Unnamed: 0
int64
0
932
title
stringlengths
20
337
abstract
stringlengths
48
2.83k
fulltext
stringlengths
22
144k
0
Clinical features of culture-proven Mycoplasma pneumoniae infections at King Abdulaziz University Hospital, Jeddah, Saudi Arabia
OBJECTIVE: This retrospective chart review describes the epidemiology and clinical features of 40 patients with culture-proven Mycoplasma pneumoniae infections at King Abdulaziz University Hospital, Jeddah, Saudi Arabia. METHODS: Patients with positive M. pneumoniae cultures from respiratory specimens from January 1997 through December 1998 were identified through the Microbiology records. Charts of patients were reviewed. RESULTS: 40 patients were identified, 33 (82.5%) of whom required admission. Most infections (92.5%) were community-acquired. The infection affected all age groups but was most common in infants (32.5%) and pre-school children (22.5%). It occurred year-round but was most common in the fall (35%) and spring (30%). More than three-quarters of patients (77.5%) had comorbidities. Twenty-four isolates (60%) were associated with pneumonia, 14 (35%) with upper respiratory tract infections, and 2 (5%) with bronchiolitis. Cough (82.5%), fever (75%), and malaise (58.8%) were the most common symptoms, and crepitations (60%), and wheezes (40%) were the most common signs. Most patients with pneumonia had crepitations (79.2%) but only 25% had bronchial breathing. Immunocompromised patients were more likely than non-immunocompromised patients to present with pneumonia (8/9 versus 16/31, P = 0.05). Of the 24 patients with pneumonia, 14 (58.3%) had uneventful recovery, 4 (16.7%) recovered following some complications, 3 (12.5%) died because of M pneumoniae infection, and 3 (12.5%) died due to underlying comorbidities. The 3 patients who died of M pneumoniae pneumonia had other comorbidities. CONCLUSION: our results were similar to published data except for the finding that infections were more common in infants and preschool children and that the mortality rate of pneumonia in patients with comorbidities was high.
Mycoplasma pneumoniae is a common cause of upper and lower respiratory tract infections. It remains one of the most frequent causes of atypical pneumonia particu-larly among young adults. [1, 2, 3, 4, 5] Although it is highly transmissible, most infections caused by this organism are relatively minor and include pharyngitis, tracheobronchitis, bronchiolitis, and croup with one fifth of in-fections being asymptomatic. [6, 7] Only 3 -10% of infected subjects develop symptoms consistent with bronchopneumonia and mortality from infection is rare. [6, 7] The organism is fastidious and difficult to grow on cultures. Therefore, diagnosis of infections caused by this organism is usually confirmed with serological tests or polymerase chain reaction-gene amplification techniques. At King Abdulaziz University Hospital (KAUH), Jeddah, Saudi Arabia, the facility to perform Mycoplasma culture has been available since January 1997. As published information concerning M. pneumoniae infections in Saudi Arabia is scarce, [8, 9, 10] we wished to study the epidemiology and clinical features of cultureproven infections caused by this organism at this hospital. KAUH is a tertiary care teaching hospital with a bed capacity of 265 beds and annual admissions of 18000 to 19000 patients. Patients with M. pneumoniae positive cultures from respiratory specimens were identified over a 24-months" period from January, 1997 through December, 1998 for this review. During the study period, respiratory specimens (sputum, nasopharyngeal aspiration, endotracheal secretion, and bronchoalveolar lavage) for M. pneumoniae culture were obtained from patients with upper or lower respiratory tract infections seen as inpatients or in the outpatient or emergency departments. Respiratory specimens were aslo Gram-stained and cultured for bacteria and viruses. M. pneumoniae serological tests for IgG or IgM were not available at KAUH during the study period. All positive culture results were obtained from the Microbiology laboratory records. Charts of patients were reviewed with standardized data collection. Information collected included patients' demographics, comorbidities, clinical manifestations, complications, and outcome. M. pneumoniae was cultured using the classic M. pneumoniae agar medium (M.P. agar) and the Pneumofast tray (Pneumofast ® , International Microbio, Signes, France). Specimens were processed according to the instructions of the manufacturer. The M.P. agars and Pneumofast trays were incubated anaerobically at 37°C and inspected daily for 4 weeks. The organism was identified based on typical colonial morphology (granular colonies, rarely fried-egg-like, 10-150 ∝ in diameter) on the M.P. agar medium and the change in the Pneumofast broth color from red to orange then to yellow (glucose fermentation) in the absence of turbidity of the broth. Antibiotic sensitivity profile on the Pneumofast tray was also used for identification according to the instructions of the manufacturer. Bacterial and viral cultures were performed using standard methods. M. pneumoniae isolates were considered community-acquired if they were recovered from unhospitalized patients or within 72 hours of admission to the hospital, and nosocomial if they were recovered beyond that period. Pneumonia was diagnosed based on clinical symptoms and signs, along with radiographic evidence of pneumonia when possible. Severe pneumonia was defined as pneumonia associated with tachycardia (>140 /minute), tachypnoea (>30/minute), hypotension (Systolic blood pressure <90 mmHg), hypoxemia (arterial oxygen partial pressure <8 kPa or oxygen saturation <90%), and/or more than 2 areas of consolidation. Outcome of patients with M. pneumoniae infection was classified into 4 categories; uneventful recovery, recovery following complications, death due to M. pneumoniae infection, or death unrelated to M. pneumoniae infection. The Statistical Package for Social Sciences (SPSS) program was used for data analysis. Comparison of categorical data was by Chi-square statistic or Fisher's exact test for small expected values. A total of 40 respiratory specimens from 40 patients were positive for M. pneumoniae over the 24-months study period. The demographic and epidemiological characteristics of the patients are summarized in Table 1 . Of all isolates, 37 (92.5%) were community-acquired and 3 (7.5%) were nosocomial. Thirty-three (82.5%) patients required admission to the hospital and the remaining 7 (17.5%) were treated as outpatients. Twenty-four isolates (60%) were associated with pneumonia, 14 (35%) with upper respiratory tract infections, and 2 (5%) with bronchiolitis. Of the 24 cases of pneumonia, 21 were confirmed radiologically and the remaining 3 were diagnosed clinically. The two cases of bronchiolitis occurred in 2 children, one and three years old. Thirty-one patients (77.5%) had comorbidities. Eleven patients (27.5%) had cardiopulmonary comorbidities (asthma, 8, lung fibrosis, 1, congestive heart failure, 1, congenial heart disease, 1), 9 patients (22.5%) were immunocompromised (malignancy, 7, steroid therapy, 3, Human immunodeficiency virus infection, 1), and 11 patients (27.5%) had other comorbidities (premature newborns, 2, and one each of myelodysplastic syndrome, myelopro-liferative disorder, sickle cell anemia, Evan's syndrome, Down syndrome, sarcoidosis, demyelinating disease, cerebral palsy, and spinal muscle atrophy). Organisms concomitantly isolated with M. pneumoniae from the respiratory tract included herpes simplex virus type 1 (2 occasions), adenovirus (2 occasions), cytomegalo virus (1 occasion), respiratory syncytial virus (1 occasion), and bacterial isolates (2 occasions: Acinetobacter species, 1, and Enter obacter cloacae, 1). Clinical manifestations associated with M. pneumoniae infections are summarized in Table 2 . Pneumonia