing physical and social environment, transportation means and travel modes change quickly as do epidemic transmission modes, overturning traditional ways of conceptualizing and measuring distance [6] . A commonly-used distance metric is the Euclidean distance, a straight line distance measurement between two points, 'as the crow flies' [7, 8] . This method is simple and intuitive, but very few are the applications where it can yield accurate distance estimates. An alternative, well-known distance metric is the Manhattan, or taxi-cab distance: as its name suggests, it is most appropriate for grid-like road networks, typical of many North American cities, characterized by a rectangular city block pattern. The Manhattan metric measures distance between points along a rectangular path with right angle turns [9, 10] . Most commonly, travel along road networks involves a mixture of Euclidean, Manhattan, and curvilinear trajectories. There is no firm consensus on methods for selecting a distance metric [11] , nor is there much published information on the extent to which Euclidean, Manhattan, and road distances relate to one another in applied distance analysis [12, 13] . Travel along a complex, or mixed network can be usefully modeled by a class of distance metrics, known as Minkowski distance [14] , which is a general distance metric, of which the Euclidean and Manhattan metrics are special cases. This array of metrics provides flexibility and generality, in that, within a single class of metrics, a range of parameters can be selected; therefore, a single yet flexible method for measuring distance can be defined for the optimal estimation of distance on a variety of empirical road networks. One further important aspect is the funda-mental role of time in accessing health care services: if distance is a crude estimate of accessibility, travel time is a more relevant estimate. Travel time computation is no longer a prohibitively time consuming and computationally intensive task, thanks to powerful GIS software, hardware, and rich road network datasets [14] [15] [16] [17] . However, actual travel time on a road network is highly variable due to local (spatial and temporal) conditions which are hardly predictable and controllable. Because of this characteristic, travel time computations lack general validity, requiring adjustments to account for specific temporal conditions, e.g., weekend vs. weekdays, rush vs. non-rush hours, season, and weather, as well as local spatial conditions, e.g., local traffic congestion, lane closures, or proximity to amenities or popular destinations. All these reasons hamper the implementation of travel time computations in spatial analytical models, since even local analytical models require the definition of a single rule for distance measurement. A crude solution to this problem is the use of average travel time in spatial models; a more realistic solution can be obtained through the use of Minkowski distance: an optimal value of Minkowski distance can be selected to model travel time on a complex road network. Spatial data tend to exhibit characteristics that negatively impact the statistical properties of quantitative models, decreasing their reliability: spatial analytical models are designed to mitigate these negative effects. The most crucial properties of spatial data are spatial dependence (near things tend to be more similar than distant things) and non-stationarity (inconstant variability of phenomena across space) [18] . Two broad categories of spatial analytical models include spatially autoregressive (SAR) methods, which deal with spatial dependence [19] , and geographically weighted (GWR) methods, which deal with spatial non-stationarity [20] . In empirical situations, spatial dependencies and non-stationarities take up specific forms, which are a function of many factors, including the nature of the phenomena under investigation and the representation of space underpinning the model. For this reason, a simplistic application of spatial analysis, one that does not carefully model the salient aspects of phenomena, often fails to fulfill the model's primary objective, which is to enhance the model reliability. The transition from a simplistic to a customized implementation of spatial analysis requires the calibration of each parameter defining the analysis: one of the most crucial parameters, affecting the analytical results, is the distance measurement method. Cardiac catheterization is a procedure that is performed to determine presence or absence of coronary artery blockages. The procedure involves the percutaneous insertion of a catheter into the arterial system, after which it is guided into the aorta where the coronary arteries are positioned. Contrast dye is then injected into the coronary arteries so that blockages can be located and identified. In some instances, cardiac catheterization can lead to immediate use of percutaneous coronary intervention with balloon angioplasty and the insertion of coronary stents that open up partially or completely blocked arteries to restore blood flow. In some instances, this procedure is performed in stable patients where distance and travel times are a minor concern. In other instances, however, the procedure is done urgently, and for such situations, consideration of distances and travel times become a central consideration in the planning of health services. In the context of an applied study of distance between patient residence and a tertiary cardiac catheterization facility in a large city, this paper analyzes the effectiveness of a selection of distance metrics in providing a useful model of travel distance and travel time along an urban road network. The comparison of different metrics leads to the identification of a metric that is conceptually sound and computationally effective. The metric thus identified is experimentally used in a spatial autoregressive model analyzing the spatial distribution of cardiac catheterization cases in the city. The study area encompasses the City of Calgary, one of the largest Canadian cities, with approximately 1 million residents [21], distributed over a large geographic area (roughly 750 Km 2 ), characterized by diversity of population, housing type, residential density, and accessibility to heath services. Cardiac catheterization is an invasive procedure for patients experiencing cardiovascular symptoms and defines coronary anatomy, left ventricular and valvular function; it provides important prognostic information for individuals affected by cardiovascular conditions [22] . During the study period the procedure was only performed at the Foothills Medical Centre, located in the northwest of the city. Three types of data are used in this study: cardiac catheterization patient database, postal code locations, and the Calgary road network. Cardiac catheterization patient data were obtained from the Alberta Provincial Project for Outcome Assessment in Coronary Heart Disease (APPROACH), an ongoing data collection initiative, begun in 1995, producing information on all patients undergoing catheterization in Alberta [22] . The data are released at the postal code spatial aggregation level. Data were extracted for Calgary residents only and catheterizations performed over the year 2002, resulting in a total of 2, 445 catheterization cases, distributed over 2, 138 postal codes. A postal code conversion file (PCCF) [23] was obtained from Statistics Canada. Only postal codes that have at least one catheterization case are retained for the analysis. It shall be observed that postal code locations refer to the primary residence of catheterization patients, not to the place where symptoms were felt or where emergency care was first administered. The Calgary road network data were obtained from the University of Calgary data holdings, based on street information collected and compiled in 2005 by DMTI Spatial [24] . This road network was used to calculate shortest road distances from patient residence location to hospital for cardiac catheterization services. Straight line (Euclidean) distance and Manhattan distance are often used in health service research [25] . Each of these distance metrics may appropriately estimate distance in some parts of a study area, but their application at the city level tends to yield large errors in areas that depart from the dominant pattern, and may lead to highly inaccurate distance estimations. One of the reasons for using Euclidean and/or Manhattan distance is the relative ease of their implementation; in contrast, it is more problematic to design algorithms implementing actual road network distance in spatial analytical models. In order to reduce the error associated with the Euclidean and Manhattan metrics while maintaining the computational simplicity of a single, intuitive mathematical formula, the general Minkowski metric is examined, to devise a single method that best approximates the average pattern of an empirical road network. Optimizing values of the Minkowski formula are calculated for road distance as well as travel time; the results are compared with more traditional distance measures in the context of assessing geographic accessibility to cardiac facilities. The Minkowski distance has the potential to provide a more accurate estimate of road network distance and travel time than the Euclidean and Manhattan metrics. A set of 2, 138 distances between each patient's postal code of residence and the Foothills hospital are calculated according to each of the distance measurement methods considered. The geographic locations of each postal code from the PCCF and the hospital are recorded in latitude and longitude; therefore, in order to implement distance computations, the road network is projected using an equidistant projection system, which is chosen in order to preserve distance and produce consistent distance measurements [26] . Latitude and longitude coordinates are then converted into Eastings and Northings, i.e., x and y values, expressed in kilometers. Alternative methods could have been used, for example the great circle distance formula [27] , which, however, provides rougher estimations. The ArcGIS 9.3 [28] Geometry calculator was used to calculate the x and y coordinates based on the projected dataset and the resulting x and y values were used in the distance formulas defined below. Euclidean [7, 8] , Manhattan [9, 10] , and Minkowski [14] distance can be calculated by the formula: where, for this application: d is the distance between a patient's residence and the hospital; x i , y i are the geographic coordinates of the centroid of each postal code of residence; x j , y j are the geographic coordinates of the Foothills hospital. The generic p parameter in Equation 1 can be replaced by the value 2 to yield the well known Euclidean distance; the value 1 would yield the Manhattan distance, and all the intermediate values in the in the [1 <p < 2] interval yield an array of Minkowski distances ( Figure 1 ). The computation of road network distances (shortest distance between two locations along a road network) is implemented directly in GIS software [28] using a shortest path algorithm. It is recognized a priori that this method depicts the actual travel trajectory and is likely to produce the most accurate distance estimates for patient travel routes to a hospital. Likewise, travel time calculated over the road network using an appropriate algorithm is likely to yield the most accurate travel time estimates. The basic need for road distance calculation is a road network with information on all the segments constituting the network, as well as on all existing constraints such as prohibited left or right turns, one way streets, etc. Travel time calculations, likewise, are implemented in GIS software [28] . Additional information required for travel time calculation include estimated speed and length of road segments, along with an algorithm capable of taking all these factors into consideration [15] . The Network Analyst extension within ArcGIS [28] enables the modeling of spatial networks and provides the tools for road distance calculations [28, 29] and travel time calculations [15, 16] from multiple postal code locations to a hospital. Figure 1 illustrates road network distance, along with the distance metrics considered in this study. A simple procedure can be implemented to select, within the [1 ≤ p ≤ 2] interval, the value of the parameter p in the distance formula that best approximates distance along a given road network. In light of the applied focus of this analysis, an empirical solution to this problem is sought for the set of 2, 138 patient-to-hospital distances. As a first step for determining the best Minkowski distance, Equation 1 is transformed into Equation 2: Two new quantities are defined as X = (x i -x j ) and Y = (y iy j ), and replaced in Equation 2, which is also further modified by means of a logarithmic transformation. Equation 2 is solved for p values in the [1 ≤ p ≤ 2] interval, sampled at regular intervals. The set of distances thus calculated are considered approximations of the road distance. A simple regression model is defined, where the dependent variable is the road distance, and the independent variable is, in turn, the distance obtained by each p value. Goodness-offit, residuals, and other regression diagnostics are then compared over the entire interval [30, 31] . This simple method helps assess and rank the various coefficients, identifying the one that produces the highest R 2 value, which is considered the best Minkowski coefficient. The procedure is then modified to determine the optimal Minkowski coefficient for travel time: conceptually, this experiment is less straightforward, because travel time is a measure of time, whereas Minkowski remains a measurement of distance in space. Once travel time is computed, in order to make the transition between time and space, in terms that are valid both conceptually and computationally, the concept of speed, a simple ratio between space and time, is introduced. Average speed over the city is calculated, yielding the following values: the average distance traveled in one hour is 58.71 km; conversely, the average time required to travel 1 kilometer is 1.02 minutes. For the sake of simplicity, in order to make the argument more intuitive, these values were rounded to an average speed of 60 km/h or 1 minute to travel 1 kilometer. It shall be observed that these values are obtained under optimal conditions, i.e., without considering delays due to rush hour, traffic congestion, traffic lights, stop signs, road closures, weather, etc. Travel time calculations are then used as the dependent variable, while the independent variables remain unchanged. To this end, actual travel times are converted to distances, using the average travel time, and replaced in the procedure for the calculation of the optimal Minkowski p value to approximate travel time. Once the two optimal Minkowski coefficients are identified, standard descriptive statistics are used to analyze and compare the four distance metrics and the two empirical distance measurements. The results of the regression models estimated to identify the optimal Minkowski coefficients are summarized in Figure 2 . Figure 2a summarizes the regressions for the road distance coefficient optimization, and Figure 2b for the optimization of the travel time coefficient. Figure 2 shows different values of a goodness-of-fit indicator (R 2 ) obtained from a regular sample of p values throughout the [ Visual illustration of road distance and distance metrics the methods used to measure distance, it displays the second largest standard deviation and range. Travel time can be immediately compared with all the distance metrics, as 1 minute corresponds to 1 kilometer: in comparison with road network distance, travel time displays a lower standard deviation and a shorter range, with a minimum travel time of 1.51 minutes, and a maximum of 26.76 minutes, under optimal conditions. Conversely, its mean displays the highest value, 12.11 minutes. Shorter range and higher mean are likely due to the opposite effects of speed limits on road segments of different length: in general, longer travel paths include large segments that occur on major roads, where speed limits are higher, whereas shorter trips tend to occur on minor roads, which are associated with lower speed limits. As a consequence, a model that takes speed into consideration produces a shorter travel range. Consistently, the standard deviation is lower. The higher mean value suggests that overall speed limits tend to slow down travel; that is, segments with low speed limits have a large impact on the overall travel pattern throughout the road network. Euclidean distance tends to underestimate road distance, as shown by the mean and range; the standard deviation is lower than for road distance, with values that are fairly close to those for travel time. This is probably due to the smoothing effect of a uniform distance model, which produces lower values than actual road distance for curvilinear segments and the most complex paths. Manhattan distance tends to overestimate road distance and produces values consistently larger than those of Euclidean distance. Its mean value is very close to the road distance mean, but it also presents the largest standard deviation and the largest range. This may suggest that the Calgary road network contains several parts that follow the Manhattan pattern (hence a similar mean value), but the presence of different patterns in the same network increases its error (large standard deviation). In comparing Euclidean and Manhattan distances, Manhattan distance produces a close approximation of the mean, and only slightly overestimates the standard deviation, whereas Euclidean largely underestimates both values. For the range, both metrics produce approximately the same error, though with opposite sign. This suggests that, overall, Manhattan is a better model than Euclidean for the Calgary road network. The value of p = 1.54 best approximates road distance in the Minkowski formula. Indeed minimum and maximum values are close approximations, whereas mean and standard deviation underestimate road distance. This result can be considered satisfactory, as it indicates that the error is minimized for individual measurements, but overall the method displays the aforementioned smoothing effect, whereby mean measurements tend to be slightly smaller than actual ones, with overall lower variations around the mean. It shall be observed that this distance metric is approximately half-way between Euclidean and Manhattan; however, all the descriptive statistics present values that are closer to the Euclidean than the Manhattan results. The value of p = 1.31 best approximates travel time in the Minkowski formula. Interestingly, the descriptive statistics suggest that this metric is the best approximation of road distance; indeed a close approximation. As noted, Determination of optimal p values for Minkowski distance Figure 2 Determination of optimal p values for Minkowski distance. travel time presents features that differ from distance, i.e., larger mean and lower range: this combination is hard to achieve by the class of metrics considered, because p values closer to 1 (Manhattan) produce larger means, whereas p values closer to 2 (Euclidean) produce lower ranges. In light of these considerations, a relatively low p value, such as p = 1.31 most closely approximates this pattern, rendering a model that is less extreme than Manhattan, as shown by a lower standard deviation. Overall the metric p = 1.31 provides a model of travel that is intermediate between measured road distance and travel time measured under optimal conditions. This is supported by almost all the descriptive statistics; the greatest shortcoming of this metric is its poor approximation of the mean. This value produces the best of all the metrics examined. Finally, it shall be noted that the travel time pattern would be very different if measurements were to consider less favorable conditions. The most common impediments to fast travel in Calgary include traffic congestion, e.g., rush hour, and severe winter weather conditions. Traffic congestion tends to affect major roads more heavily, where feasible speed can easily be reduced by 20-25% of the speed limit, whereas its effect is generally lesser on minor roads. Severe winter weather conditions are likely to have a comparable effect on major roads, but they will also have comparable or worse effects on minor roads. However, a speed reduction from 80 to 60 km/h on long road segments significantly impacts travel time, lowering the maximum and range values, and increasing the mean value. Conversely, an equivalent or greater speed reduction from 50 to 40 or 35 km/h on shorter road segments is likely to have only a minor impact on the overall travel time. Seeking to approximate such travel pattern via a Minkowski p value is therefore unlikely to produce better results. Table 2 presents a selection of summary statistics of the differences between each empirical measurement and the distance metrics considered. These are only global results, overshadowing the performance of each distance model throughout the city. These results suggest that, globally, the differences produced by each model are modest. For road distance, the smallest differences are achieved by the Manhattan and Minkowski (p = 1.31) metrics, with Manhattan producing the overall best result. The same metrics produce the best results for travel time, with Minkowski (p = 1.31) producing the lowest standard deviation. Median and mean differences are lower than 2 kilometers and just over 2 minutes, respectively. However, standard deviations tend to be quite high, relative to the mean. Figure 3 and Figure 4 illustrate graphically these differences, providing greater spatial detail. Differences between road distance and each distance metric are presented in Figure 3 , and those between travel time and each metric in Figure 4 . The figures display spatially the magnitude of the error associated with each distance metric in each part of the city. With respect to road distance (Figure 3 ), Euclidean distance produces only negative errors, but in some peripheral areas of the city these errors are large in absolute value. Manhattan distance produces mostly positive errors, and overall it produces better results, as the areas characterized by the highest absolute errors are reduced to a triangle west of the hospital and the far southeast corner of the city. The Minkowski metric with p = 1.54 improves over the Manhattan metric results, by resolving the area of large residuals in the southeast corner; conversely, the area of high residuals west of the hospital is moderately larger. The best results are produced by the Minkowski metric with p = 1.31, as for the entire eastern part of the city errors are contained in the interval 0-2.5 km, and greater errors remain only in some peripheral northwest areas. The area west of the hospital consistently emerges as an outlier, despite its close distance to the hospital. Careful observation of the topography of the area reveals that the hospital is located on the east side of a hill, with a river running to the west of the hill. Therefore, no immediate access is possible from the west side of the river to the hospital, and patients are left with no option but to travel a considerable distance either northbound or southbound With respect to travel time ( Figure 4 ) the improvement obtained by the two Minkowski metrics is more evident throughout the city. Areas of large residuals remain in the far southeast and southwest corners, and, in general each metric performs worse in the west than in the east part of the city. The latter observation is counterintuitive, since the hospital is located in the northwest; however, that part of the city, closer to the foothills, is characterized by a more complex topography. Moreover, in a city like Calgary, urban design and age of communities play an important role, as, over the decades, rectangular patterns have alternated with such patterns as crescent and cul-de-sac, and other typologies of urban connectivity. Likewise, it shall be observed that the most peripheral areas are very recent developments, and likely the planned road connections had not been completed during the study period. Within the scope of a larger project, the association between cardiovascular disease and socio-economic variables was recently analyzed [32] . Specifically, the relationship between cardiac catheterization and socio-economic variables was analyzed by means of a multivariate spatially autoregressive model. While distance does not explicitly enter in these models, it is one of the key parameters defining the spatial weight matrix (19) , which represents the neighborhood definition, hence the model's ability to cope with spatial dependencies, and ultimately the reliability of the model estimates. Figure 5 provides an example of how different distance metrics affect the neighborhood configuration defined by a spatial weight matrix. Locations in the figure represent census tract centroids, as these relatively larger spatial units are used in these regression models. The lines connecting these locations indicate whether or not 2 close locations are considered neighbors and included in the estimation of spatial dependence. Other parameters contribute to the neighborhood definition: generally the most influential param-eter is the number of nearest neighbors, complemented by a distance decay function and a weight [33] . Figure 5a shows the connectivity defined by the Euclidean metric, while Figure 5b corresponds to the Minkowski metric optimizing travel time (p = 1.31). A close comparison of the two plots (aided by the superimposed circles) reveals how the modification of the distance metric substantially alters the neighborhood configuration. Each neighborhood configuration, such as the ones presented in Figure 5 , forms the basis for the definition of a spatial weight matrix, which is one of the crucial elements that define a spatial regression model, differentiating it from a standard (non-spatial) model. Through the spatial weight matrix, the neighborhood configuration ultimately affects the variance of the model estimates, hence their reliability. As an example, Table 3 shows the variation in the parameter estimates and the regression diagnostics determined by the two alternative neighborhood configurations depicted in Figure 5 . The regression model analyzes the socio-economic variables significantly associated with cardiac catheterization: the main predictors are family status, income, and educational attainments; the spatial distribution of these variables is used to identify areas of social and economic concern. This model thus provides an effective analytical tool to support policy decisions, providing guidance for the initiation of targeted, localized preventative health measures [32] . A succinct summary of the parameters associated with each independent variable is presented in Table 3 , along with a small selection of regression diagnostics. While the coefficients (beta) linking each independent to the dependent variable remain substantially unchanged, the Minkowski distance leads to increased values of their associated t test: the t values increase thanks to a reduction of the variance associated with the estimates. All the independent variables benefit, in varying degrees, from the modified distance model. Likewise, the regression diagnostics indicate that the distance model does not appreciably affect the overall goodness of fit (represented by the Differences between road distance and distance metrics Figure 3 Differences between road distance and distance metrics. pseudo-R 2 ), but it does noticeably impact the autoregressive coefficient and, more importantly, the spatial dependence in the regression residuals. These results confirm the importance of an accurate distance model to enhance the reliability of spatial analysis for health service research. This study compares four different distance metrics, i.e., Euclidean, Manhattan, and Minkowski distance (the latter for two different coefficients), and contrasts them with road distance and travel time, respectively, in the context of applied health services research. The Euclidean metric is the most common and intuitive measure of distance; Manhattan is another common distance measure, but very rarely does either of them provide a close approximation of an empirical road network, such as an urban network. Road distance, conversely, provides an accurate measurement, but it is prone to local features, and does not provide a single model of travel throughout the urban network. Travel time is, arguably, the most relevant estimate of distance, but its calculation introduces further specificities, as temporal anomalies are added to the local features, further reducing its generality. For these reasons, Minkowski distance is a promising solution: it provides a general model of travel throughout an empirical network; it possesses a large range of parameters, which enhance its flexibility; it can be easily calculated; and it can provide a less crude approximation of travel along a road network. Euclidean distance is widely used in distance analyses in the literature [25] but it tends to underestimate road distance and travel time. Manhattan distance, on the contrary, tends to overestimate road distance and travel time. The use of either of these two metrics in any spatial anal-Neighborhood configurations determined by different distance metrics ysis may result in inaccurate results [10] . Ideally, travel time provides the most accurate estimate of a patient's travel from their place of residence to a given health facility [15] . While accessibility studies may profitably employ this method, its use in spatial analytical models is inhibited by the particularity of its calculation, which tends to be heavily affected by local features, both in space and in time. A simple, empirical optimization procedure led to the identification of the coefficients, in the Minkowski formula, that best approximate road distance and travel time, respectively. Summary statistics and cartographic representations consistently indicate the value p = 1.31 as the best coefficient, i.e., the one that leads to the most accurate approximation of both road distance and travel time. The model of distance based on this coefficient was experimentally introduced in spatial analytical routines, to define neighborhood connectivity and determine the spatial weight matrix for multivariate spatial regression analysis. The enhanced reliability of the spatial analytical model based on the optimal distance model far outweighs the cost of the computational procedure that leads to the coefficient selection. The advantage of the procedure discussed in this paper can be best appreciated by considering that empirical measurements of distance and travel time cannot be implemented in spatial analytical modeling, exactly because of their empirical nature and the great impact they receive from local features and conditions. For this reason, the optimized Minkowski coefficient represents a valuable compromise between an approach that is often simplistic (i.e., Euclidean and Manhattan metrics) and the ideal, but impractical sophistication of empirical measurements (i.e., road distance and travel time, respectively). This study has some limitations. Most importantly, it is limited to one class of distance metrics, i.e., Minkowski, whereas other metrics could be considered: Mahalanobis distance is just one example [34] . There are locational inaccuracies in the patient data [35] as well as in the road network; additional errors are likely to derive from the algorithm used for the distance calculations. Travel time calculations are based on assumptions, referred to as optimal conditions that tend to represent an ideal, but unlikely situation. The hypothesis of optimal conditions should be lifted, and more realistic conditions should be entered in the model, e.g., rush hour, or severe winter weather, and should be considered not just individually, but also jointly. An array of optimal Minkowski coefficients should consequently be calculated for the varying conditions, leading to the final identification of a stochastic optimum. The entire procedure is also based on the strong assumption of a single transportation mode: it can be argued that optimal conditions approximate ambulance travel, but the catheterization registry used in this analysis was not limited to patients who were transported directly from their residence to the tertiary catheterization facility. The APPROACH registry also includes patients initially admitted to hospitals or emergency facilities without catheterization facilities who were subsequently transferred to a tertiary center for this procedure under less urgent circumstances. Accuracy in travel route selection and in travel time estimates are most relevant to patients with more emergent cardiac conditions requiring rapid transportation from their residence to the catheterization facility. This limitation can be addressed by a twofold model, optimizing for ambulance and private vehicle travel. Still, road network travel is a reasonable assumption for urban environments, but is unlikely to be the sole mode of transportation to emergency care from rural and remote locations. The proposed approach was tested, as an example, on a spatial autoregressive model; however, virtually all spatial analytical techniques involve some distance measurement. The impact of alternative distance measurements was examined in this implementation through an analysis of the spatial weight matrix, but distance is likely to impact other analytical techniques in further, different ways [19, 20] . Given its advantages and limitations, Minkowski distance appears to be most usefully implemented in spatial analytical modeling; however, other useful applications can be envisaged, particularly in geographic areas characterized by paucity or unreliability of spatial data, or by high dynamism. Examples include urban or regional road networks of countries with poor spatial digital records, or characterized by high population mobility or varying transportations routes, for example due to seasonal variations. In all such cases, a model of travel in the area, obtained by the method discussed in this paper, can provide rough, but reasonable distance estimates, potentially useful for facility planning, or as initial input for more sophisticated analyses. The proposed method provides a single model of travel on an urban road network, via the identification of an optimal coefficient within a class of distance metrics, known as Minkowski metrics. The coefficient can be optimized to approximate different distances, e.g., road distance or travel time, under varying conditions. The resulting distance model can be usefully input in spatial analytical models, providing a method for the estimation of less simplistic spatial analytical models, by means of a more accurate representation of distance. Such models yield more reliable estimates, hence more effective tools for health service planning and management.

Evaluation of Five Decontamination Methods for Filtering Facepiece Respirators

Concerns have been raised regarding the availability of National Institute for Occupational Safety and Health (NIOSH)-certified N95 filtering facepiece respirators (FFRs) during an influenza pandemic. One possible strategy to mitigate a respirator shortage is to reuse FFRs following a biological decontamination process to render infectious material on the FFR inactive. However, little data exist on the effects of decontamination methods on respirator integrity and performance. This study evaluated five decontamination methods [ultraviolet germicidal irradiation (UVGI), ethylene oxide, vaporized hydrogen peroxide (VHP), microwave oven irradiation, and bleach] using nine models of NIOSH-certified respirators (three models each of N95 FFRs, surgical N95 respirators, and P100 FFRs) to determine which methods should be considered for future research studies. Following treatment by each decontamination method, the FFRs were evaluated for changes in physical appearance, odor, and laboratory performance (filter aerosol penetration and filter airflow resistance). Additional experiments (dry heat laboratory oven exposures, off-gassing, and FFR hydrophobicity) were subsequently conducted to better understand material properties and possible health risks to the respirator user following decontamination. However, this study did not assess the efficiency of the decontamination methods to inactivate viable microorganisms. Microwave oven irradiation melted samples from two FFR models. The remainder of the FFR samples that had been decontaminated had expected levels of filter aerosol penetration and filter airflow resistance. The scent of bleach remained noticeable following overnight drying and low levels of chlorine gas were found to off-gas from bleach-decontaminated FFRs when rehydrated with deionized water. UVGI, ethylene oxide (EtO), and VHP were found to be the most promising decontamination methods; however, concerns remain about the throughput capabilities for EtO and VHP. Further research is needed before any specific decontamination methods can be recommended.

During an influenza pandemic, a shortage of filtering facepiece respirators (FFRs) may occur if manufacturing production is unable to meet the demand or if FFR stockpiles become depleted. According to a 2006 report from the National Academies' Institute of Medicine, over 90 million N95 FFRs will be needed to protect workers in the healthcare sector during a 42-day influenza pandemic outbreak (Bailar et al., 2006) . Guidance provided by the Centers for Disease Control and Prevention (CDC) states that once an FFR is worn in the presence of an infected patient, it should be considered potentially contaminated and not be reused by the same person or a coworker (CDC, 2007) . A contaminated FFR could potentially serve as a fomite and lead to self-inoculation or spread of the organism to patients and other healthcare workers. Guidance from the Occupational Safety and Health Administration (OSHA) considers *Author to whom correspondence should be addressed. Tel: +412-386-4001; fax: +412-386-6864; e-mail: rshaffer@cdc.gov FFRs to be one-time-use devices when used in the presence of infected patients and advises employers and employees to only reuse FFRs during a pandemic if FFRs are in short supply and the device has not been obviously soiled or damaged (e.g. creased or torn), and it retains its ability to function properly (OSHA, 2007) . One possible strategy to reduce the impact of a respirator shortage would be to apply a biological decontamination process (e.g. such as those used in hospital settings for infection control) to inactivate the influenza virus that may be on the FFR. If the treatment did not deteriorate the FFR or leave potentially toxic residues on the FFR, then it could be available for subsequent reuse by the original user. Until recently, no data were published on the effects of decontamination on FFR performance. Viscusi et al. (2007) measured the laboratory filtration performance of one N95 model and one P100 model FFR that were exposed to 20 different biological decontamination treatments. They found that filtration performance after onetime decontamination treatments using bleach, ethylene oxide (EtO), microwave oven irradiation, ultraviolet germicidal irradiation (UVGI), and hydrogen peroxide (vaporized and liquid forms) was observed to have filter aerosol penetration values that remained less than the National Institute for Occupational Safety and Health (NIOSH) certification criteria. It was also found that decontamination using an autoclave, 160°C dry heat, 70% isopropyl alcohol, and soap and water (20-min soak) caused significant degradation to filtration efficiency. Expanding on that research, the goal of this study was to further evaluate five of the decontamination methods examined in the previous study using a more diverse set of nine models of NIOSH-certified FFRs to determine which decontamination methods should be considered for future research studies. The biological decontamination methods used in this study include: (i) UVGI, (ii) EtO, (iii) vaporized hydrogen peroxide (VHP), (iv) microwave oven irradiation, and (v) 0.6% aqueous solution of sodium hypochlorite (hereafter referred to as 'bleach'). Following treatment by each decontamination method, FFRs were evaluated for changes in physical appearance/odor (observational analysis) and laboratory performance (filter aerosol penetration and filter airflow resistance). Additional experiments were then conducted to examine the material properties of the FFRs in an attempt to rationalize some of the findings in the laboratory performance evaluation and observational analysis. The advantages and disadvantages of the various decontamination methods (including throughput capacity and possible health risk to the user) were also assessed. Nine respirator models were used in this study, of which six models [three N95 FFR models (N95-A, N95-B, and N95-C) and three surgical N95 respirator models (SN95-D, SN95-E, and SN95-F)] constitute a random sampling from those N95 FFR models present in the US Strategic National Stockpile (SNS). Healthcare workers often use surgical N95 respirators, which are NIOSH-approved N95 FFRs that also have been cleared by the US Food and Drug Administration (FDA) for marketing as medical devices. Surgical N95 respirators are designed to be fluid resistant to splash and spatter of blood and other infectious materials and thus may respond differently to the decontamination processes than N95 FFRs. Three models of P100 FFRs (P100-G, P100-H, and P100-I) were randomly selected from models commercially available at the time of the study and included because they were considered likely to be more resistant to filtration efficiency degradation and thus offer a more rigorous basis of comparison. All respirators were purchased and verified to be from the same respective manufacturing lot at the beginning of the study to minimize any lot-to-lot variation as well as to ensure consistency during FFR filtration performance testing. FFRs used in this study consisted of electrostatically charged polypropylene filters (electret filter media). The experimental conditions and parameters for the five decontamination methods and the 'asreceived' (control) method are summarized in Table 1 . All laboratory experiments were conducted under standard laboratory conditions (21 -2°C and relative humidity of 50 -10%) on triplicate sets of FFRs. Observational analysis. All post-decontamination and control FFR samples were inspected and scrutinized carefully for any visible sign of degradation or changes that could be noted in texture or 'feel' of the respirator (softness, pliability, coarseness, roughness, etc.). All samples were sniffed for any discernible odor or smell. Filter aerosol penetration. A Model 8130 Automated Filter Tester (AFT) (TSI, Inc., St Paul, MN, USA) was used to measure initial filter aerosol penetration for all post-decontamination and control FFR samples. All tests were conducted at room temperature with a continuous airflow of 85 -4 l min À1 in accordance with NIOSH certification test procedures (NIOSH, 2007) for challenging N-series filters, with two exceptions: all filters were tested for filter aerosol penetration without any relative humidity pretreatment or NaCl aerosol loading. Collecting the data in this manner allows consistency with previous work (Viscusi et al., 2007) . Filter aerosol penetration levels were determined using a Plexiglas test box as previously used and described by Viscusi et al. (2007) or an appropriately sized test fixture supplied by the respective FFR manufacturer, as was the case for models N95-C, SN95-D, and P100-H. Filter airflow resistance. For all control and postdecontamination FFR samples, a TSI Model 8130 AFT was also used to measure initial filter airflow resistance in millimeters of water column height pressure (mmH 2 O). It must be clarified that the NIOSH certification test for inhalation airflow resistance for FFRs is not performed using the TSI 8130 AFT but is executed in accordance with NIOSH Standard Test Procedure RCT-APR-STP-0007, which specifies the use of a different calibrated apparatus incorporating a vacuum source and manometer (NIOSH, 2005) . For this evaluation, it was convenient to report the filter airflow resistance obtained from the TSI Model 8130 AFT because filter aerosol penetration and filter airflow resistance results are generated simultaneously and the intent is to determine changes in filter airflow resistance. This methodology was used previously by the National Personal Protective Technology Laboratory (NPPTL) . The primary experimental design called for 162 FFRs (nine different FFR models  six test conditions  three samples per test condition) to be tested by observational analysis, for filter airflow resistance and for filter aerosol penetration. The 162 FFRs in the design included 135 post-decontamination FFRs and 27 control FFRs (no decontamination). For statistical analysis, the six test conditions (see Table 1 ) comprised one control group and five decontamination treatments. A one-way analysis of variance (ANOVA) test was performed for each of the nine FFR models for filter aerosol penetration and filter airflow resistance (for 18 total tests). Thus, each model was treated independently due to its inherent uniqueness (difference in number of filter layers, hydrophobicity, materials of construction, etc.). Results were considered statistically significant if the P-value was ,0.05. Statistical analyses were performed using Microsoft Excel (Microsoft Corporation, part of Microsoft Office Professional Edition 2003) . No statistical analysis of the subjective observational analysis data was done. Additional secondary experiments were subsequently conducted on the FFRs to understand better their material properties. This information can be . FFRs and a chemical indicator placed in an individual standard poly/paper pouch. EtO exposure for 1 h followed by 4 h of aeration. FFRs were shipped to and from a commercial facility specializing in low-temperature sterilization methods and were tested within 72 h of receipt. VHP STERRADÒ 100S H 2 O 2 Gas Plasma Sterilizer (Advanced Sterilization Products, Irvine, CA, USA), single 55-min standard cycle. FFRs and a chemical indicator placed in an individual Mylar/Tyvekä self-seal pouch. FFRs were shipped to and from a commercial facility specializing in low-temperature sterilization methods and were tested within 72 h of receipt. Commercially available 2450 MHz, Sharp Model R-305KS (Sharp Electronics, Mahwah, NJ, USA) microwave oven with revolving glass carousel, 1100 W (manufacturer rated); 750 W ft À3 experimentally measured; 2-min total exposure (1 min each side of FFR). A paper towel was placed on the revolving glass plate for insulation to protect the FFRs from melting onto the glass plate. Using a power setting of 10 (maximum power), FFRs were placed faceseal-side down, initially, to reduce the risk of faceseal component materials melting onto the paper towel due to elevated temperatures reached by the glass plate when microwaved for 2 min. Ambient cooling of the glass plate was maintained between trials. Thirty minutes submersion in 0.6% (one part bleach to nine parts of deionized water) aqueous solution of sodium hypochlorite (original concentration 5 6% available as Cl 2 ). Manufacturing specification: 6.00 -0.06% (w/w) available chlorine; Cat no. 7495.7-1, CAS no. 7732-18-5 (Ricca Chemical Company, Pequannock, NJ, USA). After treatment, FFRs were hung on a laboratory pegboard and allowed to air-dry overnight with assistance from a freestanding fan. Evaluation of decontamination methods for FFRs 817 used to further optimize the decontamination methods and/or explain some of the findings from the observational analyses or laboratory performance evaluation experiments. Dry oven experiments. To investigate the effects on filter aerosol penetration at various dry heat temperatures and to determine if these effects were similar to those of FFRs that underwent microwave oven irradiation, new FFRs were placed in a Fisher Scientific Isotemp 500 Series laboratory oven (Fisher Scientific, Pittsburgh, PA, USA) for 1 h at temperatures ranging from 80 to 120°C. Filter aerosol penetration was measured after samples cooled to ambient temperature. Hydrophobicity testing. A qualitative assessment of water affinity for each FFR filter media layer was performed to determine the hydrophobic/ hydrophilic nature of the various layers for the nine different FFR models. For this experiment, it was hypothesized that the number of layers and the nature of the outer layer (surface of the FFR most distant from the wearer) and the inner layer (surface of the FFR closest to the breathing zone of the wearer) would provide insight into any model-specific effects associated with liquid chemical-based decontamination methods. A circular swatch ($5 cm in diameter) was cut from additional, new as-received samples of each FFR model. Following layer separations, a 100 ll aliquot of deionized water was pipetted onto the surface of each side of each layer (front and back). Two FFR models incorporated layers of plastic webbing, presumably to support shape; these layers were not tested because they are not filtering layers. A layer was noted as hydrophilic when it absorbed the water droplet. A layer was noted as hydrophobic when the water droplet beaded on the layer's surface. Chlorine off-gassing experiments. To quantify observations of discernable odor from FFRs following bleach decontamination, a series of off-gassing experiments was conducted using a Model 4340 Chlorine Gas Analyzer (Interscan Corp., Chatsworth, CA, USA). Chlorine off-gassing was measured from FFRs after bleach treatment as described in Table 1 . A subset of four FFR models was chosen for testing based on the various combinations of water repellency discerned from the hydrophobicity experiments described previously: N95-A (outer hydrophobic layer/inner hydrophilic layer), N95-B (outer and inner hydrophilic layers), SN95-E (outer and inner hydrophobic layers), and SN95-F (outer hydrophobic layer/inner hydrophilic layer). Bleach off-gassing tests were conducted after a bleach decontamination treatment by immediately placing the FFR face up inside a plastic bag which was open to room air on one side. This setup was designed to minimize air fluctuation within the bag. The detector's sample tube inlet was positioned under the inside of the FFR and all tests were conducted at a flow rate of 0.5 l min À1 . FFRs were tested under four conditions: (i) immediately after a 30-min submersion in bleach, (ii) dried overnight after a 30-min submersion in bleach, (iii) a 30-min submersion in bleach, immediately rinsed (under a flowing stream of deionized water for $1 min) and then dried overnight, and (iv) a 30-min submersion in bleach, then dried overnight followed by rinsing with deionized water. Changes to the FFR materials of construction caused by each decontamination treatment are summarized in Table 2 . Respirator component materials melted on all six FFRs from two models (SN95-E and P100-I) during microwave oven irradiation. EtO and UVGI were the only methods that did not cause any observable physical changes to the FFRs. For each 'FFR model/decontamination treatment' combination, the average initial filter aerosol penetrations are summarized in Table 3 . Not all the 135 All three physical samples of two different models (SN95-E and P100-I) melted partially. SN95-E filtration material melted in areas adjacent to the metallic nosebands. P100-I melted in various locations of the inner foam faceseal comfort lining. Both models were considered unwearable following treatment and subsequently were not evaluated for filter aerosol penetration or filter airflow resistance. These results indicate that for all tested FFR samples that did not melt, FFR filtration performance was not adversely affected by the decontamination process. Most of the ANOVA tests for initial filter aerosol penetration were nonsignificant (P . 0.05), (Table 4 ). In terms of average initial filter aerosol penetration, only P100-I yielded a significant difference by treatment (P 5 0.0438), which appeared to be primarily driven by the increased filter aerosol penetration levels for the UVGI treatment (0.012 versus 0.008% for the control). Although statistically significant, this difference in levels of filter aerosol penetration is practically insignificant because the penetration levels still are far less than expected levels for this class of FFRs (,0.03%). For each 'FFR model/decontamination treatment' combination, the average initial filter airflow resistances are summarized in Table 3 . The six FFRs in which melting occurred could not be tested for filter airflow resistance. For the remaining 129 postdecontamination samples tested, average initial filter airflow resistance measurements were 17.0 mm H 2 O. Previous studies using the same test method on 21 models of NIOSH-approved N95 FFRs observed filter airflow resistance levels between 7 and 30 mmH 2 O . For filter airflow resistance, three of the nine ANOVA tests, including N95-B (P 5 0.0035), SN95-D (P 5 0.0170), and SN95-F (P 5 0.0014), showed significantly different means (see Table 4 ). Although statistically significant, the levels of differences in filter airflow resistance between treatments are not practically meaningful as small changes in filter airflow resistance are unlikely to be noticed by the user (Vojtko et al., 2008) . The degree to which temperature affects initial filter aerosol penetration and component melting was observed to be model specific (Figs 1 and 2) . The average initial penetration (n 5 3) for each N95 model is shown in Fig. 1 . Only three tested N95 FFR samples had filter aerosol penetrations .5% (therefore failed to maintain their expected filtration efficiency level of !95%). These three failing samples were one SN95-D (5.37% at 110°C) and two N95-C (5.18 and 5.37%, both at 120°C). Five of the SN95-D samples could not be analyzed following treatments of 100°C (one sample), 110°C (two samples), and 120°C (two samples) because their inner moisture barrier melted into the filtration media rendering those samples unsuitable for testing. For the three P100 FFR models, average initial filter aerosol pene-tration values for P100-G and P100-H exceeded 0.03% beginning at 100°C for P100-G and beginning at 90°C for P100-H (Fig. 2) . P100-I averaged an initial filter aerosol penetration value ,0.03% for all evaluated temperature increments with the exception of one 110°C temperature experiment. This unexpectedly high average result was due to a single test (%P 5 0.096). All nine FFR models demonstrated differences in their number of media layers and the hydrophobicity of their filter media (Table 5 ). Common to all three models of surgical N95 respirator was the fact that their outer layer was hydrophobic. This is not surprising since surgical N95 respirators cleared by the US FDA undergo fluid resistance testing and are used as barriers against disease transmission by airborne respiratory fluids, including blood, and other small infectious droplets (Bailar et al., 2006) . The N95 FFRs and P100 FFRs varied by having either hydrophobic or hydrophilic outer and inner layers. All middle layers, with the exception of those that were plastic webbing, were hydrophobic on both sides. Initial concentrations of chlorine gas (2-12 p.p.m.) were measured on FFRs wet with bleach immediately following submersion for 30 min (Fig. 3) . FFRs that were treated using bleach and allowed to air-dry overnight (as described in Table 1 ) had initial concentrations of $0.05 p.p.m. followed by no detectable off-gassing (0 p.p.m.) after the initial data point. FFRs which were submerged in bleach, immediately rinsed (entirely under a stream of deionized water for $1 min) and then allowed to air-dry overnight had concentrations similar to FFRs which were not rinsed, indicating that the water rinse had no effect. When FFRs were rehydrated by rinsing with deionized water following overnight air-drying, low-level chlorine off-gassing concentrations were measured at $0.1 p.p.m. (Fig. 4) . The goal of this study was to evaluate five decontamination methods using nine FFR models from three FFR types (three N95 models, three surgical N95 respirator models, and three P100 models) to determine which methods should be considered for future research studies. The five decontamination methods were selected based on previous research from the NPPTL laboratory (Viscusi et al., 2007) . Criteria for assessing methods of decontaminating disposable N95 FFRs have been suggested by the National Academies (Bailar et al., 2006) ; the decontamination method must remove the viral threat, be harmless to the user, and not compromise the integrity of the various elements of the respirator. This manuscript utilizes and expands upon the second and third criteria. For purposes of discussion, a suc-cessful FFR decontamination method is considered to be a physical or chemical treatment which does not degrade laboratory performance (filter aerosol penetration and filter airflow resistance) beyond expected performance levels, is able to be performed on enough FFRs in a short period of time to be practical in the event of a pandemic-induced shortage, and should not pose any additional health risk to the user. In this study, assessment of potential health risks (e.g. possible dermal contact with residuals and/or inhalation of off-gassing residuals) was done using the observational analysis data, off-gassing test results, and general knowledge of the physical/ chemical characteristics of the decontamination method. Chemical off-gassing is of particular concern because of the close proximity of the FFR to the wearer's face and breathing zone. A limited assessment of the throughput capability was also done Evaluation of decontamination methods for FFRs 823 using general knowledge of the various decontamination methods. Additional studies on dry heat laboratory oven exposure and FFR media layer hydrophobicity were conducted to collect data on various aspects of FFR resilience and construction in order to further optimize decontamination strategies and assess the practicality for FFR decontamination during a shortage. In the following sections, the results of laboratory performance testing and observational analysis, additional testing, and assessment of throughput and health concerns will be discussed for each of the five decontamination methods evaluated in order to provide recommendations on which decontamination methods should be considered in future research studies. Bleach is available as an aqueous solutions containing 5-15% sodium hypochlorite (active ingredient) which is a highly active oxidizing agent known to be effective against a broad spectrum of bacteria and viruses (Rutala and Weber, 1997; McDonnell and Russell, 1999) . Bleach decontamination did not affect the FFRs' filter aerosol penetration and filter airflow resistance. The metallic nosebands of all models that had them were slightly tarnished following decontamination and the inner nose cushion on the SN95-E FFRs was discolored. Throughput capability of a bleach method similar to the one used in this study is likely to be high; the main limiting factors are the size of the vessel containing the bleach and FFRs, adequate space to dry the FFRs, and sufficient time for air-drying. All FFR models had a scent of bleach following overnight air-drying. Residual bleach remaining on FFRs is of concern given its known health effects. Hypochlorite powder, solutions, and vapor can be irritating and corrosive to the eyes, skin, and respiratory tract. For example, Nixon et al. (1975) reported that a 5.25% sodium hypochlorite solution caused severe irritation to human skin over a 4-h exposure. Other studies also reported skin irritation for long-term exposure down to a 1% solution (Eun et al., 1984; Habetes et al., 1986; Hostynek et al., 1990) . Low concentrations of bleach have been shown to trigger respiratory events in asthmatics and sensitized individuals (Medina-Ramon, 2005; Mirabelli et al., 2007) . The chlorine off-gassing measurements showed that overnight air-drying significantly reduced off-gassing; however, when the FFR was rehydrated with deionized water, an increase in offgassing was measured. This observation may be significant when viewed in light of the moisture in the exhaled breath of an individual; it gives rise to the possibility of an individual being exposed to low levels of chlorine (,0.2 p.p.m.) from a bleachdecontaminated FFR. Comparing Table 5 and data shown in Fig. 4 , a relationship between hydrophobic-ity of outer and inner respirator surface layers to offgassing concentration could not be established. Considering the potential health risks, the bleach method evaluated in this study is not recommended for further study without modification. Possible modifications worth further investigation would include reduced initial bleach concentration, chemical methods for neutralizing residuals, additional rinse steps, and more aggressive air-drying procedures. EtO is used in a wide range of work settings as a sterilant or fumigant, including healthcare, diagnosis, and treatment facilities; medical products manufacturing; and libraries and museums (NIOSH, 1981) . EtO decontamination did not affect the filter aerosol penetration, filter airflow resistance, or physical appearance of the FFRs in this study. The EtO process used in this study has a 5-h total processing cycle (1-h EtO exposure followed by 4 h of aeration) and has a 4.8 ft 3 (0.14 m 3 ) chamber volume (3M, 2007) . The 5-h total processing time may be a limiting factor in the timely processing of a large volume of FFRs. Residual EtO remaining on FFRs following EtO vapor-phase decontamination is not believed to be a concern because the sterilization process includes a final aeration cycle of 4 h to remove residual EtO gas. VHP has been shown to be sporicidal at temperatures ranging from 4 to 80°C, with sterilant concentrations ranging from 0.5 to ,10 mg l À1 (Joslyn, 1991) . VHP decontamination for a single warm cycle did not significantly affect FFR filter aerosol penetration or filter airflow resistance. The only visible physical effect on the FFRs was a slight tarnishing of the metallic nosebands. The VHP process used in this study has a short cycle time (55 min) and a usable processing volume of 3.5 ft 3 (0.1 m 3 ) (Advanced Sterilization Products, 2007) . Although the 55-min cycle time is short compared to the lengthy EtO total process time, the throughput capability of VHP processing is limited by the fact that cellulose-based products (e.g. cotton, which may be present in some head straps or some FFR layers) absorb hydrogen peroxide and can cause the STERRADÒ cycle to abort due to low hydrogen peroxide vapor concentration. Significant levels of residual hydrogen peroxide vapors off-gassing from FFR materials following the STERRADÒ process are unlikely and not of concern because the vapors decompose readily into water vapor and oxygen, both of which are environmentally benign (Advanced Sterilization Products, 2007) . Biological decontamination of FFRs using a domestic microwave oven is an attractive idea since it has the advantages of convenience and short treatment times. The decontamination method used here treats the microwave oven as a source of dry heat, similar to other studies. Elhafi et al. (2004) demonstrated that four avian viruses (infectious bronchitis virus, avian pneumovirus, Newcastle disease virus, and avian influenza virus) were inactivated on dried cotton swab samples using a domestic microwave oven for as little as 20 s. Rosaspina et al. (1994) demonstrated destruction of Mycobacterium bovis dried onto scalpel blades after 4 min of microwave exposure. Of the nine FFR models that underwent microwave oven irradiation, filter aerosol penetration and filter airflow resistance were not affected for seven models. Material components melted on the two remaining models. Correlation could not be established for filter aerosol penetration results between dry oven-treated and microwave oven-irradiated samples. In microwave oven irradiation tests, all three SN95-D samples had penetration values ,5% and did not melt; however, some SN95-D samples partially melted at 100, 110, and 120°C during dry oven treatment (Fig. 1) . All SN95-E samples and all P100-I samples partially melted in the microwave oven, but no melting was observed for these two models, even at 120°C following dry oven treatment (Table 3 , Figs 1 and 2) . The throughput capability of a method similar to the one in this study was limited by microwaving one FFR at a time; however, the 2-min treatment time per FFR was relatively short. Although it is likely that processing more than one FFR at a time is feasible (limited only by the internal volume of the oven), maximizing throughput was beyond the scope of this investigation. No known health risks to the user were identified. The data presented here suggest that the dry microwave oven irradiation method requires improvement before it could be recommended for decontamination and subsequent reuse. Possible modifications worth further investigation would include microwave irradiation of wet FFRs, shorter exposure times, and lower power settings. UVGI has been demonstrated to be effective for the disinfection of drinking water and wastewater (Sykes, 1965; Angehrn, 1984; Lazarova et al., 1999; Craik et al., 2001; Lazarova and Savoye, 2004; Wu et al., 2005) and for hospital air disinfection as a method for controlling airborne infectious disease (Macher et al., 1992; Nardell, 1993; CDC, 1994; Gorsuch et al., 1998; Miller and Macher, 2000) . This study found that UVGI treatment did not affect the filter aerosol penetration, filter airflow resistance, or physical appearance of the FFRs. Throughput capability of a method similar to the one in this study is benefited by a relatively short ir-radiation time (30 min); however, it is limited by the available working surface area of a biosafety cabinet equipped with a UV-C source or other area being irradiated by a UVGI source. No known health risks to the user were identified. These findings are exploratory and the data presented in this study are applicable only to the FFRs and decontamination methods tested; other FFRs may be more easily degraded while others may be less affected and slight modifications to the decontamination methods could result in different findings. Future studies are still needed to evaluate whether the decontamination processes evaluated in this study will inactivate infectious microorganisms (or appropriate surrogates), if FFR decontamination influences respirator fit, and the effect of multiple decontamination treatments on FFR performance. Future studies should also investigate the depths that infectious organisms (or appropriate surrogates) penetrate into each FFR layer, assess the relative cost of various decontamination strategies, and determine how effective various decontamination methods are at reducing the number of viable virus in all layers of the FFRs. Recent work in the NPPTL laboratory toward developing a system for studying the virucidal capability of decontamination methods for FFRs appears promising (Fisher et al., 2009) . The effects of the various decontamination methods on the laboratory performance (filter aerosol penetration and filter airflow resistance) and physical appearance of FFRs were found to be model specific. The respirators tested have differences in their design, materials of construction, and hydrophobicity of their layers (including the filter media layers). Microwave oven irradiation melted all six samples from two FFR models. The remainder of the FFR samples that were evaluated exhibited average initial filter airflow resistances 17.0 mmH 2 O and average initial sodium chloride filter aerosol penetration values 1.86% for N95 FFRs and 0.012% for P100 FFRs. Although there were statistically significant differences found between control respirators and those that have undergone decontamination for both filter aerosol penetration and filter airflow resistance, the practical significance is minimal as the range of numerical differences is quite small. The scent of bleach remained noticeable on all FFR models following overnight drying and low levels of chlorine were found to off-gas from bleach-decontaminated FFRs when rehydrated with deionized water, thus giving rise to the possibility of low-level exposure to a subsequent wearer. In light of these results, the microwave oven irradiation and bleach decontamination methods investigated in this study were determined to be the least desirable among the five methods tested for consideration in future studies. UVGI, EtO, and VHP were found to be the most promising decontamination methods; however, concerns remain about the throughput capabilities for EtO and VHP. Further research is needed before any specific decontamination methods can be recommended.