was more common than upper respiratory tract infections (57.5 % versus 27.5%, respectively). Immunocompromised patients were more likely to present with pneumonia as opposed to upper respiratory tract infection or bronchiolitis than non-immunocompromised patients (8/9 versus 16/31, P = 0.05). Similarly, there was a tendency for patients 60 years of age or older to present with pneumonia more frequently than those below 60 (4/4 versus 20/36, P = 0.1). Of the 24 patients with clinically or radiologically confirmed pneumonia, 19 (79.2%) had crepitations and only 6 (25%) had bronchial breath sounds on physical examination. Of the 16 patients in whom wheezes were detected, 9 (56.3%) were not known to have asthma or other obstructive airway disease. Table 3 . Of the 24 patients with pneumonia, 21 (87.5%) were admitted to the hospital, and 20 (83.3%) had comorbidities. All patients with upper respiratory tract infections (11 patients) or bronchiolitis (2 patients) had uneventful recovery. Of the 24 patients with pneumonia, 14 (58.3%) had uneventful recovery, 4 (16.7%) recovered following some complications (acute respiratory distress syndrome, 2, respiratory failure, 1, septic shock, 1), 3 (12.5%) died because of M pneumoniae infection, and 3 (12.5%) died due to underlying comorbidities. The 3 patients who died of M pneumoniae pneumonia had other comorbidities; one had congestive heart failure, the second had congenital heart disease, and the third was a 3months old infant born prematurely at 32 weeks of gestation who previously had 3 episodes of pneumonia due to other pathogens. Mycoplasma pneumoniae is one of the most common causes of atypical pneumonia accounting for 5-23% of community-acquired pneumonia, [1, 2, 3, 4, 5] In a study of 511 children with acute respiratory tract infection in Riyadh, Saudi Arabia, Mycoplasma pneumoniae was found to be the second most common causative agent after Respiratory syncytial virus (RSV) accounting for 9% of all cases, [8] In a study of 112 adult patients with community acquired pneumonia admitted to a military hospital in Riyadh, Saudi Arabia, this organism accounted for 6% of all cases, [9] In another retrospective study of 567 pneumonic episodes in adult patients from Al-Qassim area, the organism accounted for 23% of all episodes, [10] The organism also causes other relatively minor infections such as pharyngitis, tracheobronchitis, bronchiolitis, and croup. It is transmitted from person-to-person by infected respiratory droplets during close contact. It is most common in school-aged children, military recruits, and college students. [11] Most cases occur singly or as family outbreaks. Larger outbreaks can also occur in closed populations such as military recruit camps or boarding schools, [12] Infection occurs most frequently during the fall and winter in temperate climates but may develop year-round, [13] The average incubation period is 3 weeks following exposure, [6] Although rare, complications are protean and may involve virtually any organ system such as the respiratory system (e.g.: pleurisy, pneumothorax, acute respiratory distress syndrome, lung abscess), the hematologic system (e.g.: hemolytic anemia, intravascular coagulation, thrombocytopenia), the dermatologic system (e.g.: maculopapular or urticarial rashes, erythema multiforme, erythema nodosum), the musculoskeletal system (e.g.: myalgias, arthralgias, arthritis), the cardiovascular system (e.g.: pericarditis, myocarditis), the nervous system (e.g.: meningoencephalitis, Guillain-Barre syndrome, neuropathies, acute psychosis), or the eye (optic disc edema, optic nerve atrophy, retinal exudation and hemorrhages). [6, 7, 14, 15, 16, 17, 18] Immunity following infection is not long lasting. [11] In our study, the infection affected all age groups but was most common in infants (32.5%) and preschool children (22.5%), and least common in adults aged 15 to 30 years (2.5%) and elderly above 70 years of age (5%). This contrasts with data from temperate countries where the infection is most common in school-aged children, and young adults. [11] One possible explanation for this difference is that infants and preschool children perhaps had more severe infections than did school-aged children, and young adults which prompted presentation of the former group to the hospital. The infection occurred year-round but was most common in the fall (35%), and spring (30%), and least common in the summer (10%). Most infections were community-acquired (92.5%). More than one half of patients (57.5%) presented with pneumonia, and about a third (27.5%) presented with upper respiratory tract infection, Immunocompromised patients and patients 60 years of age or older were more likely to present with pneumonia as opposed to upper respiratory tract infection than non-immunocompromised patients or those below 60 years of age. Cough (82.5%), fever (75%), and malaise (58.8%) were the most common presenting symptoms. Cough was usually dry or slightly productive of white sputum and mild to moderate in severity. Most febrile patients had mild to mod- erate fever of 39°C or less; high-grade fever of more than 39°C was rare. Crepitations (60%), and wheezes (40%) were the most common signs. Wheezes were as common in patients with no history of obstructive airway disease (9 patients) as it was in those with such a history (7 patients). Bronchial breathing as a sign of consolidation was detected in only one fourth of patients with pneumonia, which is consistent with the known disparity between clinical and radiological signs of M pneumoniae pneumonia. Crepitations, however, were detected in the majority (79.2%) of patients. Pleuritic chest pain and pleural effusion were rare. More than half (56.5%) of the patients with pneumonia had uneventful recovery. Mortality from M. pneumoniae pneumonia was high (12.5%) and occurred only in patients with underlying comorbidities. None of the 9 patients with no underlying comorbidities died of M pneumoniae pneumonia. The relatively high complications rate (16.7%) and mortality (12.5%) related to M. pneumoniae pneumonia are likely due to selection bias as most patients with pneumonia were sick enough to require admission to the hospital (21/24 or 87.5%) and most of them had comorbidities (20/24 or 83.3%). In conclusion, our data shed some light on the epidemiology and clinical features of M pneumoniae infections in one of the Saudi tertiary care centers. The data are comparable to those of other countries except for the finding that infections were more common in infants and preschool children than in school children and young adults. Additionally, mortality attributable to M. pneumoniae pneumonia was relatively high in patients with comorbidities. It is hoped this information will assist clinicians in their approach and management of respiratory tract infections.
1
Nitric oxide: a pro-inflammatory mediator in lung disease?
Inflammatory diseases of the respiratory tract are commonly associated with elevated production of nitric oxide (NO•) and increased indices of NO• -dependent oxidative stress. Although NO• is known to have anti-microbial, anti-inflammatory and anti-oxidant properties, various lines of evidence support the contribution of NO• to lung injury in several disease models. On the basis of biochemical evidence, it is often presumed that such NO• -dependent oxidations are due to the formation of the oxidant peroxynitrite, although alternative mechanisms involving the phagocyte-derived heme proteins myeloperoxidase and eosinophil peroxidase might be operative during conditions of inflammation. Because of the overwhelming literature on NO• generation and activities in the respiratory tract, it would be beyond the scope of this commentary to review this area comprehensively. Instead, it focuses on recent evidence and concepts of the presumed contribution of NO• to inflammatory diseases of the lung.
Since its discovery as a biological messenger molecule more than 10 years ago, the gaseous molecule nitric oxide (NO • ) is now well recognized for its involvement in diverse biological processes, including vasodilation, bronchodilation, neurotransmission, tumor surveillance, antimicrobial defense and regulation of inflammatory-immune processes [1] [2] [3] . In the respiratory tract, NO • is generated enzymically by three distinct isoforms of NO • synthase (NOS-1, NOS-2 and NOS-3) that are present to different extents in numerous cell types, including airway and alveolar epithelial cells, neuronal cells, macrophages, neutrophils, mast cells, and endothelial and smoothmuscle cells. In contrast with the other two NOS isoforms (NOS-1 and NOS-3), which are expressed constitutively and activated by mediator-induced or stress-induced cell activation, NOS-2 activity is primarily regulated transcriptionally and is commonly induced by bacterial products and pro-inflammatory cytokines. As such, inflammatory diseases of the respiratory tract, such as asthma, acute respiratory distress syndrome (ARDS) and bronchiectasis, are commonly characterized by an increased expression of NOS-2 within respiratory epithelial and inflammatory-immune cells, and a markedly elevated local production of NO • , presumably as an additional host defense mechanism against bacterial or viral infections. The drawback of such excessive NO • production is its accelerated metabolism to a family of potentially harmful reactive nitrogen species (RNS), including peroxynitrite (ONOO -) and nitrogen dioxide (NO 2 • ), especially in the presence of phagocyte-generated oxidants. The formation of such RNS is thought to be the prime reason why NO • can in many cases contribute to the etiology of inflammatory lung disease [4] [5] [6] . Despite extensive research into both pro-inflammatory and anti-inflammatory actions of NO • , the overall contribution of NO • to inflammatory conditions of the lung is not easily predicted and seems to depend on many factors, such as the site, time and degree of NO • production in relation to the local redox status, and the acute or chronic nature of the immune response. In addition, our current understanding of the pro-inflammatory or pro-injurious mechanisms of NO • or related RNS is incomplete; this commentary will focus primarily on these latter aspects. To explore a role for NO • (or NOS) in infectious or inflammatory diseases, two general research approaches have been taken: the use of pharmacological inhibitors of NOS isoenzymes, and the targeted deletion of individual NOS enzymes in mice. Both approaches suffer from the shortcoming that animal models of respiratory tract diseases generally do not faithfully reflect human disease. The use of NOS inhibitors to determine the contribution of individual NOS isoenzymes is also hindered by problems related to specificity and pharmacokinetic concerns. However, the unconditional gene disruption of one or more NOS isoforms, leading to lifelong deficiency, can have a markedly different outcome from pharmacological inhibition at a certain stage of disease, as the involvement of individual NOS isoenzymes can be different depending on disease stage and severity. Despite these inherent limitations, studies with the targeted deletion of NOS isoforms have led to some insights, indicating a role for NO • and NOS-2 in the etiology of some inflammatory lung diseases. For instance, mice deficient in NOS-2 are less susceptible to lethality after intranasal inoculation with influenza A virus, suffer less lung injury after administration of endotoxin, and display reduced allergic eosinophilia in airways and lung injury in a model of asthma, than their wild-type counterparts [7] [8] [9] . However, although the contribution of NOS-2 is expected in inflammatory conditions, recent studies have determined that NOS-1, rather than NOS-2, seems to be primarily involved in the development of airway hyper-reactivity in a similar asthma model [10] . The linkage of NOS-1 to the etiology of asthma was more recently supported in asthmatic humans by an association of a NOS-1 gene polymorphism with this disease, although the physiological basis for this association remains unclear [11] . Despite the potential contribution of NOS-2-derived NO • to lung injury after endotoxemia, the sequestration of neutrophils in the lung and their adhesion to postcapillary and postsinusoidal venules after administration of endotoxin were found to be markedly increased in NOS-2-deficient mice, and NOS-2 deficiency did not alleviate endotoxininduced mortality. It therefore seems that the 'harmful' and 'protective' effects of NOS-2 might contend with each other within the same model, which makes the assessment of the potential role of NOS in human disease even more difficult. In this context, it is interesting to note that humans or animals with cystic fibrosis have subnormal levels of NOS-2 in their respiratory epithelium, related to a gene mutation in the cystic fibrosis transmembrane conductance regulator [12] . This relative absence of epithelial NOS-2 might be one of the contributing factors behind the excessively exuberant respiratory tract inflammatory response in patients with cystic fibrosis, even in the absence of detectable respiratory infections. Overall, the apparently contrasting findings associated with NOS deficiency, together with concerns about animal disease models used, make interpretations and conclusions with regard to human lung disease all the more difficult. Pharmacological inhibitors of NOS have also been found to reduce oxidative injury in several animal models of lung injury, such as ischemia/reperfusion, radiation, paraquat toxicity, and endotoxemia (see, for example, [13] [14] [15] ). However, results are again not always consistent, and in some cases NOS inhibition has been found to worsen lung injury, indicating anti-inflammatory or protective roles for NO • . All in all, despite these inconsistencies, there is ample evidence from such studies to suggest a contributing role of NO • in various respiratory disease conditions, which continues to stimulate research into mechanistic aspects underlying such pro-inflammatory roles and modulation of NO • generation as a potential therapeutic target. Although the pro-inflammatory and injurious effects of NO • might be mediated by a number of diverse mechanisms, it is commonly assumed that such actions are largely due to the generation of reactive by-products generated during the oxidative metabolism of NO • ; these are collectively termed RNS. One of the prime suspects commonly implicated in the adverse or injurious properties of NO • is ONOO -, a potent oxidative species formed by its almost diffusion-limited reaction with superoxide (O 2 •-), which is a product of activated phagocytes and of endothelial or epithelial cells [4, 5, 13] . The formation of ONOOseems highly feasible under conditions of elevated production of both NO • and O 2 •in vivo, and its oxidative and cytotoxic potential is well documented [5, 6] . However, because the direct detection of ONOOunder inflammatory conditions is virtually impossible because of its instability and high reactivity, the formation of ONOOin vivo can be demonstrated only by indirect methods. Thus, many investigators have relied on the analysis of characteristic oxidation products in biological molecules, such as proteins and DNA, most notably free or protein-associated 3-nitrotyrosine, a product of tyrosine oxidation that can be formed by ONOO -(and several other RNS) but not by NO • itself (see, for commentary review reports primary research http://respiratory-research.com/content/1/2/067 example, [5] ). Indeed, elevated levels of 3-nitrotyrosine have been observed in many different inflammatory conditions of the respiratory tract [16] , which illustrates the endogenous formation of ONOOor related RNS in these cases. However, without known evidence for functional consequences of (protein) tyrosine nitration, the detection of 3-nitrotyrosine should not be regarded as direct proof of a pro-inflammatory role of NO • . Moreover, although the detection of 3-nitrotyrosine has in most cases been interpreted as conclusive evidence for the formation of ONOOin vivo (see, for example, [17] ), it should be realized that other RNS formed by alternative mechanisms might also contribute to endogenous tyrosine nitration. Indeed, it has recently become clear that the presence of inflammatory-immune cells, and specifically their heme peroxidases