Antiviral Activity of Some Plants Used in Nepalese Traditional Medicine

Methanolic extracts of 41 plant species belonging to 27 families used in the traditional medicine in Nepal have been investigated for in vitro antiviral activity against Herpes simplex virus type 1 (HSV-1) and influenza virus A by dye uptake assay in the systems HSV-1/Vero cells and influenza virus A/MDCK cells. The extracts of Astilbe rivularis, Bergenia ciliata, Cassiope fastigiata and Thymus linearis showed potent anti-herpes viral activity. The extracts of Allium oreoprasum, Androsace strigilosa, Asparagus filicinus, Astilbe rivularis, Bergenia ciliata and Verbascum thapsus exhibited strong anti-influenza viral activity. Only the extracts of A. rivularis and B. ciliata demonstrated remarkable activity against both viruses.

Plants have long been used as a source of medicine from ancient time to today all over the world. In developing countries the availability of modern medicines is limited. So traditional medicine is still the mainstay of health care and most drugs come from plants. Although many plants have long been recognized and widely used in Nepalese traditional medicine, some are relatively unexplored and not arrived to mainstream medicine (1) . Therefore, the search on new drugs must be continued and natural products from plants, microorganisms, fungi and animals can be the source of innovative and powerful therapeutic agents for newer, safer and affordable medicines (2, 3) . On the other hand the screening of plants as a possible source of antiviral drugs has led to the discovery of potent inhibitors of in vitro viral growth (4) (5) (6) (7) (8) (9) (10) (11) . Therefore, the present investigation was carried out to assess the antiviral effects of some native plants used by the local people belonging to Gurungs and Thakalis of Manang and Mustang districts that lie in the Annapurna Conservation Area Project (ACAP). Permission for the field study as well as the collection of voucher specimens was received from the headquarters of ACAP in Pokhara. The plants were selected on the basis of ethnopharmacological records, so the prospect of finding new bioactive compounds is always promising. The name of the plants, respective families, the parts used for the extract preparation and traditional uses of the plants are listed in Table 1 . The dried and powdered plant material (each 10 g) was extracted successively with n-hexane, dichloromethane and methanol in a soxhlet extractor for each 8 h. Evaporation of the solvent followed by drying in vacuum gave the respective crude dry extract. Only methanol extract was used for the antiviral assay, n-hexane and dichloromethane extracts were not included because of their insolubility in medium and high toxicity to the cells. Each 2 mg of the extract was dissolved in 10 ml dimethylsulfoxide (DMSO) before adding tissue culture medium supplemented with 2% fetal calf serum (FCS, GIBCO Life science technologies, Paisley, UK) and stocked at a concentration of 2 mg ml À1 . Madine-darby canine kidney (MDCK) and African green monkey kidney (Vero) cells (cell bank of the Friedrich-Loeffler-Institute, Federal Research Institute for Animal Health, Greifswald-Insel Riems, Germany) were maintained in Eagle's minimal essential medium (MEM) supplemented with 5% FCS (GIBCO, Paisley, UK). The exponentially growing cells were harvested and seeded at a cell density of 60 000/well in a 96 well microtiter plate (8 mm diameter, Falcon Plastic, NJ) and incubated for 24 h at 37 C with 5% carbondioxide in a 90% humidified chamber so as to form confluent monolayers. Human influenza virus A/WSN/33 (H1N1) London was obtained from the strain collection of the Institute of Medical Microbiology, University Greifswald, Germany, and propagated in embryonated hen eggs for 72 h. The infected allantoic fluids were harvested, the hemagglutination (HA) titer and virus infectivity were determined on MDCK cells and the virus stock was stored at À70 C. Herpes simplex virus type 1 (HSV-1, strain KOS) was obtained from the strain collection of the Consiliar and Reference Center for Alpha Herpes Virus Infection, Institute of Virology and Antiviral Therapy, University Jena, Germany and propagated in Vero cells. The virus infected cells were frozen and thawed and the virus suspension was titrated on Vero cells and stored at À70 C (7). The cellular toxicity of extracts on Vero and on MDCK cells was assessed by dye uptake method using neutral red (12) in 96-well tissue culture plates (8 mm diameter, Falcon Plastic, NJ). Only living cells are able to manage the active uptake of neutral red. Confluent monolayers of cells were treated with 100 ml 2-fold serial dilutions of extracts prepared at concentrations of 200, 100, 50 and 25 mg ml À1 in four replicates and incubated at 37 C in a humidified atmosphere of 5% CO 2 for 72 h. The supernatant was removed and 200 ml neutral red solution (0.005%) in optimum was added. The microtiter plate was further incubated for 3 h at 37 C. After removal of the supernatant, the dye incorporated by the viable cells was extracted with 100 ml ethanol/water/glacial acetic acid solution (50 : 50 : 1) by shaking for 15 min. The absorbance was measured on an ELISA reader using Ascent software at 540 nm. The cytotoxic concentration that caused the reduction of viable cells by 50% [CC 50 ] was calculated from dose-response curve. Antiviral activity was determined by dye uptake assay using neutral red as described by Mothana et al. (7) . Non-cytotoxic extracts were tested in concentrations of 100, 50, 25, 12.5 and 6.25 mg ml À1 . The antiviral tests of cytotoxic extracts started with the half of the individual CC 50 . The extracts were diluted 1 : 2 by medium. Confluent monolayers of Vero and MDCK cells were treated with 100 ml of extracts in four replicates for 30 min. After that Vero cells were infected with 30 TCID 50 of HSV-1 and MDCK cells with 30 TCID 50 of influenza virus A and incubated for 72 h at 37 C. TCID 50 (tissue culture infectious dose) is the virus dose that leads to the infection of 50% of the cells. The virus suspension and dilution medium without samples were added, respectively, to the cell cultures to serve as the virus control and cell control. The supernatant was replaced by 200 ml neutral red solution (0.005%) and the cells were incubated for 3 h at 37 C. After removal of the supernatant, the dye incorporated by viable cells was eluted with 100 ml ethanol/water/glacial acetic acid solution (50 : 50 : 1) by shaking for 15 min. The absorbance was measured at 540 nm and the percentage protection was calculated by the following formula (13): where, (OD T ) V , (OD C ) V and (OD C ) M correspond to absorbances in virus infected cells with test compounds, virus infected cells without test compounds and the mock infected control (assay without viruses), respectively. Amantadine HCl and acyclovir were used as reference compounds in concentrations of 0.1, 1, 10 and 100 mg ml À1 . In this study, 43 methanolic extracts from 41 different plant species belonging to 27 families (Table 1) were screened for their antiviral activity against herpes simplex virus and influenza virus A by dye uptake assay. By methanolic extraction, a broad spectrum of compounds with different polarity can be obtained. As prerequisite for antiviral tests, the cytotoxicity of the extracts against virus-host cells was investigated. The results are summarized in Table 2 . The extracts of Androsace strigilosa, Anemone rivularis, Delphinium brunonianum, Euphorbia longifolia and Thalictrum cultratum exhibited strong cytotoxicity in Vero cells with CC 50 (the concentration that causes the reduction of viable cells by 50%) ranging from 12.5 to 25 mg ml À1 . A moderate cytotoxicity was observed for the extracts of Asparagus filicinus, Bergenia ciliata, Primula involucrata and Saussurea auriculata with CC 50 ranging from 30 to 50 mg ml À1 . Other eight extracts showed very mild toxicity while rest of the extracts were non-toxic at 100 mg ml À1 . Similarly, in MDCK cells extracts of Artemisia caruifolia, D. brunonianum and E. longifolia showed strong toxicity with CC 50 ranging from 19 to 25 mg ml À1 . A moderate toxicity was exhibited by the extracts of A. strigilosa, A. rivularis, Asparagus filicinus, Dicranostigma lactucoides, Hyoscyamus niger, Thymus linearis and Zanthoxylum armatum with CC 50 ranging from 30 to 50 mg ml À1 . Other three extracts demonstrated very low toxicity while rest of the extracts were non-toxic at 100 mg ml À1 . Antiviral activity against HSV-1 was shown by 11 extracts at non-cytotoxic concentrations. The IC 50 values (the concentration that protects 50% of the cells against destruction by viruses) ranged from <6.25 to 82 mg ml À1 . The highest activity against HSV-1 with IC 50 values <6.25 mg ml À1 was observed for the extracts of A. rivularis, B ciliata, Cassiope fastigiata and T. linearis. Moderate activity was shown by Cotoneaster integrifolius (IC 50 18 mg ml À1 ) and Clinopodium umbrosum (IC 50 19 mg ml À1 ). Weak activity (IC 50 50-82 mg ml À1 ) was found in the extracts of Bistorta affinis, Juniperus squamata, Oxytropis williamsii, Rhododendron anthopogon and Rubus foliolosus. Antiviral activity against influenza virus A was shown by 20 extracts at non-cytotoxic concentrations. The IC 50 values ranged from <6.25 to 97 mg ml À1 . The highest activity was shown by the extracts of A. filicinus, A. rivularis and Verbascum thapsus with IC 50 < 6.25 mg ml À1 . In addition, the extracts of Allium oreoprasum, A. strigilosa and B. ciliata also exhibited high activity (IC 50 values from 8 to 10 mg ml À1 ). Moderate activity (IC 50 values from 17 to 50 mg ml À1 ) was demonstrated by 11 extracts. Weak activity (IC 50 values from 78 to 97 mg ml À1 ) was shown by three extracts ( Table 2 ). The extracts of A. rivularis and B. ciliata were found to be highly active against both viruses. The results of this work justify the potential of some of the investigated plants for the production of bioactive compounds. The phytochemical knowledge about these plants is so far very limited. The active principles present in A. rivularis are still unknown. Phytochemical investigation of A. rivularis revealed the presence of flavonoids, terpenoids and bergenin (14, 15) . Bergnia ciliata is known to contain phenolic compounds (16) . Polyphenols, especially high polymeric procyanidines possess strong anti-influenza viral activity (17) , which is in agreement with our previous study (18) . In our previous study (19) , methanol-water extract of Bergenia ligulata, which is taxonomically closely related to B. ciliata, inhibited the growth of influenza virus A in cell culture with IC 50 of 10 mg ml À1 . The extract also inhibited the viral protein and nucleic acid synthesis (18) . In the present study, the methanol extract of B. ciliata inhibited the influenza virus A and HSV-1 indicating that the genus Bergenia could be the source of potent antiviral drugs. Again potent activity of A. rivularis against both viruses indicated the high prospect of finding antiviral drugs in Saxifragaceae family. No antiviral compounds have previously been isolated from A. filicinus. The plant is known to contain steroidal saponins (20, 21) , furostanol glycosides (22) and furostanosides (23, 24) . The phytochemicals possibly responsible for the high activity of C. fastigiata against HSV are not described. Some Cassiope species are reported to contain flavonoid glycosides (25) . Similarly, the compounds responsible for the high anti-influenza viral activity of A. oreoprasum and A. strigilosa are not reported elsewhere. Likewise, no antiviral constituents have been isolated from C. integrifolius, C. umbrosum and T. linearis. Other members of the genus Cotoneaster, have been found to possess phenolic glycosides (Cotoneaster orbicularis, 26), flavonols and isoflavones (Cotoneaster simonsii, 27). From the other member of the genus Clinopodium, C. chinensis var. parviflorum, oleanane triterpene saponins have been isolated (28) . Whereas for the extract of V. thapsus, antiherpes activity has been reported (29); our study revealed only the strong anti-influenza viral activity. However, no antiviral compounds have previously been isolated. The plant is known to contain phenylethanoid and lignan glycosides (30) . On the other hand, the phytochemicals responsible for anti-influenza viral activity could be different from anti-herpes activity and also the amount of active constituents present in the plants depends on the geographical distribution, season of collection and climatic and ecological condition at the collection site. Looking at the chemical structures of the already identified compounds, most of these substances should be extracted by methanol. The foregoing extraction by more lipophilic solvents (n-hexane and dichlormethane) alleviates the methanolic extraction and the planned fractionation. Comparing the use of plants in traditional medicine and their antiviral activity, a direct correlation could be established for some plants, e.g. A. oreoprasum, A. strigilosa (anti-influenza activity) and T. linearis (antiherpes activity). For other plants, e.g. C. fastigiata, which exhibited potent anti-herpes activity, this cannot be recognized till now. The extracts that exhibited only medium and low activity, could also be the source of potential antiviral drugs because the bioactive compounds may be present in too low concentrations to show effective antiviral activity at non-toxic concentration. Further fractionation and separation of extract(s) may reveal potent antiviral activity (31) . Our results indicate that several plants used in Nepalese traditional medicine could be the lead to potential antiviral drugs, which possibly provide molecules with drug-like properties and with incredible structural diversity. Besides, the results are useful for rationalizing the use of medicinal plants in primary health care in Nepal. The phytochemical characterization of the extracts, the identification of the responsible bioactive compounds and the elucidation of the mode of action and quality standards are necessary.

Predicting intention to treat HIV-infected patients among Tanzanian and Sudanese medical and dental students using the theory of planned behaviour - a cross sectional study

BACKGROUND: The HIV epidemic poses significant challenges to the low income countries in sub Saharan Africa (SSA), affecting the attrition rate among health care workers, their level of motivation, and absenteeism from work. Little is known about how to deal with deterioration of human resources in the health care systems. This study aimed to predict the intention to provide surgical treatment to HIV infected patients among medical- and dental students in Tanzania and Sudan using an extended version of the Theory of Planned Behaviour (TPB). METHODS: Four hundred and seventy five medical- and dental students at the University of Dar es Salaam (mean age, 25 yr) and 642 dental students attending 6 public and private dental faculties in Khartoum (mean age 21.7 yr) completed self-administered TPB questionnaires in 2005 and 2007, respectively. RESULTS: Both Tanzanian and Sudanese students demonstrated strong intentions to provide care for people with HIV and AIDS. Stepwise linear regression revealed that the TPB accounted for 51% (43% in Tanzania and Sudan) of the variance in intention across study sites. After having controlled for country and past behaviour, the TPB in terms of attitudes, subjective norms and perceived behavioural control accounted for 34% and moral norms for an additional 2,3% of the explainable variance in intention. Across both study sites, attitudes were the strongest predictor of intention followed in descending order by subjective norms, moral norms and perceived behavioural control. CONCLUSION: The TPB is applicable to students' care delivery intentions in the context of HIV and AIDS across the two SSA countries investigated. It is suggested that attitudes, subjective norms, moral norms and perceived behavioural control are key factors in students' willingness to treat AIDS and HIV infected patients and should be targets of interventions aimed at improving the quality of health care delivery in this context.

is currently recognized by the World Health Organization to suffer an intermediate HIV and AIDS prevalence of 1.6% [1] . The HIV epidemic poses significant development challenges to the low income countries in SSA [2] . It affects the attrition rate among health care workers, their level of motivation, professional practices and absenteeism from work [2] . To date, little is known about how to deal with the deterioration of human resources in the health care systems. Due to the method of transmission of HIV virus through direct contact with blood, the risk for cross-infection comes into particular focus in medical and dental practices [3] . Over 90% of the HIV infections occurring among health care workers annually stem from developing countries where occupational safety is a neglected issue [4] . Although the risk of transmission in health care settings has been recognized to be low, fear of illness, contagion and death has influenced health workers' attitudes and thus the quality of care provided towards patients with HIV [5] [6] [7] . Increased personal risk, lack of necessary skills, knowledge gaps, difficulties in dealing with staff worries and concern about loosing other patients are the most frequent complaints [8] [9] [10] [11] [12] . A recent publication focusing on dental students in Khartoum, Sudan, revealed that half of the participants reported a need for further education across HIV and AIDS related issues, suggesting they are not adequately prepared for treating HIV infected patients [8] . Unacceptable knowledge and practice as well as gaps in the availability and access to policies and protocols on the part of health care workers have been observed in several sub-Sahara African countries [11, [13] [14] [15] [16] [17] . According to the World Health Organization (WHO), dentists have a professional and ethical responsibility to treat patients with HIV and AIDS [18] [19] [20] . The importance of training dental and medical staff to provide health care to HIV infected patients at the same level as non-infected people have been widely recognized [17, 19] . As future health care workers, the attitudes of dental and medical students towards delivering high quality care for HIV infected patients are of particular concern. Effective promotion of quality health care delivery in the context of HIV and AIDS requires a thorough understanding of the psychosocial determinants of students' intention to provide care to HIV infected patients. Ajzen's theory of planned behaviour (TPB) is a valuable model for identifying the determinants of health behaviours, including quality health care provision for patients with HIV and AIDS [21] . The TPB [21] constitutes a promising framework for understanding and predicting social behaviours ( Figure 1 ). The TPB includes perceived behavioural control on a level with attitude and subjective norm as predictors of behavioural intention. This theory implies that the three predictors influence subsequent behaviour indirectly through behavioural intention. The TPB posits that behavioural intention is a function of attitude, reflecting a favourable or unfavourable evaluation of the particular behaviour and subjective norm, referring to the perceived social pressure to perform the behaviour. Perceived behavioural control reflects the ease or difficulty associated with performance. Attitudes, subjective norms and perceived behavioural control are underpinned by behavioural, normative and control beliefs, respectively. A number of studies have suggested that past behaviour has a residual effect on behavioural intention after the TPB has been taken into account [22, 23] . Thus, Ajzen [21] suggested that the TPB is open to the inclusion of additional variables if it can be shown that they capture a significant proportion of the outcome variance. The TPB has been applied successfully to a range of domains, including HIV related behaviours, particularly condom use [24] [25] [26] [27] [28] . With respect to occupational behaviour, the TPB has predicted health workers' use of gloves, their intention to provide home-care for HIV infected patients, their adherence to universal precautions for venipuncture and their intention to provide professional labour support [29] [30] [31] [32] [33] . The TPB has been used previously in sub-Saharan African settings to predict HIV protective behaviours, however mostly by small scale studies [27] . It has been advocated that the applicability of socio-cognitive models to the African context should be systematically addressed considering the need for theory-based research in the planning of effective HIV and AIDS related educational programs [34, 35] . This study extends analyses of external variables within the TPB in the context of health care delivery by adding past behaviour and moral norms from Triandis' Theory of Interpersonal Behaviour [36] . Personal normative belief or moral norms represents a measure of the personal feel- The theory of planned behaviour (Ajzen 1991) Figure 1 The theory of planned behaviour (Ajzen 1991). Subjective norm Attitude Intention Behaviour ings of moral obligations or responsibility to perform or refuse to perform a given behaviour. Studies of moral norms in the context of the TPB were reviewed by Conner and Armitage [24] who estimated that across investigations moral norms predicted an additional 4% of the variance in intention after controlling for the TPB. Moreover, studies on the Theory of Interpersonal Behaviour have consistently shown that moral considerations are significant predictors of behavioural intentions in the presence of the TPB [37] . Focusing on medical and dental students from Tanzania and Sudan, this study aims to predict the intention to provide surgical treatment to patients with HIV and AIDS as part of future professional work, using the TPB and moral norms. The hypotheses of the present study were: -attitudes, subjective norms and perceived behavioural control will each contribute positively and statistically significantly to the prediction of intention to provide surgery treatment to HIV and AIDS infected patients -moral norms will add significantly to the prediction of behavioural intention over and above the TPB A cross-sectional survey was carried out from June to September 2005 at Muhimbili University College of Health and Allied Sciences (MUHAS) at the University of Dar-es-Salaam. The target population consisted of students attending the faculties of dentistry and medicine. A total of 1,021 (862 medical and 159 dental) students were enrolled at the college in 2005. Six hundred students (100 students in each study year) attending the 1 st to the 5 th study year were invited to participate. A total of 476 students agreed to participate and 454 students (response rate 75.6%), mean age 25.6 (sd 2.6), 22.5% females completed supervised self-administered structured questionnaires at the faculty in class-room settings. A cross-sectional study was carried out from April to May 2007 among a census of dental students attending dental faculties in Khartoum, the capital city of Sudan. A list of all the dental faculties was obtained from the Ministry of Higher Education and lists of all registered students in their 3 rd , 4 th and 5 th years were obtained from all faculties through the Dean's office. The faculties included in this study were both publicly and privately funded. Moreover, they represented all available dental faculties in Sudan, making potentiating admission from all over the country. The total number of dental students registered by the time of the survey was 782 out of which 642 students (response rate 82%), mean age 21.7, (sd 1.9), 82% females completed self-administered, anonymous questionnaires in supervised (by teaching assistants) class-room settings. The main reason for non-participation was absenteeism on the day of the data collection. More information regarding the sampling, recruitment and data collection is available in a previous publication [8] . Written informed consent was obtained from all participants in both countries. A formal ethics waiver was received from the research committee at the University of Science and Technology in Sudan, the Research Committee at MUHAS, Tanzania and from the Regional research committee of Norway (REK VEST). Identical survey instruments were used at both sites. Before administration in the field, the questionnaire was reviewed by experienced local researchers, dental academics and health administrators. The survey instrument was adapted from instruments previously employed in SSA [38] . In Tanzania, the survey instrument was constructed and completed in English. In Sudan, the survey instrument was constructed in English, translated into Arabic (the Sudanese national language) and then back translated into English by independent language experts. The survey instrument covered socio-demographic factors and each component of the TPB developed according to the guidelines proposed by Ajzen and Fishbein [39] . Intention to provide surgical treatment to patients infected with HIV/AIDS as part of future professional work was measured by three items. E.g. "How likely is it that you will provide--------". The respondents indicated their subjective probability along a five point response scale from (1) very unlikely to (5) very likely. A sum score was constructed from the three items. Attitudes-were measured by 5 items, three positively worded and two negatively worded. Responses were recorded on a five-point scale ranging from (1) strongly disagree to (5) strongly agree. A sum score was constructed after positively worded items were reversibly scored. Subjective norms were measured by 4 items in relation to all my patients, community leaders, my family and my teacher at the college. E.g. "My teacher at my college wants me to provide surgical treatment to HIV and AIDS infected patients as part of future professional work". Responses were indicated on five-point scales ranging from (1) strongly disagree to (5) strongly agree. A sum score was constructed from the 4 items. Perceived behavioural control was assessed using one item, "How easy or difficult is it for you to provide surgical treatment to patients infected with HIV and AIDS as part of future professional work? Responses were rated on a scale ranging from (1) very difficult to (5) very easy. Moral norms were assessed using 3 items "I I feel guilty if I do not provide ----------". "I get a bad conscience if I do not provide surgical treatment and "It is morally wrong for me not to provide surgical treatment ". Responses were indicated on a scale ranging from (1) strongly disagree to (5) strongly agree. A sum score was constructed from the 3 items. Past behaviour was assessed using a dummy variable in terms of "Have you ever participated in the clinical treatment of HIV and/AIDS infected patients". Answers were provides as (1) yes and (0) no. Confirmatory factor analysis, CFA, with AMOS 16 was employed to test the hypothesized measurement model with respect to intention, attitudes, subjective norms and moral norms, specifying the relationship between the observed variables (indicators) and the underlying latent variables (concepts). Thus CFA was used to test whether the Tanzanian and Sudanese data were consistent with an a priori hypothesized 4-factor model. The parameters of the model were estimated with maximum likelihood (ML) estimation. Missing data were assumed to be missing at random and was handled using the direct approach in AMOS 16 [40] . The adequacy of overall model fit was estimated using chisquare test statistics and the following supplemental fit indices, root-mean-squared error of approximation (RMSEA), the comparative fit index (CFI) and Akaike's information criteria, AIC. In line with the conventional recommendations of Hu and Bentler [41] , a good model fit was indicated by a RMSEA less or equal to .06, a CFI greater or equal to .90 and with models having lower AIC being more plausible. The statistical significance of parameter estimates are the critical ratio (CR) representing the parameter estimate divided by its standard error. Based on a level of 0.05, the test statistics need to be <± 1.96 before the null hypothesis can be rejected. All other analyses were performed using SPSS 15.0 (SPSS, Inc, Chicago, Illinois, USA). Internal consistency reliability of-and bivariate associations among the theoretical constructs were assessed using Cronbach's alpha and Pearson's correlation coefficients, respectively. Unadjusted and adjusted marginal means and 95% confidence intervals for the components of the TPB and the construct of moral norm were estimated using General Linear Models, GLM (ANOVA). Linear multiple regression analysis was applied to predict intention from the TPB and external variables. The effect of the independent variables were expressed in terms of standardized regression coefficients (betas) and tested for statistical significance by means of F-test. The fit of the model was reported in terms of the squared multiple correlation coefficient (R 2 ). Of the 475 Tanzanian students, 17% were in the younger age group (below or equal to 22 yr), 77.5% (368) were males, and 25.4% (121) were in their 5 th study year. The corresponding figures pertaining to the Sudanese participants were 69.6% (438), 28% (177) and 36% (231), respectively. Table 1 depicts the percentage distribution of participants according to socio-demographics and country of residence. All socio-demographic characteristics varied substantially and statistically significantly across the two cultures considered. . All factor loadings (standardized regression weights) were in the expected direction and were statistically significant at CR>1.96, with inter-factor correlations ranging from 0.46-0.78 (Tanzania) and from 0.50-0.68 (Sudan). Thus, all inter-factor correlations were below the threshold of 0.80 which is set as cut off to indicate poor discriminative validity [41] . Table 2 depicts unadjusted and adjusted marginal means and 95% confidence intervals for the components of the TPB and the construct of moral norm by country of residence. Country differences were estimated after controlling for potential confounding effect from sociodemographic variables (age, gender, parental education, study year) using General Linear Models, GLM (ANOVA). In both countries, students had on average positive attitudes, strong moral norms and strong intentions regarding care delivery to patients with HIV and AIDS. Both groups of students had on average moderately strong subjective norms and perceived less control regarding this behaviour. Tanzanian students had on average more positive attitudes and stronger intentions, perceived control, moral norms and subjective norms compared to their Sudanese counterparts. Internal consistency reliability in terms of Cronbach's alpha ranged from 0.83 (moral norms/subjective norms) to 0.44 (attitudes) in Tanzania and from 0.81 (attitudes/intention) to 0.45 (subjective norms) in Sudan. In Tanzanian and Sudanese students, the TPB components, moral norms, and past behaviour were statistically significantly associated with intention. In Tanzanian students, Pearson's correlation coefficients ranged from r = .52 between attitudes and intention to r = .14 between past behaviour and intention. In Sudanese students, Pearson's correlation ranged from r = .54 between intention and attitudes to r = .03 between intention and past behaviour. Table 3 depicts the results from hierarchical linear regression analysis assessing the fit of the extended TPB model among Tanzanian and Sudanese students. Country and past behaviour were entered in the first step explaining 15.7% (R 2 change = 0.157, p < 0.001) of the variance in intention. Adding attitudes, subjective norms and perceived behavioural control in step 2 increased the explained variance by 33.4% (R 2 change = 0.334 p < 0.001). Moral norm added in step 3 raised the explained variance by 2.3% (R 2 change = 0.023 p < 0.001). A total of 6 variables accounted for 51.4% of the variance in intention (43.7% and 43.9% in Tanzania and Sudan, respectively). In the final equation, attitude was by far the strongest predictor of intention (beta = 0.35), followed in descending order by subjective norms (beta= 0.22), moral norm (beta= 0.17), perceived behavioural control (beta = 0.16) and past experience (beta = 0.06). The effect of country (beta = 0.35, p < 0.001) in step 1 was reduced to (beta = 0.04, p = 0.099) in the final third step, whereas the strength of the effect from past behaviour was maintained from step 1 (beta= 0.08, p < 0.001) to step 3 (beta= 0.06, p < 0.001). Statistically significant two-way interactions occurred, in terms of country × attitudes (R 2 change 0.007, F change = 13.7, p < 0.001) and country × subjective norms (R 2 change 0.009, F change = 17.0, p < 0.001). Stratified analyses suggested that the relationship between attitude and intention and between subjective norms and intention were statistically significantly stronger in Sudanese-than in Tanzanian students. This study supports the applicability of an extended version of the TPB to students' intended care delivery for patients with HIV and AIDS in two culturally different sub-Saharan African countries. According to the CFA, the extended TPB questionnaire reflected four concepts across the study sites in terms of attitudes, subjective norms, moral norms and intention, thus lending support to it's within construct validity, formally. This appears to imply that the four constructs underlying the TPB questionnaire are discrete measures that can be reliably assessed in Tanzanian and Sudanese students and that those measures can be reported as four summary scores. A total of six variables explained 51% of the variance in health care delivery intentions across study sites. After having controlled for country and past behaviour, attitudes, subjective norms and perceived behavioural control accounted for an additional 34% of the explainable variance in intention. This finding is consistent with those of previous studies, whereby the TPB has explained 68% of nurses' intention to adhere to universal precautions, 48% of health workers' intention to provide home care for HIV infected individuals, 35% of nurses' intention to care for SARS patients and 70% of nurses intended labour support [29] [30] [31] [32] [33] . In line with a growing body of research supporting the role of perceived moral obligations as an independent predictor of intention, moral norms contributed 2.3% to the explained variance in students' intentions after controlling for the TPB variables. This factor which has moral connotations and represents personal feelings of responsibility has been considered to be important in the adoption of several health related behaviours [36, 37] . Conner and Armitage [25] found that moral norms contributed an additional 4% of the variance in intention after controlling for the TPB across various behaviours. The present results suggest that the TPB and its extended version is useful in the SSA context in terms of identifying correlates of health care delivery that can be targets in interventions aimed at improving health care delivery to HIV infected patients. A major shortcoming of the TPB model has been its inability to account for the influence of past behaviour [42, 43] . Evidence from meta-analytical reviews suggests that the addition of past behaviour to the TPB explains on average 7% of the variance in intention [25] . In this study, past experience significantly predicted intention after controlling for the TPB and moral norms and left the cognitive variables of the model almost unaffected. This suggests that the TPB provides a fairly accurate description of the intention formation process considering HIV and AIDS related care delivery among Tanzanian and Sudanese students. Students decide upon care delivery for HIV infected patients mainly as a consequence of situation specific expectations of the behaviour itself and to a lesser extent because they have engaged in similar care delivery previously. Tanzanian and Sudanese students had on average strong intentions to provide surgical treatment to HIV infected patients as part of their future professional work. In contrast, in a study of Taiwanese nurse students' care intentions, almost all stated that they did not intend to care for HIV infected patients [44] . In this study, intended health care delivery was primarily driven by attitudes followed in descending order of importance by social norms, moral norms and perceived behavioural control in both countries. This finding is congruent with findings in other studies of treatment delivery and compliance with precautions in health care workers [29] [30] [31] [32] . This finding is also similar to that reported by Sauls [33] , who identified attitudes as more influential in determining health care delivery intentions than subjective norms. In contrast, Vermette and Godin [30] found perceived behavioural control to be the strongest influencing factor of nurses' intention to provide home care for HIV infected people. Students who did not express confidence in their ability to circumvent difficulties associated with care delivery, who evaluated care delivery negatively and who felt less normative pressure from colleagues at the faculty and less moral obligations to act were less likely to have strong intentions. These findings add insight to faculty administrators and educators to further develop strategies to increase students' intention to care for patients with HIV and AIDS. Providing sufficient and adequate protective equipments, routinely practicing infection control measures and protocols and providing up to date continuing education and training exercises should improve students' ability to overcome perceived obstacles related to care delivery. Even more important is the enhancement and reinforcement of students' positive attitudes through verbal expression of approval, persuasive messages based on strong arguments, substantial rewards and psychological support by faculty staff to encourage and acknowledge their efforts. Students' decision to provide surgical treatment to HIV infected patients was also influenced by their expectations that faculty colleagues would approve their provision of such treatment. In planning intervention programs, one approach which could be particularly important among students having less care delivery experience is to train peer leaders to communicate the importance of quality care for HIV infected patients. Interventions might also benefit from making students' focus on moral obligations by increasing their awareness of others' needs and their perception that providing quality care could help relieve such needs. Some limitations of this study should be considered when drawing inferences based upon its results. In Tanzania, dental and medical students were recruited from one university, thus the representativeness of the findings to other undergraduate-and post-graduate students is unknown. Self-selection might also be a potential limitation since the students who chose to participate might differ from those who did not implying that only students with interest in health care delivery for HIV infected patients responded to the survey invitation. Students might have over reported intention to provide surgical treatment to HIV infected patients because of social desirability bias. However, a general effect of low reliability is weak associations between variables. Thus, the magnitude of the correlations presented, and the findings harmonizing with the TPB indicate acceptable reliability as well as validity of the results. Past experience as assessed in this study was limited to the extent that it provided no information about the level and frequency of HIV and AIDS related care delivery. Whereas the present findings support the notion that the TPB model is applicable in the sub-Saharan African context, it says nothing about the validity of the model per se, only that the TPB is just as useful in sub Saharan Africa as in other industrialized country contexts. The TPB is applicable to students' intention to provide health care to patients with HIV and AIDS across two sub-Saharan African countries. It is suggested that attitudes, subjective norms, moral norms and perceived behavioural control are key factors in students' decision to treat HIV infected patients and should be targeted in interventions aimed at improving health care quality in the context of HIV and AIDS. ful to Dr. Elizabeth Lyimo who was responsible for collecting data in Tanzania.

Pattern Recognition Receptor–Dependent Mechanisms of Acute Lung Injury

Acute lung injury (ALI) that clinically manifests as acute respiratory distress syndrome is caused by an uncontrolled systemic inflammatory response resulting from clinical events including sepsis, major surgery and trauma. Innate immunity activation plays a central role in the development of ALI. Innate immunity is activated through families of related pattern recognition receptors (PRRs), which recognize conserved microbial motifs or pathogen-associated molecular patterns (PAMPs). Toll-like receptors were the first major family of PRRs discovered in mammals. Recently, NACHT–leucine-rich repeat (LRR) receptors and retinoic acid–inducible gene–like receptors have been added to the list. It is now understood that in addition to recognizing infectious stimuli, both Toll-like receptors and NACHT-LRR receptors can also respond to endogenous molecules released in response to stress, trauma and cell damage. These molecules have been termed damage-associated molecular patterns (DAMPs). It has been clinically observed for a long time that infectious and noninfectious insults initiate inflammation, so confirmation of overlapping receptor-signal pathways of activation between PAMPs and DAMPs is no surprise. This review provides an overview of the PRR-dependent mechanisms of ALI and clinical implication. Modification of PRR pathways is likely to be a logical therapeutic target for ALI/acute respiratory distress syndrome.

(TLRs 1-9 and TLRs 11-13) in mice have been defined (14) . TLRs 3, 7, 8 and 9 are expressed intracellularly, whereas TLRs 1, 2, 4, 5, 6 and 10 are expressed on the cell surface. TLRs are expressed on a range of immune cells including macrophages, dendritic cells, B cells and certain types of T cells, as well as on certain nonimmune cells, such as endothelial cells, smooth muscle cells and epithelial cells that lie at potential sites of entry, including the skin and the respiratory, intestinal and genitourinary tracts. The expression of TLRs is modulated by activation, matu-ration or differentiation of the different cell types (15, 16) . TLR proteins are a family of type I transmembrane receptors characterized by an NH2-terminal extracellular leucine-rich repeat (LRR) domain, which mediate the recognition of their respective PAMPs, and a COOH-terminal intracellular tail containing a conserved region called the Toll/interleukin 1 (IL-1) receptor (TIR) homology domain. The TIR domain is the defining motif of the TLR/IL-1 superfamily, and it is likely to be one of the earliest signaling domains to have evolved (17) . TLRs can recognize a diverse range of PAMPs, generate inflammatory signals to coordinate innate immune responses and modulate adaptive immune responses. The list of TLR ligands is growing. However, the ligand for TLR10 and mouse TLR8 remains unknown at present. Activation of TLRs initiates two major pathways: the MyD88-dependent pathway, which is used by all TLRs except TLR3, resulting in the activation of nuclear factor (NF)-κB and activator protein-1 (AP-1); and the TRIF-dependent pathway, which is initiated by TLR3 and TLR4, resulting in the activation of type I interferons (IFNs) (13, 18, 19) . Expression of numerous proinflammatory cytokines, such as tumor necrosis factor (TNF)-α, IL-6, IL-12 and IFNs, is one of the major outcomes of the activation of the pathways (15) . TLR signaling is summarized and shown in Figure 2 . RLRs as DExD/H-containing RNA helicases are expressed in the cytoplasm in a variety of cells, including immune and nonimmune cells. Unlike membranebond TLR3, TLR7 and TLR9, which are localized on the endosome and recognize viral double stranded RNA, singlestranded RNA and DNA, respectively, RLRs are cytoplasmic proteins that recognize viral RNA produced as a consequence of viral replication (20, 21) . RLRs consist of three family members: RIG-I, melanoma differentiation-associated gene 5 (MDA5) and laboratory of genetics and physiology 2 (LGP2) (21, 22) . Role of PRRs in mediating inflammation and organ injury. Infection causes PAMP release, but also causes tissue and cell damage and subsequent DAMP release. Similarly, injury caused by trauma or various other factors not only leads to DAMP release but also renders the patient more susceptible to infection and therefore PAMP release. In turn, the PAMPs and DAMPs act through PRRs, which include TLRs, NLRs and RLRs, to activate the innate immune system, yet they can also contribute to persistent and deleterious systemic inflammation and organ injury, including ALI. Structurally, RIG-I and MDA5 contain a DExD/H box RNA helicase domain and two caspase-recruiting domain (CARD)like domains required for eliciting downstream signaling pathways (23, 24) . The C-terminal region of RIG-I contains a repressor domain (RD), which inhibits downstream signaling. The MDA5 C-terminal region is similar to the RD of RIG-I; however, its function is not clear. LGP2 contains a DExD/H helicase domain and an RD, but lacks the CARD-like region. LGP2 was suggested to play an inhibitory role in virus-induced response, because the LGP2 RD binds the RIG-I RD and suppresses signaling as a consequence of interfering with the self-association of RIG-I (20, 25, 26) . RIG-I is essential for the recognition of a series of RNA viruses, which include Sendai virus, Newcastle disease virus, influenza virus, vesicular stomatitis virus and Japanese encephalitis virus (27) . MDA5 is required for the recognition of other RNA viruses, including picornaviruses such as encephalomyocarditis virus, Mengo virus and Theiler virus (13, 27) . Thus, RIG-I and MDA5 have specificities in their detection of RNA viruses, presumably through recognition of distinct structures of viral RNA (24, 28) . Recent studies revealed a pathway of RLR regulation of NF-κB. RIG-I/MDA5 CARD domains, through a CARD-containing adaptor protein, IFNβ promoter stimulator 1, also known as mitochondrial antiviral signaling protein, and CARD adaptor inducing IFNβ, ultimately activate IRF3 and NF-κB (29, 30) . The role of RLRs in the mechanism of ALI has not been elucidated. The NLR family is a group of recently identified cytoplasmic PRRs that contain more than 23 members in humans (15) . The major role of NLRs is to recognize cytoplasmic microbial PAMPs and/or endogenous danger signals and initiate immunological responses, although the physiological function of most NLRs is poorly understood at present (31) . Members of the NLR family are categorized into at least five subfamilies according to their N-terminal structure, including NODs (nucleotidebinding oligomerization domain-1), NALPs (NACHT-, LRR-and pyrindomain-containing proteins), IPAF (ICE-protease activating factor), NAIPs (neuronal apoptosis inhibitor factors) and class II transactivator (CIITA) (32) . The NLR family shares a domain organization consisting of a C-terminal LRR domain, a central nucleotide-binding NACHT domain, and an N-terminal protein-protein interaction domain composed of a CARD, pyrin domain (PYD) or baculovirus inhibitor of apoptosis repeat (BIR) domain (33) . NODs and IPAF contain CARD effector domains, whereas NALPs have PYD domains, and NAIPs possess BIR domains. The functions of NOD1, NOD2, IPAF and NALP3 are more studied. NOD1 and NOD2 are the first NLRs that are reported (TLR2 in association with TLR1 or TLR6), TLR4, TLR5 and TLR11 are localized on the cell surface for ligand recognition. TLR3, TLR7 and TLR9 are localized in the endosome for ligand recognition in the lumen of endosome. All TLRs, except TLR3, recruit MyD88, and TLR1, TLR2, TLR4 and TLR6 recruit the additional adaptor TIRAP, which links the TIR domain with MyD88. TLR3 and TLR4 recruit TRIF. TLR4 requires the additional linker adaptor TRAM, which links the TIR domain of TLR4 with TRIF. Stimulation of the cells with TLR1, TLR2, TLR5, TLR6 and TLR11 ligands initiates the MyD88-dependent pathway whereas TLR3 ligands initiate the TRIF-dependent pathway. TLR4 activates both MyD88-dependent and TRIF-dependent pathways. In the MyD88dependent pathway, MyD88 recruits the IRAK family of proteins and TRAF6. In turn, TRAF6 activates TAK1. The activated TAK1 activates the IKK complex, which activates NF-κB subunits. The activated TAK1 also activates the MAPK pathway. In the TRIF-dependent pathway, TRIF interacts with RIP1 and TRAF6. Activated TRAF6 and RIP1 activate NF-κB and MAPKs. TRIF also interacts with TRAF3 and activates TBK1/IKKi, which activate IRF3 and IRF7. Cells stimulated with TLR7 and TLR9 ligands activate NF-κB and MAPKs via the MyD88-dependent pathway. To induce type I IFNs, MyD88 associates with the IRAK family of proteins. IRAK1 and IKKα activate IRF7. IRAK1 also interacts with TRAF3 and activates IRF7. The activated NF-κB subunits and IRFs are translocated to the nucleus. NF-κB and MAPKs initiate the transcription of inflammatory cytokine genes whereas IRFs initiate the transcription of type I interferons. (The figure is adapted from [13] and used with permission from Elsevier.) to have a direct function as PRRs in the recognition of peptidoglycan (PGN)derived peptides. NOD1 senses γ-D-glutamyl-meso-diaminopimelic acid (that is,-DAP) that is derived from Gramnegative bacteria (34) , whereas NOD2 senses muramyl dipeptide, which is from both Gram-positive and Gram-negative bacteria (35, 36) . When NODs bind with PGN-derived peptides, they rapidly form oligomers, which lead to the recruitment of the receptor-interacting protein 2 (RIP2) kinase through CARD-CARD interactions (37) . This complex of NOD1-RIP2 or NOD2-RIP2 then recruits the inhibitor of NK-κB kinase complex (IKK), leading to the activation of NF-κB. Activation NOD1 and NOD2 can also initiate an MAPK pathway that leads to the activation of p38 and ERK. In addition, NOD1 signaling can activate JNK as well (38) . NALP proteins are characterized by the presence of an N-terminal pyrin effector domain (39) . Several NLRs, namely NALP1, NALP2 and NALP3, have an important role in activation of proinflammatory caspases through formation of inflammasome (37, 40) . The inflammasome is a multiprotein complex of more than 700 kDa that is responsible for the activation of caspases 1 and 5, leading to the processing and secretion of the proinflammatory cytokines IL-1β and IL-18 (41, 42) . Martinon et al. have presented different caspase activation platforms in which different components constitute the various inflammasomes (43) . Two types of NALP inflammasome are better studied: the NALP1 inflammasome that is composed of NALP1, the adaptor protein ASC, caspase-1 and caspase-5 (41) , and the NALP2/NALP3 inflammasome that contains NALP2 or NALP3 CARDI-NAL, ASC and caspase-1 (44) . TLRs 1-10 are expressed in lung tissue (45) , and individual TLRs are differentially regulated in specific lung cell populations in response to microbial stimula-tion. TLR2, TLR4, TLR5 and TLR9 are the most likely to be involved in recognition of bacteria in the lungs (46) (47) (48) . A study by Bernard et al. has shown that ALI/ ARDS induced by lipopolysaccharide (LPS) is a major cause of mortality among humans (49) . The LPS membrane receptor complex is composed of several accessory molecules, which include phosphatidylinositol-anchored CD14, TLR4, MD2 and MD1 (50) . The LPSbinding protein (LBP) enhances the binding of LPS to its receptor (51) . Absence of CD14, MD2 or LBP abrogates most LPS responses (51, 52) . Expression of functional TLR4 has been found in many cell types in the lung (45) , and LPS-induced lethal shock and ALI have been shown to be TLR4 dependent (53) (54) (55) . Thus, TLR4 plays a critical role in the mechanism of infection-related ALI. A recent study has shown that respiratory infections in the human lung initiated by TLR2 agonist lipoteichoic acid (LTA, a component of Gram-positive bacteria) and TLR4 agonist LPS (a component of Gram-negative bacteria) exhibit different inflammatory responses (56) . The study was performed on healthy subjects with LPS or LTA instillation into the contralateral lung. Alveolar macrophages (AMφ) isolated from bronchoalveolar lavage fluid were analyzed by multiplex ligation-dependent probe amplification. The results show that whereas both LPS and LTA elicited neutrophil recruitment, only LPS instillation was associated with activation of neutrophils (PMN) (CD11b surface expression and degranulation) and consistent rises of chemokine/cytokine levels. Moreover, LPS but not LTA activated AM, as reflected by enhanced expression of proinflammatory mediators and increased spontaneous cytokine release upon incubation ex vivo. Remarkably, only LTA induced C5a release. These data suggest that stimulation of TLR2 or TLR4 results in differential pulmonary inflammation, which may be of relevance for understanding the differences during Gram-positive and Gram-negative respiratory tract infection (56) . Pulmonary endothelium is a major component of the alveolar-capillary unit, and is susceptible to injury from noxious agents that are either inhaled or delivered to the lung through the pulmonary circulation (57) . In ALI, pulmonary endothelium plays a major role by (a) altering metabolic activity to affect pulmonary and systemic homeostasis, (b) mediating polymorphonuclear PMN adhesion to promote PMN infiltration, (c) changing PMN barrier permeability to cause pulmonary edema and (d) secreting cytokines and chemokines to induce lung inflammation (58) . A recent study by Andonegui and colleagues has shown that endothelial cells (ECs) are the key sentinel cells for detecting infection by Gram-negative bacteria and recruiting PMN to peripheral tissues (53, 59) . Indeed, previous studies have shown that direct activation of circulating PMN with LPS is not sufficient to induce their sequestration within the lung (53) . Interactions of PMN with ECs seems important for the process of PMN sequestration into the lung (53, 60) . LPS stimulates the CD14 and TLR4 complex, which in turn activates NF-κB (61) and increases the expression of adhesion molecule E-selection, intercellular adhesion molecule-1 (ICAM-1) and vascular cell adhesion molecule 1 (62, 63) . The TLR4 signaling also leads to production and release of various bioactive molecules, including IL-1β, IL-6, TNF-α, chemokines and nitric oxide (64) , all of which are actively involved in the development of ALI. More importantly, TLR4 signaling can also upregulate other TLRs, such as TLR2, and thus amplify the inflammatory responses. Although TLR2 is predominantly expressed in the first-line host defense cells (monocytes, macrophages, dendritic cells and PMN) (65, 66) , its expression is low in ECs and epithelial cells (67) . Studies from our laboratory showed that LPS through TLR4-and MyD88-dependent signaling induced TLR2 upregulation in ECs (68) . We have demonstrated that TLR4 signaling, through activating NF-κB, upregulates TLR2 expression in ECs, and this process is enhanced by oxidant signaling generated by PMN NAD(P)H oxidase. The functional relevance of NAD(P)H oxidase in mediating TLR4-induced TLR2 expression in ECs is evident by markedly elevated and stable ICAM-1 expression as well as augmented PMN migration in response to sequential challenge with LPS and peptidoglycan (68) . Thus, TLR2 activation, signaled by TLR4 and as regulated by PMN NAD(P)H oxidase, is an important mechanism responsible for amplifying PMN transmigration to sites of infection ( Figure 3 ). The TLR4-TLR2 interaction suggests a highly coordinated, oxidant-mediated upregulation of TLR2 in response to LPS. When one considers the interactions of the innate immune system as microbes are first encountered, the value of such temporal organization is significant. For example, Gram-negative bacteria persist in tissues and, if they are not immediately killed through the activation of PMN, complement and other antimicrobial factors, they may spill out systemically and result in septic shock. Survival in the face of such infections depends on the innate immune system, which must be able to monitor and respond to pathogens over a prolonged period of time. Given the need for a prolonged response to bacterial infection, it has always seemed somewhat surprising that response to LPS is temporally finite. This endotoxin tolerance means that within hours after exposure to LPS, innate immune cells are incapable of responding again to a rechallenge (69) . But it is now clear that as LPS sensitivity wanes, the immune system has at its disposal the capability of marshaling responses via oxidative metabolites and their ability to upregulate other TLRs (70) . The subsequent means of responding to bacteria depend on the ability of the innate immune system to destroy microbes and enhance the release of alternative immune stimuli. The TLRs that are utilized are the ones that bind the constituents of degrading bacteria, such as lipopeptides, PGN, heat shock proteins and CpG DNA. It seems plausible that activated PMN may even alter the phenomenon of LPS tolerance, at least in a localized context, by setting into action a positive feedback loop at sites to which PMN are chemoattracted (71) . This would enhance inflammatory responses locally and help fight infection. However, in a setting of posttrauma SIRS or ALI, the primed PMN activation serves as an amplifier to cause enhanced PMN infiltration and organ injury. Extensive PMN influx into the lungs is one of the characteristics of ALI. Studies have shown that excessive induction of proinflammatory cytokines in PMN and delayed PMN apoptosis are associated with higher mortality and more severe organ dysfunction in sepsis patients (72, 73) . Human PMN express all TLR mRNA except TLR3. TLR2 are more abundant than TLR4 on PMN (74, 75) . PMN recruitment to the lung after LPS inhalation is primarily dependent on TLR4-NF-κB signaling (76, 77) . The PI3K/Akt pathway was also reported to be involved in TLR4-induced expression of IL-1β, TNF-α and chemokine macro phage inflammatory protein (MIP)-2 (78) . We have reported that TLR4, through regulating G-protein-coupled receptor kinases (GRKs), promotes PMN migration. We demonstrated that MIP-2 induces GRK2 and GRK5 expression in PMNs through phosphoinositide-3-kinase (PI3K)-γ signaling, and LPS-activated signaling through the TLR4 pathway transcriptionally downregulates the expression of GRK2 and GRK5 in response to MIP-2. The reduced expression of GRKs lowers chemokine receptor desensitization and markedly augments the PMN migratory response. These data indicate that TLR4 modulation of PMN surface chemokine receptor expression after the downregulation of GRK2 and GRK5 expression is a critical determinant of PMN migration (79) (Figure 4) . A recent study explored a novel role of mTOR complex 1 (mTORC1) in TLR2and TLR4-induced PMN activation (80) . Administration of rapamycin, an inhibitor of mTORC1, decreased the severity of lung injury after intratracheal LPS or PAM (a TLR2 ligand) administration, as determined by diminished neutrophil accumulation in the lungs, reduced interstitial pulmonary edema, and diminished levels of TNF-α and IL-6 in bronchoalveolar lavage fluid. These results indicate that mTORC1 activation is essential in TLR2-and TLR4-induced PMN activation, as well as in the development and severity of ALI. The E3 ubiquitin ligase Cblb has a crucial role in the prevention of chronic inflammation and autoimmunity. However, a recent study showed that Cblb also has an unexpected function in acute lung inflammation (81) . Cblb attenuates the sequestration of PMN in the lungs after administration of LPS. In a model of polymicrobial sepsis in which acute lung inflammation depends on TLR4, the loss of Cblb expression accentuates acute lung inflammation and reduces survival. Cblb controls the association between TLR4 and the intracellular adaptor MyD88. Expression of WT Cblb, but not expression of a Cblb mutant that lacks E3 ubiquitin ligase function, prevents the activity of a reporter gene for NF-κB in monocytes that have been challenged with LPS. The downregulation of TLR4 expression on the cell surface of PMN is impaired in the absence of Cblb. These data reveal that Cblb regulates the TLR4mediated acute inflammatory response that is induced by sepsis (81) . PMN apoptosis is a crucial injurylimiting mechanism of inflammatory resolution. Circulating PMN undergo constitutive apoptosis that results in the shutdown of secretary capacity and allows PMN recognition and removal by macrophages (82, 83) . Several inflammatory agents, such as LPS, TNF, IL-8, IL-6, IL-1 and granulocyte colony-stimulating factor (G-CSF), can delay apoptotic response, providing PMN with a longer life span, which in turn allows the PMN to accumulate at local tissue sites of inflammation/infection (84, 85) . Protein 53 (p53) is a transcription factor that is important in multicellular organisms, where it regulates the cell cycle and promotes apoptosis. Modulation of p53 by nutlin-3α diminished the response of PMN and macrophages to stimulation through TLR2 or TLR4 as well as attenuated LPS-induced ALI. NF-κB has been reported as a modulator of apoptosis in inflammatory cells (86) . p53 can negatively regulate NF-κB activity by decreasing binding of NF-κB to the promoters of genes for proinflammatory cytokines. In p53 -/mice, the inflammatory process and severity of ALI in response to LPS are enhanced (87) . AMφ account for approximately 95% of airspace leukocytes (88) . Tissue damage induced by LPS is mediated mainly by inflammatory products released from AMφ (89, 90) , thus activated AMφ play a critical role in the development of ALI (91) . LPS inhalation induces AMφ to produce and release inflammatory mediators TNF-α, IL-1β and MIP-2 in a TLR4dependent manner (92) , which further result in the recruitment of PMN into the lower respiratory tract and activate other cell types, including epithelia and endothelia (93) . TLR4 is constitutively expressed in AMφ, and TLR2 can be induced in response to LPS or proinflammatory cytokines. The inducible TLR2 expression might be important in responding to other bacterial components from Grampositive bacteria (48) . The role of CD44 in the regulation of LPS-TLR signaling in macrophages has recently been reported (94) . CD44 is a transmembrane adhesion molecule and hemopoietic CD44 has an essential role in hyaluronan clearance and resolution of noninfectious lung injury. Following intratracheal LPS treatment, CD44 -/mice demonstrated an exaggerated inflammatory response characterized by increased inflammatory cell recruitment, elevated chemokine expression in bronchoalveolar lavage fluid and a marked increase in NF-κB DNA-binding activity in lung tissue in vivo and in macrophages in vitro. Furthermore, CD44 -/mice were more susceptible to LPS-induced shock. The study further found that the induction of the negative regulators of TLR signaling IL-1R-associated kinase-M, Toll-interacting protein and A20 by intratracheal LPS in vivo and in macrophages in vitro was significantly reduced in CD44 -/mice. Collectively, these data suggest that CD44 plays a role in preventing exaggerated inflammatory responses to LPS by promoting the expression of negative regulators of TLR-4 signaling (94). Impairment of the alveolar epithelial barrier is important in the development of ALI. Under physiologic conditions the epithelial barrier is less permeable than the endothelial barrier; thus, destruction of epithelial barrier integrity prompts a progressive influx of protein-rich fluid into the alveoli (95) . On the other hand, the loss of epithelial integrity represents an impairment of physiologic transepithelial fluid transport and further inhibits the reabsorption of alveolar edema (96) . TLRs are critical for airway epithe-lial cell recognition of inhaled pathogens and for innate immune signaling. In cultured human lung epithelial cells, mRNA of all TLRs has been detected (97, 98) . TLR2, TLR3, TLR5 and TLR6 have the highest expression, and the ligands for these TLRs increased IL-8 and vascular endothelial growth factor (VEGF) production in normal human bronchial epithelial cells (99) . TLR2 is a heterodimer with TLR1 and TLR6, and each of these is present on the airway epithelial surface. TLR3, which recognizes dsRNA or poly(I:C), is located in endosomes in unstimulated human bronchial epithelial cells (100) . TLR5 is also present on the airway epithelial surface where it can interact with epidermal growth factor receptor (EGFR). Studies have shown that TLR ligands stimulated IL-8 and VEGF production via EGFR and the downstream signaling that might include MAP kinases and NF-κB (101, 102) . Interestingly, heat shock proteins (Hsp), such as Hsp72 and Hsp90, appear to be intimately involved in the recognition of LPS. Extracellular Hsp72 released from virally infected airway epithelial cells induces IL-8 express in human bronchial epithelial cells, resulting in the recruitment and activation of PMN via TLR4 (103, 104) . Type II alveolar epithelial cells can also be activated by LPS mediated through TLR4 signaling and in turn promote pulmonary inflammatory processes (105, 106) . A study by Togbe et al. of whether TLR gene dosage contributes to infection has demonstrated that overexpression of TLR4 augmented an LPS-induced bronchoconstrictive effect, as well as TNF-α and CXC chemokine ligand 1 (keratinocyte-derived chemokine) production (55) . The study further showed that PMN recruitment, microvascular and alveolar epithelial injury with protein leak in the airways, and damage of the lung microarchitecture were dependent on TLR4 gene dose. Therefore, the TLR4 expression level determines the extent of acute pulmonary response to inhaled LPS, and TLR4 may thus be a valuable target for immunointervention in acute lung inflammation as a result of infection. In addition to the TLRs, NLRs are critically involved in the sensing of bacterial pathogens. NOD1 senses diaminopimelic acid-containing peptidoglycan present in Gram-negative bacteria, whereas NOD2 senses the muramyl dipeptide present in most organisms. Because of the apparent lack of direct effects on cell signaling induced by activators of NLRs, it is suggested that their role in pathogen sensing is one of cooperation with the TLRs (107) . However, studies suggested that the actions of NOD1 vary between cell types and, unlike those seen with LPS, the in vivo effects may be independent of leukocyte activation. For instance, Cartwright and colleagues have shown that although selective activation of NOD1 in macrophages has no apparent effect, in vascular cells NOD1 activation results in the profound induction of NOSII and shock in vivo (108) . NOD2 is thought to be important in the maintenance of a healthy gut barrier because individuals who carry a defective NOD2 have an increased risk of Crohn disease and other intestinal disorders (109) . Although the importance of TLR family in sensing pathogens is well recognized, it is also plausible that they may function in noninfectious diseases because TLR expression is also regulated in conditions other than infection. Indeed, growing evidence has shown that PRRs play a central role in the mechanisms of noninfectious ALI. Several TLRs not only have the ability to recognize more than one ligand, but often recognize ligands with completely different chemical structures. Such TLR activation does not occur under normal circumstances but only when there is a change in the environment that either leads to the release of endogenous ligands from a cellular compartment, or leads to the modification of endogenous mediator that gives them the ability to activate TLRs (110, 111) . Because of the association of many endogenous ligands with tissue injury, the nomenclature of DAMPs has been suggested. The best characterized DAMPs include those products released from cells in response to stress or undergoing abnormal death, including Hsp60, Hsp70, the extra domain A of fibronectin, oligosaccharides of hyaluronic acid and high-mobility group box 1 (HMGB1). Most of these ligands act as agonists of TLR2 or TLR4, or both receptors (112) . Resuscitated hemorrhagic shock (HS) often promotes the development of lung injury by priming the immune system for an exaggerated inflammatory response to a second, often trivial, stimulus, the so-called "two hit hypothesis" (113) . We used a simplified animal model of the two-hit paradigm to address the mechanisms of HS-primed PMN migration and lung inflammation (114) . In this model, animals are subjected to a nonsevere resuscitated HS (hypotension at 40 mmHg for 1 hour, followed by a small dose of intratracheal LPS. Although neither shock nor LPS alone induces injury, the combination caused lung PMN accumulation and increased 125 I-albumin transpulmonary flux. Findings from this model have suggested that the mechanisms underlying the priming of PMN and inflammation involve a complicated receptor cross-talk process and interaction between PMN and AMφ, which is described below. We demonstrated that LPS-TLR4 signaling upregulates TLR2 expression in AMφ, and HS-activated PMN play a critical role in the mechanism of TLR2 upregulation (115) . This cross-talk between TLR4 and TLR2 in AMφ results in the amplification of expression of cytokines and chemokines in response to the bacterial products LPS and PGN, and subsequently leads to enhanced PMN sequestration in the lung. These findings reveal a novel mechanism underlying HSprimed lung injury, namely that HS-activated PMN that were initially sequestered into the alveoli can instruct AMφ to upregulate TLR2, thereby sensitizing AMφ to TLR2 ligands and promoting enhanced lung inflammation. How does HS-activated PMN enhance TLR4 upregulation of TLR2 in AMφ? We found that reactive oxygen species (ROS) derived from PMN NAD(P)H oxidase play an important role in amplifying the TLR2 upregulation (116) . Studies have also shown that lack of endogenous NAD(P)H oxidase in the AMφ caused a decrease in TLR2 expression in response to LPS stimulation; however, the decrease was restored when the AMφ was coincubated with PMN isolated from wild-type mice subjected to HS. These results indicate that although the endogenous NAD(P)H oxidase in AMφ is also involved in the signaling, the exogenous oxidants from PMN NAD(P)H oxidase are essential for inducing amplified TLR2 expression in AMφ in response to LPS. The TLR2 gene promoter contains multiple binding sites for transcriptional factors, which include NF-κB, CCAAT/enhancer binding protein, cAMP response element-binding protein and STAT (signal transducer and activator of transcription) (117) . Of these, NF-κB has been reported to regulate TLR2 expression in response to cytokines and mycobacterial infection (117, 118) . It has been demonstrated that LPS-TLR4-induced TLR2 upregulation in AMφ is largely mediated through the NF-κB signaling pathway, because the NF-κB inhibitor IKK-NBD significantly decreased LPS-induced TLR2 expression in AMφ (115) . Although oxidants are involved in the NF-κB signal transduction pathway (119) (120) (121) , their molecular targets have not yet been defined. The contribution of redox regulation and location of potential redox-sensitive sites within the NF-κB activation pathway are the subjects of controversy (119) . Recently we reported that HMGB1/TLR4 signaling mediates the HS-induced increase in TLR2 surface expression and decrease in TLR4 surface expression in the lung as well as in mouse lung vascular ECs (MLVEC) (122) . These alterations in TLR4 and TLR2 expression result in HMGB1-mediated activation of NAD(P)H oxidase and expression of ICAM-1 in MLVEC that is TLR4 dependent in the early phase and switches to being TLR2 dependent in the late phase following HS. More importantly, the HS-induced surface expression of TLR2 contributes to an enhanced activation of MLVEC and augmented pulmonary PMN infiltration in response to the TLR2 agonist PGN. Thus, the study demonstrates a novel mechanism underlying HS-augmented lung inflammation, namely that induction of increased TLR2 surface expression in lung endothelial cells, which is induced by HS/R and mediated by HMGB1 activation of TLR4 signaling, is an important mechanism responsible for EC-mediated inflammation and organ injury following HS (122) . We have shown that HS-induced PMN NAD(P)H oxidase activation is mediated by HMGB1-TLR4 signaling. HMGB1 was originally defined as a nuclear protein that functions to stabilize nucleosome formation, and it also acts as a transcription factor that regulates the expression of several genes (71) . HMGB1 can be secreted by innate immune cells in response to microbial products or other inflammatory stimuli (123, 124) . HMGB1 is also released by injured cells and is known as one of the main prototypes of the emerging DAMPs (125) (126) (127) . HMGB1 was initially identified as an inflammatory cytokine that is a late mediator of lethality in sepsis (123, 124) . However, recent studies suggest that HMGB1 acts as an early mediator of inflammation, contributing to the development of ALI after hemorrhage (128) , and hepatic injury after liver ischemia-reperfusion (129) . We found in our study that HS/R activates the TLR4-MyD88-IRAK4 signaling pathway through HMGB1, and further activates p38 MAPK and Akt pathways to initiate PMN NAD(P)H oxidase activation. PMN NAD(P)H oxidase-derived oxidants, in turn, mediate TLR4-TLR2 cross-talk in AMφ and sensitize AMφ response to TLR2 ligands, which act in a positive feedback manner to amplify pulmonary PMN infiltration and inflammation (130) . We have also addressed a fundamental question regarding how HS globally regulates PMN infiltration in the lungs. We have shown that HS, through alarmin HMGB1, induced IL-23 secretion from macrophages in an autocrine and TLR4 signaling-dependent manner. In turn, IL-23, through an IL-17-G-CSF-mediated mechanism, induced PMN egress from bone marrow. Therefore a sustained and HS-primed migration of PMN was maintained. We have also shown that β-adrenergic-receptor activation by catecholamine of macrophages mediated the HS-induced release of HMGB1. These data indicate that HS, a global ischemia/ reperfusion stimulus, regulates PMN mobilization through a series of interacting pathways that include neuroendocrine and both innate and acquired immune systems (131) . Hyaluronan (HA) is a massive sugar polymer in the extracellular matrix. Under physiologic conditions, HA exists as a high-molecular-weight polymer (>106 D) and undergoes dynamic regulation resulting in accumulation of lower molecular weight species (10-500 kD) after tissue injury. HA fragments can trigger innate immune responses in a manner that overlaps with both Grampositive and Gram-negative organism recognition pathways (132) . It has been demonstrated that fragmented HA accumulates during tissue injury (133) (134) (135) . CD44 is required to clear HA during tissue injury, and impaired clearance of HA results in unremitting inflammation. Additionally, fragmented HA stimulates the expression of inflammatory genes by inflammatory cells at the injury site (136) . Recently, Jiang et al. demonstrated that HA fragments require both TLR2 and TLR4 to stimulate mouse macrophages to produce inflammatory chemokines and cytokines. In a noninfectious lung injury model, mice deficient in both TLR2 and TLR4 showed an impaired transepithelial migration of inflammatory cells, increased tissue injury, elevated lung epithelial cell apoptosis and decreased survival (132) . Lung epithelial cell overexpression of high molecular mass HA protected mice against ALI and apoptosis, in part through TLR-dependent basal activation of NF-κB. The exaggerated injury in TLR2-and TLR4-deficient mice appears to be due to impaired HA-TLR interactions on epithelial cells. These studies demonstrate that host-matrix component HA and TLR interactions provide signals that initiate inflammatory responses, maintain epithelial cell integrity and promote recovery from ALI (136) . Mechanical ventilation (MV) provides life-saving support for many patients with respiratory failure (137) . However, mechanical stresses produced by MV can induce lung injury, termed ventilator-induced lung injury (138) . Evidence from animal experimental studies has demonstrated that MV per se can induce inflammatory responses (139) (140) (141) . A recent report has demonstrated that TLR4, but not TLR2, played a role in development of the inflammatory response after shorttime MV (142) . MV not only causes ventilator-induced lung injury in healthy animals, but also exacerbates damage in the injured lung (143) . TLR4 blockade reduces pulmonary inflammation caused by the combination of LPS and MV (144) . In the mechanism of hyperoxia-induced ALI, TLR3 expression and activation seems impotent. Exposure of human epithelial cells to hyperoxia in the absence of an exogenous viral pathogen significantly increased TLR3 expression. In vivo studies showed that both the absence of TLR3 via gene deletion in mice and the presence of an anti-TLR3 antibody in wild-type mice conferred significant protection in a hyperoxia-mediated lung injury model (145) . The inflammasome is a multiprotein complex that mediates the activation of caspase-1, which promotes secretion of the proinflammatory cytokines IL-1β and IL-18, as well as pyroptosis, a form of cell death induced by bacterial pathogens. Members of the NLR family, including NLRP1, NLRP3 and NLRC4, and the adaptor ASC are critical components of the inflammasome that link microbial and endogenous danger signals to caspase-1 activation. The role of NALP3 in mediating noninfectious ALI has been revealed by two recent studies. Gasse and colleagues reported that uric acid locally produced in the lung upon bleomycin (BLM)-induced DNA damage and degradation triggers NALP3 inflammasome activation, and in turn causes lung injury (146) . Reduction of uric acid levels using the inhibitor of uric acid synthesis allopurinol or uricase leads to a decrease in BLM-induced IL-1β production, lung inflammation, repair and fibrosis. Local administration of exogenous uric acid crystals recapitulates lung inflammation and repair, which depend on the NALP3 inflammasome, MyD88 and IL-1R1 pathways and TLR2 and TLR4 for optimal inflammation but are independent of the IL-18 receptor (146) . Babelova et al., however, reported the role of biglycan, a ubiquitous LRR proteoglycan of the extracellular matrix, in mediating ALI through interacting with TLR2 and TLR4 on macrophages (147) . The study showed that in macrophages soluble biglycan induces the NLRP3/ASC inflammasome and subsequent activation of caspase-1 and release of mature IL-1β without need for additional costimulatory factors. This is caused by the interaction of biglycan with TLR2/4 and purinergic P2 × 4/P2 × 7 receptors, which induces receptor cooperativity. Furthermore, ROS formation is involved in biglycan-mediated activation of the inflammasome. By signaling through TLR2/4 biglycan stimulates the expression of NLRP3 and pro-IL-1β mRNA. These results provide evidence for direct activation of the NLRP3 inflammasome by biglycan and suggest a fundamental paradigm of how tissue stress and injury are monitored by innate immune receptors detecting the release of the extracellular matrix components and turning such a signal into a robust inflammatory response (147) . Susceptibility and response to infectious disease is, in part, heritable. Potential associations between clinical outcome from sepsis and many inflammatory cytokine gene polymorphisms, innate immunity pathway gene polymorphisms and coagulation cascade polymorphisms have been observed. We may yet be able to tease out the complex influence of genetic variation on susceptibility and response to infectious disease (148) . In 2000, Arbour and colleagues reported that two polymorphisms of the TLR4 gene were present in a higher proportion of individuals who are hyporesponsive to inhaled LPS (149) . This finding led to a number of studies investigating the potential impact of these TLR4 polymorphisms on the course of infectious diseases and the development of septic shock and TLR4 polymorphisms (150) (151) (152) . These polymorphisms do not seem to confer sus-ceptibility to all Gram-negative infections, because other groups (153, 154) have shown no correlation in other infectious diseases such as meningococcal disease, and again one should be mindful that very significant hypofunctioning TLR4 alleles are likely to have a very strong negative selection pressure across the generations (155) . TLR-2 polymorphisms have also been linked to susceptibility to staphylococcal infection (156) and lepromatous leprosy (157, 158) . Although some of these studies are still relatively small in scale, they reinforce the important role of TLRs in pathogen recognition and immune response in humans. CD14 polymorphisms, key accessory molecules for TLR signaling, have been associated with increased prevalence of positive bacterial culture findings and sepsis attributed to Gramnegative infections in a critically ill population (159) , as well as the susceptibility to chronic Chlamydia pneumoniae infection in patients with coronary artery disease (160) . The discovery of the importance of PRRs in the pathogenesis of SIRS and organ injury, including ALI, has led to a therapeutic strategy targeting PRRs. TLRs are the most extensively studied family of PRRs, and thus recently developed new drugs mainly target TLRs and are either agonists of TLRs to enhance immune responses against infectious agents or antagonists designed to reduce inflammation due to infection or autoimmune responses (161) . The approaches to modulating TLR activity have focused on the following aspects: (a) ligands or analogues such as Eritoran (E5564) from Eisai (Woodcliff Lake, NJ, USA), a synthetic analogue of bacterial lipid A that inhibits LPS from activating cells through the TLR4/CD14/MD2 complex (162) (163) (164) ; (b) monoclonal antibodies, soluble receptors and other accessory proteins, such as a natural soluble form of TLR2, found in mouse plasma and breast milk, which acts to block TLR2 ligand stimulation (165) , and a member of the TLR/IL-1 receptor family (TIR8 or SIGIRR) that inhibits NF-κB signaling and may be an endogenous inhibitor of the TLR system (166); (c) signal transduction blockers; many of the key molecules in the signaling pathways for each TLR have been identified and are considered potential drug targets (167) (168) (169) . The structural bases of TIR domain interactions between TLRs and adapters such as MyD88, Mal, TRAM and TRIF have been modeled, and small peptidic sequences based on the TIR domain BB loop or peptidomimetics of this region have been made that can block the interactions (110, 169, 170) . (d) siRNA and antisense; studies with knockout mice suggest that deficiency in individual TLRs has limited consequences for animals under normal conditions, but exhibits impact under conditions of specific infectious challenge (110, 171, 172) . However, siRNA sequences themselves may be ligands for intracellular TLRs (173) ; thus attention to the design of appropriate sequences is necessary. Drugs targeting TLRs have not yet been clinically applied to the treatment of ALI, but because of the critical role of PRRs in the development of ALI, targeting of PRRs has opened up a productive area for the therapy of ALI. Current pharmacotherapy has not been highly successful in increasing patient survival in cases of ALI/ARDS. Since PRRs were recognized, their significance in the mechanisms of ALI has been quickly identified. The combined activation of these different receptors may result in complementary, synergistic or antagonistic effects that modulate the process of ALI. Therefore, modification of PRR pathways is likely to be a logical therapeutic target for ALI/ARDS. However, a complete understanding of the role of PRRs in the mechanism of ALI requires further "decoding" of these multiple receptor interactions. This work was supported by the National Institutes of Health Grant R01-HL-079669, National Institutes of Health Center Grant P50-GM-53789, and a VA Merit Award. The authors declare that they have no competing interests as defined by Molecular Medicine, or other interests that might be perceived to influence the results and discussion reported in this paper.

Exposure of cats to low doses of FeLV: seroconversion as the sole parameter of infection

In felids, feline leukemia virus (FeLV) infection results in a variety of outcomes that range from abortive (virus readily eliminated and never detectable) to progressive infection (persistent viremia and viral shedding). Recently, a novel outcome was postulated for low FeLV infectious doses. Naïve cats exposed to faeces of persistently infected cats seroconverted, indicating infection, but remained negative for provirus and p27 antigen in blood. FeLV provirus was found in some tissues but not in the bone marrow, infection of which is usually considered a necessary stage for disease progression. To investigate the impact of low FeLV doses on young cats and to test the hypothesis that low dose exposure may lead to an unknown pathogenesis of infection without involvement of the bone marrow, 21 cats were infected oronasally with variable viral doses. Blood p27, proviral and viral loads were followed until week 20 post-infection. Tissue proviral loads were determined as well. The immune response was monitored by measuring FeLV whole virus and p45 antibodies; and feline oncornavirus-associated cell membrane antigen (FOCMA) assay. One cat showed regressive infection (transient antigenemia, persistent provirus-positivity, and seroconversion) with provirus only found in some organs at sacrifice. In 7 of the 20 remaining cats FOCMA assay positivity was the only sign of infection, while all other tests were negative. Overall, the results show that FeLV low dose exposure can result in seroconversion during a presumed abortive infection. Therefore, commonly used detection methods do not detect all FeLV-infected animals, possibly leading to an underestimation of the prevalence of infection.