myeloperoxidase (MPO) and eosinophil peroxidase (EPO), can catalyze the oxidization of NO • and/or its metabolite NO 2 to more reactive RNS and thereby contribute to protein nitration [16, 18, 19] . This notion is further supported by the fact that 3-nitrotyrosine is commonly detected in tissues affected by active inflammation, mostly in and around these phagocytic cells and macrophages, which can also contain active peroxidases originating from apoptotic neutrophils or eosinophils. Hence, the detection of 3-nitrotyrosine in vivo cannot be used as direct proof of the formation of ONOO -, but merely indicates the formation of RNS by multiple oxidative pathways, possibly including ONOObut more probably involving the activity of phagocyte peroxidases [16, 20] . In this regard, a preliminary study with EPO-deficient mice has recently demonstrated the critical importance of EPO in the formation of 3-nitrotyrosine in a mouse model of asthma [21] . Future studies with animals deficient in MPO and/or EPO will undoubtedly help to clarify this issue. Given the considerable interest in 3-nitrotyrosine as a collective marker of the endogenous formation of NO •derived RNS, the crucial question remains of whether the detection of 3-nitrotyrosine adequately reflects the toxic or injurious properties of NO • . The formation of ONOO -(or of other RNS that can induce tyrosine nitration) might in fact represent a mechanism of decreasing excessive levels of NO • that might exert pro-inflammatory actions by other mechanisms. For instance, NO • can promote the expression of pro-inflammatory cytokines or cyclo-oxygenase (responsible for the formation of inflammatory prostanoids) by mechanisms independent of ONOO - [22, 23] , and the removal of NO • would minimize these responses. Furthermore, although ONOOor related NO •derived oxidants can be cytotoxic or induce apoptosis, these effects might not necessarily relate to their ability to cause protein nitration (see, for example, [16]). For instance, the bactericidal and cytotoxic properties of ONOOare minimized by the presence of CO 2 , even though aromatic nitration and other radical-induced modifications are enhanced [5] . Similarly, the presence of NO 2 in the incubation medium decreases the cytotoxicity of MPO-derived hypochlorous acid (HOCl) toward epithelial cells or bacteria, despite increased tyrosine nitration of cellular proteins (A van der Vliet and M Syvanen, unpublished data). Thus, it would seem that the cytotoxic properties of NO • and/or its metabolites might instead be mediated through preferred reactions with other biological targets, and these might not necessarily be correlated with the degree of tyrosine nitration. The extent of nitrotyrosine immunoreactivity in bronchial biopsies of asthmatic patients was correlated directly with measured levels of exhaled NO • and inversely with the provocation concentration for methacholine (PC 20 ) and forced expiratory volume in 1 s [24] . However, an immunohistochemical analysis of nitrotyrosine and apoptosis in pulmonary tissue samples from lung transplant recipients did not identify patients with an imminent risk of developing obliterative bronchiolitis [25] . It is therefore still unclear to what degree tyrosine nitration relates to disease progression. Several studies with purified enzymes have suggested that nitration of critical tyrosine residues adversely affects enzyme activity, but there is as yet no conclusive evidence in vivo for biological or cellular changes as a direct result of tyrosine nitration [16, 20] . For instance, tyrosine nitration was suggested to have an effect on cellular pathways by affecting cytoskeletal proteins or tyrosine phosphorylation, thereby affecting processes involved in, for example, cell proliferation or differentiation [16, 26] . Recent studies have provided support for selective tyrosine nitration within certain proteins [27, 28] and of selective cellular targets for nitration by RNS (see, for example, [29, 30] ), and such specificity might indicate a potential physiological role for this protein modification. However, in none of these cases could tyrosine nitration be linked directly to changes in enzyme function. Chemical studies have indicated that tyrosine nitration by RNS accounts for only a minor fraction of oxidant involved, and reactions with other biological targets (thiols, selenoproteins, or transition metal ions) are much more prominent [5, 6] . Indeed, the extent of tyrosine nitration in vivo is very low (1-1000 per 10 6 tyrosine residues according to best estimates [16]), although different analytical methods used to detect 3-nitrotyrosine in biological systems have often given inconsistent results. It is important to note that recent rigorous studies have unveiled substantial sources of artifact during sample preparation, which might frequently have led to an overestimation of tyrosine nitration in vivo in previous studies [31] . On the basis of current knowledge, the formation of 3-nitrotyrosine seems to be merely a marker of NO •derived oxidants, with as yet questionable pathophysiological significance. In view of the low efficiency of tyrosine nitration by biological RNS, and the endogenous presence of variable factors that influence protein nitration (antioxidants or other RNS scavengers), it seems unlikely that tyrosine nitration is a reliable mechanism of, for example, enzyme regulation. Nevertheless, the recent discovery of enzymic 'denitration' mechanisms that can reverse tyrosine nitration [32] merits further investigation of the possibility that tyrosine nitration might reflect a signaling pathway, for example analogous to tyrosine phosphorylation or sulfation. The biological effects of NO • are mediated by various actions, either by NO • itself or by secondary RNS, and the overall biochemistry of NO • is deceptively complex. Moreover, the metabolism and chemistry of NO • depend importantly on local concentrations and pH; the recently described acidification of the airway surface in asthmatics [33] might significantly affect NO • metabolism in these patients. It is well known that interactions with the ion centers of iron or other transition metals are responsible for many of the signaling properties of NO • ; the activation of the heme enzyme guanylyl cyclase and the consequent formation of cGMP is involved not only in smooth-muscle relaxation but also in the activation of certain transcription factors, the expression of several pro-inflammatory and anti-inflammatory genes (including cytokines and cyclo-oxygenase), and the production of respiratory mucus [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] . In addition to such direct signaling properties, many actions of NO • might be due largely to secondary RNS that can react with multiple additional targets, in some cases forming nitroso or nitro adducts as potentially unique NO • -mediated signaling mechanisms. As discussed, the formation of protein nitrotyrosine has been postulated as a potential RNSspecific signaling pathway. Even more interest has been given to the reversible S-nitros(yl)ation of protein cysteine residues, which has been proposed to affect a number of redox-sensitive signaling pathways, for example by the activation of p21 ras or the inhibition of protein tyrosine phosphatases [35, 36] . Similar modifications of reactive cysteine residues in transcription factors such as nuclear factor-κB or of caspases contribute to the regulation of gene expression and apoptosis [37] [38] [39] . The precise mechanisms leading to protein S-nitrosylation in vivo are still not clarified, but might involve dinitrogen trioxide (formed during the autoxidation of NO • ), iron-nitrosyl complexes, and perhaps ONOO - [16] ; changes in NO • metabolism during inflammatory lung diseases undoubtedly affect such NO • -dependent signaling pathways. In addition, S-nitrosylation can be reversed by either enzymic (thioredoxin or glutaredoxin) or chemical (metals or oxidants) mechanisms, and evidence is increasing that this reversible modification is complementary to more widely accepted oxidant-dependent redox signaling pathways [40] . The reported alterations in S-nitrosothiol levels in tracheal secretions of patients with asthma or cystic fibrosis further point to altered NO • metabolism in these cases, and might provide new clues to the role of S-nitrosylation in controlling such disease processes [41, 42] . Unfortunately, technical limitations to detect S-nitrosylation in specific protein targets in vivo have limited a full