Feline leukemia virus (FeLV) is a retrovirus of great veterinary importance that was discovered more than 40 years ago [20] and infects domestic cats and some related small felids worldwide [18, 22, 32] . An infection with FeLV may cause disorders of hematopoietic cells, a state of immunodeficiency and fatal neoplasia. Common symptoms are fever, anaemia, anorexia and weight loss. The prevalence of FeLV infection has been decreasing in the past few years. In Switzerland, FeLV prevalence was found to be 3% in healthy and of up to 13% in ill cats [29] . Decreasing prevalence is the consequence of identification and segregation of infected cats, and of extensive vaccination programs. However, cases of recurrence keep occurring and emphasise the importance of an accurate diagnosis [24] . FeLV is mainly transmitted directly from cat to cat. Transmission occurs through contact with saliva via licking, mutual grooming and sharing of food or water dishes or trough bites [5, 6] . Infectious FeLV can also be transmitted via faeces and milk [8, 34] . FeLV RNA was also detected in urine, but, its infection potential has not been demonstrated so far [1] . Pathogenesis is well understood if infectious pressure is high. Infection starts in the oropharynx where the virus first replicates in tonsils and local lymph nodes [37] , from which the virus spreads to the bone marrow, thymus, spleen and intestine through infected lymphocytes. Bone marrow infection is an important hallmark of FeLV pathogenesis as the virus replicates extensively in bone marrow and infects blood precursor cells first, and many organs and tissues thereafter, including salivary gland, tonsils, pharyngeal, urinary bladder, gastric, intestinal, colonic, pancreatic and endometrial epithelia, lymph nodes throughout the body, bone marrow, and spleen [36, 37] . FeLV infection presents a variety of outcomes [11, 16-18, 25, 27, 35, 37, 38] , which are influenced by both host and virus factors. Known host resistance factors include age and immune system status. Known virologic factors are virus strain, subtype and viral infection dose [18] . The outcome of infection is still a rather controversial issue. In the past, infection outcome was classified as viremic, non-viremic, and transiently viremic based on results of virus isolation, immunofluorescence assays and/or antigen detection [27] . Viremic cats are characterized by continuous expression of p27 viral antigen. FeLV infection is not contained due to a lack of FeLV specific immunity [3, 4, 18] . Viremic cats continuously shed virus, thereby posing a risk of infection to susceptible cats, and usually succumb to FeLV-associated diseases (anaemia, immunosuppression, and neoplasia). In transiently viremic cats, viremia is overcome after a few weeks post-infection (p.i.). However, transiently viremic cats remain provirus positive [15] . In addition, some nonviremic cats were shown to have localized infection characterized by virus replication in certain tissues, such as mammary, salivary and urinary epithelium [8, 12, 34] . This additional form of FeLV infection was termed atypical or sequestered infection. New sensitive molecular assays have been described recently for the use in detection and quantification of FeLV provirus DNA and viral RNA [14, 39, 42] , resulting in a more sensitive measure for FeLV exposure. The spectrum of host response categories was refined accordingly. Especially in p27 negative cats, the outcome should be reevaluated based on the presence or absence of FeLV proviral DNA in blood or bone marrow. The spectrum of host response categories was thus reclassified into: abortive (no virus detected after exposure), regressive (p27-negative, provirus positive after or without transient antigenemia) and progressive (persistently p27 positive, virus isolation, provirus positive, FeLV RNA positive) [16] . Whether cats that show no signs of infection (abortive) are truly immune or just resistant, is still open to debate. In this context, a very interesting finding was that FeLV infection can be transmitted by contact with faeces [8] with no apparent viremia and infection involving the bone marrow, and constant negativity for proviral DNA in blood -which would be classified as abortive infection. However, in the study of Gomes-Keller et al. [8] FeLV provirus could be detected in several organs and the cats showed seroconversion, indicating that infection indeed had occurred. It was concluded that under low infectious pressure a different pathogenesis may take place in which bone marrow is not involved. Since FeLV transmission at very low infection levels appears to be the most natural way of infection, many cats may show this yet uncharacterized outcome. Cats that test negative for FeLV by conventional diagnostic methods could still have been infected and constitute a relevant portion of the cat population affected by the virus. The prevalence of FeLV infection may therefore have been underestimated, a factor that can have important consequences for FeLV control management, e.g. in Iberian lynxes, which seem to be particularly susceptible for FeLV infection and for which a correct estimation of FeLV spreading potential is of great importance [32] . It was thus the aim of this study to further characterize the course of infection after exposure to low infectious pressure and to test the hypothesis that low FeLV doses applied to young cats may lead to seroconversion even when the infection does not progress through bone marrow. The virus may not be completely eliminated and a limited viral replication in tissues may lead to a development of a weak immune response. Three groups of 7 cats were exposed to different, low doses of FeLV. Blood p27 antigen, proviral DNA and viral RNA were measured in blood at regular intervals until week 20 p.i. and in popliteal and mesenteric lymph nodes, bone marrow, spleen, kidney, urinary bladder, lungs, thymus, myocardium, parotid gland, and pancreas after euthanasia. Immune response against FeLV was assessed by the detection of antibodies by ELISA to FeLV whole virus and to FeLV p45 (the recombinant env-gene product), by immunofluorescence assay to feline oncornavirus-associated cell membrane antigen (FOCMA), and by Western blot analysis to the separated FeLV components. Twenty-six, 9 weeks-old Specific Pathogen-Free (SPF) male kittens were obtained from Liberty Research, Inc. (Waverly, NY, USA). Animals were kept under barrier conditions and under optimal ethological and hygienic conditions in one group of 5 cats (uninfected control group) and three groups of 7 cats each (group 1K, 10K, and 100K, based on viral infectious dose). Prior to the beginning of the experiment the cats were tested by PCR, RT-PCR and serology and shown to be negative for FeLV, FIV, Herpes-, Corona-, Calici-, Parvovirus and feline hemotropic mycoplasmas. At the age of 16 weeks, each kitten of groups 1K, 10K, and 100K were infected once oronasally with FeLV-A/Glasgow-1 [19] by introducing 0.2 mL of virus suspension into each nostril and 0.6 mL into the mouth. The virus suspension for group 1K contained a total of 1 000 focus-forming units (FFU), the one for group 10K 10 000 FFU and group 100K 100 0000 FFU. The viral stock origin and infectivity was the same as described in earlier works [15, 41] . Blood samples were collected under sedation (0.01 mg/kg midazolam (Dormicum Ò , Roche Pharma AG, Reinach, Switzerland) and 10 mg/kg ketamine (Narketan Ò , Vétoquinol AG, Belp, Switzerland)) prior to challenge at week À7, À5, À3, and then weekly, starting from week 0 until week 6 p.i., later in biweekly intervals until week 20 p.i. Blood samples were obtained by jugular venipuncture using 5 mL syringes and blood was immediately transferred into EDTAtubes. 400 lL of EDTA blood was submitted for haematology analysis, 200 lL of EDTA anticoagulated whole blood were aliquoted for DNA extraction. Plasma was obtained by centrifuging approximately 2 mL of EDTA blood at 1 700· g for 10 min. Blood and plasma samples were immediately frozen at À80°C until they were processed. For determination of FeLV proviral loads, total nucleic acids were extracted from a blood volume containing 10 6 white blood cells using the MagNa Pure LC Total Nucleid Acid Isolation Kit (Roche Diagnostics AG, Rotkreuz, Switzerland). The extracted total nucleic acids were analyzed by real-time TaqMan PCR as described in [39] using the 2· TaqMan Ò Fast Universal PCR Master Mix (Applied Biosystems, Foster City, CA, USA) on a ABI 7500 sequence detection system (Applied Biosystems) and under the following cycling conditions: an initial denaturation of 20 s at 95°C was followed by 45 cycles of 95°C for 3 s and 60°C for 30 s. For each run, a glyceraldehyde-3-phophate dehydrogenase (GAPDH) pseudogene of which one copy is present in the genomic DNA of feline cells [33] was also quantified as described [39] using the 2· TaqMan Ò Fast Universal PCR Master Mix and the same PCR run conditions as for FeLV provirus. FeLV proviral DNA amounts were normalized to feline GAPDH by dividing FeLV copy numbers by fGAPDH copy numbers to calculate FeLV copies per cell. Viral RNA in plasma samples was extracted from 200 lL of plasma (either from 5-sample pools or from single samples) using the MagNa Pure LC Total Nucleic Isolation Kit and quantified by real-time TaqMan reverse transcriptase (RT)-PCR as described [39] using a ABI 7500 sequence detection system. The presence of plasma FeLV p27 antigen was determined using a sandwich ELISA as previously described [26] . FL-74 feline lymphoblastoid cell line permanently expressing FeLV), which was considered 100%. Samples reaching > 5% of the positive control signal were considered positive [14] . The plasma samples were also analysed for the presence of antibodies to FeLV whole virus, to FeLV p45 (the non-glycosylated form of gp70 surface unit of the envelope glycoprotein), and to FOCMA. Anti-FeLV p45 and anti-FeLV whole virus antibodies were measured by ELISA as described [21, 25] , using 100 ng of p45/well and 100 ng of gradient purified FL-74 FeLV, respectively. Plasma was used at a dilution of 1:200 and antibody levels assessed by comparison with predefined control antisera [25] . Antibody to FOCMA was measured at week 0 and week 20 p.i., by indirect cell membrane immunofluorescence as described [2] . FL-74 cell culture medium was tested for the absence of FCV, FHV, FPV, FCoV, FIV, hemotropic mycoplasma and presence of FeLV by RT-PCR/PCR as described [9, 13, 23, 31, 39, 44, 45] . The culture was consistently free of the unwanted contaminants. The cat sera were titrated at 4-fold dilutions from 1:4 to 1:256. Samples showing a minimal titre of 1:4 were considered to be FOCMA positive. In addition, samples from week À3 and week 20 p.i. were examined for the presence of antibodies to FeLV gp70, p27 and p15(E) [26, 30] by Western blot analysis as described [28] . Cats of group 10K and group 100K were euthanized at week 20, and tissue samples from popliteal and mesenteric lymph nodes, bone marrow, spleen, kidney, urinary bladder, lungs, thymus, myocardium, parotid gland, and pancreas were collected within 30 min post-mortem. Samples were snap-frozen in liquid nitrogen. Ten mg of tissue were used for the extraction of DNA using MagNA Pure LC DNA Isolation Kit II (Tissue kit, Roche Diagnostics). FeLV provirus was quantified by real-time PCR as described [39, 40] . To guarantee a comparable sensitivity, samples yielding less than 15 000 fGAPDH copies per reaction were re-extracted until a suitable concentration was reached. In addition, samples were collected under sterile conditions from mesenteric lymph node, urinary bladder, lungs, thymus, and myocardium for virus isolation. Tissue samples were co-cultured with FEA cells and supernatants of the co-cultures collected at days 4, 8, 12, 16, 20, 24 , and 28 post-inoculation were tested for p27 antigen by ELISA as described [7] . Bone marrow samples were cultured in medium and supernatants collected at days 18 and 21 were used for total nucleic acid isolation (MagNA Pure LC Total Nucleic Acid Isolation Kit, Roche Diagnostics) and samples were tested for the presence of FeLV RNA. Additionally, supernatants were tested by p27 ELISA [26] . Statistical analyses were carried out using R software version 2.9.0. (The R Foundation for Statistical Computing, Vienna, Austria). Longitudinal effects (time) on antibody titres of the different groups were compared to each other by multivariate analysis of variance (MANOVA). A p value < 0.05 was considered to be statistically significant. With one exception, cats did not become provirus positive at any time point tested. The one cat (JCR2) that was found to be positive belonged to the group challenged with the highest of the three doses of FeLV (100 000 FFU). Proviral DNA was detected from week 2 p.i. on with a peak at week 3 (3.33 copies/cell) (Fig. 1) . Similarly, viral RNA was consistently detected only in samples of the provirus-positive cat (JCR2) until week 12 and additionally at week 18 once more. The highest viral RNA load measured was at week 2 p.i. (Ct-value of 22.95, Fig. 1 ). Viral RNA was not detectable in pooled plasma samples (5 samples each pool) of any of the provirus-negative cats at week 0, 10, and 20 p.i. In addition, samples from week 2, 3, and 4 were tested individually, as the viral RNA peak usually occurs in this time period; the results were negative, too. p27 antigen was detected in plasma of one (JCR2) of the 26 cats tested. This cat became transiently positive from week 3 p.i. to week 4 p.i. (Fig. 1) . p27 was not detectable at any time point in all other cats. By measuring antibodies against FeLV p45 and whole virus by ELISA, seroconversion was observed in one cat. In the p45 ELISA, this cat (JCR2) had a positive response throughout weeks 8 to 20 with a peak of 26% of the positive control at week 14, and throughout weeks 4 to 20 with a peak of 54% of the positive control at week 20 in the FeLV whole virus ELISA (data not shown). Analysis of the remaining animals showed no statistically relevant difference between the groups in p45 as well as whole virus ELISA (P MANOVA 0.53 and 0.47, respectively Figs. 2 and 3). Prior to challenge, all samples were negative for the presence of antibodies directed against FOCMA. At week 20 p.i., antibodies to FOCMA were detected in 4 cats (KCT1, JCU3, JCO2, JCP1, 57%) of group 10K and 4 (KDA1, JCR2, KCQ1, JCS1, 57%) cats of group 100K. Of these 8 cats, the PCR-positive cat (JCR2) reached the highest titre (1:64), the other 7 cats developed antibody titres which ranged from 1:4 to 1:16. No negative control group and group 1K animal had a detectable titre (Tab. I). In Western blot analysis, sera from 2 cats of group 10K (JCU3, JCS2) showed a weak reactivity to the upper band of the p15(E) protein at a serum dilution of 1:40 as well as at 1:100. However, the same cats showed some reactivity to the upper band of p15(E) already at week 0 p.i. (Fig. 4) . In the PCR-positive cat (JCR2) the presence of antibodies against FeLV was confirmed at a dilution of 1:100. All other cats from group 100K, 10K and 1K had either a completely negative Western blot pattern or showed cross-reactive bands already present at week À3 (cats KDA1 and JCS2, Fig. 4 ). At week 20 post-challenge, cats of groups 10K and 100K were euthanized and different tissue samples were collected and analysed for the presence of FeLV DNA sequences. Popliteal and mesenteric lymph nodes, bone marrow, spleen, kidney, urinary bladder, lungs, thymus, myocardium, parotid gland, and pancreas were tested. FeLV DNA sequences were found in organs of the cat (JCR2) positive for FeLV provirus in blood. The copy numbers were in average very low and ranged from 5.6 to 704.25 copies/reaction (3.4 · 10 À5 to 0.0024 copies/cell, data not shown). The popliteal lymph node showed the highest FeLV proviral load (0.0024 copies/cell). No FeLV DNA sequences were detected in tissue samples of all other cats. Virus isolation from urinary bladder, myocardium, lungs, mesenteric lymph node, thymus, and bone marrow was negative for all cats. The aim of this study was to investigate to what degree cats can become infected by low dose oral FeLV infection without detectable involvement of the bone marrow. Even though FeLV prevalence has reportedly decreased during the past years, FeLV is still among the most important infectious diseases of cats. Hence, a deep knowledge of all aspects of FeLV pathogenesis and biology is essential. Recently, a novel course of infection was postulated that Week 0 Week 20 Week 0 Week 20 Week 0 Week 20 Week deviates from the classical course involving a phase in which bone marrow is infected. In this alternative course, cats exposed to low loads of virus via faecal route seroconverted without involvement of bone marrow and without showing any other sign of infection [8] . [16] . In this study, we demonstrated that a single oronasal exposure to FeLV as low as 10 000 FFU is sufficient to elicit an antibody response and to lead to seroconversion in at least 18% (lower 95% confidence interval of the distribution) of the cats, demonstrating that the postulated alternative course of infection may occur in a considerable portion of the animals. We propose to designate this outcome of infection as abortive infection with seroconversion. No hallmark of infection was observed in the cats that received the smallest of the three doses (1 000 FFU), for which individual local innate immunity might have been enough to contain viral replication and to lead to abortive infection without a detectable adaptive immune response. However, under field conditions, multiple exposures in the range of 1 000 FFU or lower may occur giving rise to seroconversion or even provirus positivity. Considering that the assay for proviral DNA is able to detect one copy per reaction, we believe that the consistent PCR-negativity in blood of the cats that seroconverted is a strong indicator for the absence of bone marrow infection, a confirmation that an alternative course of infection may indeed exist, where the bone marrow is not involved [1] . FeLV provirus may still be integrated into the genome of the cells of some tissues but, in most cases, at levels beyond detection. In the study presented here, FeLV DNA sequences were not detectable in the organs tested. However, the presence of FeLV provirus in tissues other than the ones we analysed can not be excluded and should be investigated further. In localized infection where FeLV replication is confined to certain tissues, presence of plasma viral RNA loads is regarded a sensitive parameter for infection. RNA loads reflect ongoing viral replication somewhere in the cat's body even in sequestered places. However, detection of RNA is generally less frequent than that of proviral DNA [39] . Seroconversion was observed by use of the FL-74 cell membrane immunofluorescence test. This assay originally had been designated FO-CMA test and was thought to detect antibodies at a non-virion tumour-associated antigen [2] . Later, it became clear that the FOCMA phenomenon can be explained not by a tumour specific antigen but by viral proteins including FeLV p15(E), gag-polyproteins and gp70 of FeLV subtype C present on the FL-74 cell membrane [43] . The FL-74 cell membrane immunofluorescence test seems to be the most sensitive method to detect seroconversion compared to ELISA and Western blotting. This may be due in part to the less stringent test conditions, which allow antibodies with lower affinity to bind to their epitopes, and in part to the fact that in the FOCMA assay the binding targets are presented in a native conformation, thus allowing a better binding of conformation-specific antibodies. Induction of cytotoxic T-cells or virus-neutralizing antibodies by oral low dose infection, or Th1 cytokine expression that might have been induced by the experimental infection were not measured. Instead, antibody response was used as marker of infection. It is unclear to what degree the alternative pathway may be relevant to the spread of FeLV infection. We speculate that in most cases cats will overcome the infection for good. In very few cases -when the infected cats undergo immunosuppression due e.g. to concomitant FIV infection -full-blown FeLV infection with bone marrow involvement and viral shedding may develop. It will be important to determine to what extent this may occur. In conclusion, the present study provides additional evidence for the existence of an alternative FeLV infection course that leads to development of antibodies without involvement of the bone marrow. The results may be important for the surveillance of the FeLV status of a cattery or a cat population. Most likely a cattery cannot be considered free of FeLV as long as cats with FeLV-reactive antibodies are present. The potential for overt infection cannot be neglected.

A large and accurate collection of peptidase cleavages in the MEROPS database

Peptidases are enzymes that hydrolyse peptide bonds in proteins and peptides. Peptidases are important in pathological conditions such as Alzheimer's disease, tumour and parasite invasion, and for processing viral polyproteins. The MEROPS database is an Internet resource containing information on peptidases, their substrates and inhibitors. The database now includes details of cleavage positions in substrates, both physiological and non-physiological, natural and synthetic. There are 39 118 cleavages in the collection; including 34 606 from a total of 10 513 different proteins and 2677 cleavages in synthetic substrates. The number of cleavages designated as ‘physiological’ is 13 307. The data are derived from 6095 publications. At least one substrate cleavage is known for 45% of the 2415 different peptidases recognized in the MEROPS database. The website now has three new displays: two showing peptidase specificity as a logo and a frequency matrix, the third showing a dynamically generated alignment between each protein substrate and its most closely related homologues. Many of the proteins described in the literature as peptidase substrates have been studied only in vitro. On the assumption that a physiologically relevant cleavage site would be conserved between species, the conservation of every site in terms of peptidase preference has been examined and a number have been identified that are not conserved. There are a number of cogent reasons why a site might not be conserved. Each poorly conserved site has been examined and a reason postulated. Some sites are identified that are very poorly conserved where cleavage is more likely to be fortuitous than of physiological relevance. This data-set is freely available via the Internet and is a useful training set for algorithms to predict substrates for peptidases and cleavage positions within those substrates. The data may also be useful for the design of inhibitors and for engineering novel specificities into peptidases. Database URL: http://merops.sanger.ac.uk

Peptidases (proteases or proteinases) are enzymes that hydrolyse the peptide bonds between amino acids in a protein or peptide chain. Hydrolysis of such bonds is required for removal of targeting signals (signal and transit peptides (1), ubiquitin (2), SUMO (3) and NEDD8 (4) peptides), the release of a mature protein from its precursor (5) , the switching off of a biological signal by degradation of the signal protein (6) , and for widespread catabolism of proteins for recycling of the amino acids. When proteolysis occurs unchecked, then diseases can result, such as Alzheimer's (7) , osteoarthritis (8) , emphysema (9) , tumour invasion (10) and acute pancreatitis (11) . Pathogens use peptidases to enter the host and to degrade host proteins for food (12) . Peptides and proteins have been widely used to characterize the specificity of peptidases, but frequently the substrates chosen have been physiologically irrelevant. One of the most popular substrates has been the oxidized insulin B-chain, because this is a peptide without tertiary structure, and cleavage depends solely on the preference of the peptidase (13) . (The terms 'specificity', 'selectivity' and 'preference' are used interchangeably here.) However, peptidase preference is exactly that: a preference only. Researchers often find that after prolonged exposure to a peptidase other bonds are degraded, albeit slowly, once none of the preferred bonds remain. If the peptidase preparation is not pure, then there is the danger that some of the observed cleavages are due to contaminating peptidases. The bond in a substrate where hydrolysis occurs is known as the 'scissile bond'. In the Schechter and Berger nomenclature (14) , residues on the left-hand side of the scissile bond (towards the N-terminus) are numbered P1, P2, P3, etc. and residues on the right-hand side (towards the C-terminus) are numbered P1 0 , P2 0 , P3 0 , etc. with cleavage occurring between P1 and P1 0 . The substrate-binding pocket in the peptidase that accommodates the P1 residue is known as the S1 pocket, and that accommodating the P1 0 residue is the S1 0 pocket. Predicting where a peptidase will cleave in a native protein is difficult. Where cleavage does occur in a protein is due to a combination of the preference of the peptidase and the availability of bonds in the substrate. Although the preference of the peptidase can be quite simple, for example trypsin (MEROPS identifier S01.151) cleaves lysyl and arginyl bonds (15) and caspase-3 (C14.003) cleaves only aspartyl bonds (16) , very often peptidase preference is cryptic. It is relatively easy to predict trypsin cleavages in a denatured protein, but few lysyl and arginyl bonds will be cleaved in a native protein. This has proved useful for researchers wishing to separate structural domains in a multidomain protein using limited proteolysis (17) . It is not possible to predict where in a peptide cathepsin B (C01.060) will cleave, for example, despite its known preferences for a hydrophobic residue in the S2 pocket and arginine in S1 (18) . Even though for some peptidases the specificity has been clearly defined, in all probability only a few bonds will be susceptible to cleavage in a mature protein. A protein will have few bonds flexible enough to thread into a peptidase active site if the protein is in a native state, because of the stabilizing interactions within and between secondary structure elements within the substrate. It is widely assumed that the susceptible bonds will be within surface loops and interdomain connectors. However, once a bond is cleaved and the tertiary structure perturbed, further bonds may become susceptible. Most studies of the action of a peptidase on a supposed physiological substrate are performed in vitro. It may be, however, that in vivo peptidase and substrate do not meet, either because of a physical boundary, such as being in different intracellular (or extracellular) compartments, because inhibitors inactivate the peptidase, the cleavage sites are inaccessible because the substrate is bound to another protein, or the environment is unsuitable and the peptidase is not active. Despite the importance of protein cleavage, there has been no centralized repository for cleavage data collection and no attempt to curate these cleavages by mapping them to residue positions in protein primary sequence databases. Given that nearly all proteins are eventually degraded, and that any one protein can be degraded by several different peptidases often by cleavages at multiple peptide bonds, the potential total number of cleavages will always exceed the number of known proteins. Up until recently each cleavage had to be characterized biochemically, which meant N-terminal sequencing of the products, a timeconsuming and labour-intensive task. Now that proteomic analyses are possible, where cell lysates or similar samples are subjected to cleavage by a peptidase, peptides isolated, composition determined by mass spectroscopy, and possible source protein(s) determined from the composition (19) , the amount of data is set to rise exponentially. This makes it vitally important that the information be accurately stored and curated. Such a collection made readily available would provide a comprehensive training set for algorithms and software for the prediction of physiological substrates and cleavage positions. The classification of peptidases into clans and families was first published in 1993 (20) , and this was converted into an Internet resource, the MEROPS database (21), in 1996. The database was extended to include nomenclature and bibliographies, and has been developed over the years to be a one-stop shop for researchers with an interest in proteolysis. The collection of known cleavages in substrates which was started in 1998 (22) has now been added to the MEROPS database. For each peptidase there is a page listing known substrates, and, where enough substrates are known, the peptidase summary has displays to show peptidase specificity. For each protein substrate, the sequence is displayed showing where cleavage occurs and which peptidase performs that cleavage. In addition to the MEROPS collection, there is also a collection of physiologically relevant protein cleavages assembled by the CutDB database (23) and more specialist collections of substrates for individual peptidases or peptidase families, such as CASBAH for caspases (24) . The primary source of protein cleavage information is the published literature. Search profiles have been developed for use at PubMed (25) and Scopus (http://info.scopus.com/). These are updated regularly and currently include $500 names that are known to be used for peptidases. These retrieve a set of $250 potentially interesting abstracts each week. There is much redundancy, in that a single article may be retrieved by several search terms. Once a nonredundant list of articles has been obtained, the abstracts are reviewed to select the subset that is to be included in MEROPS. In a typical week, 50-60 references come through this filter. Keywords, including the MEROPS identifiers for the relevant peptidases, are manually attached to each and these determine which pages in MEROPS the reference will appear on. If from the abstract it is clear that the paper contains substrate-cleavage data these are entered immediately into the MEROPS collection. Periodically, the bibliography in MEROPS for a peptidase is reviewed to find substrate cleavages with a preference for peptidases without any substrates in the MEROPS collection. If the substrate is a protein, it is mapped to a UniProt protein sequence database entry (26) initially by name and species. Each cleavage in the protein is mapped to a specific residue in the UniProt entry. Frequently the residue number reported in the paper refers to a position in the mature protein, and to map this to the UniProt sequence the length of any signal peptide and/or propeptide has to be added. The UniProt accession, the P1 residue number, the CRC64 checksum for the sequence and the MEROPS identifier for the peptidase are stored. In addition other information may be retained, including whether the cleavage is deemed by the authors of the source paper to be physiological or not, whether the substrate was in native conformation, the pH of the reaction, and the method used to identify the cleavage. The four residues either side of the scissile bond are also stored so that the cleavage position can be recalculated should the UniProt protein sequence change, and to provide the data for what amino acids are acceptable in the binding pockets S4-S4 0 for each peptidase. A bespoke program (in Perl) was written to add each cleavage in a protein substrate to ensure consistency; the program connects to the locally installed version of UniProt so that each cleavage position can be confirmed as the data are entered. Some data were acquired from proteomics studies. Again a bespoke program was written to parse the data from the Excel spreadsheets available as Supplementary Data to the published papers. Some cleavages were acquired from the CutDB database, but these have been manually checked against the original reference and the UniProt sequence. Once again a bespoke program was written to collect the data, translate the provided substrate Protein Identifier to a Uniprot accession, check that a cleavage event was not already present in the MEROPS collection (and add the CutDB accession number if it were), and add new cleavage events to the MEROPS collection, reporting any inconsistency between the P4-P4 0 residues and the sequence in the UniProt entry. The data collected are non-redundant. If more than one paper reports the cleavage at the same position in the same protein by the same peptidase, then only data from the paper published first is retained. If several peptidases cleave the same protein in the same position, each is considered a different cleavage event and each is entered. There is no attempt to map cleavages to isoforms of a protein, unless different isoforms were used by the original researchers. Synthetic substrates that differ only in leader (for example benzyloxycarbonyl, succinyl or tosyl) or reporter groups (for example aminomethylcoumarin, naphthylamide or nitroanilide) are considered different cleavages even if all are performed by the same peptidase. The MEROPS website has two displays to show peptidase specificity. Both use the data from natural and synthetic substrates, but show only naturally occurring amino acids. The first display is a logo which uses the WebLogo software (27) . Residues P4-P4 0 from all the substrates for a peptidase are treated as an alignment. The observed frequency for each amino acid in each position is calculated as a bit score, the maximum possible being 4.32 bits. An amino acid is shown in the logo (in single-letter code) if the bit score exceeds 0.1. The logo is also shown as a text string, where if a single amino acid predominates at one position (i.e. the bit score exceeds 0.4) the letter is shown in uppercase, and if more than one amino acid predominates in any position a letter is shown in uppercase when the bit score exceeds 0.7 and in lower case if the bit score is between 0.1 and 0.7. The second display is a frequency matrix, which is an 8 Â 20 matrix with residues P4-P4 0 along the x-axis and all amino acids along the y-axis. The amino acids are ordered so that those with similar properties are adjacent. The order is Gly, Pro, Ala, Val, Leu, Ile, Met, Phe, Tyr, Trp, Ser, Thr, Cys, Asn, Gln, Asp, Glu, Lys, Arg and His. Preference is calculated in terms of the percentage of substrates with each amino acid in each position, and a different shade of green is used for each tenth percentile interval. The number of times a residue occurs at each position is shown. The UniProt accession for each substrate with a known cleavage site was used to search the UniRef50 database (clusters of sequences that have at least 50% sequence identity to the longest sequence) (28) . If a UniRef50 entry was found, then all the UniProt accessions included in that entry were extracted and the sequences retrieved from the UniProt database in FastA format. Short fragments were excluded and the remaining sequences were then aligned with MUSCLE (29), using the default parameters and performing two iterations over the complete alignment to minimize gaps. Because each UniRef50 entry contains sequences sharing 50% or more sequence identity, the program is very quick, and the resulting alignment approximates to an alignment of orthologues. However, some UniRef50 entries will also contain closely related paralogues. Sequence alignments were generated and highlighted to show not just conserved residues but also peptidase preference. For each cleavage the residues P4-P4 0 were highlighted to indicate whether the residue in each sequence had been observed in any substrate at that position for the peptidase in question. Residues identical to the sequence where the cleavage is known are shown with a pink background. Replacements observed in other substrates are shown with an orange background. Where no substrate for this peptidase is known with this amino acid in this position the residue is shown with a black background. The term 'atypical' is used for an amino acid that has not been observed in a particular binding pocket in any known substrate for a peptidase. The MEROPS cleavage collection The MEROPS cleavage collection contains 39 118 cleavages (as of 7 August 2009). The number of cleavages that can be mapped to entries in the UniProt database is 34 606, the remaining 4512 consisting mainly of synthetic substrates. The number of different entries in the UniProt database to which cleavages are mapped is 10513. The number of cleavages that are designated as 'physiological' is 13 307; whereas 20 187 cleavages are designated 'nonphysiological' and 2677 cleavages are designated 'synthetic'. The remaining 2947 are cleavages in peptides that can not be mapped to UniProt because: they are too short; they are significantly modified, such as the non-alpha peptide bond between ubiquitin and its target protein; they are derived from phage displays; they are theoretical cleavages or because it is unclear whether the cleavage is physiological or not. The data are derived from 6095 publications. The number of cleavages in common between the MEROPS and CutDB collections is 5876, of which 3424 were originally found in the literature by the CutDB researchers. The number of cleavages from the CutDB database that failed to make the MEROPS collection, excluding 892 isoforms and 35 duplicates, was 560 (9.5%), mostly due to being mapped to the wrong residue or sequence. The CutDB curators have been informed of these discrepancies. Because the CutDB database includes only cleavages thought to be of physiological relevance and those that can be included in their proteolytic pathways, it has fewer cleavages than in the MEROPS collection. It does not include cleavages in synthetic substrates and those peptides used solely to map peptidase specificities, or general purpose processing enzymes such as signal peptidases and methionyl aminopeptidases. There are 2415 different peptidases recognized in the MEROPS database (excluding hypothetical peptidases from model organisms). Substrate cleavages have been collected for 1086 peptidases (45%); for the remainder any cleavage positions in substrates are either unknown or have not yet been found in the literature. Only 312 peptidases have had ten or more cleavages collected, and it is only these for which there is enough data for further analysis. The total number of cleavages for these 312 The residues from P4-P4 0 were collected for each substrate cleavage. Figure 1 shows the number of peptidases showing some selectivity for one or two residues in each binding pocket from S4 to S4 0 . Clearly, many peptidases have extended substrate-binding sites with preferences beyond S1, with 52 showing a preference in the S4 pocket. There are a few peptidases that have a preference at S5 (30), but a preference so far from the scissile bond is rare. It is conceivable that mitochondrial intermediate peptidase (M03.006), which removes a transit octapeptide from the N-terminus of proteins synthesized in the cytoplasm but destined for the mitochondrial matrix, may have a preference as far away from the cleavage site as S8 (31) . Preference on the prime side of the scissile bond is thought to rarely extend beyond the S1 0 pocket, but Figure 1 shows that 32 different peptidases have a preference in S4 0 . Exopeptidases cleave near protein termini, and because the binding pockets do not exist are unable to accept any amino acids in some positions. Dipeptidases can only accept residues in the S1 and S1 0 pockets; aminopeptidases are unable to accept any residue in S4-S2, carboxypeptidases in S2 0 -S4 0 , dipeptidyl-peptidases in S4 and S3, tripeptidylpeptidases in S4 and peptidyl-dipeptidases in S3 0 and S4 0 . Some omega peptidases (peptidases which do not cleave normal peptide bonds but release substituted amino acids such as pyroglutamate or cleave isopeptide bonds, such as many deubiquitinating enzymes) may also be unable to accept any residue in certain positions, or it is not possible to interpret cleavages in terms of P4-P4 0 , for example for isopeptidases. There are 36 peptidases with 10 or more cleavages that cannot accept any residue in S4, 35 for S3, 26 for S2, 15 for S2 0 , 22 for S3 0 and 25 for S4 0 . Table 1 shows the number of peptidases showing some selectivity in each binding pocket from S4 to S4 0 for amino acid properties (where 'acidic' is Asp or Glu; 'basic' is Arg, His or Lys; 'aliphatic' is Ile, Leu or Val; 'aromatic' is Phe, Trp or Tyr; and 'small' is Ala, Cys, Gly or Ser). Properties are taken from Livingstone and Barton (32) . Only the categories with the fewest amino acids, and those that do not overlap (with the exception of His, which can also be considered aromatic) have been used. If categories such as 'hydrophobic' and 'polar' are used then nearly every binding pocket is highlighted because each category contains more than half of the amino acids. Most preference is directed towards the S1 (201 different peptidases) and S1 0 (160 different peptidases) pockets. The commonest preferences are for a basic amino acid in the P1 position, small amino acids in P1 and P1 0 , and an aliphatic amino acid in P1 0 . No aromatic amino acids were observed in P4 0 in any of the substrates of these 312 peptidases. For each amino acid category preference was most pronounced in the S1 pocket with the exception of aliphatic amino acids, where most peptidases have a preference in the S1 0 pocket. Preference for acidic amino acids is very rare except in the S1 pocket, and similarly aromatic amino acids are rarely preferred except in S1 and S1 0 . The preference for individual amino acids is shown in Table 2 . It is clear from the table that cysteine is an S3 S2 S1 S1' S2' S3' A count is made whenever an amino acid occurs in one binding pocket in 40% or more of the substrates. There are 15 peptidases that have a preference for two amino acids in a binding pocket: walleye dermal sarcoma virus retropepsin (A02.063, Asn or Gln in S2), sapovirus 3C-like peptidase (C24.003, Glu or Gln in S1), SARS coronavirus picornain 3C-like peptidase (C30.005, Gly or Gln in S1), peptidyl-peptidase Acer (M02.002, Gly or Pro in S1), vimelysin (M04.010, Phe or Leu in S1), carboxypeptidase M (M14.006, Arg or Lys in S1 0 ), carboxypeptidase U (M14.009, Arg or Lys in S1 0 ), dactylysin (M9G.026, Leu or Phe in S1 0 ), chymase (S01.140, Phe or Tyr in S1), tryptase alpha (S01.143, Lys or Arg in S1), trypsin 1 (S01.151, Lys or Arg in S1), plasmin (S01.233, Lys or Arg in S1), flavivirin (S07.001, Lys or Arg in S2), dipeptidyl aminopeptidase A (S09.005, Ala or Pro in S1) and kumamolisin (S53, 004, Glu or Gly in S3). Many peptidases show a preference in more than one binding pocket. There are 13 peptidases with a preference for all eight binding pockets, another 13 with a preference in seven, five peptidases in six, three in five, eight in four, 24 in three, 47 in two and 89 in only one. The number of peptidases showing a preference for an amino acid in a binding site is shown. Only those 312 peptidases with 10 or more known substrate cleavages are included. An amino acid must occur at that position in 40% or more of substrates. Therefore, it is possible for two amino acids to be preferred in any one binding pocket, as is the case for trypsin 1 where there is a preference for either Lys (59% of substrates) or Arg (41%) in S1. There are 202 peptidases that show a preference, of which 13 show a preference at all eight sites, 13 for seven sites, five for six sites, three for five sites, eight for four sites, 23 for three sites, 49 for two sites and 88 for one site. unwelcome amino acid near a cleavage site. Only the peroxisomal transit peptide peptidase shows a preference for cysteine binding to the S2 and S1 pockets. Tryptophan is also rare around cleavage sites, with only tryptophanyl aminopeptidase (M9A.008, a preference in the S1 pocket) and mast cell peptidase 4 (Rattus) (S01.005, in the S2 pocket) showing a preference; however, this may have more to do with the fact that tryptophan is the rarest of the amino acids. Asparagine is also very rare in the proximity of a cleavage site, one of the few examples being the specialist peptidase legumain (C13.006) which only cleaves asparaginyl bonds (33) . Histidine is also a rare preference, with only three peptidases showing any preference for it, namely chymosin (A01.006; S4), carnosine dipeptidase I (M20.006; S1 0 ) and Xaa-methyl-His dipeptidase (M20.013; S1 0 ). Methionine is also not preferred by most peptidases, exceptions being methionyl aminopeptidases (M24.001, M24.002), where the preference is as expected for methionine binding in the S1 pocket, some members of the peptidase Clp family (S14) and the unsequenced Met-Xaa dipeptidase (M9B.004). The gpr peptidase (A25.001) shows a preference for Met binding to S4. The commonest preference is for arginine binding to the S1 pocket, which occurs in over fifty peptidases. However, arginine is relatively rare outside the P1 position. There are peptidases that show a preference for Gly, Pro and Val for every binding pocket in the range S4-S4 0 . Peptidases showing unique preferences are listed in Table 3 . Despite there being a large number of substrates collected, the specificity of some peptidases can not be explained in terms of S4-S4 0 preferences. These peptidases Specificity logos and frequency matrices present the user with a visual representation of peptidase specificity. An example specificity logo is shown in Figure 2 . From the logo and the cleavage pattern string it is clear that caspase-3 has an absolute requirement for Asp in the S1 pocket (position 4, only one cleavage after Glu is known) and a preference for Asp in S4. There are minor preferences for Glu in S3 and Gly or Ser in S1 0 . While the logo indicates which amino acids are acceptable in each position, it does not indicate which amino acids are unobserved. These are shown in the frequency matrix, and an example is also shown in Figure 2 for caspase-3. In this example Asp occurs in the P1 position in all 413 substrates, Asp occurs in P4 in almost half the substrates, while Glu occurs in P3 in 27% of substrates. Note that in this frequency matrix every amino acid occurs in positions P4-P2 and P1 0 -P4 0 , but tryptophan is observed only once in P4, P2, P1 0 and P4 0 . This gives an indication of the minimum number of substrate cleavages that has to be collected for a peptidase before definite conclusions about specificity in all binding pockets can be drawn. A substrate alignment is shown in Figure 3 . The density of residues highlighted in black is high, implying that this cleavage position is very poorly conserved and thus may not be physiologically relevant. The logo is shown at the top with the frequency matrix below. The cleavage pattern is a textual representation of the logo, where the scissile bond is shown as a red cross, and the binding pockets separated by forward slashes. The preferred residue is shown in uppercase if the preference is strong. The number of cleavages on which these data are based is given in parentheses. For the logo, the binding pockets S4-S4 0 are shown along the x-axis, where 1 is S4, 2 is S3, etc. The bit score is shown on the y-axis. The height of the letter is proportional to the bit score. The letters are coloured to indicate amino acid properties: blue for basic, red for acidic, black for hydrophobic and green for any other. In the frequency matrix below the logo, each cell shows the number of substrates with an amino acid occupying one of the positions P4-P4 0 . Cells in the matrix are highlighted in shades of green where the greater the preference, i.e. the more often an amino acid occurs at that position, the brighter the shade. Cells are highlighted in black if the amino acid is unknown at that position for any substrate. . ........................................................................................................................................................................................................................................................................................ ........................................................................................................................................................................................................................................................................................ .. Protein sequence alignments were constructed for every substrate where the cleavage had been assumed in the literature to be of physiological significance. The total number of alignments generated was 3141. A selection of cleavage sites which were not conserved in all homologues included in the same UniRef50 database entry are listed in Table 4 . Only those cleavages by peptidases with at least 20 known substrates are included. There are a number of possible causes for a cleavage site not to be conserved which are listed below. (1) The UniRef50 entry might include paralogous sequences which although at least 50% identical to the sequence with the known cleavage, might be processed or degraded differently and there is no evolutionary pressure to maintain the known cleavage site. Where a cleavage site was not conserved, a paralogue was identified in an alignment as a second protein from the same species that was clearly not a splice variant. (2) UniRef50 entries contain many translated genes from genome sequencing projects; gene finding in eukaryote genomes is notoriously difficult and it is possible that erroneous gene building has resulted, for example, in the loss of the exon encoding the cleavage site or the inclusion of part of an intron in its place. (3) It is also probable that for some peptidases there are not enough substrates known to be sure that any amino acid is really excluded from a particular binding site. The number of substrates known for each peptidase is included in Table 4 , because the greater the number of substrates the more likely that an amino acid is really atypical and not just unobserved. (4) The alignment is incorrect. This is unlikely given the close relationship between the sequences, which are all 50% or more identical; however there are situations where an insert or deletion occurs within the range P4-P4 0 . (5) Some endogenous cleavages (for example removal of signal and transit peptides) may be the result of more than one cleavage, because aminopeptidases nibble away the N-terminus (1), and may thus be incorrectly mapped to the specificity of the leader peptidase. (6) It is theoretically possible that if the substrate and peptidase are from the same organism both will have evolved to accommodate a change in the cleavage position. (7) A single residue mismatch may also be due to a single-base sequencing error. Potential errors of this kind can be identified using a codon dictionary, provided the atypical residue could be the result of a single base change, and that it is the only residue not conserved, regardless of the number of sequences in the alignment. (8) Some cleavages regarded as 'physiological' are actually fortuitous. If a cleavage site is extremely poorly conserved it is unlikely to be physiologically relevant. ---P----(4) Cytochrome C Where it is possible to suggest a cause why a cleavage site is not conserved this is indicated in Table 4 by the letters a-h. Included in category d, where insertions and or deletions occur in the homologous cleavage sites, is 50S ribosomal protein L7Ae (UniProt accession P12743). There are N-terminal extensions to most homologues so that the known methionyl aminopeptidase 2-cleavage site is not aligned. Five of these sequences may be derived from erroneous gene builds (point b). The UniRef50 database entry for 60S ribosomal protein L10 (P27635) includes a wide range of species (the cleavage is known in the human protein) and the peptidase performing the cleavage (granzyme B) is not present in Paracoccidiodes brasiliensis, where the substrate cleavage is also not conserved. The replacements that are reported as atypical in hemoglobin subunit alpha (P69905) by Schistosoma cathepsin D (A01.009) (34) are the rarest naturally occurring amino acids, tryptophan and cysteine, and despite there being 109 known cleavages for this peptidase, this may still not be enough to properly exclude these rare amino acids. On the other hand, this is the cleavage of a host protein by a parasite peptidase and the specificity may have adapted to limit the availability of hosts. None of the cleavages listed in Table 4 has been assigned to cause f above, namely where changes in the substrate cleavage site may be mirrored by changes in peptidase specificity. Without modelling the substrate binding sites, if that were possible, detecting this situation is difficult. However, the autolytic processing of cathepsin E (P43159) may be such an example (35) . In some cases, a poorly conserved cleavage site may represent a pathological condition in the species where the cleavage was first identified. For example, despite there being few cleavages for cathepsin H, the reported cleavages in the BID protein (36) are in particularly poorly conserved regions (see Figure 3 ). Cleavage of the BID protein leads to the induction of apoptosis. That the cleavage sites are not well conserved amongst mammalian orthologues is not surprising given that the cytoplasmic substrate and the lysosomal peptidase should not meet under normal circumstances. The mouse protein in which the cleavage was identified may therefore be unusually susceptible to cleavage should the lysosomal membrane be ruptured. The specificity logos and frequency matrices for all peptidases with 10 or more known substrate cleavages are already available in the MEROPS database. Alignments are also available for all protein substrates that have a corresponding UniRef50 entry, showing conservation of both physiological and non-physiological cleavages. The next release of the database will include tables showing comparative peptidase specificity in terms of preference for both amino acid and amino acid type. The MEROPS database includes over 39 000 cleavages in substrates (synthetic and naturally occurring) which have been collected from the literature. These are classified as physiological or non-physiological, depending on whether the substrate is naturally occurring and if it is in native conformation. At least one substrate is known for 45% of the different peptidases identified in the MEROPS database. Displays in the database give insights into peptidase specificity and to the conservation of cleavage sites amongst orthologous proteins. The data provide a substantial training set for algorithms to predict peptidase substrates and cleavage positions in those substrates. The data may also be useful for the design of inhibitors and engineering novel specificities into peptidases. By examining the conservation of cleavage sites in protein substrates in terms of peptidase substrate binding sites, it is clear that there are a number of cleavages where atypical replacements occur. Many of these can be explained by gene build or sequencing errors, inserts or deletions in the region around the cleavage site, or the alignments contain one or more paralogues in which cleavage may be absent or different. In a few cases it is possible that more than one peptidase is involved in processing, or there may not be enough known substrates for some peptidases to be sure that an atypical replacement is really unacceptable. A number of substrate cleavages that may be fortuitous and not of any physiological relevance have been identified. This cleavage set is freely available and can be downloaded from the MEROPS FTP site (ftp://ftp.sanger.ac.uk/ pub/MEROPS/current_release/database_files/ Substrate_search.txt).

Efficient Assembly and Secretion of Recombinant Subviral Particles of the Four Dengue Serotypes Using Native prM and E Proteins

BACKGROUND: Flavivirus infected cells produce infectious virions and subviral particles, both of which are formed by the assembly of prM and E envelope proteins and are believed to undergo the same maturation process. Dengue recombinant subviral particles have been produced in cell cultures with either modified or chimeric proteins but not using the native forms of prM and E. METHODOLOGY/PRINCIPAL FINDINGS: We have used a codon optimization strategy to obtain an efficient expression of native viral proteins and production of recombinant subviral particles (RSPs) for all four dengue virus (DV) serotypes. A stable HeLa cell line expressing DV1 prME was established (HeLa-prME) and RSPs were analyzed by immunofluorescence and transmission electron microscopy. We found that E protein is mainly present in the endoplasmic reticulum (ER) where assembly of RSPs could be observed. Biochemical characterization of DV1 RSPs secretion revealed both prM protein cleavage and homodimerization of E proteins before their release into the supernatant, indicating that RSPs undergo a similar maturation process as dengue virus. Pulse chase experiment showed that 8 hours are required for the secretion of DV1 RSPs. We have used HeLa-prME to develop a semi-quantitative assay and screened a human siRNA library targeting genes involved in membrane trafficking. Knockdown of 23 genes resulted in a significant reduction in DV RSP secretion, whereas for 22 others we observed an increase of RSP levels in cell supernatant. CONCLUSIONS/SIGNIFICANCE: Our data describe the efficient production of RSPs containing native prM and E envelope proteins for all dengue serotypes. Dengue RSPs and corresponding producing cell lines are safe and novel tools that can be used in the study of viral egress as well as in the development of vaccine and drugs against dengue virus.