understanding of this potential signaling pathway; further research in these areas can be expected to establish more clearly its significance in the pathophysiological properties of NO • . Despite the by now overwhelming evidence for the increased formation of NO • and NO • -derived oxidants in many different lung diseases, the exact contribution of NO • or its metabolites to inflammatory lung disease is still unclear. Indeed, NO • might have distinctly different roles in different stages of respiratory tract inflammatory diseases, being pro-inflammatory or pro-injurious in acute and severe stages but perhaps being protective and antiinflammatory in more stable conditions; it is uncertain whether NOS is a suitable therapeutic target in the management of inflammatory lung disease. Caution is clearly needed when interpreting observations of tyrosine nitration in animal models of disease or in human tissues, which does not automatically implicate ONOO -(as often thought), but rather indicates the formation of RNS by various mechanisms. Furthermore, animal models of chronic lung disease that usually reflect short-term or acute inflammation might not always be applicable to chronic airway diseases in humans. For instance, phagocyte degranulation, a common feature observed in association with human airway inflammatory diseases such as asthma, does not seem to occur in mouse models of asthma [43] . Therefore the importance of granule proteins, such as heme peroxidases, in the pathology of human airway diseases might not be adequately reflected in such animal models. More work with animal models more characteristic of human diseases or with biopsy materials from human subjects will be required to unravel the precise role of NO • in inflammatory lung disease, and might establish more clearly whether the pharmacological inhibition of NOS isoenzymes can be beneficial. This brings up the interesting paradox that, despite presumed adverse roles of NO • in such inflammatory lung diseases as septic shock and ARDS, NO • inhalation has been suggested as a potential therapeutic strategy to improve overall gas exchange [44] . Intriguingly, in a rat model of endotoxemia, inhalation of NO • was found to reduce neutrophilic inflammation and protein nitration [45] , again supporting the crucial involvement of inflammatory-immune cells in this protein modification. For a better assessment of the role of NO • in respiratory tract diseases in humans, the production of RNS and/or characteristic markers would need to be more carefully monitored during various disease stages. Care should be given to analytical techniques, their quantitative capacity and the possibility of artifacts. The monitoring of exhaled NO • , although convenient and non-invasive, does not reflect the actual production or fate of NO • in the respiratory tract and is not well correlated with NOS activity in the lung [46] . We therefore need to continue research into the local biochemistry of NO • in the lung, taking into account the presence of secreted or phagocyte peroxidases and possible changes in local pH, as in asthmatic airways [33] , that might modulate NO • activity and metabolism. This might result in a better understanding of relationships between the various metabolic endproducts of NO • (NO 2 -, NO 3 -, or nitroso and nitro adducts) and its pro-inflammatory or injurious properties.
2
Surfactant protein-D and pulmonary host defense
Surfactant protein-D (SP-D) participates in the innate response to inhaled microorganisms and organic antigens, and contributes to immune and inflammatory regulation within the lung. SP-D is synthesized and secreted by alveolar and bronchiolar epithelial cells, but is also expressed by epithelial cells lining various exocrine ducts and the mucosa of the gastrointestinal and genitourinary tracts. SP-D, a collagenous calcium-dependent lectin (or collectin), binds to surface glycoconjugates expressed by a wide variety of microorganisms, and to oligosaccharides associated with the surface of various complex organic antigens. SP-D also specifically interacts with glycoconjugates and other molecules expressed on the surface of macrophages, neutrophils, and lymphocytes. In addition, SP-D binds to specific surfactant-associated lipids and can influence the organization of lipid mixtures containing phosphatidylinositol in vitro. Consistent with these diverse in vitro activities is the observation that SP-D-deficient transgenic mice show abnormal accumulations of surfactant lipids, and respond abnormally to challenge with respiratory viruses and bacterial lipopolysaccharides. The phenotype of macrophages isolated from the lungs of SP-D-deficient mice is altered, and there is circumstantial evidence that abnormal oxidant metabolism and/or increased metalloproteinase expression contributes to the development of emphysema. The expression of SP-D is increased in response to many forms of lung injury, and deficient accumulation of appropriately oligomerized SP-D might contribute to the pathogenesis of a variety of human lung diseases.
Surfactant protein-D (SP-D) is a member of the collagenous subfamily of calcium-dependent lectins (collectins) that includes pulmonary surfactant protein A (SP-A) and the serum mannose-binding lectin [1] [2] [3] . Collectins inter-act with a wide variety of microorganisms, lipids, and organic particulate antigens, and can modulate the function of immune effector cells and their responses to these ligands. This article reviews what is currently known about the sites of production, structure, function, and regulated expression of SP-D. Emphasis will be placed on functional attributes, known ligand interactions, and structure-function relationships believed to be important for host defense. For additional information on SP-A and other members of the collectin family, the reader is referred to other recent reviews [4] [5] [6] . SP-D is synthesized and secreted into the airspaces of the lung by the respiratory epithelium [1] . At the alveolar level, SP-D is constitutively synthesized and secreted by alveolar type II cells. More proximally in the lung, SP-D is secreted by a subset of bronchiolar epithelial cells, the non-ciliated Clara cells. Because SP-D is stored within the secretory granules of Clara cells [7, 8] , it seems likely that SP-D is subject to regulated secretion via granule exocy-tosis at this level of the respiratory tract. In some species, SP-D is also synthesized by epithelial cells and/or submucosal glands associated with the bronchi and trachea [9] . Although many alveolar macrophages show strong cytoplasmic and/or membrane labeling with antibody against SP-D, they do not contain detectable SP-D message. The lung seems to be the major site of SP-D production. However, there is increasing evidence for extrapulmonary sites of expression as assessed with monoclonal or affinity-purified antibodies, reverse-transcriptase-mediated PCR (RT-PCR), and/or hybridization assays of tissues from humans and other large mammals [10 • ,11-14] (summarized in Table 1 ). It is difficult to entirely exclude crossreactions or amplification of related sequences; however, localization to many of these sites in human tissues was confirmed by using monoclonal antibodies in combination with RT-PCR with sequencing of the amplified products [10 • ]. Non-pulmonary expression seems to be largely restricted to cells lining epithelial surfaces or ducts and certain glandular epithelial cells that are in direct or indirect continuity with the environment. Notable exceptions to this generalization might include heart, brain, pancreatic islets, and testicular Leydig cells. SP-D has also been identified in amnionic epithelial cells by immunohistochemistry [15] ; however, it is unclear whether this is synthesized locally or derived from the lung by way of the amniotic fluid. Interestingly, in many of these sites SP-D microscopically co-localizes with gp-340, an SP-D binding protein and putative SP-D receptor [10 • ]. Sites of extrapulmonary expression have also been described in small mammals. In the rat, SP-D message was identified in RNA extracted from skin and blood vessel [16] , and both protein and message were identified in gastric mucosa [17] and mesentery [13] . Using RT-PCR, SP-D message has also been identified in mouse stomach, heart, and kidney [14] . SP-D (43 kDa, reduced) consists of at least four discrete structural domains: a short, N-terminal domain; a relatively long collagenous domain, a short amphipathic connecting peptide or coiled-coil neck domain, and a C-terminal, Ctype lectin carbohydrate recognition domain (CRD). Each molecule consists of trimeric subunits (3 × 43 kDa), which associate at their N-termini (Fig. 1) . Although most preparations of SP-D contain a predominance of dodecamers (that is, four trimeric subunits), the proportions of various oligomers vary between species. For example, rat lavage and recombinant rat SP-D are almost exclusively assembled as dodecamers (four trimers), whereas recombinant human SP-D is secreted as trimers, dodecamers and higher-order multimers [18] . SP-D isolated from the lavage of some patients with alveolar proteinosis consists predominantly of higher-order multimers, which can contain up to 32 (or more) trimeric subunits (Fig. 1 ). Recent crystallographic and mutagenesis studies suggest that the structural determinants of saccharide binding are similar to those originally described for mannose-binding lectin [19,20,21 • ,22 • ]. At least two bound calcium ions and two intrachain disulfide crosslinks stabilize the required tertiary structure, and Glu321 and Asn323 within the CRD participate in glucose/mannose type recognition. Interactions with at least one glycolipid ligand, phosphatidylinositol (PI), require the participation of the C-terminal end of the protein [23, 24] . A trimeric cluster of CRDs is necessary for high-affinity binding to carbohydrate ligands [21 • ,25]. The crystal structure of human SP-D suggests that the spatial distribution of CRDs within a trimeric subunit permits simultaneous and cooperative interactions with two or three glycoconjugates displayed on the surface of a particulate ligand [21 • ]. Furthermore, solid-phase binding studies have shown that monomeric CRDs have an approximately 10-fold lower binding affinity for multivalent ligands than trimeric CRDs. Crystallographic studies of human SP-D further suggest that the spatial organization of CRDs within a trimer is stabilized by interactions of the C-terminal sequence with the trimeric neck domain [21 • ,26]. Interestingly, the three CRDs show a deviation from threefold asymmetry, suggesting some flexibility of the CRDs in relation to the neck. Thus, the dependence of the binding of PI on the C-terminal sequence could reflect conformational effects, rather than the direct participation of this sequence in ligand interactions. The collagen domain length of SP-D is highly conserved and lacks interruptions in the repeating Gly-X-Y sequence (in which X and Y are different amino acids). As for other collagenous proteins, this domain is enriched in imino acids and contains hydroxyproline. Unlike SP-A, SP-D also contains hydroxylysine. Although the collagen domain of rat, human, bovine, and mouse SP-D lacks cysteine residues, cDNA sequencing has identified a codon for cysteine within the collagen domain of pig SP-D [27 • ]; this suggests the possibility of alternative patterns of chain association and oligomeric assembly for pig SP-D. The first translated exon of SP-D contains a highly conserved and unusually hydrophilic Gly-X-Y sequence that shows little homology with the remainder of the collagen sequence. The functional significance of this region is unknown. However, it has been suggested that this region contributes to oligomer assembly or mediates interactions with cellular receptors. The collagen domain determines the maximal spatial separation of trimeric, C-terminal lectin domains within SP-D molecules, but might also contribute to normal oligomeric assembly and secretion. For example, deletion of the entire collagen domain of rat SP-D results in the secretion of trimers rather than dodecamers [28] . In addition, 2,2-dipyridyl, an inhibitor of prolyl hydroxylation that interferes with the formation of a stable collagen helix, causes the intracellular accumulation of 43 kDa monomers and dimers [29] . In any case, the complete conservation of the number of Gly-X-Y triplets suggests that the spatial separation of trimeric CRDs is critical for normal SP-D function. The N-terminal peptide of the mature protein contains two conserved cysteine residues at positions 15 and 20. These residues participate in interchain disulfide crosslinks that stabilize the trimer, as well as the N-terminal association of four or more trimeric subunits. Stable oligomerization of trimeric subunits permits cooperative or bridging interactions between spatially separated binding sites on the same surface or on different particles. The process of forming interchain disulfide bonds is complex, and appropriate crosslinking of the N-terminal domains might be rate limiting for secretion [30] . Subcellular fractionation studies suggest that interchain bonds form initially between the three chains of a trimeric subunit. Subsequent rearrangements within the rough endoplasmic reticulum might allow the covalent crosslinking of a single chain from one subunit and two crosslinked chains of another, with the associated elimination of free thiol groups. Mutant proteins that contain unpaired N-terminal cysteine residues are not secreted. However, it is unclear whether this results from abnormalities in disulfide bonding itself, or the failure to stabilize the required N-terminal conformation. The collagen domain contains hydroxylysyl-derived glycosides and a single N-linked oligosaccharide. In most species (human, rat, mouse, and cow) the site of N-linked glycosylation is located near the N-terminal end of the collagenous domain. Recently, it was shown that pig SP-D has an additional potential site of N-linked glycosylation within the CRD [27 • ]. Although rat and human lung lavage SP-D seem to be sialylated, as suggested by charge heterogeneity and cleavage with highly purified neuraminidase, preparations of human amniotic fluid and bovine lavage SP-D recovered from amniotic fluid showed predominantly complex type biantennary structures and no sialic acid [31] . A variant form of SP-D (50 kDa) has been identified in lavage from a subset of human lavage samples; this protein shows O-linked glycosylation of threonyl residues within the N-terminal peptide domain [32 • ]. At present, the functional significance of these sugars is not known. The presence of O-linked glycosylation within the N-terminal domain might be predicted to interfere with normal dodecamer assembly. In this regard, the O-glycosylated 50 kDa form of human SP-D is recovered as trimeric subunits or smaller species. As for many glycoproteins, the functional role of the attached carbohydrate is unknown. Mutational analysis has shown that the N-linked sugar on rat SP-D is not required for secretion, for dodecamer formation, or for interactions with a variety of microorganisms [29,33]. Consistent with its designation as a 'mannose-type' C-type lectin, SP-D preferentially binds to simple and complex saccharides containing mannose, glucose, or inositol [34, 35] . SP-D also interacts with specific constituents of pulmonary surfactant including PI [36-38] and glucosylceramide [39] . Binding to glucosylceramide involves interactions of the carbohydrate-binding sequences of the CRD with the glucosyl moiety. However, the interaction of SP-D with PI involves interactions with the lipid, as well as CRD-dependent interactions with the inositol moiety [24, 40] . Microorganisms are surfaced with a diverse and complex array of polysaccharides and glycoconjugates, and most classes of microorganism contain one or more sugars recognized by SP-D. However, the outcome of this interaction depends on the specific organism and can be modified by the conditions of microbial growth. The potential consequences of this interaction include the following: varying degrees of lectin-dependent aggregation (namely, microbial agglutination), enhanced binding of microorganisms or microbial aggregates to their 'receptors' on host cells, phagocyte activation, and opsonic enhancement of phagocytosis and killing, potentially involving one or more cellular receptors for SP-D. Binding to organisms in suspension is often -but not always -accompanied by some degree of aggregation. SP-D binds to purified lipopolysaccharide (LPS) isolated from a variety of Gram-negative organisms [35, 41] . In addition, LPS is the major cell wall component that is labeled on lectin blotting of outer membranes isolated from Escherichia coli [41] . The latter interactions involve the recognition of the core oligosaccharide domain, which contains glucose and heptose [41] . SP-D interacts preferentially with purified LPS molecules characterized by short or absent O-antigens and