Dengue is one of the most important vector-borne viral diseases in humans. However, the interaction between dengue virus (DV) and host cells is only partly understood. Therefore, there is an urgent need to develop new tools to gain insight into the viral journey through host cells. As a member of the Flavivirus genus in the Flaviridae family, DV is a small, positive strand RNA enveloped virus. There are four serotypes of dengue virus (DV1-4). Their genome encodes a polyprotein precursor of at least seven non-structural proteins and three structural proteins which are the capsid protein (C), the membrane protein (M) and the envelope glycoprotein (E) [1] . The polyprotein is processed co-and post-translationally by cellular signalase in the lumen of the rough endoplasmic reticulum (ER) and by a viral protease in the cytosol [1, 2, 3] . The nascent C protein contains a C-terminal hydrophobic domain that acts as a signal sequence for translocation of the immature form of M, the prM, into the lumen of the rough ER. Two adjacent prM C-terminal transmembrane domains are responsible for prM membrane anchoring and E translocation into the ER [2] . prM and E associate into heterodimers at ER membranes [4, 5] where they assemble with the viral RNA/C complex to form progeny virions [1] . During the egress of virions through the secretory pathway, prM protein is cleaved by the trans-Golgi resident furin protease to form the M envelope protein and the soluble pr segment, which is released into the extracellular medium upon particle secretion [6] . prM cleavage marks maturation of flavivirus virions [7, 8] . Cleavage of prM is intimately correlated to change of conformation of envelope protein complexes. Although it was thought that prM cleavage is a prerequisite for E dimerization, recent studies show that change of conformation most probably occurs at low pH in the TGN and allows cleavage of prM by furin [6, 9, 10] . prM and E proteins from flaviviruses, such as yellow fever virus [11] , Japanese encephalitis virus (JEV) [12, 13] , West Nile virus (WNV) [14] and tick-bone encephalitis virus (TBEV) [15, 16] , are able to assemble into subviral particles in the absence of any other viral component. Subviral particles and infectious virions are coproduced in infected cells, assemble in an immature form, and subsequently undergo the same maturation process and display similar fusion activity as infectious viruses [17, 18] . Therefore, subviral particles could be a precious tool for research on cell biology of DVs. Although there have been attempts by several groups to obtain DV RSPs, either their production was inefficient or the sequence of DVs structural proteins had to be substantially modified [19] before they could efficiently generate RSPs. For example, the furin cleavage site on prM had to be mutated to establish the DV2 RSPs producing CHO cell line because it was found that the DV2 RSPs cause cell-cell fusion [20, 21] . Others replaced a portion of the carboxy-terminal region of DV1 and DV2 E genes with the corresponding sequence of JEV to observe significant RSP secretion [22, 23] . In either case, these modifications may interfere with the interactions with cytosolic proteins and, possibly, with the maturation and folding of the structural proteins. Here we describe the efficient production of native RSPs of the four DV serotypes by using a codon optimization strategy. Codons of the DV prME gene were replaced with those preferentially used in mammalian cells. Optimization was first applied to the DV1 prME gene. We found that this experimental approach could efficiently increase the intracellular expression level of the E glycoprotein in human cells and therefore enhance the production of DV1 RSPs. We have established a DV1 RSPs producing stable cell line (HeLa-prME). Our data by immunofluorescence and electron microscopy show that most of E protein co-localizes with markers of the ER, where RSPs accumulate. Biochemical analysis of secreted RSPs demonstrates that they contain E homodimers and M, indicating that RSPs have trafficked through the secretory pathway and that maturation process has occurred. Using the HeLa-prME cell line, we have developed a semi-quantitative assay to screen a 122 genes cellular membrane trafficking siRNA library and identified genes that may be involved in the secretion of DV. From our screen, 23 genes show a significant reduction in DV1 RSPs secretion whereas for 22 genes RSPs levels were increased in cell supernatants. These data provide a proof of concept that RSPs of DVs and the producing cell line are safe tools that can be used in the study of viral egress. Finally, the optimization strategy was applied to DV2, DV3 and DV4 prME, and RSPs for all four DV serotypes were efficiently produced. These native RSPs are, therefore, a safe mimicry of virions that can be used to study viral and cellular requirements for virus assembly and egress. To generate DV RSPs, we first transfected HeLa cells with native DV1 prME gene derived from the DV1 FGA/NA d1d [24] strain, which contained the signal sequence of vesicular stomatitis virus G (VSV-G) envelope glycoprotein inserted in frame upstream of the prME cDNA. 48 hours post-transfection, expression of dengue E protein was monitored by flow cytometry on permeabilized transfected HeLa cells with the 4E11 monoclo-nal antibody against dengue E protein. E protein was detected in transfected cells at low expression level ( Figure 1A) . Analysis of the prME gene sequence revealed that 15% of nucleotide triplets were rarely used in mammalian cells so that its codon usage was not adapted for efficient expression in this cell system. Thus we designed and synthesized an optimized DV1 prME (prMEopt) gene in which the rarely used codons were replaced with those preferred by mammalian cells, without changing the amino acid sequence ( Figure S1 ). More than 70% codons were modified for prME optimization. In addition, we also removed the negative cisacting sites (such as splice sites, poly(A) signals, etc) which might have negatively influenced expression. The optimized gene resulted in high and stable expression levels in transfected mammalian cells ( Figure 1B ). Both the mean fluorescence intensity (MFI) and percentage of positive cells increased by almost 3 fold. E protein was hardly detected without cell permeabilization, indicating that its localization was intracellular and was not expressed at cell surface (data not shown). A HeLa cell line that stably expresses DV1 prM and E envelope proteins was then established by transducing HeLa cells with a retroviral vector pCHMWS-prMEopt-IRES-Hygromycin. Hygromycin-selected cells were designated HeLa-prME. Several single colonies of HeLa-prME were used in this study and they showed similar results. The HeLa-prME cell line has already been propagated for more than thirty passages in the presence of hygromycin and has kept a stable prME expression level. Culture of HeLa-prME cells in the absence of hygromycin resulted in a significant reduction of prME expression (data not shown). The intracellular distribution of DV1 E glycoprotein was analyzed in fixed HeLa-prME cells by co-labeling with Erp72, ERGIC-53 and Golgin-97, which are proteins of the endoplasmic reticulum (ER), the ER-Golgi intermediate compartment (ERGIC) and the Golgi apparatus, respectively. As expected, the DV1 E glycoprotein mainly distributed in the ER in HeLa-prME cells, as shown by immunofluorescence, where it colocalized with Erp72 ( Figure 2A -D), whereas it showed no colocalization with ERGIC-53 ( Figure 2E -H) and Golgin 97 ( Figure 2I -L). Interestingly, E and Erp72 were also enriched in a perinuclear compartment, which is not usually labeled by anti-Erp72 antibody and other ER markers (data not shown), suggesting that it has been induced by prM and E viral protein expression. In addition, no staining was observed at the plasma membrane, confirming the previous flow cytometry result. Altogether, our data demonstrate that the DV1 E glycoprotein is efficiently expressed and enriched in the endoplasmic reticulum in the HeLa-prME stable cell line. The endoplasmic reticulum is the assembly site for flaviviruses. In order to verify whether this ER staining reflects the presence of E-containing assembled particles, we analyzed the HeLa-prME stable cell line by transmission electron microscopy (TEM) on thin sections of fixed, epon-embedded cells ( Figure 3A ). We found that HeLa-prME, but not control HeLa cells (data not shown), displayed dilated ER membranes with aligned round particles and elongated parallel tubules. Particles were ,20 nm in diameter and homogenous in size and shape. The aligned particles that we have observed in HeLa-prME cells are similar to the stacked viral particles which have been described in cisternae of the rough ER in infected insect and mammalian cells [25, 26, 27, 28] . To confirm that these structures contain viral proteins, we labeled thawed cryosections with the 4G2 antibody that recognizes the E protein ( Figure 3B -D). We found antibody binding for the E protein Figure 2 . Subcellular localization of the DV1 E protein in HeLa-prME cells. HeLa-prME cells were fixed, permeabilized, and stained for E protein (green) and for cellular marker antigens (red). DAPI staining was used to label cell nuclei (blue). Erp72, ERGIC-53 and Golgin-97 are proteins of the endoplasmic reticulum (ER), the ER-Golgi intermediate compartment (ERGIC) and Golgi apparatus, respectively. The scale bar represents 10 mm. doi:10.1371/journal.pone.0008325.g002 present in the lumen of cisternae, where it is localized to electron lucent, round particles and tubular structures. In addition, E protein is also found on amorphous, electron dense material present in these cisternae ( Figure 3B ). To confirm the nature of these cisternae by other than morphological criteria we performed a double labeling of the 4G2 antibody with the ER resident protein calreticulin. We found antibody binding for calreticulin present on the 4G2 positive cisternae ( Figure 3C ), unequivocally identifying them as ER. In addition, we found a lower but significant amount of label for the E protein present in the Golgi stack and in vesicles in the Golgi area ( Figure 3D ). Thus, our observations indicate that expression of prM and E DV1 envelope proteins induces the formation of RSPs in an ERderived compartment. We then investigated if the RSPs observed by EM in HeLa-prME cells could undergo maturation and be secreted like DV. Cell lysate (CL) and clarified supernatant (SN) concentrated by ultra-centrifugation were analyzed by Western blotting followed by incubation with either the anti-E monoclonal antibody 4E11 ( Figure 4A -B) or an anti-DV1 serum from a human patient ( Figure 4C -D). In CL samples, a 50 kDa monomeric form of the E glycoprotein was predominantly observed ( Figure 4A ,C), whereas high levels of 100 kDa E glycoprotein homodimers were readily detected in SN samples ( Figure 4A -D). This suggests that whereas in HeLa-prME cells the majority of viral proteins are localized in the early pre-Golgi secretory pathway in an immature state, secreted RSPs have passed through maturation stages in the Golgi apparatus thus resulting in homodimerization of E proteins. To further confirm the presence of E dimers in SN and exclude the possibility that dimerization resulted from treatment for electrophoresis, the SN were incubated in the presence or absence of a cross linker (3, 39-dithiobis [sulfosuccinimidylpropionate]; DTSSP) and then subjected to Western blotting using the anti-E 4E11 antibody. Under these conditions, we observed that E dimer/ monomer ratio was lower in non-treated RSPs compared to the Figure 3 . Electron microscopy analysis of the HeLa-prME cell line. HeLa-prME (DV1) cells were fixed and either prepared for epon embedding (A), or for immuno labeling on thawed cryosections (B-D). A), Round particles are found aligned in the lumen of the ER (arrows) together with tubular structures (arrowheads). B) Labeling for the E protein is present in the lumen of cisternae. It is localized to electron lucent, round particles (arrows) and tubular structures (arrowheads). In addition E protein is also found on amorphous, electron dense material present in these cisternae (asterix). C) Double labeling of E protein (10 nm gold, black arrows) and the ER marker calreticulin (15 nm gold, white arrows). Calreticulin is present on the limiting membrane of the cisternae, which contain round particles positive for E protein labeling. D) Label for the E protein is also found within the Golgi stack (G) and on vesicles in the Golgi area. N = nucleus, all scale bars represent 200 nm. doi:10.1371/journal.pone.0008325.g003 DTSSP treated (data not shown). These data indicate that, in SN, a significant proportion of E protein is present as a homodimer and detection of E monomers most likely results from partial dissociation under the denaturing conditions of electrophoresis. E homodimers were no longer detected when samples were heated in presence of dithiothreitol ( Figure 4B ). In addition, the prM protein could be readily detected with the anti-DV1 serum in CL but not SN samples ( Figure 4C ). By contrast, when the HeLa-prME cells were treated with NH 4 Cl, which inhibits acidification of the trans-Golgi compartment and, hence, the activity of furin protease and prM cleavage, prM was also found in SN ( Figure 4D -E). To our surprise E dimers were also detected in these conditions. Interestingly, the E monomers from SN exhibited slower electrophoretic mobility than those of CL samples, which suggests that E glycoproteins have acquired complex N-glycans in the Golgi apparatus. This was confirmed by N-Glycosidase F (PNGase F) and endoglycosidase H (EndoH) digestion ( Figure 4 F-G), which cleaves nearly all types of N-glycan chains or only high mannose and some hybrid N-glycan chains from glycoproteins, respectively. Whereas E protein in CL was sensitive to both treatments, only PNGase F digestion increased the electrophoretic mobility of E protein in SN, demonstrating that RSPs had acquired EndoH-resistant, complex N-glycans in the Golgi apparatus before secretion. Moreover, we found that secretion of DV RSPs was altered by incubation of cells at 15uC and 20uC, which block cargo transport from ERGIC to cis-Golgi and exit from the trans-Golgi, respectively, and by brefeldin A treatment, which inhibits exit from ER and induces fusion of Golgi membranes with ER (data not shown). Altogether, homodimerization of E, acquisition of complex sugars and efficient cleavage of prM indicate that prM and E viral proteins have correctly assembled in the ER into RSPs which have trafficked through the secretory pathway before secretion into the cell medium. To study the dynamic of RSP production, we performed a pulse chase experiment on overnight starved HeLa-prME cells that were metabolically labeled with S 35 -methionine for 1 hour and subsequently chased for 24 hours in normal medium. Supernatant and cell lysate were collected at the specified time points after chase; RSPs were immuno-precipitated using the 4E11 anti-E monoclonal antibody and then analyzed by autoradiography. Detection of prM or M in this assay results from co-immunoprecipitation and, therefore, indicates efficient interaction with E protein. Secretion of E and M proteins was detected from the 8 hours time point on, with a significant increase at 24 hours postchase ( Figure 5A ). Consistently, high levels of E and M proteins were found in cell lysates at all time points with a substantial Figure 5B ). Two major protein products were found in cell lysates, with an apparent molecular weight corresponding to E and prM proteins. Bands of higher molecular weight were also present and could be oligomeric forms of E and prM or M. Interestingly, the pattern of immunoprecipitated proteins from cell supernatant was slightly different. Higher proportions of E dimers were found and a band corresponding to the molecular weight of M but not prM was detected at the 24 hours time point, indicating that most secreted RSPs are mature. Our data demonstrate that newly synthesized proteins need 8 hours to be translocated through the secretory pathway and released into the supernatant as mature RSPs. To further characterize secreted DV1 RSPs, we performed sucrose gradient fractionation on RSPs concentrated from supernatant of HeLa-prME cells. Supernatants were ultracentrifuged and RSPs-containing pellets either resuspended in PBS or in 0.5% Triton-X100 containing PBS to solubilize E from lipid membranes, as described [29] . As expected, Triton-X100solubilized E protein and RSP-associated E protein appeared in distinct fractions of a discontinuous 20 to 60% sucrose gradient ( Figure 6 ). The amount of E glycoprotein in each fraction was quantified. E glycoprotein in non-treated sample sedimented in fractions containing 20% to 30% sucrose ( Figure 6 , black bars) whereas in Triton-X100 treated samples, E protein was solubilized and detected in fractions at the top of the gradient (Figure 6 , white bars). These data further confirm that the RSPs formed in HeLa-prME are secreted into culture medium from which they can be easily purified by ultracentrifugation. Moreover, we have obtained preliminary evidence that DV1 RSPs and DV1 viral particles isolated from HeLa-prME and mammalian infected cells, respectively, sediment at similar densities on 10-60% continuous gradient of sucrose (Dr Philippe Despres, Institut Pasteur, personal communication). Our results show that the DV1 RSPs produced by HeLa-prME cell line mimic maturation and secretion of DV1, thus providing a useful tool to study the interaction between DV and host cells during viral egress. To identify host factors that could either enhance or reduce production of DV RSPs, we first developed a quantitative assay to relatively quantify levels of secreted particles in supernatant of HeLa-prME cells. The chemiluminescence dotblot (CLDB) assay is based on the concentration of RSPs from cell supernatant on PVDF membranes, followed by detection of E with a specific horseradish peroxidase (HRP)-conjugated antibody and quantification of substrate-induced luminescence using a luminometer. Our data using purified E protein of known concentrations showed that, when ranging between 400 pg to 40 ng, E protein on PVDF membrane displayed a very good linear correlation with the luminescence density in CLDB assay ( Figure 7A ). The CLDB was first used to estimate the RSPs yield from HeLa-prME cell line, and we found that the concentration of E protein in supernatant of HeLa-prME cell line was around 500-1000 ng/ml under the culture condition used for siRNA transfection (data not shown). We then screened a siRNA library that consisted of 122 genes which target cellular membrane trafficking using the HeLa-prME cell line. Non-targeting siRNA (NT) and siRNA targeting DV1 prME were added as controls. Library and control siRNAs were transfected in triplicates on 96-well plates. Levels of RSPs secreted by siRNA transfected HeLa-prME were measured by CLDB assay from 40 microliters of supernatant from each well. Levels of E protein in cell supernatant were expressed in relative luminescence units and ratios to that of NT controls were shown in Figure 7B . T test was used to assess the statistical significance of differences between each sample and NT. Differences were considered statistically significant when P,0.05. We observed that targeting of 23 genes resulted in significant reduction of DV1 RSPs amounts in supernatants whereas targeting of 22 other genes induced a significant increase. For instance, our results showed that downregulation of ADP-ribosylation factor 1 (ARF1), which regulates secretory membrane transport [30] , resulted in 3 fold decrease of DV1 RSPs in cell supernatant, suggesting its involvement in the secretion of DV. Some genes whose down-regulation enhanced levels of DV RSPs in the supernatant are involved in endocytosis, such as the three dynamins which show a 2 to 4 fold increase in dengue RSPs secretion. These proteins are involved in the budding process or in the transport of vesicles [31] . Such an enhancement might have been due, at least in part, to a blockade in the re-internalization of secreted RSPs by HeLa-prME cells. Although further experiments are required to confirm the screening data, our study validates the use of DV RSPs-producing HeLa-prME cell line in combination with the CLDB-based quantification strategy as a promising system to facilitate the identification of cellular factors involved in DV secretion. The codon usage of DV2, DV3 and DV4 prME genes differed substantially from that of mammalian cells and, therefore, was optimized as described for DV1 by replacing more than 70% of the codons ( Figure S1 ). RSPs of the four serotypes were first produced in 293T cells by transient expression of prME genes. Particles were purified and detected by Western blotting using either the anti-E 4E11 monoclonal antibody or a mixture of sera from four patients who were infected by all dengue serotype ( Figure 8A-B) . With the anti-E 4E11 antibody, monomeric and dimeric forms of the DV1 E protein could be readily detected whereas weaker signals were observed for the other three serotypes, because of limited affinity of the antibody ( Figure 8A ) [32] . E proteins of all DVs were detected using the mixture of sera ( Figure 8B ) with a signal of similar intensity, which suggested that RSPs of four serotypes could be generated to comparable levels. Monomeric E proteins of four serotypes displayed slightly different electrophoresis mobility, which could be due to a differential level of N-glycosylation. E glycoproteins of DV1 and DV3 have two N-glycosylation sites at Asn 67 and Asn 153, whereas those of DV2 and DV4 have only one at Asn 67 [33] . Dimeric E protein was observed in DV1, DV3 and DV4 but not in DV2. Besides E, the prM protein was also detected in RSPs produced by transient transfected 293T cells. Recently, we have established HeLa-prME and 293T-prME cell lines of DV1, DV2 and DV3 using the same procedure described for DV1 HeLa-prME. We have compared the maturation of RSPs produced by both cell types using SDS-PAGE and silver staining of the gel ( Figure 8C ). Supernatants from parental HeLa and 293T cells were used as controls in the experiment. We found that, whereas only a small fraction of prM was cleaved in the RSPs produced by 293T-prME cell lines, cleavage of prM was much more effective in RSPs from the HeLa-prME cell lines. This result indicates that efficacy of maturation is cell type dependent. RSPs of the four serotypes were further analyzed by sucrose gradient. As for DV1, RSPs of DV2, DV3 and DV4 were concentrated in fractions containing 20% to 30% sucrose ( Figure 8D) . Altogether, our results demonstrated that the strategy of codon optimization was successful in leading to the production of RSPs of all serotypes with comparable efficacy and similar sedimentation properties. In this study, we have used a codon-optimization strategy to establish and characterize stable cell lines that produce RSPs for the four dengue serotypes. Previous studies had reported the production of DV RSPs in which the glycoproteins were mutated either to avoid host cell fusion or to delete the ER retention signal. For example, DV1 RSPs were not secreted effectively and DV2 RSPs could not form unless the transmembrane domain of the E glycoprotein, which contains an ER retention motif was replaced with that of JEV [22, 23, 34] . Our ability to obtain the efficient production of RSPs of all serotypes without any change in the amino acid sequence is likely the result of codon optimization for expression in mammalian cells, which has been proven to be an Figure 6 . Sucrose gradient analysis of DV1 RSPs. The supernatant from HeLa-prME cells was concentrated and resuspended in PBS or 0.5% Triton-X 100 containing PBS. RSPs were then centrifuged in a 20 to 60% sucrose gradient at 28,000 rpm (Beckman SW-41Ti rotor) for 2.5 hours at 4uC. Fractions of 0.5 ml were collected and E content was measured using CLDB. The percentage of E protein in each fraction is displayed on the Y axis. doi:10.1371/journal.pone.0008325.g006 Figure 7 . Application of the DV RSP-producing HeLa-prME cell line and CLDB to screen a small library of siRNA which targets 122 genes involved in membrane trafficking. A) The correlation between luminescence density and amount of E protein on PVDF membrane in CLDB assay. B) The screen results of the siRNA library which targets 122 genes involved in membrane trafficking. Non-targeting siRNA (NT) and siRNA targeting DV1 prME were used as controls. Level of E protein in cell supernatant of each siRNA was expressed as its ratio to that of NT controls. Two-sided Student's t test was used to assess the statistical significance of differences between each sample and NT. Differences were considered statistically significant when P,0.05. Genes inducing either a significant decrease or increase in RSPs production are shown in gray and black columns, respectively. doi:10.1371/journal.pone.0008325.g007 , respectively (B) . C) Production of RSPs by 293T-prME and HeLa-prME stable cell lines. DV1-DV3 RSPs in supernatants were analyzed by SDS-PAGE and silver staining of polyacrylamide gels. Bands corresponding to the approximate molecular weight of E monomers and dimers, as well as prM and M are indicated by arrows. Supernatants from parental 293T and HeLa cells were used as controls. D) Analysis of RSPs of 4 serotypes by sucrose gradient. RSPs were centrifuged in a 20 to 60% sucrose gradient at 28,000 rpm for 2.5 hours in 4uC. Fractions of 0.5 ml were collected and measured using CLDB. The percentage of E protein in each fraction is displayed on the Y axis. doi:10.1371/journal.pone.0008325.g008 effective method to increase the expression level of glycoproteins from various viruses [35, 36] . Thus, gene optimization significantly enhanced expression of prME glycoproteins in transfected cells, and this over-expression increased both RSP production and secretion. However, it is also possible that using other viral strains may result in different efficacy of RSPs production, regardless of codon optimization. To our knowledge, this is the first report of RSPs using the native prM and E envelope proteins for the four serotypes of DVs in mammalian cells. We have characterized the maturation of DV1 RSPs produced by the HeLa-prME cell line. Sucrose gradient and sedimentation analysis demonstrated that the DV1 RSPs are concentrated in fractions containing 20-30% sucrose and are sensitive to detergent treatment. Although prM/E heterodimers were not found in our experiments, possibly because prM/E interaction is weak, we observed processing of prM into M protein in HeLa-prME cells. This cleavage was sensitive to NH 4 Cl treatment, which inhibits acidification of the trans-Golgi compartment and, hence, the activity of furin protease. Newly synthesized E proteins first form heterodimers with prM proteins in the ER of host cells and then rearrange as homodimers during the process of secretion [37, 38] . The rearrangement from heterodimer to homodimer was first thought to require cleavage of prM by furin in the trans-Golgi to form M and soluble pr proteins [39] . A recent study, however, has shown that such rearrangement is mainly caused by the progressive acidification of the milieu along the secretory pathway, which facilitates prM cleavage by furin [6, 10] . If rearrangement is a pre-requisite for prM cleavage to occur, it could explain, at least in part, why E homodimers were still observed in our study following treatment with the acidotropic reagent NH 4 Cl. The precise mechanism underlying their presence in RSPs under these experimental conditions, however, requires further investigation. Finally, we found by pulse-chase experiments that 8 hours are necessary for production and secretion of RSPs. Altogether, these results show that mature DV1 RSPs are efficiently produced by the stable HeLa-prME cell line. We have studied the subcellular localization of DV1 E and RSPs in the HeLa-prME cell line. By immuno-staining of fixed cells and fluorescence microscopy, we have shown that most of the E protein is localized in the ER compartment, where dengue glycoproteins are synthesized [1] . The fact that we did not detect any significant signal of E in ERGIC and Golgi apparatus could be explained by a very low amount of protein in these organelles. In these conditions, saturating signals in the ER would mask its detection in other compartments along the secretory pathway. The fluorescent microscopy results were in accordance with our electron microscopy data. Analysis of cell sections by transmission electron microscopy has revealed that E proteins in the ER are associated with both round particles and tubular structures. The tubules could either be intermediate forms of RSPs assembly or result from accumulation of prM and E proteins in the ER. Budding of virions from cellular membranes depends on assembly of viral structural proteins that generate pushing and/or pulling forces simultaneously to induce curvature of membranes, which is necessary for particle formation and membrane fission. It is possible that assembly of overexpressed prM and E proteins in the ER allows formation of long tubules but that the fission event is a limiting factor. The secretion pathway was also investigated by temperature block or pharmacological experiments. Incubation at either 15uC, to stop traffic between the intermediate compartment and the cis-Golgi, or at 20uC, to block exit from the trans-Golgi network, or BFA treatment, to block the exit from ER, significantly reduced secretion of RSPs. These results demonstrate that RSPs traffic through ER, ERGIC and Golgi compartments before being secreted out of the cell. The efficient production of DV1 RSPs by HeLa-prME cell line and the fact that RSPs could mimic the maturation and secretion processes allowed us setting up an assay to study the interaction between DV and host cells during egress, a step which has received little attention in comparison to DV viral entry and replication [40, 41, 42, 43] . We have validated our assay by using a siRNA library preferentially targeting genes involved in cellular transport. As expected, a number of genes involved in the secretory pathway resulted in reduced release of RSPs in the cell supernatant, whereas other factors were associated with an opposite effect. Clearly, the identification of cellular factors involved in the egress process will be helpful to understand the maturation of DV and its pathogenicity. Moreover, our system will allow comparing the effect of cellular factors using RSPs assembled from the four dengue serotypes to test whether there are strain-specific interactions with host proteins. HeLa and 293T cells maintained in our lab [44] were cultured in DMEM containing 10% fetal bovine serum. Purified anti-E 4E11 and 4G2 monoclonal antibodies were provided by Dr. A. Amara and P. Despres (Institut Pasteur, France), respectively. Purified anti-DV1 mouse IgG and sera from four patients infected by the four dengue serotypes (1) (2) (3) (4) , respectively, were kindly provided by Dr. Philippe Buchy (Institut Pasteur, Cambodia). The native (non-codon optimized) DV1 (strain FGA/NA d1d) prME gene containing construct was provided by Dr A. Amara. The prME sequences of DV1 (strain FGA/NA d1d), DV2 (strain FGA/02), DV3 (strain PAH881/88) and DV4 (strain 63632) were codon-optimized and synthesized by the Geneart Company (Regensburg, Germany) and subcloned into pcDNA or retroviral vector pCHMWS-IRES-Hygromycin (kindly provided by Dr. Rik Gijsbers, from Molecular Medicine at Katholieke Universiteit Leuven) using BamHI and XhoI restriction sites. A nucleotide sequence encoding for the signal peptide of VSV-G (MKCLLY-LAFLFIGVNC) was included upstream of each prME gene. Sequences of codon-optimized prME genes are provided as Supporting Information ( Figure S1 ). To produce the retroviral vector for delivery of the prME-opt gene into HeLa cells, the pCHMWS-prME-opt-IRES-Hygromycin, pcDNA-VSV-G and p8.71 (modified HIV provirus coding for gag and polymerase) plasmids were co-transfected into 293T cells. The cell supernatant containing infectious VSV-G-pseudotyped retroviral particles was harvested 48 hours post-transfection and used to infect HeLa cells. Two days after infection, cells were selected in culture medium containing 500 mg/ml of hygromycin for two weeks. Selected cells (HeLa-prME) were tested by flow cytometry for E expression and were maintained in DMEM +10% FBS +500 mg/ml of hygromycin. HeLa cells were transfected with pcDNA-prME, pcDNA-prME-opt or a pcDNA empty vector using calcium phosphate precipitate method. Two days after transfection, cells were detached by incubation in 10 mM EDTA at 37uC for 10 min, fixed in 2% paraformaldehyde, and then permeabilized in 0.1% Triton X-100. After washing, the cells were incubated with a mouse anti-E antibody (4E11 For fluorescence microscopy on fixed cells, HeLa-prME cells were grown on glass coverslips. Cells were fixed, permeabilized, and incubated with anti-DV human sera (1:200) and anti-Erp72 Rabbit polyclonal Ab (1:200, Stressgen Bioreagents, Ann Arbor, MI, USA), anti-ERGIC-53 mAb (1:1000, Alexis Biochemicals, Farmingdale, NY, USA), or anti-Golgin-97 mAb (1:50, Invitrogen, Carlsbad, CA, USA), followed by the incubation with corresponding secondary antibody conjugated with FITC or TRITC. Nuclei were stained with DAPI and coverslips were mounted on glass slides for analysis. Fixed cells were visualized under AxioObserver Z1 inverted motorized fluorescent microscope using the ApoTome module and piloted through the Axiovision 4.6 software and images were acquired through the MRm AxioCam high resolution CCD camera (Carl Zeiss, Germany). HeLa-prME cell line was fixed at 28 hours post-trypsination and processed for EM and immuno-EM. For conventional TEM, cells were fixed in 2.5% glutaraldehyde in 0.1 M cacodylate buffer (pH 7.2) for 1.5 hr at room temperature and post-fixed with 1% osmium tetroxide in 0.1 M cacodylate buffer (pH 7.2) for 1 hour at room temperature. Then cells were embedded in 2% agarose and cell blocks were post-fixed with 2% uranyl acetate in 30% ethanol, dehydrated in graded series of ethanol and embedded in epoxy resin. Ultrathin sections were cut with a Leica Ultramicrotome UCT (Leica Microsystems; Vienna, Austria) and collected on 400-mesh formvar coated copper grids. Sections were stained 45 minutes with 4% aqueous uranyl acetate and 5 minutes with lead citrate. For immuno-EM, cells were fixed with 4% formaldehyde in 0.1 M phosphate buffer (pH 7.4), and embedded in 12% gelatin. Blocks were infiltrated with 2.3 M sucrose for cryoprotection, mounted on specimen holders and frozen in liquid nitrogen. Cryosections were cut with a Leica EM UC6/FC6 Microtome (Leica Microsystems, Vienna, Austria). Thawed cryosections were labeled with anti-E 4G2 antibody, rabbit polyclonal antibody against ER resident protein calreticulin (Abcam, Cambridge, MA, USA) and protein-A gold (10 nm and 15 nm) obtained from Utrecht University (Utrecht, The Netherlands) and used as described before [45] . Double labeling was performed sequentially, using for each antibody a different size of protein A gold. Unspecific binding of the second protein A to the first antibody was blocked by incubation with 1% glutaraldehyde as described before [46] . The grids were viewed with Philip CM10 electron microscope at 80 kV and images were taken with KeenView camera (Soft Imaging System, Lakewood, CO, USA) using iTEM 5.0 software (Soft Imaging System GmbH). Supernatants of HeLa-prME or its parent HeLa cells were harvested and cleared by centrifugation at 3,000 rpm for 15 minutes and 10,000 rpm for 30 minutes. Clarified supernatants were then concentrated by ultracentrifugation at 28,000 rpm for 2.5 hours. Pellets were then resuspended in 100 ml of Phosphate buffered saline (PBS). For the production of immature RSPs, the HeLa-prME cells were cultured in medium containing 20 mM of ammonium chloride (NH 4 Cl). To generate RSPs of dengue 2, 3 and 4, pcDNA constructs containing the optimized prME genes (10 mg each) were transfected into 293T cells separately. The DV1 pcDNA-prME construct was transiently transfected as control. Supernatants of transfected 293T cells were harvested, clarified and concentrated as mentioned above. After ultracentrifugation, RSPs were resuspended in 100 ml PBS to which 33 ml of 4X NuPAGE LDS (lithium dodecyl sulfate) sample buffer (Invitrogen) was added. RSPs were then analyzed by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) using 4,12% NuPAGE Novex Bis-Tris Gels (Invitrogen). For analysis of cell lysates by SDS-PAGE, HeLa-prME cells were detached by incubation in 10 mM EDTA at 37uC for 10 min and resuspended in 500 ml of PBS. Cell suspensions were then mixed with 167 ml 4X of sample buffer and then sonicated before electrophoresis of samples as described above. For immunodetection, proteins were blotted from gels onto polyvinylidene difluoride (PVDF) membranes. The membrane was blocked overnight in 5% milk in PBST solution and then incubated with anti-E antibody (4E11, 1:1000) for 1 hour. After washes, the membrane was incubated for 1 hour at room temperature with a horseradish peroxidase-labeled goat anti mouse IgG polyclonal antibody. The membrane was finally visualized using ECL Western blot detection reagents (Invitrogen) and Amersham Hyperfilm ECL (GE Healthcare, Waukesha, WI, USA). For silver staining, the gel was fixed for 30 minutes and incubated with sodium thiosulfate for 30 minutes at room temperature. After three washes, the gel was incubated with Silver Nitrate for 40 minutes and developed for 15 minutes in sodium carbonate solution (25 g/L). EDTA solution (40 mM) was used to stop the development. For the endoglycosidase treatment, RSPs or HeLa-prME cell lysates were treated with 500 U of EndoH or PNGase F at 37uC for 3 hours according to the manufacturer's instructions (New England Biolabs, Beverly, MA, USA), and subsequently analyzed by Western blot. A chemiluminescence dot-blot (CLDB) method was developed to quantify RSPs. Briefly, 40 ml of supernatant of either HeLa-prME cell line or purified E protein solution at known concentration were blotted onto PVDF membrane through a Dot Blot 96 System (Biometra, Goettingen, Germany). The membrane was blocked overnight in 5% milk in PBST solution, incubated with anti-E antibody (4E11, 1:10,000) for 1 hour and then for 1 additional hour with a peroxidase-labeled goat anti mouse IgG polyclonal antibody (1:10,000; ZYMED). ECL Western blot detection reagents (Invitrogen), diluted five times, were mixed and added to the membrane and the luminescence intensity was measured using the Microbeta luminometer (PerkinElmer, Waltham, MA, USA). For sucrose gradient analysis, concentrated RSPs were ultracentrifuged in a 20 to 60% discontinuous sucrose gradient at 28,000 rpm (Beckman SW-41Ti rotor) for 2.5 hours in 4uC. All sucrose solutions were prepared with HEPES buffer (20 mM). Fractions of 0.5 ml were collected and levels of E were measured using the CLDB assay. Alternatively, RSPs were treated with 0.5% Triton X-100 for 1 hour before sucrose gradient fractionation, for RSPs denaturation. The human membrane trafficking siRNA library targeting 122 genes and the corresponding transfection reagents were purchased from Dharmacon (#G-005500; Dharmacon Research Inc, Lafayette, CO, USA). For the screen, the HeLa-prME cells were seeded in eight 96-well plates with 10,000 cells per well. 24 hours later, 10 pmol of each siRNA was added to each well together with the transfection reagent. siRNA targeting DV1 prME (target sequence: AGATC-CAGCTGACCGATT) and non-targeting siRNA (NT) (Dharmacon Research Inc) were used as controls. Each plate contained in triplicates 15-16 siRNAs from the library, as well as positive (DV1 prME) and negative (NT) siRNA controls. Culture medium was changed two days post-transfection, the supernatant from each well was harvested after an additional 48 hour incubation and then cleared by centrifugation at 4000 rpm for 15 min. 40 ml of supernatant from each well were used to measure by CLDB the levels of RSPs secreted by siRNA transfected HeLa-prME. Two-sided Student's t test was used to assess the statistical significance of differences between each sample and the non-targeting control. Differences were considered statistically significant when P,0.05. Levels of E protein in cell supernatant were expressed in relative luminescence units and ratios of experimental conditions to controls, set as unity, were calculated. Figure S1 The four optimized DV prME sequences. Each optimized prME gene has a BamH I restriction enzyme site, a kozak sequence GCCACC, a signal sequences from VSV-G, and a Xho I restriction enzyme site. Found at: doi:10.1371/journal.pone.0008325.s001 (0.03 MB DOC)

Breaking the Waves: Modelling the Potential Impact of Public Health Measures to Defer the Epidemic Peak of Novel Influenza A/H1N1

BACKGROUND: On June 11, 2009, the World Health Organization declared phase 6 of the novel influenza A/H1N1 pandemic. Although by the end of September 2009, the novel virus had been reported from all continents, the impact in most countries of the northern hemisphere has been limited. The return of the virus in a second wave would encounter populations that are still nonimmune and not vaccinated yet. We modelled the effect of control strategies to reduce the spread with the goal to defer the epidemic wave in a country where it is detected in a very early stage. METHODOLOGY/PRINCIPAL FINDINGS: We constructed a deterministic SEIR model using the age distribution and size of the population of Germany based on the observed number of imported cases and the early findings for the epidemiologic characteristics described by Fraser (Science, 2009). We propose a two-step control strategy with an initial effort to trace, quarantine, and selectively give prophylactic treatment to contacts of the first 100 to 500 cases. In the second step, the same measures are focused on the households of the next 5,000 to 10,000 cases. As a result, the peak of the epidemic could be delayed up to 7.6 weeks if up to 30% of cases are detected. However, the cumulative attack rates would not change. Necessary doses of antivirals would be less than the number of treatment courses for 0.1% of the population. In a sensitivity analysis, both case detection rate and the variation of R0 have major effects on the resulting delay. CONCLUSIONS/SIGNIFICANCE: Control strategies that reduce the spread of the disease during the early phase of a pandemic wave may lead to a substantial delay of the epidemic. Since prophylactic treatment is only offered to the contacts of the first 10,000 cases, the amount of antivirals needed is still very limited.