preferentially agglutinates bacterial strains expressing a predominance of rough (O-antigen-deficient) LPS [41, 44] . Although the core oligosaccharide domain of LPS constitutes the major ligand for SP-D on at least some Gram-negative bacteria, the mechanism of interaction with this group of microorganisms is probably heterogeneous. SP-D binds to some smooth, unencapsulated strains of Gram-negative bacteria by immunofluorescence. The mechanism is uncertain; the quantity or quality of binding differs from that observed for rough strains and does not necessarily result in agglutination. LPS molecules on the surfaces of bacteria show heterogeneity in the extent of maturation, so it is possible that this interaction is mediated by a subpopulation of LPS with deficient O-antigens and that the density of binding sites is too low for high-affinity binding. The recognition of the surface glycoconjugates on Gramnegative bacteria by SP-D depends not only on the expression of lectin-specific residues by a given strain or species, but also on the accessibility of these residues [1, 45] . For example, SP-D binds inefficiently to the core region of LPS of encapsulated Klebsiella, but efficiently agglutinates the corresponding unencapsulated phase variants. Interactions of SP-D with the core oligosaccharides of Gram-negative organisms are also influenced by the number of repeating saccharide units associated with the terminal O-antigen of the LPS [41,44]. Other potential ligands include the O-antigen domain of LPS, certain capsular polysaccharides, and membraneassociated glycoproteins. In this regard, SP-D can bind to di-mannose containing O-antigens expressed by a subset of Klebsiella serotypes (I Ofek, H Sahly and EC Crouch, unpublished data). Although other C-type lectins, specifically SP-A and the mannose receptor, can interact with specific capsular polysaccharides [46], a specific interaction of SP-D with capsular glycoconjugates or exopolysaccharides has not been described. The mechanism of interaction with Gram-positive organisms has not been elucidated. Lipoteichoic acids, which are the major glycolipids associated with the Gram-positive cell wall, do not detectably compete with LPS for binding to SP-D (I Ofek, A Mesika, M Kalina, Y Keisari, D Chang, D McGregor and EC Crouch, manuscript submitted). In preliminary studies we observed that binding was competed only partly with maltose and/or EDTA, raising the possibility that binding might be more complex than for some Gram-negative organisms. . However, similar effects were observed when the neutrophils were preincubated with SP-D, and there was only a slight enhancement of uptake when bacteria were incubated with human SP-D and washed before their addition to neutrophils. Notably, the extent of binding and internalization was dependent on the extent of multimerization, with human SP-D multimers demonstrating the highest potency. Differences in cell type, the extent of SP-D multimerization, or differences in size or organization of bacterial aggregates could account for some of the apparent inconsistencies. Although LPS mediates the binding of SP-D to at least some Gram-negative bacteria, SP-D can also bind to spe- In the latter study the authors suggested that fungal aggregation inhibits phagocytosis. Interestingly, SP-D binding directly inhibited fungal growth and decreased the outgrowth of pseudohyphae, the invasive form of the fungus, in the absence of phagocytic cells [57] . It is possible that these effects are also secondary to agglutination, possibly as result of nutrient deprivation. Purified rat and human SP-D inhibit the infectivity and hemagglutination activity of influenza SP-D can interact with host cells, both directly and indirectly. As indicated above, SP-D can enhance the phagocytosis and killing of certain microorganisms and enhance the oxidant response to microbial binding. However, at present there is only one study that suggests that the enhancement of phagocytosis by SP-D might involve the participation of an opsonic receptor. Furthermore, the enhanced uptake of IAV seems to be mediated by viral aggregation, with enhanced interactions of the virus with its natural receptors on the host cell. In any case, SP-D can interact directly with host cells, and in some cases can influence their behavior. SP-D is chemotactic and haptotactic for neutrophils and certain mononuclear phagocytes [59 • ,67-69] and can elicit directional actin polymerization in alveolar macrophages [69] . In this regard, SP-D is considerably more potent than SP-A. Early studies with natural proteins isolated from silicotic animals reported directed effects on the oxidant metabolism of isolated alveolar macrophages [70] . However, such effects can probably be attributed to endotoxin contamination and/or aggregation. Purified dodecamers do not significantly increase the production of nitric oxide [71] or of proinflammatory cytokines such as tumor necrosis factor-α (Y Kesari, H Wang, A Mesika, E Crouch and I Ofek, unpublished data). Interestingly, purified SP-D has been reported to increase the production of several metalloproteinases in the absence of a significant effect on proinflammatory cytokine production [72] . Despite the ability of SP-D to modulate a variety of cellular functions, little is currently known about potential cellular receptors for this protein. compartments [73] , but it is unclear whether the uptake is receptor dependent and whether SP-D is being internalized in association with specific ligands. There are at least two classes of binding to host cells: CRD-dependent and CRD-independent. Some studies have demonstrated CRD-dependent binding to phagocytes that can be inhibited with EDTA or competing saccharides, both in vitro and in vivo. As indicated above, the ability of SP-D to elicit the chemotaxis of neutrophilic and monocytic cells depends on the lectin activity of SP-D [68] . In addition, Kuan and coworkers reported that extracting formaldehyde-fixed alveolar macrophages with detergents largely eliminates the binding of purified SP-D, suggesting a membrane-associated ligand or glycolipid receptor [73] . Dong and Wright have extended these findings and suggest that PI can contribute to SP-D binding by alveolar macrophages [74] . It is of interest that SP-D can bind to recombinant sCD14 through interactions with N-linked oligosaccharides [51 • ]. Given that the membrane-associated form of CD14 is widely expressed on host cells, it is possible that CD14 can serve as a binding site on macrophages and other cell types. The phagocytic uptake of certain bacteria by neutrophils is also inhibited by calcium chelation or competing sugars [42]; however, this could result from the inhibition of microbial agglutination rather than lectin-dependent interactions with the phagocyte. Wang et al suggested that SP-D can bind to lymphocytic cells by a lectin-dependent mechanism [75 •• ] . In this regard, it is interesting to note that glucosylceramide, a ligand for SP-D in vitro, is one of the most abundant neutral glycolipids expressed by lymphoid cells. Reid and co-workers were the first to present evidence for lectin-independent binding [76] . These and other studies suggested that binding does not involve known C1q or collectin receptors. The only putative receptor protein, gp-340, is a widely expressed member of the scavenger receptor superfamily [77,78 • ]. It binds to the CRD of SP-D in a calcium-dependent manner that does not require the lectin activity of SP-D. Although the protein has been immunolocalized to alveolar macrophage membranes and distributes together with SP-D in many different human tissues [10 • ,77], it has not yet been shown to mediate the binding of SP-D to these cells or to participate in signal transduction events. The cDNAs isolated from lung have not shown a membrane-spanning region [77] , and the protein is abundant as a soluble component in BAL. Given that gp-340 is a highly multimerized protein that contains numerous potential ligand binding domains (Fig. 1b) , it is possible that the protein cooperates with SP-D in the neu-tralization or clearance of certain ligands rather than specifically mediating the interactions of SP-D with host cells. Wright and co-workers have demonstrated the binding of SP-D to isolated type II pneumocytes. The mechanism seems distinct from the binding to macrophages [79 • ]. The binding was dependent on concentration, time, and temperature and required calcium; it was not sensitive to protease treatment or to PI-phospholipase C. Although