On June 11, 2009 , the World Health Organization (WHO) declared Pandemic Phase 6 based on sustained community transmission in more than one WHO region [1] . By the end of September 2009, in most European countries the impact has been limited, which is likely due to the dampening effect of the summer months. What the future holds is difficult to predict as past pandemics have been rather variable because of the complex interactions of immunity, pathogenicity, season and other factors [2] . Many countries have established a stockpile of antivirals for the treatment of the (severely) sick. We examine how a control strategy in the very beginning could defer the peak of the epidemic by using only a very small part of the stored antivirals for prophylactic treatment. There are a number of possible public health measures that may be used to stop or slow down the spread. Border control has been extensively discussed in this context to delay the international spread of influenza. However, in order to achieve a significant delay, more than 99% of air travel would have to be stopped [3] . As has been shown for SARS, entry screening methods are unlikely to detect more than 10% of imported infections and the positive predictive value of temperature screening is low especially at the beginning of a pandemic [4, 5] . It is therefore inevitable that importation occurs. For the management of imported cases, other measures may be used and include contact tracing, isolation and quarantine, as well as post exposure prophylaxis. It is unclear; however, which public health strategy could be effective in preventing the spill-over from imported cases and slow-down the transmission within the general population. Even less clear is how long these measures should be maintained, particularly once that domestic transmission has started. Mathematical models can be used to aid in decision making and have been increasingly applied to analyse the potential impact of containment strategies, pharmaceutical interventions and public health measures on the course of a novel influenza pandemic [6] [7] [8] [9] . Before the emergence of the novel influenza A/H1N1, virus modelling studies have suffered from the high number and uncertainty of necessary assumptions [10] . Even during the first months of the pandemic with the novel virus A/H1N1 only limited knowledge has emerged about the characteristics of the new virus [11, 12] . Therefore, we do not stress the particular timing and severity of a certain baseline scenario, but rather concentrate on the effect of the intervention in the model that should allow an estimation of the possible effects of these interventions. To construct possible baseline scenarios we used some of the estimates of properties of the novel virus that have been published based on data from the initial outbreak in Mexico by Fraser [11] . Similar estimates for R0 and a little higher estimates for the generation time were found in the USA [13] . We constructed a deterministic model with the following goals: (1) To model a possible evolution of an epidemic in Germany including assumptions about importation and domestic spread using the present knowledge about the virus; (2) to quantify the potential impact of public health measures, such as case detection, case isolation, quarantine of contacts, and the use of antiviral medication for therapy and post exposure prophylaxis, with a given effectiveness, on the initial evolution of the epidemic; (3) to identify possible conditions, which -if known -favour the adaptation of measures or the termination of the control strategies. Until June 5th, the rate of imported cases has been stable in Germany. The first identified cases in Germany were confirmed on April 29 th , and until June 5 th a total number of 49 confirmed cases had been reported ( Figure S1 (A)). Of these 41 (84%) had a known travel history to Mexico, the United Kingdom or the USA; of the remaining 8 cases 7 had been contacts to one of the imported cases and for one case the source case was unknown but might have been related to contacts with travellers (airport worker). The average number of imported cases during this time period (38 days) corresponds to 1.1 cases per day, however, 28 of the 41 imported cases were reported during the 10 days prior to June 5 th (corresponding to 2.8 imported cases per day). We considered two different scenarios of case importation ( Figure S1 (B)): firstly, a constant number of imported cases per day (five or ten cases per day), and secondly, an exponentially growing number of imported cases limited to 120 imported cases per day. The exponential growth rate was determined by assuming an ''import-R0'' of 1.1 and the same generation time as for the transmission inside Germany. An example for the prevalence of symptomatic cases resulting from our SEIR model is shown in Figure S2 for three different values of R0 (1.34, 1.58, and 2.04). For an R0 of 1.58 the point estimate of the timing of the peak after introduction of the first case would be 10 to 11 weeks (depending on the number of imported cases, compare Table 1 ), the point estimate for the peak prevalence of the population infected would be 4.3%, for the total attack rate of the population infected 44.8% and for the population diseased 38.5% (Table 1) . Depending on the three R0 the cumulative proportions of children that develop symptoms are 48%, 67%, and 79%, and the cumulative proportion of symptomatic adults 17%, 34%, and 54%, respectively. As a result of the higher susceptibility of children in the model the peak proportion of infected children is reached 18, 12 or 8 weeks (126, 82, or 56 days) after the first infected case, roughly 1-1.5 weeks earlier (i.e. 11, 8 or 6 days) than in adults (data not shown). Figure S3 shows how the peak is delayed for an assumed R0 of 1.58 and 5 imported cases per day when the first 500 cases are managed with a combination of intensive case-based measures (CCM1), followed by 10,000 with management mainly restricted to members of the household (CCM2; for details see Methods section). The effect of the number of household focused interventions (CCM2) on the delay of the peak is dependant on the basic reproduction number R0 and the sensitivity of the surveillance system ( Figure S4 ). When R0 is at least 1.58 and not more than 30% of cases can be detected, saturation occurs relatively early. Even if 50% of the cases can be detected or R0 is as small as 1.34 the peak of the epidemic can not be deferred any more after management of approximately 10,000 cases with CCM2. In general can be said: the higher R0 the earlier management with CCM2 becomes ineffective. This can only be balanced to a certain degree by a higher sensitivity of the surveillance system. The following considerations are done on the basis of 5 imported cases per day and R0 equal to 1.58. Effect of sensitivity of the surveillance system. The delay of the peak increases with the proportion of detected cases. When 10% of cases are detected and these are followed-up with CCM1 for the initial 500 cases (but no CCM2) the delay is 6 days, but can be raised to 20 days (gain of 2 weeks) when case detection is improved to 30%. The combined approach of 500 cases targeted with CCM1 and additional 10,000 cases with CCM2 the gain based on the improved case detection results in an increase for up to 6 weeks (11 to 50 days; Figure S3 , Table 2 ). Separate analysis of the effect of CCM1 and CCM2. In the example above 55% (6 of 11 days; 10% case detection rate) or 40% (20 of 50 days; 30% case detection rate) of the delay, respectively, is already achieved through CCM1 alone ( Table 2) . Table 1 . Characteristics of the baseline scenario without preventive interventions under the assumption that each day 5 cases were imported to Germany. This holds similarly in the scenario of the exponentially growing number imported cases -of course under the assumption that already in the beginning the surveillance system detects a percentage of the imported cases. Effect of R0. All of the above calculations are very sensitive to the value of R0. If R0 is as low as 1.34 it is much easier to delay the peak. If it is high, such as 2.0 or more, and the number of imported cases is at least 5 per day, then the effect of CCM1 and CCM2 can delay the peak at most by 8 weeks (57 days; Table 2 ), but only when case detection rate is high (50%). The personnel and personnel time that is needed to implement such measures depends on the number of local health departments involved, their capacity and many other factors. The maximum number of antivirals, however, can be estimated. Assuming that each case has 15 contacts that merit attention for CCM1 and would get antiviral post exposure prophylaxis, this would result in a maximum of 500615 ( = 7,500) treatment courses (as the number of doses for treatment equals the number of doses for prophylaxis). Further, if the household members of 10,000 cases are given antiviral prophylaxis this would amount to another 10,00063 ( = 30,000) treatment courses, in total 37,500 treatment courses. We present here a model how public health measures can contribute to the delay of an epidemic wave with the novel influenza virus A/H1N1 when the epidemic is detected in a very early stage. Delaying the pandemic spread is an important achievement because it gains time for other measures and preparations, such as early assessment of the virus' characteristics, activation of surge capacities or vaccine production and the development of a vaccination strategy [10] . The transmission parameters of our model were derived from the initial analysis of the epidemic in Mexico by Fraser [11] and a constant or slowly increasing influx of imported cases as observed for one month after the identification of the first case in Germany. This is also in agreement with European data showing a constant proportion of travel-related cases over time (the travel related cases within Europe between April 16 and June 2 also seem to remain constant over time [14] ). It has been shown before that so called targeted layered containment strategies, a combination of antiviral prophylaxis and non-pharmaceutical interventions, can be effective in reducing the transmission of pandemic influenza [8] . We extended this approach by analyzing the effect of a more intensive phase including contact tracing, identification and management of contacts outside of the household (CCM1), followed by household centred measures (CCM2). The assumed effectiveness for CCM1 indicates that a corresponding strategy would be very effective in reducing the spread of the epidemic, if R0 is moderate or low and if the number of imported cases does not increase rapidly. In contrast, the household centred measures (CCM2) continue to gain time, even when larger amounts of cases are imported per day. In the model, the effect of CCM1 and CCM2 rapidly decreased for higher Table 2 . Delay (in days) of the peak of the epidemic wave for adults in Germany as a function of the number of imported cases (column A), the number of CCM1 and CCM2 treatments, R0 and the case detection rate (10%, 30% or 50%). values of R0, indicating that this intervention should be stopped when the delay becomes only marginal. To our knowledge, the question of conditions that would lead to stopping interventions (or the lowered effects of specific measures) has so far been studied only in the context of community mitigation strategies and interventions, such as school closures [15] . Modelling studies analyzing preventive measures often assume that the intervention will be available for a large part or the whole population [6, 8, 16] . E.g. Carrat et al. analysed the effect of combined interventions targeting up to 70% of households in their ''small-world-like'' model [17] . In contrast, our model focuses on the initial stage and the first few thousand cases in Germany, and on the delay that can be achieved. We have restricted the analysis of maintaining the less intensive CCM2 strategy for 10,000 courses. The maximum amount of antivirals used in this approach corresponds to treatment courses for less than 0.1% of the German population and 5% of the amount available for seasonal influenza in the winter season. After the end of isolation, the treated contacts were assumed to remain susceptible to the infection. However, these conservative assumptions result only in a shift (depending on R0 significant) of the epidemic curve and virtually no reduction of the overall attack rate. As we have shown, a number of factors are important when effects of public health interventions are considered. First, the rate of seeding has significant impact on the delay that can be achieved by the interventions. Other published models start their simulation with one infection, a random number each day or an increasing number at a reduced R0. In contrast, our analysis is based on the observed number of cases. Second, the sensitivity of the surveillance system to detect cases is important. In reality, detection of cases without travel history to countries with community transmission of novel influenza A/H1N1, i.e. domestic cases, may be difficult as the clinical picture of the disease has proven to be often non-specific. Therefore, sensitivity to detect cases may be relatively low. Given these limitations we have provided a number of different scenarios (case detection rates of 10, 30 and 50%) addressing how the described interventions may impact the epidemic. As a note of caution it must be mentioned that these calculations should of course not be mistaken as a prediction. It was for example not possible to validate the assumptions about the effectiveness of the interventions using real data. Other limitations are: (1) While before the start of the novel A/ H1N1 epidemic it was difficult to make realistic assumptions, it is now easier to do so now, since first estimates can be drawn for a number of parameters of the novel influenza virus. Nevertheless, many pieces of information are uncertain and may change due to more information coming to light, or due to real changes of the virus and its epidemiology. (2) The effect of season is not taken into account, we expect that the virus is more easily transmitted in the fall or winter time [18, 19] . (3) The proportion of asymptomatic cases and their contribution to transmission is still unknown. (4) Lacking realistic alternative information we distinguished only two age groups, children and adults. (5) The evolution of the number of cases imported is unknown and will probably change over time. (6) The age dependence of susceptibility in Germany is unknown and is likely to differ from the one in Mexico. (7) The sensitivity of surveillance is unknown, and therefore the true proportion of cases detected is also unknown. If persons were infected in Germany from a source without travel history, then they may have been more easily missed, especially since initial surveillance efforts usually focus on diseased persons with travel history. (8) The rigor of public health measures is likely to vary among different local health departments. However, it is plausible that the measures taken contributed to the delay of the initial spread of the infection, because, until mid June at latest, virtually no tertiary cases had been detected by close surveillance of contacts and the surrounding of cases. In the model the change from CCM1 to CCM2 was suggested on a population level. In reality, of course, there might not be a real threshold and the strategy might change depending on the individual resources of local health authorities. Of course with the change of the epidemiologic picture more rigorous measures of social distancing, such as school closures may be implemented. When leaving the intensive phase of contact tracing and case management (CCM1), we believe that CCM2, or a strategy with similar effect, might be well suited to follow after because it focuses on the household. This is a much more amenable unit and is based on the knowledge that being a member of a household with a confirmed case is the highest single risk factor for influenza infection [6] . In conclusion, despite the many possible pitfalls and caveats of our study we believe that we have demonstrated the possible impact of a sequential strategy on the spread of the novel influenza virus A/ H1N1 in a country where imported cases start the epidemic. Cased-based information was used to assign reported confirmed cases in Germany and status of either imported or domestic. Cases with travel history of more than 7 days before onset of symptoms (two times the maximal incubation period) were considered domestic. (a) Assumptions. We assumed that at the outset of the epidemic the entire population is fully susceptible to infection with the influenza A/H1N1 virus. Infectiousness was assumed to be equal in symptomatic and asymptomatic persons. This is based on the rationale that a lower degree of infectiousness is coupled with unrestricted mobility resulting in a higher number of potentially infectious contacts. In comparison, a higher degree of infectiousness in symptomatic patients is compensated by the reduction of the number of contacts, because patients are isolated and bedbound. Assuming that the epidemiologic and virologic characteristics are similar to the epidemic in Mexico allowed us to use the values as described by Fraser [11] . They found the ''most likely'' basic reproductive number was 1.58, range 1.34 to 2.04, and estimated a generation time of 1.91 days (95% confidence interval 1.3-2.71). They distinguished children (,15 years of age) and adults (. = 15 years of age) and found that children were 2.06 (95% confidence interval 1.60-3.31) as susceptible as adults. The assortativity of mixing between children and adults was estimated as 0.5 (95% confidence interval 0.00-0.72) -an assortativity of 0 corresponds to a completely random mixing, whereas 1 corresponds to fully assortative groups. Finally, they found that 86% (95% confidence interval 69%-100%) of the infected persons become symptomatic. We considered three different scenarios of R0, namely R0 equal to 1.34, 1.58 or 2.04 and used the point estimates for all other parameters. The model does not incorporate assumptions about the severity of disease or how severity might alter infectiousness. Lastly we needed to make assumptions about the effectiveness of the public health measures and the sensitivity of the surveillance system. Assumptions are made for the effectiveness of two approaches that combined several case-based public health measures (combination of case-based measures; CCM): (1) Intensive measures, including isolation and therapy of cases, contact tracing, quarantine and post-exposure prophylaxis of selected contacts in-and outside of the household (CCM1). Because CCM1 consumes many resources it is assumed that CCM1 will only be sustained in the first phase. We model two scenarios with the first 100 or up to 500 cases cared for by CCM1 countrywide. (2) Less-intensive measures, focusing on the household, including isolation and therapy of cases, quarantine and post-exposure prophylaxis of household contacts (CCM2). Thus, we assume that following the initial 100-500 confirmed index cases there will be no contact tracing any more, i.e. no more post exposure prophylaxis for non-household contacts. CCM1 and CCM2 are set to be 75% and 50% effective in reducing secondary cases, respectively. We modelled four scenarios: in the first and second, CCM1 is maintained for the first 100 cases followed by 5,000 or 10,000 cases cared for with a CCM2 strategy, in the third and forth, CCM1 is maintained for 500 cases followed by 5,000 and 10,000 cases with a CCM2 strategy, respectively. To include the effect of surveillance we made assumptions about the number of imported cases that are recognized. For this purpose we varied the proportion of recognized imported cases from 0%, 10% and 30% to 50%. The assumed sensitivity of the surveillance system reflects the probability (10%, 30% or 50%) to detect domestic cases. A higher probability to detect imported cases would have led effectively to a reduced total number of imported cases per day in the model. (b) Construction. We used -similar to Fraser [11] in their description of the outbreak in La Gloria -a generalised agestratified deterministic SEIR model to describe the spread of the disease [20, 21] . We used the following assumptions about the age distribution and size of the population of Germany: 71,000,000 adult population (. = 15 years of age), 11,000,000 children (,15 years of age). The complete model is described in the Appendix S1. Figure S1 Confirmed imported (red) and domestic (blue) cases in Germany by date of onset of symptoms (A Figure S3 Delay of epidemic curve. The ''most likely'' Ro of 1.58 from the study of Fraser et al. (Science, 2007) was used and case detection rates of symptomatic cases were set to 10% and 30%, respectively. Ro is assumed to be 1.58, and each day five cases were imported. The household and non-household contacts of the first 500 detected cases were treated with a combination of case-based measures that include contact tracing, quarantine, and post-exposure prophylaxis (CCM1); and the household contacts of the next 10,000 cases were managed with strategy CCM2, which includes only preventive measures in the household of the cases.

Expression of the VP2 Protein of Murine Norovirus by a Translation Termination-Reinitiation Strategy

BACKGROUND: Expression of the minor virion structural protein VP2 of the calicivirus murine norovirus (MNV) is believed to occur by the unusual mechanism of termination codon-dependent reinitiation of translation. In this process, following translation of an upstream open reading frame (ORF) and termination at the stop codon, a proportion of 40S subunits remain associated with the mRNA and reinitiate at the AUG of a downstream ORF, which is typically in close proximity. Consistent with this, the VP2 start codon (AUG) of MNV overlaps the stop codon of the upstream VP1 ORF (UAA) in the pentanucleotide UAA UG. PRINCIPAL FINDINGS: Here, we confirm that MNV VP2 expression is regulated by termination-reinitiation and define the mRNA sequence requirements. Efficient reintiation is dependent upon 43 nt of RNA immediately upstream of the UAA UG site. Chemical and enzymatic probing revealed that the RNA in this region is not highly structured and includes an essential stretch of bases complementary to 18S rRNA helix 26 (Motif 1). The relative position of Motif 1 with respect to the UAA UG site impacts upon the efficiency of the process. Termination-reinitiation in MNV was also found to be relatively insensitive to the initiation inhibitor edeine. CONCLUSIONS: The termination-reinitiation signal of MNV most closely resembles that of influenza BM2. Similar to other viruses that use this strategy, base-pairing between mRNA and rRNA is likely to play a role in tethering the 40S subunit to the mRNA following termination at the VP1 stop codon. Our data also indicate that accurate recognition of the VP2 ORF AUG is not a pre-requisite for efficient reinitiation of translation in this system.

For most eukaryotic mRNAs, translation initiation is a 59-enddependent process beginning with recognition of the cap structure by the cap-binding complex eIF4F [1] and (usually) recognition of the AUG codon of the first open reading frame (ORF) on the mRNA by the scanning ribosome complex [2] . This 59-end dependence is a problem faced by many RNA viruses with polycistronic genomes and elaborate strategies have been developed to facilitate access of ribosomes to downstream open reading frames (ORFs). Amongst these, a number of unconventional translation strategies have been described [3] . These include leaky scanning of 40S subunits past the start codon of the first ORF [4] , the possession of intercistronic internal ribosome entry signal [5] , programmed ribosomal frameshifting during elongation [6] and stop codon suppression at the termination step [7] [8] . Another strategy that has evolved to allow expression of a downstream ORF is termination-reinitiation (also referred to here as stop-start). In this process, ribosomes translate the upstream ORF but following termination, a proportion of 40S subunits remain tethered to the mRNA and go on to reinitiate at the start codon of the downstream ORF. This termination-dependent reinitiation strategy allows the coupled expression of products from adjacent ORFs and thus the production of a defined ratio of gene products. Termination-reinitiation in virus systems [9] was first described in the synthesis of the BM2 protein of the orthomyxovirus influenza B virus [10] and subsequently in expression of VP2 of feline calicivirus (FCV) of the genus Vesivirus [11] [12] [13] and VP10 of the calicivirus rabbit haemorrhagic disease virus (RHDV) of the genus Lagovirus [14] . A related phenomenon is also seen in expression of the M2-2 protein [15] [16] of the paramyxovirus respiratory syncytial virus (RSV) and the M2-2 protein [17] of pneumovirus of mice (PVM). In FCV, the stop codon (UGA) of the major capsid stop-start protein VP1 overlaps the start codon of the minor capsid protein VP2 (AUGA) (the stop-start ''window''). Efficient termination-reinitiation depends upon several factors, including the close proximity of the stop and start codons, the transit of ribosomes along the VP1 mRNA up to the stop codon and a stretch of some 70-80 nucleotides (nt) of mRNA upstream of the stop-start window whose primary sequence, rather than the encoded protein, is key. This region of the mRNA, termed the termination upstream ribosomal binding site (TURBS), is needed for the retention of post-termination 40S subunits [11] . A short sequence of the TURBS (termed Motif 1) that is complementary to part of helix 26 of 18S rRNA likely acts to tether the 40S ribosomal subunit to the mRNA post-termination, allowing time for the ribosome to acquire the factors necessary to initiate on the downstream ORF [12, 14, 18] . The TURBS may also act by recruitment of eukaryotic initiation factor 3 (eIF3) or eIF3/40S complexes [13] . Recent studies of termination-reinitiation in the expression of the orthomyxovirus influenza BM2 protein have revealed a requirement for a shorter stretch of mRNA (45 nt) upstream of the stop-start window, but nevertheless, the RNA contains a similar TURBS Motif 1 [19] . From RNA secondary structure probing, it has been proposed that this stretch may be displayed on the apical loop of a stem-loop structure that may form following transit of the ribosome through the region and termination at the upstream ORF stop codon [9, 19] . In this paper, we describe an analysis of termination-reinitiation in the expression of the VP2 protein of murine norovirus (MNV), a calicivirus of the genus Norovirus. The VP2 start codon (AUG) of MNV overlaps the stop codon of the upstream VP1 ORF (UAA) in the pentanucleotide UAAUG, consistent with a terminationreinitiation strategy, and a stretch of bases (59 UAUGGGAA 39) complementary to 18S rRNA helix 26 is present upstream. Using a luciferase-based reporter plasmid, we show that VP2 is expressed by termination-reinitiation and provide evidence consistent with a functional interaction between the coding region of the VP1 mRNA and 18S rRNA. The formation of mRNA secondary structure within the TURBS is also investigated. Overall, our data suggest that the mechanism of VP2 expression is broadly similar to that of the other caliciviruses and influenza B. However, in contrast to what was observed with the FCV signal [13] and seen here with influenza BM2, termination-reinitiation at the MNV signal shows resistance to the initation inhibitor edeine. Thus the mechanism by which the AUG of the downstream ORF is recognised may differ. To investigate termination-reinitiation in the synthesis of the MNV VP2 protein, a 255 bp fragment of viral cDNA was cloned between the SalI and BamHI sites of the dual-luciferase reporter vector p2luc [20] . The cloned fragment, which contained 203 bp of sequence information upstream of the UAAUG stop-start window, and 52 bp downstream was suspected, on the basis of work with other viruses (see Introduction), to contain all of the required sequences for termination-reinitiation. The cDNA fragment was cloned in such a way that the Renilla and Firefly luciferase ORFs were in frame with the stop and start codons respectively of the termination-reinitiation motif to give an ORF configuration 59 rlucVP1-VP2fluc 39 ( Figure 1 ). This vector, named p2luc-MNVwt, contains a T7 RNA polymerase promoter allowing synthetic mRNAs to be generated to investigate the stopstart process in in vitro translation reactions. The translation of in vitro synthesised wild-type (wt) mRNA from p2luc-MNVwt was carried out in FlexiH rabbit reticulocyte lysate (FlexiHRRL) supplemented with 140 mM KCl (see Materials and Methods) and gave products of the expected sizes (upstream rlucVP1 ORF, ,42 kDa, downstream VP2fluc ORF, ,64 kDa, Figure 1b ). The molar ratio of VP2fluc to rlucVP1 (taking into account the methionine content of the two proteins) was typically in the region of 1:10. Thus, initiation on the downstream ORF occurred at a frequency of about 10% of that of the upstream ORF. That this was indeed the product of the second ORF was further confirmed by comparing the migration of RRL translation products from mRNAs derived from p2luc-MNVwt that had been linearised at different points within the second ORF (data not shown). Termination-reinitiation is distinct from IRES-mediated expression of downstream ORFs as translation through the upstream ORF is an absolute requirement [11, 14, 16] . In order to establish whether this is also the case for MNV expression, a premature in-frame stop-codon was inserted close to the end of the rluc ORF but upstream of VP1 sequence information (219 bp upstream of the authentic rlucVP1 termination codon). If the expression of VP2fluc is a result of termination-reinitiation, translating ribosomes would be unable to reach the AUG start Figure 1 . Minimal sequence requirements for MNV termination-reinitiation. A) Schematic of the p2luc-MNV reporter mRNA. The termination-reinitiation region (203 nt upstream and 52 nt downstream of the UAAUG motif) was cloned into the SalI and BamHI sites of the p2luc reporter plasmid. HpaI run-off transcripts for in vitro translation were generated using T7 RNA polymerase. The location of the T3 promoter present in the structure mapping construct p2luc-MNV-T3 is indicated. B) Deletion analysis of MNV termination-reinitiation. A series of p2luc-MNV variants were prepared with stepwise, in-frame deletions from the 59 end of the inserted viral sequence. The wild-type (wt), premature stop (ps) and deletion mutant plasmids were linearised with HpaI and run-off transcripts translated in FlexiH RRL at a final RNA concentration of 50 mg/ml in the presence of [ 35 S]-methionine and 140 mM added KCl. The products were resolved by 12% SDS-PAGE and visualised by autoradiography. The number of nucleotides of viral sequence remaining up to the AUG start codon of the MNV ORF is shown below the gel. The product of the full-length or truncated versions of the rlucVP1 ORF (predicted size of MNVwt is 42 kDa) is marked rluc, and the VP2fluc product (predicted size, 62 kDa) is marked fluc. The MNV ps rluc product is the shortest (predicted size, 33 kDa). RRF denotes the relative reinitiation frequency in comparison to MNVwt (set at 100). The figure in brackets represents the ratio of the intensity of the fluc and rluc products (adjusted for methionine content and expressed as a percentage) for the MNVwt mRNA. doi:10.1371/journal.pone.0008390.g001 codon of VP2fluc in the mutant mRNA and the ORF could not be translated. As is clear in Figure 1b , the introduction of a premature stop codon into the rluc/M1 ORF abolished expression of the VP2/fluc product, but had no effect on synthesis of the upstream ORF (rlucVP1ps, ,33 kDa). These data are thus consistent with a termination-reinitiation strategy for the expression of the VP2 protein and confirm a requirement for translation through the upstream ORF. Expression of MNV VP2 Is Dependent on ,40-43 nt Upstream of the UAAUG Motif Previous work has suggested that viral termination-reinitiation events show little dependence on sequence information downstream of the ''stop-start'' window but require 45-250 nt of upstream primary sequence [11,14,16.19] . In order to determine the minimal sequence requirements for termination-reinitiation in VP2 expresssion, deletions of increasing size were made from the 59 end of the inserted viral information (Figure 1b) . The stop-start product was synthesised efficiently with up to 43 nt of VP1 information present upstream of the UAAUG motif, and to a lesser extent with 40 nt. However, deletion to 37 nt or less abolished expression of the termination-reinitiation product ( Figure 1b) . These data indicate that only 40 nucleotides of VP1 primary sequence immediately upstream of the stop-start window are required for termination-reinitiation in vitro, although 43 nt are required for full activity. Termination-Reinitiation of MNV VP2 Synthesis Is Dependent upon an mRNA Sequence with Complementarity to 18S rRNA In FCV, RHDV and influenza B, it has been shown that termination-reinitiation requires a closely conserved primary sequence element (referred to as Motif 1) that is complementary to a region of helix 26 of 18S rRNA [11] [12] 14] ). The position of Motif 1 varies somewhat, with the 59 base 73 nt (RHDV), 63 nt (FCV) or 34 nt (influenza B) upstream of the stop codon of the first ORF. Mutational analysis has revealed that this sequence is essential for the stop-start process [12] [13] [14] [18] [19] . Within the ,43 nt minimal region of the MNV VP1 RNA required for VP2 expression, a stretch of bases with a similar level of complementarity to 18S rRNA is also found (Figure 2a , complementary bases are shown in italics). To investigate whether this region plays a role in termination-reinitiation in VP2 expression, two point mutations were made to disrupt potential mRNA:rRNA pairs ( Figure 2a ). In the first, the A at -31 was mutated to a G (p2luc-MNV GU), creating a presumably slightly weaker putative U-G base pair between the rRNA and mRNA. In the second, the G at -32 was changed to a C (p2luc-MNV CC), which would act to disrupt the interaction between 18S rRNA and mRNA. As can be seen in Figure 2b , the latter mutation greatly reduced expression of the VP2fluc product, supporting the idea that an interaction between the 18S rRNA and the mRNA just upstream of the termination-reinitiation site is required. In the mutant where pairing was predicted to be maintained (p2luc-MNV GU) termination-reinitiation was clearly detectable, although the efficiency was reduced somewhat compared to that of wild-type mRNA. The experiments described above confirm the existence of Motif 1 and its role in reintiation in MNV. It was therefore of interest to determine the context of this 18S rRNA complementary region within the global RNA secondary structure of the minimal functional sequence, and to compare the structure with that determined for the influenza BM2 signal [19] . To achieve this, a bacteriophage T3 promoter was inserted upstream of the viral sequence of the p2luc-MNV.61 plasmid ( Figure 1b ). The plasmid was linearised with BamHI, T3 run-off transcripts synthesised and the RNA end-labelled with [ 33 P]-cATP. The labelled transcripts were subjected to limited chemical and enzymatic probing prior to analysis on denaturing polyacrylamide gels. The chemical probes used were imidazole and lead acetate, specific for cleavage of single stranded regions. Enzymatic probes were RNases T1, U2 and CL3, which preferentially cleave single-stranded G, A and C residues respectively, and RNase CV1, which cuts in helical regions in double-stranded or stacked conformations. A representative stucture mapping gel is shown in Figure 3 and in Figure 4 , the data are mapped onto mfold predictions of the secondary structure of the ''stop-start'' region. Structure probing analysis of the MNV signal revealed that, like the BM2 signal, the mRNA in the region essential for terminationreinitiation is not highly structured. This was especially evident from the chemical probes, with most residues sensitive to imidazole and lead cleavage. The enzymatic probes were also active against the majority of bases in the region and consistent with this, CV1 probing identified very few double-stranded or stacked bases. We also noticed a few CL3 cuts at residues other than C, although the reason for this is uncertain. Minimal free energy mfold predictions, performed using the online server of Zuker (http://mfold.bioinfo.rpi.edu/cgi-bin/rna-form1.cgi) indi- cated that the most stable RNA fold was the bulged stem-loop shown in Figure 4 . However, the correspondence between this mfold and the mapping data was not absolute. Whilst in general, the single-stranded probes displayed more activity against regions of the model predicted to be single-stranded than they did against predicted helices, there were anomalies. For example, residues G51-52 were sensitive to RNase T1, yet were predicted to be in a double-stranded region (stem 2). Generally, the predicted duplexes showed more reactivity to single-stranded probes than one would expect for stable double-stranded stretches. Therefore, it seems likely that the RNA in this region is metastable, potentially adopting a number of co-existing structures. In our model, the sequence complementary to 18S rRNA is sequestered between two putative stems (stems 2 and 3; Figure 4 ) at a location similar to that found with BM2 [19] . Given that the termination-reinitiation process requires the ribosome to translate through the VP1 ORF, secondary structure in the RNA upstream of the ''stop-start'' window would be unwound and perhaps remodelled as the ribosome transits to the termination codon. Toeprinting of ribosomes paused at initiation codons has shown that the 59 edge Sites of cleavage were identified by comparison with a ladder of bands created by limited alkaline hydrolysis of the RNA (OH-) and the position of known RNase U2 and T1 cuts, determined empirically. Products were analysed on a 10% acrylamide/7M urea gel containing formamide. Data was also collected from 6% and 15% gels (gels not shown). Enzymatic structure probing was with RNases T1, U2, CL3 and CV1. Uniquely cleaved nucleotides were identified by their absence in untreated control lanes (0). The number of units of enzyme added to each reaction is indicated. Chemical structure probing was with imidazole (I, hours) or lead acetate (Pb; mM concentration in reaction). The water lane (W) represents RNA which was dissolved in water, incubated for four hours and processed in parallel to the imidazole-treated sample. The sequence of the probed RNA and the inferred secondary structure is shown in Figure 4 . of the ribosome is some 12 to 13 nt from the first base of the AUG [21] . This would place the 59 edge of the terminating ribosome (with the UAA codon in the A-site) close to residue C78 on our mRNA. Thus a terminating ribosome would prevent formation of the secondary structure, conceivably releasing the 18S rRNA complementary region for interaction with the ribosome (see Discussion). An alternative structure can be predicted under such circumstances, shown in the inset box in Figure 4 . In this structure, the 59 arm of the original stem 2 is predicted to pair with alternative bases to generate a new stem (stem 29) with Motif 1 forming part of the apical loop. This alternative fold is attractive for a number of reasons. By displaying Motif 1 on an apical loop, this could promote 18S rRNA binding and ribosome tethering [19] . Until ribosomes transit through this region, Motif 1 would remain within a larger structure with potentially reduced access to the ribosome which could, at least in part, account for the observation that the signal does not appear to function as an IRES. The deletion analysis of Figure 1 is also consistent with a role for this alternative structure as the functional ''end-point'' maps to the start of the 59 arm of stem 29. Furthermore, most, if not all, viral TURBS have the potential for base-pairing between regions flanking Motif 1 [18] . Nevertheless, it should be noted that the structure mapping data are not fully consistent with this alternative structure, for example, there is considerable sensitivity to RNase T1 cleavage within the 59 arm of stem 29. This is considered further in the Discussion section. Efficient termination-reinitiation seems to require the close proximity of the stop and start codons [12, 14, 19] . To investigate whether this is also the case for MNV, the authentic stop codon of rlucVP1 (in the context of the fully functional MNV49; see Figure 1b ) was mutated from UAA to CAA such that the first ORF was extended by 13 amino acids (MNV49.1 Figure 5 ). The separation of stop and start codons by such a distance in BM2 is known to reduce reinitiation about 10-fold [19] , but with the MNV signal, only a three-fold reduction in flucVP2 synthesis was observed ( Figure 6 ). A possible explanation for this lies in the fact that the ''new'' stop codon is itself embedded within a second potential stop-start sequence (UGAUG) which could facilitate some reinitiation, but perhaps at a lower frequency, as it would not necessarily be spaced appropriately with respect to Motif 1. Another possibility is that 40S subunits terminating at the downstream stop-start sequence can reinitiate, at a reduced frequency, at the correct (upstream) AUG, despite the increased spacing, with the 40S subunit remaining tethered to the mRNA and ''snapping-back'' to the normal position of reinitiation. In an attempt to distinguish between these possibilities, additional constructs were prepared in which point mutations were introduced into pMNV.49 such that the termination and start codons in the two stop-start regions were changed separately and in combination (MNV49.2 to MNV49.8; see Figures 5 and 6) . From this analysis, it is evident that modification of the authentic termination-reinitiation motif reduces reinitiation, irrespective of whether the stop or start codon is eliminated. Alteration of the AUG codon had the most effect, with reinitiation reduced to 8-38% of the wild-type level. When the natural termination site was changed such that termination now took place 13 or 15 amino acids downstream, the frequency of termination-reinitiation was also reduced, to 25-49% of the wild-type level. In MNV49.8, where termination of the upstream ORF occurred 30 amino acids downstream of the natural site, very little reinitiation was seen (8% of the wild-type level), indicating that the ribosome is unable to locate the authentic AUG from such a distal termination site. In the translation of this mRNA, an additional product was seen (asterisked in Figure 6 ) whose size is consistent with a fusion of the encoded ORFs (this is considered in the Discussion section). Reinitiation events that take place at either the authentic or the downstream stop-start motifs would produce polypeptides that differ in size by only 13 amino acids, thus we would not expect to be able to distinguish them by SDS-PAGE, and this is clear in Figure 6 , where the reinitiation products show very similar electrophoretic mobilities. Thus we cannot say with confidence whether a particular AUG (or both) is used. However, the substantial reintiation activity displayed by MNV49.7, an mRNA in which both AUGs were changed, indicates that non-AUG codons can act as reinitiation codons, although probably at reduced efficiency. This is consistent with other work demonstrating that reinitiation can occur at non-AUG codons within the context of a termination-reinitiation signal [12, 14, 19] . Whilst in principle, reinitiation of translation of the MNV 49.7 VP2fluc ORF, following termination, could occur at the next available AUG, this is located 54 amino acids from the natural stop-start signal and initiation here would produce a substantially shorter product that would have been detectable by SDS-PAGE. Thus in this mRNA, a significant proportion of ribosomes (25% of the wild-type level) that terminate 13 amino acids downstream of the authentic stop-start site can reinitiate in an AUG-independent manner within the stop-start window. The precise mechanism of termination-reinitiation is not known, but the sensitivity of FCV VP2 protein expression to the translation initiation inhibitor edeine suggests that the reinitiation process bears at least some similarity to standard initiation at AUG codons [13] . To ascertain whether edeine sensitivity is a general feature of termination-reinitiation, we analysed the effect of the peptide on the activity of the MNV and BM2 signals (with FCV as a control) using translation time-courses (Figure 7 ). Reactions were programmed with the relevant mRNA and at various times an aliquot was removed, edeine added (to 5 mM) and the aliquot reincubated such that the total time of translation was 60 minutes. To determine the time of first appearance of the termination and reinitiation products, identical reactions were also performed in which the elongation inhibitor cycloheximide replaced edeine. In the edeine experiments, it was evident that for FCV and BM2, only a trace of ''stop-start'' product was synthesised at the early time points. In these experiments, the vast majority of ribosomes did not reach the stop-start window until at least 7.5 minutes had passed (as shown in the cycloheximide time course experiments [data not shown; see legend to Figure 7 ]), thus the trace of VP2fluc seen likely corresponds to the product of infrequent internal initiation at the VP2fluc AUG or is derived from those few ribosomes that had reached the stop-start window prior to edeine addition. At later time points, however, the termination-reinitiation product steadily accumulated, with the ratio of the upstream and downstream ORFs stabilising after 30 minutes (at a reinitiation frequency of ,4%). Thus for the FCV and BM2 signals, when edeine is present prior to arrival of ribosomes at the stop-start signal, it greatly inhibits termination-reinitiation, but has little effect on translation post-reinitiation. Unexpectedly, the MNV signal responded differently, with the termination-reinitiation product being more evident at early times post-edeine addition (in comparison to FCV and BM2). At these early time points, few ribosomes would have reached the stop-start window prior to edeine addition, thus the MNV signal shows increased resistance to the effects of edeine. Examination of the kinetics of synthesis of the two ORFs (Figure 7d ) reveals that in all cases, the frequency of termination-reinitiation at early time points was higher than that seen at the steady state. This is indicative of a titration effect; early in the time course, when fewer ribosomes have loaded onto the mRNA (due to the earlier addition of edeine), the greater frequency of reinitiation may reflect the increased relative abundance of a necessary factor. The molecular basis of the resistance to edeine seen with the MNV signal is difficult to explain. It may be that recognition of the stop-start motif is indeed blocked by edeine but somehow, a proportion of initiation complexes still recognise the AUG present in the second pentanucleotide motif (UGAUG; see above) on the mRNA. In this paper we show that expression in vitro of the murine norovirus VP2 protein occurs by coupled translation terminationreinitiation. The process requires the close proximity of stop and start codons, a defined region of mRNA upstream of the stop-start window that includes a functional TURBS Motif 1 and translation by the ribosome through this region up to the site of terminationreinitiation. Secondary structure mapping indicates that the RNA in this region is weakly structured, with Motif 1 loosely embedded in the 59 arm of a putative stem-loop structure. The MNV signal thus exhibits many of the features and functional characteristics of the stop-start signals of FCV, RHDV and influenza B. The Figure 1 ] which acts as the ''wild-type'' reference construct [WT] in these experiments), in which the stop and start codons of the termination-reinitiation signal were altered. The figure shows the primary sequence and three-frame translation of the relevant region of the mRNA encoded by each construct. The natural stop-start motif is shown in pink and emboldened text, the downstream fortuitous stop-start motif in pink. Mutations within the mRNA sequence are highlighted by uppercase, red emboldened characters. The upstream rlucVP1 ORF is highlighted in grey, as is the downstream VP2fluc ORF where this is known. Likely key methionines (start codons) or their replacement amino acid are highlighted in green. doi:10.1371/journal.pone.0008390.g005 molecular mechanism of termination-reinitiation remains to be fully elucidated, however. Central to the discussion is the TURBS and in this context the purpose of the identified Motifs, the role (if any) of RNA secondary structure, and the functional requirement for translation through the TURBS. Regarding Motif 1, it is clear that in all studies so far, mRNA mutations that would destabilise an interaction with 18S rRNA reduce or abolish reinitiation and changes not predicted to affect pairing having a lesser effect or none at all. Recently, the reciprocal experiment was performed, where mutations were introduced into the relevant region of (yeast) 18S rRNA. Their effect on termination-reinitiation was found to be highly consistent with a role for mRNA-18S rRNA pairing [18] . These experiments confirm a role in tethering through rRNA, although do not rule out the contribution of other factors, for example, binding of eIF3 [13] . A comparative alignment of the MNV signal with other known or suspected termination-reinitiation signals (Figure 8 ) reveals that Motif 1 is always present and that the stop and start codons of the termination-reinitiation site are in close proximity to each other. What does vary is the spacing between the two elements, from only 26 nt in the case of BM2 to 29 nt in MNV, 53 nt in FCV, 61 nt in RHDV and 62 nt (the longest) in the Lagovirus European brown hare syndrome virus. It is not clear whether the ''additional'' sequences present in viruses with longer TURBS have a role in termination-reinitiation. Deletion analysis of the FCV and RHDV TURBS has revealed some dispensible sequences -there may be some flexibility in the spacing of Motif 1 that allows other biological information to be accommodated into the TURBS without affecting function in stop-start. However, there is little sequence conservation between the signals of viruses of different genera, arguing against the presence of other primary sequence motifs. Another stretch of bases of functional consequence has been identified in FCV and RHDV, namely TURBS Motif 2, which is located closer to the stop-start window than Motif 1 and is speculated to help position ribosomes correctly at the reinitiation codon [12, 14] . Recent work has shown that the functional requirement for Motif 2 is in its participation in a basepaired region that forms between this motif and a stretch of bases immediately upstream of Motif 1 [18] ; see Figure 8 . This basepairing has previously been noted from structure predictions of the signal of FCV [13] and direct RNA secondary structure probing of BM2 stem 2 [19] and the MNV stem 2 (see Figure 4) . Based on the observations of Luttermann and Meyers [18] , the formation of this stem is likely to be important to termination-reinitiation in the BM2 and MNV systems. Indeed, it is noticeable that in the deletion analysis of the MNV signal, and that of BM2 [19] , those deletions that would affect formation of stem 2 showed reduced activity in termination-reinitiation (Figure 1b, Figure 4) . Despite this progress, the occurence and role of RNA secondary structure within viral TURBS is poorly understood. Direct structure probing and mfold analysis indicates that the RNA upstream of the stop-start window is metastable and whilst the secondary structures proposed for FCV [13] , BM2 [19] and MNV (this study) are superficially similar, the largely single-stranded nature of the TURBS weakens these models and their comparison. The insertion of a premature termination codon upstream of the TURBS blocks reinitiation, ruling out the possibility that VP2 expression occurs by ribosome recruitment to a conventional, structured, viral IRES or by shunting from the untranslated region of the upstream ORF. The requirement for translation through the TURBS may simply reflect the need to deliver ribosomes to the stop-start window, but it could also indicate a requirement to remodel the TURBS, conceivably by alteration of RNA secondary structure or displacement of a bound factor. Based on chemical and enzymatic RNA structure probing of the BM2 signal and folding predictions (mfold), it has been suggested that transit of the ribosome to the stop-start window leads to melting of one stemloop structure and the formation of an alternative structure that has Motif 1 displayed on its apical loop [19] . The position of MNV Motif 1 relative to the stop-start window is very similar to that of BM2 suggesting that the same remodelling could operate (Figures 4 and 8) . However, whilst transit and termination of the ribosome would destabilise the identified secondary structure (Figure 4) , liberating TURBS motif 1 in close proximity to helix 26, it is not clear whether this motif would subsequently be displayed as part of an alternative secondary structure. Whilst mfold analysis of the MNV region present locally upstream of the terminating ribosome does suggest an alternative secondary structure, further work will be needed to confirm this possibility. Studies on the FCV signal have revealed that the reinitiation process occurs in the standard fashion by the criterion of sensitivity to edeine, but it is distinct in being completely independent of eIF4G or the eIF4F complex [13] . Analysis of the MNV signal here provides further evidence that the process deviates from the standard mechanism. First, like BM2 [19] , there appears to be efficient use of non-AUG codons to reinitiate translation, indicating a relaxed requirement for the full complement of initiation factors, which would include eIF1 and eIF1A, thought to play important roles in locating and correct recognition of the AUG start codon [22] [23] . Secondly, in contrast to what has been observed with FCV and BM2, the MNV signal is more resistant to treatment with edeine. Edeine does not inhibit binding of the eIF2/GTP/Met-tRNAi ternary complex to the 40S ribosomal subunit, nor Met-tRNAi/40S complex scanning, but there is a complete failure of AUG codon recognition, so that scanning continues past all AUG codons, and, probably as a secondary consequence, there is no ribosomal subunit joining [24] [25] . The relative insensitivity of the MNV signal to edeine suggests that recognition of the AUG start codon during reinitiation may not require a scanning ternary complex. It is not clear why the FCV and BM2 signals respond differently to edeine, especially as the organisation of the BM2 signal (with regard to the position of Motif 1 and the primary sequence of the stop-start window) is so similar to that of MNV. Another observation that hints at nonstandard reinitiation mechanisms relates to the the translation pattern seen with the MNV49.8 transcript. In this mRNA, the two termination-reinitiation windows (the natural UAAUG and the fortuitous downstream UGAUG) were mutated to eliminate the stop codon in each case. In translations of this mRNA, where termination occurs 30 amino acids downstream of the authentic site, very little termination-reinitiation product was seen, but an additional product was synthesised whose size is consistent with that of a fusion of the two reporter ORFs (asterisked in Figure 6 ). The origin of this protein is uncertain. It could have arisen through a ribosomal frameshift event, although no obvious conventional frameshift signals are present in the region of overlap between the two ORFs [26] [27] . It could also represent the outcome of a failed attempt to terminate and subsequent resumption of translation by ribosomes on the downstream ORF. Further work will be required to elucidate the nature and origin of this product and how it relates to the mechanism of termination-reinitiation. Plasmids used to assay termination-reinitiation were based on the p2luc reporter vector [20] . Sequences encompassing the stopstart signal of MNV (203 bp of sequence information upstream of the VP1 stop codon and 52 bp downstream) and FCV (97 bp upstream of the VP1 stop codon and 14 bp downstream) were generated by PCR (using Pfu polymerase [Roche]) from, respectively, plasmids pT7:MNV (kind gift of Dr Ian Goodfellow, Imperial College, London) and pSG-2/3* [13] , a kind gift of Dr Tuiya Pöyry, University of Cambridge. The PCR products and p2luc were digested with SalI and BamHI and ligated together. Sequences were confirmed by dideoxy sequencing (using the facility at the Department of Biochemistry, University of Cambridge). The influenza B termination-reinitiation assay plasmid (p2luc-BM2wt; 250 bp upstream of M1 stop-codon, 18 bp downstream) was described previously [19] . Site-directed mutagenesis was performed using the Quikchange II site-directed mutagenesis kit (Stratagene) according to manufacturer's instructions. For large deletions (greater than 48 bp) a modification of the manufacturer's protocol was used with the primers containing ,30 bp of complementary sequence either side of the site of deletion, as described previously [28] . Mutagenesis to introduce insertions longer than 6 bp was performed in two steps [29] , by first subjecting mutagenesis reactions (containing either the sense or antisense primer) to three cycles of PCR, then mixing the reactions and performing a further 18 cycles according to manufacturer's instructions. Reporter plasmids were linearised with HpaI and capped run-off transcripts generated using T7 RNA polymerase as described [30] . Messenger RNAs were recovered by a single extraction with phenol/chloroform (1:1 v/v) followed by ethanol precipitation. Remaining unincorporated nucleotides were removed by gel filtration through a NucAway spin column (Ambion). The eluate was concentrated by ethanol precipitation, the mRNA resuspended in water, checked for integrity by agarose gel electrophoresis and quantified by spectrophotometry. Unless otherwise stated, mRNAs were translated in FlexiH rabbit reticulocyte lysate (FlexiHRRL, Promega) programmed with 50 mg/ml template mRNA. Typical reactions were of 10 ml and composed of 60% (v/v) FlexiHRRL, 20 mM amino acids (lacking methionine), 500 mM MgOAc, 2 mM DTT, 5U RNAse inhibitor (RNAguard, GE Healthcare Life Sciences), 130 mM-160 mM KCl (optimised for each batch of FlexiHRRL) and 0.2 MBq [ 35 S]methionine. Reactions were incubated for 1 h at 30uC and stopped by the addition of an equal volume of 10 mM EDTA, 100 mg/ml RNase A followed by incubation at room temperature for 20 minutes. Samples were prepared for SDS-PAGE by the addition of 10 volumes of 2X Laemmli's sample buffer [31] , boiled for 3 minutes and resolved on 12% SDS-PAGE gels. The relative abundance of products on the gels was determined by direct measurement of [ 35 S]methionine incorporation using a Packard Instant Imager 2024. A plasmid encoding the putative termination-reinitiation signal of MNV (p2luc-MNVwt) was modified by site-directed mutagenesis to include a T3 RNA polymerase promoter 30 bp upstream of the minimal required viral sequence generating plasmid p2luc-MNV-T3. RNA for structure mapping was prepared by in vitro transcription of BamHI-digested p2luc-MNV-T3 using T3 RNA polymerase. Transcription reactions were performed on a 200 ml scale essentially as described [30] . Structure mapping was performed using a 59 endlabelling procedure as described previously [30, 32] . All probing reactions were performed in a final volume of 50 ml and contained ,40,000 c.p.m. 59 33 P-end-labelled transcript, 10 mg Escherichia coli rRNA and the relevant enzymatic or chemical probe. Further details are provided in the legend to Figure 3 .

Interaction of the HIV-1 frameshift signal with the ribosome

Ribosomal frameshifting on viral RNAs relies on the mechanical properties of structural elements, often pseudoknots and more rarely stem-loops, that are unfolded by the ribosome during translation. In human immunodeficiency virus (HIV)-1 type B a long hairpin containing a three-nucleotide bulge is responsible for efficient frameshifting. This three-nucleotide bulge separates the hairpin in two domains: an unstable lower stem followed by a GC-rich upper stem. Toeprinting and chemical probing assays suggest that a hairpin-like structure is retained when ribosomes, initially bound at the slippery sequence, were allowed multiple EF-G catalyzed translocation cycles. However, while the upper stem remains intact the lower stem readily melts. After the first, and single step of translocation of deacylated tRNA to the 30 S P site, movement of the mRNA stem-loop in the 5′ direction is halted, which is consistent with the notion that the downstream secondary structure resists unfolding. Mechanical stretching of the hairpin using optical tweezers only allows clear identification of unfolding of the upper stem at a force of 12.8 ± 1.0 pN. This suggests that the lower stem is unstable and may indeed readily unfold in the presence of a translocating ribosome.

The ribosome translocates mRNA and bound tRNA molecules accurately in order to maintain the reading frame. This process results in movement of the ribosome along the mRNA by three nucleotides toward the mRNA's 3 0 -end. Translocation of mRNA and tRNAs is a property of the ribosome itself (1, 2) , however binding of elongation factor G (EF-G) and subsequent hydrolysis of GTP strongly catalyzes it (3) . Although the ribosome acts as its own helicase, stable folded structures within the coding regions of mRNA affect the rate of translocation, and more seriously, may trigger a change of reading frame (4, 5) . Such frameshifting mRNA elements play a crucial role in the translational control of viral proteins via -1 programmed ribosomal frameshifting (-1 PRF) where the reading frame has shifted by one base toward the mRNA 5 0 -end. The -1 PRF requires both the mRNA slippery sequence at the ribosome coding sites as well as a downstream structural element that resists unfolding, representing a physical barrier to the mRNA translocation machinery. While in many cases the downstream barrier constitutes a hairpin (H)-type pseudoknots (6, 7) , on a rare occasion it can also be a simple stem-loop structure (8) (9) (10) (11) . It is of interest to note that although these pseudoknots or stem-loops also trigger ribosomal pausing at the slippery site, consistent with notion of them acting as physical barriers, the extent of pausing shows no correlation with frameshift efficiency (12) . The crystal structure of the ribosome in complex with mRNA has revealed that the mRNA is in a singlestranded conformation in the narrow downstream tunnel (13, 14) . The ribosome therefore has to unwind mRNA secondary structure through its mRNA helicase activity (15, 16) . A mechanistic basis for mRNA helicase activity has been proposed involving ribosomal proteins S3, S4, S5 at the mRNA entrance and rotational movement of the head of the 30 S subunit (13, 15) . The 9 Å and the torsional restraint models (17) propose that -1 PRF is dependent on the mechanical tension induced when a pseudoknot resist unfolding by a moving translating ribosome. Possible effects of such tension were directly observed in the cryo-electron microscopic (Cryo-EM) images of eukaryotic ribosomes stalled in the process of -1 frameshifting in complex with eEF2, tRNA and a frameshifting mRNA pseudoknot (18) . The opposing actions of translocation catalyzed by eEF2 and resistance to unfolding by the mRNA strand generate strain that deforms the P-site tRNA which may weaken the codonanticodon interaction and promote the shift by one nucleotide into the 5 0 direction. The notion that mechanical stability of mRNA structural element is crucial for -1 PRF, has triggered mechanical unfolding experiments of individual mRNA pseudoknot, mutants as well as some of the constituent hairpins using optical tweezers hairpins (19) (20) (21) . Some of these experiments suggest that frameshift efficiencies correlate with unfolding forces rather than the free-energy difference between the folded and unfolded state (21) . This indicates that -1 PRF is kinetically controlled, as has been proposed previously (22) . In HIV-1, translational frameshifting leads to synthesis of the Gag-Pol fusion protein which gives rise to the viral protease, reverse transcriptase (RT) and integrase. This HIV-1 RNA frameshifting signal is a potential target for antiviral therapy (23) (24) (25) (26) . The exact structure of this RNA frameshifting signal has been the subject of debate. Jacks and collaborators initially proposed that it is a stem-loop structure downstream of the slippery sequence that is essential for efficient frameshifting (4) . Alternative structures have subsequently been proposed-reviewed by Brierley and Dos Ramos (27)in which the stimulatory RNA folds as a pseudoknot (28, 29) ; a pseudoknot with an RNA triple helix motif (30) and two-stem helix containing a three-purine bulge (8) . Recently however, two independent nuclear magnetic resonance (NMR) studies have shown that, in the absence of the ribosome, the fold is a long hairpin ( Figure 1 ) with an internal three-nucleotide bulge (9, 10) . A more recent structure-function analysis of the ribosomal -1 FS signal of two human HIV-1 isolates (31) favors the two-stem helix model of Dulude et al. (8) . The internal loop of the long stem-loop introduces a distinct bend between the lower and upper helical regions, a structural feature which, remarkably is also often seen with frameshifting pseudoknots. It has been proposed that the lower stem and the bend serve to initiate contacts between the upper stem-loop and the ribosome. Subsequently the lower stem melts allowing the slippery sequence to bind at the decoding site (10, 32) . Based on the identification of position +11 as the limit of accessibility of an RNA double helix approaching the ribosome (15), we previously proposed that the upper segment of the lower stem and the bulged region could be structured and/or contact the ribosomal surface (9) . NMR studies also pointed out that the upper stem, rich in conserved G-C Watson-Crick base pairs, is highly stable whereas the bulge region and the lower stem are much less so, and may readily unfold/melt. We therefore decided to unfold individual HIV-1 hairpins using optical tweezers, which complement existing methods, such as thermal denaturing, chemical probing or NMR spectroscopy, that address local features within the context of the entire global RNA structure. In principle, optical tweezers aid in applying forces locally to a folded RNA molecule in a way that may be more similar to in vivo conditions than, for example, thermal or chemical denaturing. In the case of a hairpin, the mechanical force will act locally on the 5 0and 3 0 -ends of the RNA unzipping the stem base-pair by base-pair toward the loop. A similar situation may be found in the ribosome where the translocation movement will generate a force pulling the 5 0 -end of the mRNA inside the ribosome. Questions, however, remain about the structure of the HIV-1 frameshift signal when ribosomes are present and bound at or downstream of the slippery sequence. Here, using Escherichia coli ribosomes, we address this question and probe possible interactions of the HIV-1 frameshift signal with the ribosome by using toeprinting and chemical probing assays. The ability of this eukaryotic mRNA frameshifting signals to promote -1 PRF in the prokaryotic translational machinery has been previously demonstrated (33, 34) . The translocation of mRNA was assayed by toeprinting as described (35, 36) . mRNA (1 mM) was annealed to primer (2 mM) in 50 mM K-Hepes (pH 7.0) and 100 mM KCl by heating to 90 C for 1 min and placing at room temperature until the temperature reached 45 C. To form the complexes, tight-couple ribosomes (2-5 mM) (37) from E. coli MRE600 were added to 0.6 mM of mRNA in 60 mM NH 4 Cl, 10 mM Tris-Acetate (pH 7.4) and 20 mM MgCl 2 and incubated at 37 C for 10 min. A first tRNA (4 mM) was added to fill the P site by incubation at 37 C for 10 min and aliquot (0.6 pmol mRNA) was removed to ice for later extension. A second tRNA (4 mM) was added to fill the A site by incubation at 37 C for 10 min and aliquot (0.6 pmol mRNA) was removed to ice for later extension. EF-G was added in buffer (50 mM Tris-HCl (pH 7.6), 20 mM MgCl 2 , 100 mM NH 4 Cl and 1 mM DTT, 1.5 mM GTP) such that the final concentrations of GTP and EF-G were 300 and 1 mM, respectively. Reactions were incubated at 37 C for 10 min, and aliquots (0.6 pmol mRNA) were removed from each reaction lacking (-G) or containing (+G) EF-G. Each of the aliquots was then extended in parallel (38) with (5 0 -CTTTATCTTCAGAAGAAAAACC-3 0 ) primer, and the product were resolved by 8% denaturing PAGE. Stepwise translocation experiments were done as previously described (15) . Tight-couple 70 S ribosomes (1 mM final concentration) from E. coli MRE600 (39) were incubated with mSP-HIV-1 mRNA (1 mM) in 30 ml binding buffer (10 mM Tris-HCl (pH 7.4), 60 mM NH 4 Cl, 10 mM Mg(OAc) 2 , 6 mM b-ME) for 10 min at 37 C, followed by addition of tRNA Phe (1 mM) and a further 10 min incubation to fill the P site. Aliquots (5 ml) of this reaction were then added to separate tubes containing either GTP (600 mM) (F), GTP + tRNA Leu (1 mM) (L), GTP + tRNA Leu , EF-G (1 mM) (L'), GTP + tRNA Leu , EF-G + tRNA Gly (1 mM) (G), GTP + tRNA Leu , EF-G + tRNA Gly + tRNA Lys (1 mM) (K), GTP + tRNA Leu + tRNA Gly + tRNA Lys (K'), in binding buffer (final volume, 10 ml). These tubes were incubated for 10 min at 37 C and then placed on ice for the primer extension in toeprinting analysis (4 ml). To footprint, binding of mRNAs was performed by incubating 70 S ribosome (in the range of 50-750 pmol according to the mRNA tested) with mRNA (10-50 pmol) in 50 ml reaction buffer A (20 mM MgCl 2 , 150 mM NH 4 Cl, 80 mM potassium cacodylate, pH 7.2) at 37 C for 10 min. A first tRNA (320 pmol) was added to fill the P site by incubation at 37 C for 10 min. A second tRNA (320 pmol) was added to fill the A site by incubation at 37 C for 10 min. mRNA-tRNA-70 S complexes were then purified by ultrafiltration (MICROCON YM-100 100 000 Da, Fisher scientific LABOSI). The ternary complex was diluted in 250 ml of buffer A and distributed into 50 ml aliquots (2 pmol mRNA). Chemical probing (38) was performed by addition of 2, 4 or 8 ml dimethyl sulfate (DMS; 1:10 dilution in 95% ethanol), 2, 4 or 8 ml kethoxal (KE; 19 mg/ml in H 2 O) on 50 ml aliquot, followed by incubation at 37 C for 10 min. All modification reactions were stopped by addition of 150 ml 95% ethanol and 5 ml 3 M sodium acetate followed by rapid mixing. KE-modified samples were adjusted to 25 mM potassium borate (pH 7.0). The pellets were resuspended in 200 ml of 0.3 M sodium acetate, 2.5 mM EDTA and 0.5% SDS (with addition of 25 mM potassium borate for KE samples), extracted three times with phenol, twice with chloroform and resuspended in 10 ml H 2 O (for DMS samples) or in 10 ml 25 mM potassium borate (for KE samples). Primer extension reactions were performed as described (38) . Purified tRNA Lys , tRNA fMet and tRNA Phe were purchased from Sigma, tRNA Gly and tRNA Leu were gracefully donated by Henry Grosjean. Messenger RNAs were prepared by in vitro transcription. Plasmid pGENE32 is pUC118 containing a region of phage T4 gene 32 (40) from nucleotide position -54 to +84 (where +1 is the translational start) downstream of an engineered T7 promoter sequence. The introduction of slippery sequence in pGENE32 was performed by sitedirected mutagenesis kit (Stratagene). Transcripts with stem-loop were obtained by in vitro transcription of synthetic genes flanked upstream by T7 RNA polymerase promoter region and downstream by a BamHI restriction site. The synthetic genes were constructed by shotgun ligation of 10 DNA fragments (24-30-mers) covering both strands and ligated in the KpnI and BamHI sites of pGENE32. All transcripts have been purified by denaturing PAGE. His-tagged EF-G was purified from pET24b-fusA in E. coli BL21(DE3) as described (41) . RNA was synthesized from a template obtained by polymerase chain reaction (PCR) from bases 3821 to 628 of the pBR322 DNA plasmid, where the frameshifting RNA signal from HIV-1 was cloned into the EcoRI and HindIII restriction sites and a T7 promoter was appended to the template in the course of the PCR reaction (42) . The DNA components of the handles were prepared by PCR from pBR322. Handle A (pBR322 bases 3821 to 3) was biotinylated, and one of the primers used to amplify handle B (pBR322 bases 30 to 628) was purchased with a 5 0 digoxigenin group. RNA and DNA handles were resuspended in 10 mM sodium phosphate buffer (pH 6.4), and incubated at a ratio of $1:1:1 at 90 C for 1 min and transferred to room temperature to cool down gradually, and subsequently diluted to a final concentration (of RNA) of $1 mM. Five microliters of anti-digoxigenin-coated polystyrene beads (0.3 nM; diam. 0.49 mm) were mixed with 1 ml of the DNA-RNA hybrid ($1 mM) in binding buffer (10 mM Tris buffer [pH 7.0], 250 mM NaCl, 10 mM MgCl 2 , 0.4% w/v BSA), and incubated at 4 C for overnight on a rotator. Sample cells were preassembled prior to use. Two thin strip spacers (thickness $200 mm) were positioned $5 mm apart on the center of a pre-cleaned microscope slide, and epoxy applied at the outer edges of the spacers. A streptavidin functionalized 24 Â 40 mm, no. 1.5 coverglass (Xenopore) was then put on the top of the slide. Before introducing the bead and RNA mixture, the sample cell was surface-coated with acetylated BSA by incubation with binding buffer for 30 min at room temperature and washed with 1 ml of the binding buffer, to prevent any sticking of beads to the surface. The bead and RNA mixture was then introduced into the sample cell, and incubated for 30 min at room temperature, and finally washed with 1 ml of binding buffer to remove any unbound beads and RNA. Molecules were stretched in 10 mM Tris buffer (pH 7.0), 250 mM NaCl, 10 mM MgCl 2 , or alternatively in 10 mM Tris-acetate (pH 7.4), 60 mM NH 4 Cl, 6 mM b-mercapto, 20 mM MgCl 2 . Unfolding/refolding parameters were statistically indistinguishable for both conditions. The spring constant of the optical tweezers was 0.1-0.2 pN nm -1 . The extension of the unfolded single stranded HIV-1 hairpin, x SS was computed as x SS (F) = Áx + L HP for convenience, with F = (F 1 +F 2 )/2 the unfolding force and Áx the increase in extension ( Figure S2 ). The increase in contour length (expressed in number of nucleotides) was subsequently computed using the worm-like chain model for polymer elasticity assuming a stretching modulus of 1000 pN, a persistence length of 1 nm (43,44) and inter-phosphate distance of 0.59 nm. Alternatively, one may choose to do the computation taking into account F 1 and F 2 explicitly, where it is found that Þ þ L HP . Then however, one needs to fit the section of the force versus extension curve up to the unfolding event with a worm-like chain model in order to compute the extension of the handles, x H F 2 ð Þ at F 2 . When performing this more laborious analysis on a subset of our data, we only find a 1-2 nucleotide difference in contour length compared to the analysis that utilizes the applied approximation Standard free-energies at zero force were computed for each trajectory according to Fdx using the worm-like chain model with persistence length of 1 nm and stretching modulus of 1000 pN (43) . We used a toeprinting assay to monitor the position and/ or structure of the HIV-1 hairpin during the movement of the ribosome along mRNA. This primer-extension inhibition assay has been shown to be a powerful tool for mapping the position of mRNA within 30 S and 70 S ribosomal complexes that contain tRNA (45) . When RT encounters the ribosome (a so-called hard-stop), it terminates cDNA synthesis thereby generating a highly specific toeprint. Alternatively, RT may also stall at mRNA structural elements further downstream from the ribosome that prove too hard for RT to unwind, resulting in what is generally called an extended toeprint. We first demonstrated that under simple experimental conditions (where tRNA was non-enzymatically delivered at the ribosomal A site), toeprinting allows localization of the ribosome on the mRNA containing the wild-type HIV-1 slippery sequence (mSP-HIV-1) but not the downstream hairpin. The sequences and secondary structures of the mRNA constructs used are derived from T4 gene 32 mRNA in which we introduced the HIV-1 frameshifting signals as shown in Figure 1 . In the case of the mSP-HIV-1 construct, two distinct toeprints are observed at +15 and +16 when tRNA Phe is bound at the P site ( Figure S1A , lane F), in correspondence with the two possible reading frames created by the slippery sequence. A third toeprint at +17 appears when tRNA Leu subsequently binds at the A site ( Figure S1A , lane F 0 ). This is thought to be due to a conformational change in the ribosomal complex following A-site binding so that a single position of the tRNA in the A site results in a doublet of bands (at positions +16 and +17) (46, 47) . It is important to note that the signal at +15 did decrease, indicating that binding of the tRNA Leu at the A site assisted in the positioning of the mRNA with tRNA Phe preferentially bound to the u +1 u +2 u +3 codon adjacent to the leucine codon u +4 u +5 a +6 . After EF-G catalyzed translocation of tRNA Leu to the P site ( Figure S1A , lane L), the toeprints at +16, +17 were moved to +20 and +21, corresponding to a 4-nt translocation event. We cannot conclude if a toeprint at +19 exists since this band is also present in the control lane (without ribosome). However, if it exists, this toeprint is very weak. When tRNA Gly is subsequently added to the EF-G containing reaction mixture ( Figure S1A , lane G), translocation proceeds further giving the expected toeprints at +22 and +23 for tRNA Gly bound to P site and paired with the g +7 g +8 g +9 codon. A weak toeprint at +24 may indicate a low population of tRNA Gly bound to P site and paired with the g +8 g +9 a +10 glycine codon in the +1 frame. The addition of tRNA Lys (Figure S1A, lane K) triggered translocation leading to a toeprint at +25, which corresponds to a post-translocation complex with tRNA Lys in the P site bound to the a +10 a +11 g +12 lysine codon and leaving an empty A site. We then applied this assay to a gene 32 message (mSP-SL-HIV-1), which contains the mRNA frameshifting signal (Figure 2A) . In order to avoid any unnecessary ambiguity in interpreting the toeprints, we substituted the phenylalanine codon (u +1 u +2 u +3 ) for a methionine codon (a +1 u +2 g +3 ) (mSP-SL-HIV-1; Figure 1 ) to avoid complications due to the presence of the slippery site. This allows tRNA binding in only a single unique reading frame. When tRNA fMet was bound at the P site, a doublet of toeprints at positions +U16/+G17 was detected ( Figures S2 and 2A) . We note the existence of a stop at position +17 in the control lane (in the absence of ribosomes). However, the signal in presence of ribosome is substantially stronger and therefore is interpreted as a ribosomal toeprinting signal. tRNA Leu4 was subsequently bound to the A site ( Figures S2, lane 3, and 2A , lane M 0 ) which resulted in the disappearance of the +16 toeprint. The +16 and +17 toeprints are hard stops, independent of the reverse transcription activity indicating that the mRNA secondary structure unfolds during cDNA synthesis as previously seen for a stem-loop (48) and pseudoknots (49) . However, in addition to these hard stops downstream extended toeprints can also exist indicating that mRNA interactions with the ribosome 5'-(N 42 ) aAGGAaauaaa aug uuu aaa cgu aaa ucu (N 68 )-3' can extend beyond the usual 15 nucleotides buried in the ribosome-mRNA track (48, 50) . Unlike hard stops, extended toeprints are generally dependent upon change of temperature, RT concentration, or the source of RT. Therefore, we varied the temperature and also tested different RTs to look for extended toeprints in the case of mSP-SL-HIV-1. Two additional reverse transcription stops were identified within the 3 0 region of the stemloop at positions +47 (very weak) and +43 (Figures 2A and 3A ). These extended toeprints are weak in comparison to the hard stop at +17 but are absent in the control lane ( Figure 2A , lane C, in the absence of the ribosome). The +43 and +47 toeprints indicate that a fraction of the RT molecules halted when the enzyme approached or encountered bulge region in the HIV-1 hairpin. In order to test the contribution of the bulge region to the extended toeprint, we changed the GGA-bulge by a CCC-bulge (mSP-SL-CCC-HIV-1, Figure 1 ). The intensities of both toeprint signals (at +43 and +47) decreased ( Figure 2B ) and were reproducibly of lower level than the RT stops found in the control lane -70 S. Interestingly, the substitution of the three purines in the bulge by pyrimidines also decreases frameshifting efficiency (8, 32) . U-G C-G C-G U-A U-A C-G C-G G-C G-C U-A C-G U-A A-U G-U A-U A-U G-C G-U A A G 12 G C A A46 We subsequently monitored the position of mSP-SL-HIV-1 within the ribosome allowing a single or multiple rounds of EF-G catalyzed translocation. The signal at positions +43 and +47 disappeared to give a new toeprint at +39 (Figure 2 , lane L, and 3B), indicating that the lower stem readily gave way upon just a single translocation step. An identical result was observed in the mSP-SL-HIV-1 RNA which has a wild-type slippery sequence (first codon is u +1 u +2 u +3 ) indicating that codon substitution to a methionine codon did not affect the observed toeprints except for the +47 signal that is ambiguous ( Figure S1B ). Furthermore upon a single translocation, the +16/+17 toeprint signals disappeared without, however, any occurrence of a new toeprint at position +19 (lane L). The disappearance of the mSP-SL-HIV-1 +16/+17 toeprints upon addition of EF-G is intriguing. We also observed the same phenomenon with the construct containing the wild-type slippery sequence ( Figure S1B ), and interestingly also with an mRNA containing a pseudoknot (BWYV) bound to ribosome (51) . This phenomenon is characteristic to mRNAs containing downstream structural elements, however its cause still need to be elucidated. Ribosomes are known to change conformation upon EF-G binding (52) (53) (54) (55) . One may speculate that this either blocks RT access directly or stabilizes the mRNA structure. Any subsequent addition of tRNA Gly (lane G) and tRNA Lys (lane K) in presence of EF-G did not further affect the position of the extended toeprint, which remained at +39, indicating that further movement of the mRNA through the ribosome was impaired. We next tested an mRNA with an additional codon between the AUG and the start of the lower stem (mSP-Tyr-SL-HIV-1 RNA) extending the spacer, which should at least allow for one round of translocation to be visualized before the ribosome stalls. The toeprints at the new positions A+16/U+17 were unambiguously identified ( Figure S3 ), which places the boundary region contacting the ribosome upstream from the bulge region. These bases are in the upper stem for the mSP-SL-HIV-1 RNA. In Figure 2C , where experimental conditions were tuned as to specifically detect the extended toeprints, the toeprint signals in the region +17 prove somewhat weaker than in Figure S3 , but are always present. As expected, with the longer spacer, a band at +19 appeared upon addition of EF-G ( Figure 2C, lane L) . Interestingly, the extended toeprints did not change and appeared at the exact same positions at A + 46 and U + 50 (corresponding to the A + 43 and U + 47 in mSP-SL-HIV-1 RNA) ( Figure 2C ). Subsequent addition of EF-G and tRNA Tyr and tRNA Gly produced the same changes in the toeprint pattern as for the mSP-SL-HIV-1 RNA (a shift of 4-nt from +46 to +42). We note that in this case the +19 toeprint remains the strongest toeprint in the upper region of the gel (by comparison with the +17 toeprint) ( Figure 2C , lanes Y and G) indicating that translocation, as before, seems to be impaired with the downstream upper stem. Toeprinting assays fail to provide information on more detailed structural changes that the mRNA might undergo upon binding to the ribosome. Therefore, a footprint assay was performed to obtain structural information upon the interaction of mSP-SL-HIV-1 mRNA with 70 S ribosome-tRNA complexes. In these complexes the ribosome is positioned with the slippery sequence at the decoding site. Results of the chemical probing with KE, 1-cyclohexyl-2-morpholino-carbo-diimide-bmetho-ptoluene sulfonate (CMCT) and DMS are shown in Figure 4 . For free mSP-SL-HIV-1 mRNA the chemical modification patterns and levels of reactivities are identical to those previously published (9) . Nucleotides in the apical ACAA teraloop as well as A +43 and A +44 from the bulge are found to be reactive to DMS (Figure 4) , while nucleotides G +41 and G +42 were reactive to KE. In the lower stem most of the guanine, adenine and uracil bases are accessible to the chemical probes indicating poor stability (9) . Subsequently, mSP-SL-HIV-1 mRNA was probed in complex with the ribosome and tRNA. For each chemical probe tested, bands that were present in a control lane of unmodified mRNA incubated with ribosome were not taken into account. In the complex, the characteristic strong protections at the guanine nucleotides in the Shine-Dalgarno sequence (G -11 , G -10 ) are clearly seen. Base A +1 , of the methionine codon is protected from chemical modifications demonstrating, as previously described (56), Watson-Crick pairing with the anticodon of tRNA fMet . Interestingly, nucleotide A +6 experienced an increase in reactivity similar to what has previously been seen for a mRNA containing a hairpin (selenocysteine incorporation sequence SECIS) in complex with 30 S subunit and tRNA fMet (57) . In the spacer and hairpin regions, most of the changes in chemical reactivities were detected in the lower stem that is supposed to be close to the ribosomal surface. Bases G +7 , G +8 and G +12 were protected from modification by KE (Figure 4) . The reactivity of G +9 toward KE slightly increased in a way similar as A +10 toward DMS. Since toeprinting experiments suggested reduced structural stability of the lower stem, we investigated the mechanical stability of the HIV-1 hairpins. Individual hairpins, sandwiched between two differentially endlabeled hybrid DNA-RNA handles were unfolded using optical tweezers (42) . Molecules were tethered between a streptavidin-coated glass cover slip and anti-digoxygenin coated polystyrene beads (diameter 1 mm), and stretched by moving the piezo-actuated microscope stage while holding the bead with optical tweezers. Force versus extension curves were computed taking into account this experimental geometry (58) . Stage velocities were 67 nm/s, so that the system remained at or close to thermodynamic equilibrium as confirmed by the overlap of the stretching and relaxation force versus extension curves ( Figure 5B ). In addition, repeated unfolding and folding can be observed within single force versus extension curves ( Figure 5B) , a further indication that the loading rate is sufficiently low as to assure thermodynamic equilibrium. The increase in contour length (expressed in number of nucleotides) upon unfolding was subsequently computed using the worm-like chain model for polymer elasticity assuming a stretching modulus of 1000 pN, a persistence length of 1 nm (43,59) and a inter-phosphate distance of 0.59 nm (see 'Materials and Methods' section). Results are summarized in Figure 5C , yielding a mean increase of 25.3 AE 3.4 nm (mean AE SD) nucleotides, consistent with the contour length of the upper stem of the HIV-1 hairpin. The mean unfolding force is 12.8 AE 1.0 pN (mean AE SD), whereas the standard free-energy change at zero force was found as ÁG = 10 AE 2 kcal/mol (mean AE SD). Refolding statistics are summarized in the Figures S4 and S5 and yield a decrease of contour length of 26.0 AE 4.7 nm (mean AE SD) at an average refolding force of 13.1 AE 1.1 pN (mean AE SD), and ÁG = 11 AE 3 kcal/mol (mean AE SD), essentially unchanged from the unfolding statistics, as one would expect when at thermal equilibrium. Extended toeprints (albeit weak ones) at +43 and +47 in the pre-translocation state indicate that a fraction of RTs was incapable of unwinding part of the lower stem and bulge region when encountering the HIV-1 hairpin bound at the ribosome. Since such toeprints do not occur with free hairpins, contacts with the ribosome (34) may be responsible for these signals. Extending the spacer sequence by an extra codon yields an identical toeprint; not surprising as it is hard to imagine that a longer spacer would stand in the way of forming ribosomal contacts. This will be discussed further below. Furthermore, we showed that these toeprints are dependent on the structure of the bulged region of the frameshifting signal. The subsequent replacement of the GGA-bulge with a CCC-bulge, known to reduce -1 PRF efficiency (8, 32) , practically erases the +43 and +47 toeprints, indicating such a mutation affects either contacts with the ribosomes or stability of the hairpin itself. We tend to favor the former possibility as mechanical unfolding of free hairpins indicates the lower stem is fairly weak to begin with. Certainly it remains possible that contact with the ribosome can stabilize part of the lower stem. If such stabilization was to occur, it does not prevent further translocation as a single cycle of EF-G catalyzed translocation moves the toeprint to +39 providing direct evidence for melting of the lower stem. Interestingly, the extended toeprint at +39 remains even under conditions where further translocation is allowed. This suggests that the upper stem is capable of inhibiting translocation of a sizeable fraction of ribosomes. Our mechanical unfolding study supports the notion of a weak lower stem since no transition in the force versus extension curves indicative of unfolding of solely the lower stem have been observed. Although on occasion we suspect such a transition (by visual inspection of the force versus extension data, typically at forces $6 pN and lower), the increase in extension in those cases should be $4-5 nm (by theory), too small to reliably distinguish from thermal fluctuations at those force levels. We note that were these transitions to occur in the steeper part of the force versus extension curve (F > 6 pN), we should be able to detect a shift of the curve toward the right consistent with unfolding of the lower stem. However, the increases in contour lengths ( Figure 5C ) observed are consistent with unfolding of solely the upper stem. Not on any occasions have we observed increases in contour lengths consistent with simultaneous unfolding of the lower and upper stems. Since unfolding is hierarchical, the lower stem has to unfold before the upper stem does. We conclude that the lower stem therefore has much reduced mechanical stability compared to the upper stem. Quantitative assessment of the stability of the lower stem is currently not possible in the existing experimental geometry, and requires higher resolution experimental designs (60) , if at all possible. The question naturally arises how these findings accord with the observed +43 and +47 toeprint and the dependence of frameshifting upon the lower stem and bulge regions. It seems unlikely that tensions as low as a mere 6 pN and below, can trigger a frameshift (61) . Therefore, it is appealing to consider that unfolding of the hairpin in the presence of ribosomes will differ significantly from that when ribosomes are absent. When adhering to a tension-dependent mechanism of -1 frameshifting this suggest that contacts with the ribosome's exterior surface may indeed stabilize the part of the lower stem of the hairpin. Single-molecule mechanical unfolding experiments in the presence of ribosomes are considered, but beyond the scope of this article. Our chemical probing experiments provide further insight into the structure of the HIV-1 frameshift signal when bound to the ribosome. The chemical reactivity pattern indicates that almost the entire structure of the hairpin is maintained. Nucleotides from the ACAA tetraloop and the GGA bulge region remained reactive to chemical probes whereas nucleotides in the upper stem were unreactive. As expected, most of the changes in chemical reactivity of the bases are concentrated in the lower stem. The changes observed in the spacer at positions +7, +8 are consistent with this segment of the RNA being engaged in the 30 S mRNA tunnel (13) . Further downstream in the spacer sequence, nucleotides G +9 , A +10 and G +12 that are expected to be located in the vicinity or at the ribosomal helicase center indeed experienced a change in reactivity. In this segment of the RNA, the DMS reactivity of nucleotide A +11 remained mostly unaffected by the presence of the ribosome. Unfortunately, our chemical probing experiments could not provide useful information on the 3 0 strand of the lower stem. In summary, the data demonstrate that the 5 0 region of the lower stem experiences structural changes when contacted by the ribosome. A number of other observations, although of no direct consequence for the interpretation of the extended toeprints, are of interest. First is the observation of the occurrence of an apparent 4-nt translocation step during transfer of tRNA Leu from A site to P site. The disappearance and appearance of respectively the +43 and +39 toeprints indicate that the addition of EF-G triggered a movement of the mSP-SL-HIV-1 RNA inside the ribosome of four nucleotides. However, we note that this 4-nt shift is detected on the 3 0 strand of the mRNA hairpin that is located outside the ribosome and therefore may not necessarily present a 1:1 reflection of what happened at the coding site. On the other hand, single translocation cycles with only the slippery sequence (mSP-HIV-1) also show a 4-nt step. In this experiment we analyzed the toeprint signals that result from A site, P site tRNA binding and subsequent EF-G catalyzed translocation. The mSP-HIV-1 toeprint at +15 corresponds to tRNA Phe that recognizes the overlapping codon, one nucleotide upstream (u -1 u +1 u +2 ). Binding of tRNA Phe to the u -1 u +1 u +2 codon results in an optimal spacing of seven bases between the SD sequence and the start codon (62) . The toeprint at +16 corresponds to tRNA Phe paired to codon 1 (u +1 u +2 u +3 ). Interestingly, overlapping codons u +2 u +3 u +4 and u +3 u +4 u +5 are not pairing with tRNA Phe because no toeprints were observed at +17 and +18. This is most likely due to unfavorable spacing between the P-site codon and the Shine-Dalgarno sequence (47, 63) . After P-site filling and the first round of translocation, the toeprint at +19 is the expected signal for a post-translocation complex with tRNA Leu in the P site bound at the u +4 u +5 a +6 codon. However, the major toeprint signal is at position +20 suggesting that in the post-translocation complex the P site-bound tRNA Leu is interacting with the u +5 a +6 g +7 codon. This corresponds to a 4-nt translocation event. Subsequent translocation cycles seem to force the toeprints back into register. It remains unclear what gives cause to this behavior. We note that a 4-nt translocation has been observed for punctuated mRNA with an extra, unpaired nucleotide between codons (64). Interestingly, Leger and collaborators proposed that slippage and repairing of Psite-bound tRNA Phe in the -1 frame would leave an unpaired nucleotide between the Phe and Leu codons (32) . But this is unlikely at equilibrium in our assay because A-site binding of tRNA Leu mostly repositioned tRNA Phe into the canonical frame removing any unpaired nucleotide between Phe and Leu codons. The P-site tRNA binding triggered a major RT stop at position +17 while only a weak signal at position +16 (the 'classical toeprint'). Previous toeprinting assays under identical conditions but with different phage T4 mRNAs showed a variety of toeprints signals confined to the +14 to +17 range (45) . Thus toeprinting assays are very sensitive to the type of mRNA tested and we note that the toeprint signals with the mSP-HIV-1 and mSP-SL-HIV-1 mRNAs respectively at +16 and +17 fall into this range. Could the observed difference between the two mRNAs be attributed to the hairpin? With unstructured mRNA, we expect a pulling force on the spacer mRNA out of the ribosome purely based on entropic arguments. Such a force presumably exposes base +16 giving rise to the classical toeprint. In the case of the HIV-1 hairpin, physical interactions of the ribosome with the hairpin may provide some sort of strain relieve allowing +16 to slightly relax back into the ribosome. On the other hand, how could such a mechanism be consistent with +16/+17 toeprint obtained when an additional codon was inserted in the spacer, between the AUG codon and the hairpin? If the hairpin provides strain relieve would one not expect the mRNA to recede further back into ribosome, not only protect +16 but also perhaps +17. However, lacking any quantitative information about tension in the spacer, such extra movement may be too small to also protect +17. These are intriguing possibilities and raise questions that require the design of new experiments, well beyond the scope of this work, for answering. Site-directed mutagenesis and amino-acid sequencing localized the site of frameshifting to the UUA codon of the HIV-1 slippery sequence (65) . The presence of the GCrich upper stem strongly enhances frameshifting efficiencies. This level of mRNA slippage is enhanced further by the bulge and lower stem. We show here that when the slippery sequence is bound at the decoding site, the ribosome directly influences the spacer sequence in the lower stem that very likely enters the mRNA track. In addition, extended toeprints localized in the upper segment of the lower stem support a physical interaction between this region and the ribosome. These results are in agreement with the importance of the lower stem and bulge regions for the -1 frameshift (8, 32, 66) . It is conceivable that this interaction enhances -1 frameshifting at the particularly slippery UUUUUUA sequence (67) .

Outdoor environments and human pathogens in air

Are pathogens in outdoor air a health issue at present or will they become a problem in the future? A working group called AirPath - Outdoor Environments and Human Pathogens in Air was set up in 2007 at University College London, UK with the aim of opening new discussion and creating a research network to investigate the science and impacts of outdoor pathogens. Our objective in this paper is to review and discuss the following areas: What is the source of human pathogens in outdoor air? What current, developing and future techniques do we need? Can we identify at-risk groups in relation to their activities and environments? How do we prepare for the anticipated challenges of environmental change and new and emerging diseases? And how can we control for and prevent pathogens in outdoor environments? We think that this work can benefit the wider research community and policy makers by providing a concise overview of various research aspects and considerations which may be important to their work.

Low moisture and nutrient levels, combined with high levels of ultraviolet (UV) radiation mean that the atmosphere is inhospitable to microbial life [1] . The huge volume of air outdoors compared to air indoors also helps to dilute the concentration of microbes and reduce the level of exposure. Nevertheless, we need to ask: are pathogens in outdoor air a health issue at present or will they become problematic in the future? The multidisciplinary working group, AirPath was organised to review and discuss the problem of pathogens. Four two-day meetings over a period of 18 months from July 2007 have generated contributions from more than 30 participants, over 20 oral presentations, and various round-table discussions, which were recorded on DVD. In order to summarise the wealth of knowledge contributed by the multidisciplinary group, the panel (the authors of the present paper) has arranged the discussion into five themes: 1. What is the source of human pathogens in outdoor air? 2. What current, developing and future techniques do we need? 3. Can we identify at-risk groups in relation to their activities and environments? 4. How do we prepare for the anticipated challenges of environmental change and new and emerging diseases? and 5. How can we control for and prevent pathogens in outdoor environments? Participants were selected on the basis of their involvement, for example, as president or executive committee members of related professional and academic bodies such as the British Aerobiology Federation, the Aerosol Society and the International Association of Aerobiology; their work in related UK organisations, such as the Health Protection Agency (HPA), hospitals and Defence Science and Technology Laboratory (DSTL); or their research in relevant fields such as epidemiology, meteorology, geoinformatics, and natural resource management. Researchers outside the UK were invited to give an international dimension and network of collaboration to the UK participants. Rather than focus on a narrow topic, the aim of the AirPath working group is to explore the complex and multidisciplinary facets of research connected with the outdoor environment and human pathogens in air, analyse their potential implications, and investigate applications of this research for the well being of society. At the first two meetings, participants were given a topic and asked to prepare a one-page literature review to support their presentations; each topic was assigned to two participants from different disciplines in order to ensure cross-disciplinarity of the discussion. These presentations now form the core of our review, and all the discussions from all the meetings were recorded. We have conducted an extended literature search and review to ensure that the content is representative, comprehensive and connected to the five main areas arranged below. Participants discussed four main areas that are known to contribute pathogens to the outdoor air and have proved to be linked with human health -1. Natural environments 2. Engineering environments 3. Agriculture and 4. Waste treatment ( Figure 1 ). As shown in Figure 1 , air pathogens from environmental sources are diverse in terms of type of source and aerosolisation factors, and it is moreover possible that some pathogens are under-reported because a number of them cause similar respiratory symptoms, e.g. coughing and sneezing [2, 3] . Because there are numerous types of pathogens released to the outdoors and given the publishing constraints, the list of pathogens from different sources can be found in the cited references in Figure 1 . In addition, although many environmental pathogens are restricted to limited geographic areas, it may be naïve to assume that pathogens previously restricted to specific locations will not shift with impending global climate change; or we may find that our current knowledge is based on inadequate data and we have yet to discover the actual distribution of a number of these pathogens [2, 4] . The pathogens cited in the references may not yet be relevant to the UK, but could pose a future threat due, for example, to climate change. Moreover, various pathogens can potentially be released from waste treatment facilities; how will the mounting levels of composted green waste, food waste and other traditional landfill materials, as well as increasing bio-solid applications to land, impact on the pathogens in air? What types and how many pathogens are we exposed to? Is our environment changing and impacting on our health? To answer these questions, we require various techniques to support our research, such as sampling, detection and identification, monitoring, transport models, and laboratory experimentation (Table 1) . Kuske reviewed the current and emerging technologies for the study of bacteria in outdoor air in 2006 [5] . We searched the literature from 2006 onwards to extend Kuske's review and paid close attention to the study of viruses. Some studies show that climatological factors can influence viral disease transmission, e.g. respiratory syncytial virus [6] . The environmental sampling of human viruses is therefore an area that we regard as important in order to prepare for new and emerging diseases in our time. Table 1 highlights the diverse skills and knowledge outside the traditional microbiology and aerobiology fields, which can be adopted to understand pathogens in the outdoor air. Kuske's review [5] contributes largely with regard to sampling and detection and identification in Table 1 . Our review has added, for example, the spatial and temporal monitoring issues, epidemiology and computational fluid dynamic models, and the application of laboratory experimental facilities. Moreover, a variety of examples were given on the study of viruses in the air. Environmental sources of pathogens in outdoor air Figure 1 Environmental sources of pathogens in outdoor air. Engineering environments Waste treatment Agriculture Urban environments [12, 16] Ventilation & air conditioning systems [8] Water/Soil/ Biosphere [16] Desert storms [14] Landfills [15] Composting [10, 15] Flooded buildings [13] Sewage treatment plants [7] Disposal of biosolids [17] Crop and livestock production [9, 15] Land application of biosolids [17] Can we identify at-risk groups in relation to their activities and environments? It is important to better understand the risk factors associated with the outdoor air transmission of infections, but the best approach is not always straightforward. Epidemiology is the study of the link between exposure, outcome and confounders, but the measurement of exposure to outdoor pathogens in the air is difficult, as it is problematic to conduct controlled laboratory experiments, for example, to expose people to pathogens within the laboratory. In addition, the technology to obtain accurate organism counts may not yet be available. Moreover, it will be a challenge to evaluate the possible environmental influences such as climatic conditions and proximity to a source, when we assess exposure levels, especially retrospectively [6] [7] [8] . Furthermore, it is not always clear how the outcomes should be measured [9, 10] . Many respiratory infections do not have a definitive causal organism [3] . Infections may be asymptomatic and most infections have numerous subtypes [3, 8] . Age, gender, social class, health, exposure to pollution, and a variety of other factors are the potential confounders that must be addressed in future epidemiology studies. From the AirPath point of view, to carry out an exposure assessment -that is, to estimate and measure the amount of infectious pathogens entering our bodies through inhalation -is already a complex and multidisciplinary science without the added component of using exposure data and disease outcomes to predict risk factors. We think that Air-Path has contributed towards forming a technical network, as shown in the output in Table 1 , as well as a medical network to further clarify a comprehensive and systematic exposure assessment. Sampling Sampling of bioaerosols (bacteria and fungi) is widely reviewed [5] . Bioaerosols can be collected on various media depending on the type of microbial detection, and can be collected according to their size to estimate their deposition on the respiratory system. All of these sampling techniques have pros and cons regarding the issues of size separation, sampling volume and time, biological recovery, and particle removal efficiency as well as the choice of subsequent analytical and detection methods. The sampling and quantification of viruses is less widely studied. One recent study has developed methods for airborne influenza and avian influenza virus, which is currently one of the biggest concerns of public health [11] . Detection and identification of pathogens has changed since the development of different molecular methods and innovative approaches other than culture methods [5] . The existing detection methods can be divided into two levels: generic and specific. Generic detection gives information about whether the particles are biological materials, microbes or living cells, e.g. bioluminescent measurement of ATP using continuous flow luminometer and mass-spectrometry. Specific methods such as micro-arrary and immuno-assays can tell us what kind of microbes are detected and identified. Other new techniques have been proposed for bio-detection, for instance, by characterising the size and shape of bioaerosols, pollens and fungal spores under microscope [18] and analysing fluorescence spectrum of bacteria [5] . It is widely recognised that background biological and chemical materials and their continuous environmental fluctuation will significantly influence monitoring. Air movement, sunlight/UV radiation, humidity, rainfall, and inversions are some of the environmental factors that need to be considered during monitoring. Another consideration is where and when to sample with regard to spatial and temporal relevance [19] . For example, the release of pathogens can cause a significant downwind hazard which requires a wide area and long period of sampling [7] . Transport/Transmission models Epidemiology studies can link disease cases together and develop a disease transport and transmission model [8] . However, it will not always explain the mechanism. Moreover, it requires a significant number of cases in order to develop a model. The use of computational fluid dynamic (CFD) models and tracer gas simulation has demonstrated that the Severe Acute Respiratory Syndrome (SARS) virus can travel and disperse outdoors through air, and became a source of pathogens to other indoor environments [12] . A similar technique has been used in the modelling of aerosols and chemical pollutants in streets, waste treatment facilities, and other pathogen sources outdoors [7] . Because pathogens travel in air, it is inevitable that the biological activity will be influenced by the environment. Data can be collected from field studies to determine the impact of environments on the fate and behaviour of pathogens. However, since the pathogens and environment vary and fluctuate frequently, it is not easy to build this scientific link using field data alone. Some studies have investigated the viability and environmental limits of airborne viruses and bacteria using a rotating drum and controlled climate environmental chambers [20] . Studies from both field and laboratory settings indicate that environments and environmental factors can significantly impact on the fate and behaviour of bioaerosols and health risk. It is generally recognised that the environment is constantly changing, either physically, climatically, socially, or a combination of all three. Not only is the environment ever changing, but the types of diseases are also changing. The presence of new and emerging diseases is one of the most urgent threats to humanity across the globe [3, 11, 12] . Climate change is a high priority issue that everyone is facing; and it may be that environmental pathogens will respond to climate change as well. Climate change not only affects the pathogens in air, but also their source, source strength and aerosolisation mechanisms (e.g. through extreme weather conditions) [13, 14] ; these are fairly unexplored at present. New and emerging diseases are a major health concern because we do not know much about them. We do not yet know how best to prevent and control them and, most importantly, we often do not know how to treat them, but it is likely that genetics and the state of the immune system of the global population plays a key role in disease prevalence [2] . With advances in medical treatment, greater numbers of disease-susceptible groups, such as immuno-compromised individuals are expected to survive longer in the overall population. This trend may change our understanding of pathogens because many unexpected microbes and infection pathways can appear within this transforming population. Once pathogens are in the air, one of the few things we can do to minimise the health risk is to source control (reduce/eliminate the source and its strength and the aerosolisation factors, as well as potential dispersal interception near the source) [15] . Modelling the movement of bioaerosols in order to advise people to prepare or even evacuate is another option for reducing the health risk [10] . Although the physical properties and transport of aerosols can be predicted in most environments, it is not always known how biological properties change in the aerosol dynamics pathway and during outdoor transport (e.g. aggregation, scavenging and deposition). Source control is not always possible, for instance, if the sources of pathogens are unknown. The longer it takes to identify the source, the higher the risk it poses to people. Because of the differences in the type of sources, the nature of the pathogens and their geographical prevalence, various research and policy approaches have been taken by different regulatory bodies. Using Legionellosis as an example, guidelines and regulations are available from various government agencies, cooling tower manufacturers and industrial trade organisations to control and prevent the growth and dispersal of these particular pathogens. Although guidelines and regulations are widely known and implemented, outbreaks of Legionnaire's disease are not uncommon. Some studies report that the pathogens can travel up to 12 km from the suspected source and still cause infections [8] , thus indicating a more complicated relationship between source and disease transmission than was previously understood. There is a significant knowledge gap in understanding and assessing the risk of outdoor pathogens in air. In order to complete the transmission pathway and set up regulation, control and prevention measures, we need to better understand and identify pathogens at their source, identify the aerosolisation mechanisms and dispersal plume, have adequate qualitative and quantitative techniques to detect the pathogens in different media, and understand the deposition process and required dose for infection. As in Skyes et al.'s analysis, a risk assessment framework was used to assess the potential health risks from bioaerosols from composting [10] . Current knowledge of pathogen risk assessment in air is far from complete; without knowing the infectious dose and actual risk to society, there is little motivation on the part of governments or manufacturers to research methods of control. In our complex, changing world, some environmental health problems are not straightforward to identify until a very serious health impact occurs. When health problems do emerge, we may only have a very short period of time and opportunity in which to react, identify and restore the damaged natural environment. As a result, we strongly assert the necessity to understand the environment and its relationship with pathogens in air before health hazards associated with the relevant pathogens can appear. In the present paper we have briefly reviewed and presented our views on various issues. Our next step is to encourage and support focused multidisciplinary research in order to fill the missing knowledge gaps and translate research into practice and policy.

Combination strategies for pandemic influenza response - a systematic review of mathematical modeling studies

BACKGROUND: Individual strategies in pandemic preparedness plans may not reduce the impact of an influenza pandemic. METHODS: We searched modeling publications through PubMed and associated references from 1990 to 30 September 2009. Inclusion criteria were modeling papers quantifying the effectiveness of combination strategies, both pharmaceutical and non-pharmaceutical. RESULTS: Nineteen modeling papers on combination strategies were selected. Four studies examined combination strategies on a global scale, 14 on single countries, and one on a small community. Stochastic individual-based modeling was used in nine studies, stochastic meta-population modeling in five, and deterministic compartmental modeling in another five. As part of combination strategies, vaccination was explored in eight studies, antiviral prophylaxis and/or treatment in 16, area or household quarantine in eight, case isolation in six, social distancing measures in 10 and air travel restriction in six studies. Two studies suggested a high probability of successful influenza epicenter containment with combination strategies under favorable conditions. During a pandemic, combination strategies delayed spread, reduced overall number of cases, and delayed and reduced peak attack rate more than individual strategies. Combination strategies remained effective at high reproductive numbers compared with single strategy. Global cooperative strategies, including redistribution of antiviral drugs, were effective in reducing the global impact and attack rates of pandemic influenza. CONCLUSION: Combination strategies increase the effectiveness of individual strategies. They include pharmaceutical (antiviral agents, antibiotics and vaccines) and non-pharmaceutical interventions (case isolation, quarantine, personal hygiene measures, social distancing and travel restriction). Local epidemiological and modeling studies are needed to validate efficacy and feasibility.

Many countries have developed pandemic preparedness plans in response to the threat from pandemic influenza [1] , to attempt containment of the virus or to reduce the pandemic's impact. The influenza A (H1N1-2009) pandemic has underscored the importance of such plans, with the World Health Organization (WHO) calling for the activation of pandemic plans worldwide [2] . Although the WHO has made public guidelines for developing pandemic plans [3] , the comprehensiveness and standards of pandemic plans differ widely across different countries and continents [4] [5] [6] . To ensure the success of these plans, it is necessary to adopt a combination of different strategies. Although there are existing historical data on the possible success of strategies used in previous pandemics such as personal hygiene, school and workplace closures, and social distancing, these are often anecdotal and difficult to interpret [7, 8] . Mathematical models provide a platform for the assessment of multiple interventions in an environment where individual parameters can be altered. The recent increase in mathematical modeling studies on pandemic interventions suggests the effectiveness of these strategies and provides guidance for policy makers. Although the 2009 pandemic has spread rapidly, these combination strategies can be applied in populations yet to be severely affected, for the second wave, or for the next pandemic [9, 10] . This systematic review aims to determine the individual components that constitute combination strategies, and the quantitative impact of these combination strategies in reducing pandemic spread and morbidity. This study explored available mathematical modeling publications on the effectiveness of combination strategies for an influenza pandemic. To obtain papers on the effectiveness of combination strategies, data for this review were identified by the authors through searches of the PubMed search engine for English language articles and articles translated into the English language. The authors used the following search terms to focus on modeling studies, and those which had a focus on pandemic preparedness and strategies -influenza and pandemic and (preparedness or strateg* or model*); influenza and modeling or modelling. The search included all published articles listed on PubMed from 1990 to 30 September 2009there were few articles on influenza pandemic planning or modeling before this period. Abstracts were reviewed where available to determine if a study met the inclusion criteria and the full manuscript was obtained for further scrutiny. Snowball searches by hand were performed on the reference lists of articles meeting the inclusion criteria to find additional studies. The inclusion criteria were primary mathematical modeling papers that compared and reported the quantitative effectiveness of combination strategies (two or more strategies used together) versus individual strategies for human pandemic influenza. Mathematical modeling papers were those which used quantitative predictive methods to determine the likely impact of strategies, and had descriptions of these methods which could be reproduced or verified. All influenza preparedness strategies were considered, including pharmaceutical and nonpharmaceutical public health strategies. These articles would allow clear comparison on the advantages of combination strategies over and above the impact of individual strategies. An explanation of some of the key strategies are found in the appendix. Mathematical modeling articles that described the effectiveness of multiple singular strategies but did not analyze the quantitative effect of combination strategies were excluded. Articles that referred to general pandemic preparedness without quantitative evidence, or provided only qualitative discussion were also excluded. Reviews without primary data, articles in abstracts without full publication, and unpublished studies were excluded as their methodology and results could not be verified. Mathematical models are based on input variables which are assumptions made based on available evidence in specific scenarios. One important assumption is the reproductive number (Ro), which is the average number of secondary infections generated by a single case in a completely susceptible population. No attempt was made in this review to homogenize data across studies for comparison; on the contrary, the heterogeneity of data provides public health professionals with evidence of the effectiveness of strategies across a wide range of assumptions and scenarios. We have instead listed the different types of models used, and the scenarios, interventions, and countries where they were applied. The search yielded a total of 1,920 papers including overlaps. Of these, 162 used mathematical modeling techniques and on closer review, 144 were excluded based on the exclusion criteria listed in Methods. The remaining 18 studies were included for analysis, together with one additional study identified from the snowball searches ( Figure 1 ). The selected modeling papers that show the effectiveness of combination strategies in increasing the impact of individual strategies are listed in Additional files 1 and 2 [11] . The following sections highlight key findings on the effectiveness of combination strategies in these modeling studies on pandemic influenza. Zoonotic influenza such as H5N1 influenza is endemic in several countries, and there is interest in containing a highly virulent pandemic at the earliest sign of localized efficient human-to-human transmission. Two key modeling studies suggested a high probability of success for rapid containment of an influenza epicenter with combination strategies under favorable conditions [9, 10] . These studies formed the basis for the epicenter containment strategies recommended by the WHO. Longini showed that antiviral prophylaxis alone could contain a pandemic influenza virus with reproductive number (Ro) less than 1.7; while 70% household quarantine alone was effective up to Ro of 1.7. A combination of quarantine and antiviral prophylaxis was effective up to Ro of 2.1; while a combination of pre-pandemic vaccination, household quarantine and antiviral prophylaxis was effective for Ro of 2.4 [9] . Ferguson found that antiviral prophylaxis for contacts only would have a 90% chance of containing a virus with a Ro less than 1.25, while antiviral prophylaxis for contacts and all individuals in a 10 km zone would have a 90% chance with Ro less than 1.7 [10] . Combined anti-viral prophylaxis and either school and workplace closures or area quarantine provided a similar chance of containment with Ro of 1.7 to 1.8, while a combination of all three strategies would contain a virus with Ro of 1.9 and allow for greater initial surveillance errors [10] . Combination strategies can be used to reduce the global spread of the influenza virus [12, 13] . Redistribution of limited antiviral drugs can help contain pandemics or reduce the global attack rate (AR) [12] . If global antiviral stockpiles are limited, non-cooperative strategies where countries keep their antiviral stockpiles for their own use can only contain a pandemic influenza virus with Ro less than 1.5; in contrast, if redistribution of 25% of stockpiles from countries that have them to countries that do not, a pandemic with Ro up to 1.9 may be contained, and overall AR reduced by 25% at higher Ro [12] . Another example of combination strategy is reduction of pandemic spread through air travel. Suspension of 99.9% of air travel can only delay individual national epidemics by up to four months, while a combination of local strategies reducing influenza transmission by 40% can delay pandemic spread by up to 10 months [13] . A combination of vaccination and travel restrictions may delay epidemic growth, allowing vaccination of susceptible individuals [14] . With a pandemic starting in July in Asia, the number of United States (US) metropolitan cases was 102.4 million -0.1% daily vaccination alone reduced this to 73.0 million, and vaccination together with travel restriction reduced this to 56.9 million [14] . Combination strategies may have substantial impact in reducing the global spread of resistant viruses. For example, if the probability of emergence of anti-viral drug resistance was 1%, antiviral monotherapy was associated with overall AR of 67% and resistant AR (RAR) of 38% [15] . In contrast, early combination chemotherapy was associated with reduced AR of 58% and RAR of 2%, while sequential multi-drug chemotherapy was associated with AR of 57% and RAR of 3%. During the pandemic, several studies found that combination strategies delayed the spread of the virus, reduced the overall number of cases, and delayed and reduced the peak AR much more than individual strategies which may be ineffective if used alone [16] [17] [18] [19] . A study using individual-based modelling in the United Kingdom and United States examined the effects of antiviral treatment and prophylaxis, vaccination, case isolation, household quarantine, school and workplace closure and travel restrictions in pandemics with Ro of 1.7 to 2.0. It found that external or internal travel restrictions alone would delay spread by two to three weeks only if more than 99% effective [16] . Reactive school and work-Flow diagram for selection of combination strategy modeling studies Figure 1 Flow diagram for selection of combination strategy modeling studies. place closures alone did not impact on overall AR, but reduced peak AR by about 40%; antiviral treatment and prophylaxis within the household reduced overall AR by 35% and peak AR by 45%.; while household quarantine alone reduced overall AR by 10% and peak AR by 20%. Combination antiviral treatment and prophylaxis, and household quarantine reduced overall AR by 40% and peak AR by 60%. Combination school and workplace closure, antiviral treatment and prophylaxis, and household quarantine reduced overall AR by more than 60% and peak AR by more than 80%. Combination antiviral treatment and prophylaxis, school closure and 20% pre-pandemic vaccination reduced overall AR by more than 60% and peak AR by more than 75%. Combination antiviral treatment and prophylaxis, household quarantine, school and workplace closure, and effective border control reduced overall AR by more than 70% and peak AR by more than 90% [16] . Similarly, another individual-based stochastic simulation model in Chicago evaluating the effects of antiviral treatment and prophylaxis, quarantine, isolation, school closure, community and workplace social distancing showed that social distancing alone may reduce overall AR by 60% for pandemic Ro of 1.9 but combination antiviral treatment and prophylaxis, quarantine, social distancing, and school closure could reduce overall AR by more than 90% for similar pandemic Ro of 1.9 [17] . Another study in France examined the effects of antiviral treatment and household prophylaxis, vaccination, household quarantine, school and workplace closure at the individual and community level [20] . Treatment only with anti-viral drugs did not affect AR substantially. Antiviral prophylaxis of 90% of household contacts reduced AR by 50%. Vaccination of 70% of the population within one day reduced AR by 80%. A combination of antiviral treatment and prophylaxis, and household quarantine reduced AR by 90% [20] . An Australian individual-based stochastic simulation model assessed the effects of non-pharmacological pandemic mitigation measures of case isolation, school closure, workplace non-attendance and community contact reduction [21] . For a pandemic with Ro of two, school closures alone reduced AR by 20%, case isolation by 40%, workplace non-attendance by 15%, and social distancing by 25%. In contrast, combination of all these measures reduced AR by more than 95% [21] . A deterministic compartment model using InfluSim based on a small community of 100,000 population assessing the effects of antiviral treatment, case isolation and social distancing showed that case isolation and social distancing could reduce overall AR by 25%, and antiviral treatment alone by 20%, compared with a reduction of 40% with a combination of case isolation, social distancing and antiviral treatment [18] . The triple combination strategy could delay the peak by one month compared with 10 days for the first two strategies [18] . Another study using a deterministic model with a stochastic simulation component based on Italy examined the effects of household antiviral prophylaxis, pre-pandemic vaccination, and social distancing via closure of all schools, public offices and public meeting places [22] . In a pandemic with an attack rate of 35%, vaccination alone reduced AR by up to 10% even at vaccine efficacy levels of 70%; antiviral prophylaxis alone for even the entire pandemic duration reduced AR by up to 6% only; and social distancing alone reduced AR by less than 1%. However, a combination of all three measures reduced AR by up to 30% [22] . The relative success of interventions depends on the transmissibility of the pandemic, which is commonly reflected in the Ro. In an influenza pandemic with higher Ro, the effectiveness of interventions is reduced and individual interventions are commonly ineffective. However, across most scenarios, combination strategies maintain some effectiveness as shown clearly in the studies on containment by Longini [9] and Ferguson [10] . A stochastic agent-based discrete-time simulation model in the United States examining the effect of antiviral prophylaxis, vaccination, school closure and travel restriction found that for a pandemic influenza virus with Ro of 2.4, unlimited antiviral prophylaxis and best vaccination program may reduce cases by 64% and 34% respectively, while school closure within seven days of pandemic onset may reduce cases by 14%, social distancing within seven days by 6%, and travel restrictions exceeding 90% was ineffective [19] . However, a combination strategy of all of these measures may reduce cases by 99.8% [19] . The effectiveness of any strategy in delaying the pandemic or reducing the AR is highly dependent on the Ro. For example, for a pandemic with Ro of 1.6, individual strategies of prophylaxis, vaccination, or school closures had very high effectiveness [19] . However, once the Ro increased beyond 2.0 (which is similar to the Ro for the 1918 pandemic), individual strategies were much less effective, whereas combination strategies still maintained effectiveness across a range of Ro. An individual-based model in Italy assessing the effects of household, school and workplace antiviral prophylaxis, vaccination, international air travel restriction, social distancing via school closure and closure of some public offices showed that without any interventions, importation of pandemic influenza would occur 37 to 77 days after the first case elsewhere in the world. Air travel restric-tion would delay introduction by one week to one month. For a pandemic with Ro of 1.7, travel restriction and social distancing did not affect overall AR, household prophylaxis reduced AR by 50%, and vaccination reduced AR by 0 to 40%. A combination of antiviral prophylaxis, social distancing, vaccination, and travel restriction reduced AR by more than 90% [23] . For a pandemic with Ro of 2.0, travel restriction in fact increased overall AR by 1% and peak AR by 20%. Household prophylaxis reduced AR by 35%, while vaccination reduced AR by 0 to 30%. A combination of antiviral prophylaxis, social distancing, vaccination, and travel restriction reduced AR by 80%. An individual-based stochastic model in Hong Kong looking at the effects of antiviral prophylaxis, case isolation and household quarantine reported that in a pandemic with Ro of 1.8 and AR of 74%, household quarantine could reduce AR to 49%; household quarantine and isolation to 43%; household quarantine with anti-viral prophylaxis to 44%; household quarantine, isolation and antiviral prophylaxis to 40% which was recommended. Although adding contact tracing and quarantine of all contacts to the latter combination strategy reduced AR to 34%, the number of people under quarantine would be excessive. Therefore, contact tracing was not recommended [24] . Another study examining the effects of antiviral treatment and prophylaxis, home quarantine and social distancing based on a community of a million population with the assumption that pandemic influenza was introduced by an undetected airline passenger, found that if a pandemic Ro was 3.0, individual interventions would result in increased transmission while combination measures may break community transmission [25] . This was similarly shown by Ciofi and colleagues for a pandemic with Ro of 2.0 [23] . A deterministic compartmental model evaluating the effects of antiviral treatment and prophylaxis, vaccination, case isolation and air traffic reduction globally demonstrated that individual strategies such as case isolation and air travel restrictions may result in higher peak AR even though overall AR could be reduced [26] . A study in Taiwan evaluated the effects of enhanced ventilation, use of respiratory mask and vaccination on pandemic influenza transmission in a school [27] . Vaccination alone of 80% of children was effective in preventing the spread of the virus but this was only if a suitable vaccine was available, which is often not the situation. A combination of masks and ventilation, or a combination of vaccination and masks achieved similar effectiveness [27] . Many modeling studies were performed as a result of H5N1 influenza threat and an impending pandemic, but all have used parameters based on historical pandemics and existing studies on the influenza transmission. In addition, these studies provided sensitivity analyses across a wide range of influenza parameters. As such, they are directly relevant to the 2009 influenza pandemic which has an Ro of between 1.2 to 1.6 [28] , similar to the 1957 and 1968 influenza pandemic [16] , and for future pandemics. At the same time, the 2009 influenza pandemic provides the opportunity to study unknown variables to validate and refine these models. All of these modeling studies in various settings, and using different models and assumptions, consistently show that combination strategies are more effective compared to individual strategies. Given the lack of good experimental, observation or controlled studies on these strategies, and the difficulties of performing trials during a pandemic, it is difficult for policy makers to know the effectiveness of their policies. These modeling studies provide policy makers with a suggestion of the effectiveness of different combination strategies. At the same time, new models will have to be developed using local data to provide realistic outcomes for local settings. The diverse methodology available from these studies provides sufficient information for countries to build and validate their results locally. Although the use of individual-based and other stochastic models provide better data resolution, deterministic models mentioned in this review show similar outcomes [18, 22, 23, 27] . These deterministic or simple stochastic compartmental models are much easier to build and may provide rapid results for policy making. This is especially true in countries where the vast amounts of data required for individual-based and complex stochastic models may not be available compared with high-income countries where most sophisticated models were built. The use of combination strategies necessitates the availability of resources and feasibility for each individual component. For example, stockpiling of pharmaceutical agents is an integral part of preparedness plans and currently widely adopted in well-resourced countries. The increase in anti-viral drug resistance underscores the importance of combination drug use and provides policy makers with recommendations for their stockpiles [15] . Combination stockpiles of sufficient amounts of different antiviral drugs such as oseltamivir, zanamivir and adamantanes will allow for early combination chemotherapy or sequential multidrug therapy which was modeled to be effective against antiviral resistance when a small secondary stockpile was used to augment a primary stockpile [15] . The United States Federal stockpile is composed of 80% oseltamivir and 20% zanamivir, and several million doses of rimantadine from previous stockpiles [29] . The United Kingdom has purchased additional antiviral drugs to ensure it has a total stockpile for 50% of its population, comprising 68% oseltamivir and 32% zanamivir [30] . Bacterial pneumonia results in substantial morbidity and mortality among pandemic influenza cases [31, 32] . Antibiotics should therefore be considered for stockpiling [31] . Stockpiles should take into account common locally circulating bacteria, and recommended amounts range from 10 to 25% of the population [33] . In contrast to antiviral drugs that are not widely used, antibiotics can be part of a rolling stockpile which ensures sufficient stockpiles without expiry issues. Vaccination against bacterial infections should likewise be considered. From the effectiveness of combination strategies in reducing global spread of influenza or resistant viruses [12] [13] [14] [15] , resource-rich countries should consider redistributing their resources for the greater global benefit and their own benefit if they have yet to be affected by the pandemic. Controlling local outbreaks through combination strategies can reduce global spread, and countries affected early during the pandemic should be provided with assistance [13] . Vaccines are part of many combination strategies and modeling has shown that introduction of a vaccine four months after the pandemic virus has arrived has limited effectiveness, while stockpiling prototype pandemic vaccines could reduce overall AR [16] . Therefore countries were stockpiling H5N1 vaccines as candidate pandemic vaccines [34, 35] . However, if the pandemic influenza virus is totally different from the vaccine virus, the vaccines would be of negligible effectiveness. Investments are needed to develop new vaccines with greater cross-protection against conserved viral regions; vaccine libraries to quickly produce candidate vaccines; better adjuvants and antigen-sparing strategies to increase production capacity; and modes of administration for improved immunogenicity and cross-protection [36, 37] . Although some individual strategies may seem very effective, they may not be feasible and models assist policy makers in avoiding potentially disastrous decisions. Social distancing has been widely used in epidemics [7] but their impact remains unclear and highly dependent on disease severity, transmission, and risk groups affected. Local interventions such as school closures may be effective if done early, decisively, and for prolonged periods [20, [38] [39] [40] . A United Kingdom model based on a 1957like pandemic showed more than 20% case reduction if the Ro were low (<2) and schools were closed early, but less than 10% case reduction in pandemics with high Ro [38] . A French study showed that prolonged closure and limiting contact among children outside school may reduce cases by 17% and peak AR by 45% [39] . However, school closures and limiting social contact may be socioeconomically difficult to achieve. Another study found that total closure of schools and workplaces reduced AR by 95%. However, the socio-economic impact would be unimaginable [20] . Similarly, most modeling studies found that travel restrictions alone did not impact overall AR [13, 16, 19, 23] . Reducing air travel has been modeled to be effective in delaying pandemic spread if nearly 100% reduction can be achieved [13, 16] , and will be difficult if not impossible to achieve [41] . If used alone, local epidemic severity may increase because restriction-induced travel delays can push local outbreaks into high epidemic season [14] . Although combination strategies are more effective than individual measures, not all combination strategies may be feasible. Active surveillance, isolation of cases, and quarantine of close contacts are important interventions during epicenter containment. These interventions may reduce the Ro of the disease to below one and contain the outbreak. However, it is often difficult to ensure total compliance with these measures and if used alone, will result in missed cases due to surveillance failures, isolation facility exposures, and quarantine failures as shown in the SARS experience [42] . A Hong Kong modeling study found that although contact tracing and quarantine of all contacts was effective, it was not feasible because the number of people under quarantine would be excessive [24] . Therefore combination strategies enable policy makers to leverage on the effectiveness of some measures and reduce potential negative impact of others. For combination strategies to work, they have to be tailored for each scenario at organizational, community, national, and international levels. To facilitate integration of interventions into effective combination strategies, more evidence is needed through targeted research, for example, the effectiveness of non-pharmaceutical interventions (e.g. personnel cohorting, school closures or reduction in air travel). In the absence of definitive studies, mathematical modeling studies provide an effective means of assessing the effectiveness of these strategies. A limitation of this study is the restriction of our searches to the PubMed database. While we have made attempts to include additional articles from snowball searches, there is the potential for other published or unpublished studies to be missed from other databases and private sources. Other intrinsic limitations of modeling studies exist, and include the fact that they are based on theoretical epidemiology and not fully based on clinical or epidemiological evidence. For example, widespread use of pandemic vaccines raises safety concerns, and widespread use of antiviral drugs raises concern for antiviral resistance. Viral transmission during treatment with anti-viral drugs is also not well understood. It is therefore important to perform clinical and epidemiological studies during pandemic or seasonal influenza to understand the effectiveness and impact of these interventions. Models are also highly dependent on the assumptions and input variables, and are specific for a local context. However, if these limitations are understood by decision makers, modeling provides a reflection of the possible outcomes, helps to delineate possible strategies for inclusion, and avoids costly errors. Modeling studies show that combination strategies increase the effectiveness of individual strategies, guard against individual failures, and may reduce socio-economic impact. In the initial phases of an influenza pandemic, combination strategies provide the opportunity to contain the novel virus or delay its spread, allowing unaffected areas within a country and other countries to activate preventive strategies. During a pandemic, combination strategies allow for different strategies to have synergistic effect in reducing the impact of pandemic influenza, and the socio-economic impact of individual interventions. Finally, combination strategies protect against failure of individual interventions and should be considered in preparedness plans.

Let the sun shine in: effects of ultraviolet radiation on invasive pneumococcal disease risk in Philadelphia, Pennsylvania

BACKGROUND: Streptococcus pneumoniae is a common cause of community acquired pneumonia and bacteremia. Excess wintertime mortality related to pneumonia has been noted for over a century, but the seasonality of invasive pneumococcal disease (IPD) has been described relatively recently and is poorly understood. Improved understanding of environmental influence on disease seasonality has taken on new urgency due to global climate change. METHODS: We evaluated 602 cases of IPD reported in Philadelphia County, Pennsylvania, from 2002 to 2007. Poisson regression models incorporating seasonal smoothers were used to identify associations between weekly weather patterns and case counts. Associations between acute (day-to-day) environmental fluctuations and IPD occurrence were evaluated using a case-crossover approach. Effect modification across age and sex strata was explored, and meta-regression models were created using stratum-specific estimates for effect. RESULTS: IPD incidence was greatest in the wintertime, and spectral decomposition revealed a peak at 51.0 weeks, consistent with annual periodicity. After adjustment for seasonality, yearly increases in reporting, and temperature, weekly incidence was found to be associated with clear-sky UV index (IRR per unit increase in index: 0.70 [95% CI 0.54-0.91]). The effect of UV index was highest among young strata and decreased with age. At shorter time scales, only an association with increases in ambient sulphur oxides was linked to disease risk (OR for highest tertile of exposure 0.75, 95% CI 0.60 to 0.93). CONCLUSION: We confirmed the wintertime predominance of IPD in a major urban center. The major predictor of IPD in Philadelphia is extended periods of low UV radiation, which may explain observed wintertime seasonality. The mechanism of action of diminished light exposure on disease occurrence may be due to direct effects on pathogen survival or host immune function via altered 1,25-(OH)(2)-vitamin-D metabolism. These findings may suggest less diminution in future IPD risk with climate change than would be expected if wintertime seasonality was driven by temperature.

Many infectious diseases of public health importance exhibit predictable periodicity, with major increases in incidence during a specific season of the year [1] . Empirical evidence of such "seasonality" has been noted by physicians for centuries, and has been prominent enough to become a part of our vernacular (e.g., "cold and flu season") [1, 2] . Despite this wealth of experiential evidence, the mechanisms underlying seasonality are poorly understood, especially diseases characterized by person-to-person transmission [1] . Invasive bacterial disease due to Streptococcus pneumoniae and other respiratory pathogens exhibits striking seasonality in its occurrence [3] [4] [5] [6] [7] . Pneumococcal infections are a common cause of severe, community-acquired illnesses, including community-acquired pneumonia requiring hospitalization, bacteremia, and meningitis [8] . While the introduction of antibiotics dramatically reduced the case fatality rate (CFR) for pneumococcal disease, the current CFR for bacteremic pneumococcal disease is still estimated at 5-10% in the United States, and may be twice as high among the elderly and in cases of meningitis [9, 10] . The emergence of antimicrobial resistance to beta-lactam agents, macrolides, and other antibiotic classes is an important clinical concern [11] . Although the introduction of conjugate pneumococcal vaccines has been associated with a reduction in disease incidence [12] , the recent increase in invasive infection by non-vaccine serotypes [13] , which may be highly resistant to commonly-used antimicrobials [14] , suggests that this microorganism will persist in challenging both the medical and public health communities. The incidence of IPD peaks in the winter months, with annual periodicity [4, 5] , but the forces that drive this characteristic seasonality are unknown. Wintertime seasonality of communicable respiratory diseases are often assumed to be driven by seasonal changes in environmental conditions (e.g., diminished ultraviolet radiation (UV) exposure, decreased temperature) but such associations may be confounded by other seasonally varying factors, including population behaviours (e.g., clustering indoors), co-occurrence of other infections (e.g., influenza) [4] , and frequency of laboratory testing [15] . A more thorough understanding of the effect of environmental factors on seasonal IPD incidence could offer significant insight into pathogenesis, improve disease forecasting, and help determine the likely direction of pneumococcal disease incidence in the face of global climate change [16] . Our objective was to investigate how environmental factors influence IPD occurrence in Philadelphia County. We used both traditional analytic methods (i.e., Poisson regression with seasonal smoothers) and a novel case-crossover method to examine the effects of acute weather fluctuations on IPD occurrence. Both methods reduce confounding by environmental, behavioural, and infectious influences that might otherwise distort the observed magnitude of environmental effects on disease risk. Philadelphia County encompasses an area of 369 km 2 in south-eastern Pennsylvania, and is coterminous with the City of Philadelphia (population 1,517,550 in the year 2000 [17] ). The population receives public-health services from the Philadelphia Department of Public Health (PDPH). IPD has been a notifiable condition in the Commonwealth of Pennsylvania since 2002; Pennsylvania uses the uniform case definition endorsed by the National Notifiable Diseases Surveillance System [18] . A case is considered "confirmed" when a consistent clinical syndrome occurs in association with the isolation of S. pneumoniae from a normally sterile site (e.g. blood, cerebrospinal or pleural fluid). Data on IPD case occurrence in Philadelphia was obtained from PDPH records, and included date of onset, age, sex, race and ethnicity of the patient, isolation site, and fatal outcome (if known). Meteorological data including temperature, relative humidity, wind speed, atmospheric pressure, and precipitation for the period from 2002 to 2006 was obtained from the weather station at Philadelphia International Airport, located eight kilometres southwest of Philadelphia's city center [19] . Information pertaining to air quality in Philadelphia County during the years of interestincluding concentrations of lead, ozone, sulphur oxides and particulate matter-was obtained from the Environmental Protection Agency [20] . Because daily readings were taken at various locations throughout the region, the arithmetic means of the air quality values were used as exposure variables. UV index forecast estimates for Philadelphia during the same period were retrieved from the National Weather Service Climate Prediction Center [21] . Clear-sky UV indices represent an integral of measured UV radiation levels weighted by the ability of the different UV wavelengths to cause skin erythema. The issued UV index is a similar measure, which accounts for the effect of clouds on radiation transmission; because of inconsistencies in cloud measurement during the study period, we used the clear-sky UV index as our exposure variable. Rates of invasive pneumococcal disease were calculated using demographic data for Philadelphia County from the year 2000 US Census, as well as 2006 population estimates from the Bureau of the Census, with linear interpolation and extrapolation used to generate estimates for population by age and sex in other years [17] . We evaluated the seasonality of disease occurrence through construction of periodograms and autocorrelograms [15, 22] for weekly case counts. As yearly periodicity was observed, we estimated seasonal and year-on-year trends in IPD occurrence using Poisson regression models that incorporated sine and cosine oscillators, with 52 week (annual) frequencies (i.e., incorporated fast Fourier transforms) [7, 22] . Using these parameters, the expression for the expected number of case counts for a given week, E (Y) is given by: where cases is an autoregressive model term reflecting the cumulative case count in the month prior to case occurrence, (i.e., cases = ). The phase-shift of the composite waveform generated by combining sine and cosine components of the above equation can be approximated as the arctangent of β 2 /β 3 , and can be used to estimate the timing of peak disease incidence [22] . We also included model terms that controlled for longer term trends in incidence of invasive pneumococcal disease, which may have reflected the initiation of surveillance, the introduction of public funding for conjugate pneumococcal vaccination [12] , changes in medical diagnostic practices, or other long-term changes in real or apparent pneumococcal epidemiology. As yearon-year trends in disease occurrence reflected a non-linear increase in disease risk, we used models that incorporated both linear and quadratic yearly terms. The quadratic model term was statistically significant, but is difficult to interpret, thus we present our final Poisson model with separate linear yearly terms for the period from 2002 to 2003, and the period from 2004 to 2007 [23] . We evaluated the impact of environmental exposures on weekly IPD incidence by incorporating exposure variables into Poisson models both individually, and using a backwards elimination algorithm (with variables retained for P < 0.2 [24] ). To explore the possibility of effect modification by subject characteristics, we evaluated stratum-specific estimates of effect for age categories and genders. Heterogeneity of effects across strata was assessed using meta-analytic techniques, including both graphical inspection and calculation of meta-analytic Q-statistics [25] . We further explored the sources of between-stratum heterogeneity through construction of meta-regression models that estimate the contribution of group-level covariates to between-stratum variation in effects [25] . We used a case-crossover approach to evaluate acute (i.e., daily) associations between environmental exposures and IPD occurrence. This approach provides a means for evaluating the association between brief, transient exposures and rare outcomes. The design is characterized by "self matching", in that cases serve as their own controls. In the context of environmental epidemiology, a "case" is a day on which a case occurred, while a "control" is an appropriately selected day on which a case did not occur [6] . We used a time-stratified 2:1 matched case-crossover design in which hazard periods were defined as the reported date of IPD onset from Philadelphia County public health. Beginning on January 1, 2002 the person-time at risk was divided into three-week time strata. Control periods were chosen by matching the hazard period by day of the week within the stratum, and could both precede, both follow, or straddle the hazard period [26, 27] . Random directionality of control selection was used in order to avoid biases that can occur with unidirectional or uniform bidirectional control selection [26] . The 1-3 day incubation period of S. pneumoniae was used to estimate the lag days between acute environmental occurrence and case onset, or plausible effect period [28] . We also evaluated effects during the period immediately preceding incubation (i.e., 4-6 day lags) to evaluate the possibility that environmental conditions might affect risk via enhanced transmission of S. pneumoniae. We evaluated the effects of both raw environmental exposures, and quantile ranks within time strata through construction of conditional logistic regression models [24] . Analyses were performed using SAS version 8.01 (SAS Institute, Cary, NC) and Stata version 9.1 (Stata Corporation, College Station, TX). Both spectral decomposition and construction of autocorrelograms identified annual periodicity of infection, with peak incidence in mid-February (phase = 6 weeks) (Figures 1 and 2) . Strong statistical evidence for seasonal oscillation was obtained from Poisson regression models (P for seasonal oscillation < 0.001). A significant annual increase in incidence was seen throughout the study period, though this was more marked prior to 2004 (IRR per year 1.34, 95% CI 1.08 to 1.66) than subsequently (IRR 1.22, per year 95% 1.16 to 1.34) (Figure 3 ). We found no clear trends in the incidence of IPD or case-fatality in individual age groups, and no significant heteroge- neity was detected between age groups with respect to year-on-year trends in incidence or case-fatality (see Additional File 1 and Additional File 2). In univariable models the risk of IPD increased with several seasonally oscillating environmental exposures, including temperature, humidity, pressure, air pollution, and UV radiation, as shown in Table 2 . Risk of IPD increased with average weekly barometric pressure, sulphur and nitrous oxides, and decreased with average weekly temperature, relative humidity, and UV index. However, after controlling for seasonal oscillation and longer term temporal trends, only cooling-degree days (i.e., average number of degrees above 18°C), maximum temperature, and clear-sky UV index were independently associated with case occurrencein a final multivariable model ( (Figure 4 ). Evaluating associations between environmental and meteorological exposures and IPD risk using a case-crossover approach, we identified an inverse association between ambient levels of sulphur oxides and disease risk during the likely incubation period (Table 3) . No other significant associations between environmental exposures and risk were identified either during the incubation period, or in the period immediately prior to incubation. In particular, occurrence of IPD was not affected by daily changes in clear-sky UV index, temperature or coolingdegree days, in contrast to associations on longer time scales described above. Notwithstanding the existence of vaccination and effective antibiotic therapy, invasive pneumococcal disease remains an important source of population morbidity and mortality. The seasonality of IPD is well recognized, but poorly understood. Epidemiological mechanisms invoked to explain this pattern have included co-occurrence of other infectious diseases [4] , wintertime social gatherings [5] , and seasonal oscillation in immune function [29] . However, concurrent seasonal changes in a variety of environmental, behavioural, and epidemiogical exposures make identification of causal associations particularly challenging [30] . We attempted to address this challenge by using analytic approaches that should control for seasonal confounders, known and unknown, at different time scales. At a weekly time scale we found increases in UV radiation to be most strongly associated with decreased numbers of invasive pneumococcal disease cases, though average temperatures also appeared to influence disease risk. At short time scales, fluctuations in ambient air quality, as manifested by differences in concentrations of sulphur oxides, were associated with changes in risk. The direction of this association was at variance with existing models relating air pollution to pneumonia occurrence [3, 31, 32] . The casual (as opposed to causal) association between low UV radiation in the winter and surges in respiratory disease has been noted previously [33] , and has been proposed as an important driver of influenza seasonality, but has not to our knowledge been evaluated in a way that accounts for coincident seasonal changes in other seasonally oscillatory factors. The degree to which such seasonal oscillation can result in "just so" stories that lead to misattribution of causation to non-specific seasonal exposures is highlighted in the univariable analyses we conducted without including seasonal oscillators. In these models, a variety of environmental conditions, including weather variables and air quality indices, were strongly associated with IPD risk. However, after controlling for non-specific seasonality, only UV radiation (and, more weakly, temperature) were associated with disease risk; indeed, the apparent protective effect of UV radiation was actually strengthened after controlling for seasonal oscillation. The interpretation of such a model is that increases in UV radiation reduce IPD risk, even after accounting for the fact that IPD risk is maximal during low-UV periods of the year. An important consideration is whether changes in UV radiation, operating at a weekly time scale, constitute a biologically plausible mechanism that explains seasonal oscillation in pneumococcal disease risk. Indeed, there are several mechanisms that may have substantial biological plausibility. Modulation of risk may occur through direct effects of UV radiation on host immune function: Dowell Figure 1 Periodogram Constructed from Spectral Decomposition of Weekly Pneumococcal Case Counts. Spectral density is represented on the y-axis, and can be conceptualized as a measure of goodness-of-fit for oscillatory regression models at different frequencies. The large peak at a frequency of 51 weeks suggests that invasive pneumococcal disease is a process that oscillates with annual periodicity (and is, in other words, compatible with wintertime seasonality). The two peaks at lower frequencies are lower harmonics illustrating bi-and tri-annual behaviour. Spectral Density reviewed a variety of immunological changes associated with diminished UV radiation exposure in experimental settings, and noted that in granulocyte and monocyte function were reduced during periods of short light exposure [33] . UV radiation also influences the production of 1,25-(OH) 2 -vitamin D, which has important immunomodulatory functions [34] . Namely, enhanced maturation of macrophages, macrophage secretion of bactericidal substances such as lysozomal enzyme phosphatase and hydrogen peroxide [33, 35] , and secretion of antimicrobial peptides (including cathelicidins and defensins) by both immune cells, and respiratory tract epithelium [36] . Vitamin D deficiency is associated with a marked increase in the risk of pneumonia [37] , and most human vitamin D is acquired via sun exposure. Thus, extended periods of low UV light could result in an increased susceptibility to S. pneumoniae infection resulting from a lack of vitamin D production, though the week-to-week fluctuations in risk described in this paper may be too rapid to represent a vitamin D effect. An alternative mechanism of action of UV radiation in reducing IPD risk could be direct effects of radiation on pathogen survival. Many bacterial respiratory pathogens, including pneumococcus, are transmitted in respiratory secretions over short distances (i.e., via "large droplet transmission"), and thus encounter the physical environment directly during transmission events. The bactericidal effects of UV-B radiation have been well known for decades; such radiation inactivates bacteria by causing harmful genetic mutations through creation of pyrimidine dimers [38] . A model of UV effect via diminished transmissibility, rather than decreased host susceptibility, is supported by our finding that UV effect is strongest in the youngest individuals in the population (i.e., toddlers), where disease risk is likely to be driven by mobility, contact with peers, prolonged carriage and carelessness with respiratory secretions. We found very little protective effect against IPD in oldest individuals, whose risk may be more strongly linked to immune senescence than to high rates of contact with infectious contemporaries [39, 40] . An unexpected finding in our case-crossover analysis was the identification of increased levels of ambient sulphur oxides with diminished risk of IPD. This may represent a chance association, due to the established association between sulphur dioxide and adverse respiratory outcomes [32] , as well as the previously described correlation between ambient sulphur dioxide and pneumonia risk [3] . Nonetheless, this association may warrant further exploration; for example, ambient air pollutants might have adverse effects on respiratory tract pathogens as well as hosts. Other findings in this study are also worthy of comment; in Philadelphia, the wintertime peak in pneumococcal incidence occurred about six weeks later than previously been described in U.S. adults [5] . In addition, a gradual increase in cases was observed during the 5-year study period. We suspect that this increase is less likely to represent a true surge in disease rates, which have actually been falling in the U.S. with the introduction of 7-valent conjugate pneumococcal vaccine [41] ; rather, we suspect that this increase, which is more attenuated after 2003, represents the gradual increase in reporting that commonly fol-lows implementation of new infectious disease reporting requirements [42] . Like any observational study, ours has several limitations. First among these is our heavy reliance on public health surveillance data which is known to suffer from underreporting of notifiable infectious diseases [43] . Thus our data set may be incomplete, consisting of only a subset of the cases of IPD in Philadelphia during the study period. However, selection bias will only be introduced if the notification of infectious diseases to public health was correlated with meteorological patterns (e.g., cases which occur during days with higher UV index have a greater likelihood of being diagnosed and reported), which seems unlikely [44] . Second, we obtained weather data from a single site at Philadelphia International Airport, and air pollution and UV data were averaged over several locations throughout the county, which may not represent the true exposure status of individual cases. This is 8 0 a n d o v e r probable non-differential misclassification, and will bias the results towards the null hypothesis. The effects seen here are most likely conservative, weakening the strength of observed associations and suggesting our estimates tend towards the lower bound [44] . In summary, we described the occurrence of IPD in a major U.S. urban center and found that incidence was associated with marked wintertime seasonality that may be partly explained by diminished exposure to UV-B radiation in winter months. Further study is needed, but this result is consistent with observed patterns of respiratory infectious disease, and would be consistent with several biologically plausible models of effect. As the nature of future changes in UV radiation related to climate change are more difficult to predict than general changes in temperatures or precipitation patterns, the implications of these findings for future pneumococcal disease epidemiology are unclear [45] . Nonetheless, we believe this obser-vation goes some way towards explaining the notable seasonality of IPD, and could conceivably lead to novel disease control strategies through improved understanding of this common and virulent infectious disease.