the internalized SP-D was degraded or recycled to lamellar bodies, SP-D binding did not alter the uptake of surfactant lipids. SP-D has demonstrated comparatively few direct effects on the metabolism of host cells, at least in situations where self-aggregation and endotoxin contamination have been excluded. One possible explanation is that modulation of cellular function requires the prior interaction of SP-D with a ligand. This would have numerous potential physiological advantages, because the presence of 'active' protein might be restricted to sites of microbial or antigenic deposition. The binding of complex, multivalent, particulate antigens to two or more CRDs could markedly alter the conformation of SP-D molecules, with respect to the spatial orientation of the arms in relation to the N-terminal crosslinking domain and/or with respect to the spatial orientation of the CRDs within a given trimeric subunit. Thus, the 'charging' of SP-D with a particulate ligand could lead to local or distant conformational changes that expose 'cryptic' binding sites for cellular receptors. There is some preliminary evidence consistent with the notion that the interaction of SP-D with a ligand alters its capacity to activate host cells. Table 3 and discussed below. SP-D can be isolated in different multimeric forms from proteinosis lavage [32 • ] and are produced by Chinese hamster ovary K1 cells transfected with human SP-D cDNA [18] . As described previously, the effects of SP-D on the neutrophil response to influenza virus are highly dependent on the ability of SP-D to agglutinate the viral particles, and the agglutination activity is directly correlated with the extent of multimerization. Trimers can bind to the virus but have little capacity to modulate neutrophil interactions. By contrast, highly multimerized proteins show greater activity than dodecamers [81] . Given these observations, factors that favor enhanced oligomerization or lead to the accumulation of trimeric subunits promote might influence SP-D function. For example, the liberation of active trimers by a hypothetical microbial protease could lead to the accumulation of molecules that might inhibit the aggregation-dependent activities of SP-D. In contrast, recombinant trimeric CRDs can stimulate chemotaxis [67] and decrease viral infectivity [65 • ]. Although higher-order oligomers of SP-D can self-aggregate and precipitate in the presence of calcium in vitro, the functional consequences are not known. The lectin activity of SP-A is decreased after the nitric oxide-dependent nitration of tyrosine residues [82] , and nitration decreases the ability of SP-A to enhance the adherence of Pneumocystis to alveolar macrophages [83] . However, similar findings have not yet been reported for SP-D. Conditions of mildly acidic pH, as might be found in endocytic compartments, are predicted to disrupt the lectin-dependent activities of SP-D [34]. Proteolytic degradation remains an important possibility. However, SP-D is highly resistant to degradation by a wide variety of neutral proteases in vitro, and degradation products have not yet been shown to accumulate under pathological conditions in vivo. Glucose concentrations at levels encountered in diabetes can interfere with SP-D's ability to interact with specific strains of IAV or other microorganisms in vitro [84 • ]. Many microorganisms release cell wall polysaccharides or glycoconjugates, which might interfere with the binding of collectins to the same or other organisms. In this regard, SP-D recovered from rats after the instillation of LPS into the airway shows decreased lectin activity, which is attributed to occupancy of the CRD with LPS [49 • ]. It seems reasonable to speculate that some organisms might compete with other organisms for binding to SP-D. Such a situation could conceivably predispose to secondary infections. Lastly, the potential inhibitory effects of competing saccharide ligands presents important methodological considerations for experiments using carbohydrate-containing cell culture medium or buffers. Non C-type lectins (such as ficolins) It is difficult to predict the functions of SP-D within the airspace. Other lectins with overlapping specificity are also present. Although the levels of mannose-binding lectin are probably low in the absence of increased vascular permeability, SP-A and the macrophage mannose receptor could conceivably interact with the same ligands in the distal airways and alveoli. Such interactions could lead to antagonistic or cooperative effects. Furthermore, we have little knowledge regarding the microanatomic distribution of these molecules in specific circumstances in vivo. Although most SP-A is probably associated with the insoluble phase of the alveolar lining material, and the macrophage mannose receptor is membrane-associated, the distribution might be altered in the setting of lung injury. Models of SP-D deficiency show no detectable anatomical or physiological abnormalities at birth. However, the animals gradually develop a patchy, subpleural alveolar lipidosis with associated type II cell hypertrophy, the accumulation of enlarged and foamy macrophages, and an apparent expansion of peribronchial lymphoid tissue [85 • ,86 • ]. Interestingly, the mice eventually develop distal-acinar emphysema and areas of subpleural fibrosis, which could reflect a continuing inflammatory reaction associated with abnormal oxidant metabolism and metalloproteinase activity [87 • ]. By contrast, SP-A-deficient mice (-/-) show essentially normal respiratory function and surfactant lipid metabolism [88, 89] but numerous apparent host defense abnormalities [90] . The capacity of SP-D to bind to specific strains of influenza A in vitro is highly correlated with the capacity of the virus to proliferate in mice in vivo [62] . Specifically, strains with more oligosaccharide attachments on the HA are preferentially neutralized by SP-D in vitro and show decreased proliferation in mice. Because the administration of mannan together with the virus increased the replication of IAV in the lung, the involvement of a mannose-type, C-type lectin was implicated. SP-D-sensitive IAV strains also replicate to higher titers in the lungs of diabetic mice than in nondiabetic controls [84 • ]. Replication of the virus is positively correlated with blood glucose level, and decreases in response to insulin treatment. Significantly, blood glucose levels comparable to those measured in the diabetic mice were sufficient to inhibit the interaction of SP-D with these viral strains in vitro. PR-8, a strain that does not interact with SP-D but does interact with SP-A, replicated to the same extent in diabetic and control mice. SP-D levels increase in association with certain infections. For example, SP-D levels, but not the levels of serum mannose-binding lectin, increase markedly after IAV infection [62] . Impressive increases in SP-D have also been observed in murine models of Pneumocystis carinii [91] and P. aeruginosa infection [92] . SP-D-deficient mice have not yet been extensively characterized with respect to host defense function. However, they show decreased viral clearance and enhanced inflammation after challenge with respiratory syncytial virus [93] and IAV (AM Levine, personal communication). In addition, they show increased inflammation, increased oxidant production, and decreased macrophage phagocytosis in response to intratracheally instilled group B streptococcus and Haemophilus influenzae (AM Levine, personal communication). Although the overexpression of wild-type SP-D in type II pneumocytes with the SP-D-deficient mice can prevent the lipidosis and inflammatory changes [94] , the ability of overexpressed wild-type SP-D or exogenous SP-D to ameliorate these abnormalities has not yet been described. The coexisting pulmonary abnormalities also complicate the interpretation of challenge models. For example, macrophage activation might enhance killing and offset any decrease that results more directly from SP-D deficiency. SP-D deficiency modifies the host response to instilled LPS with decreased lung injury and inflammatory cell recruitment [50]. Molecules that can bind to potential antigens and deliver them to macrophages and other antigen-presenting cells might contribute to the development of acquired immunity. In this regard, a few published observations suggest possible roles in the development of humoral and/or cellular immunity in response to microorganisms or complex organic antigens. For example, SP-D can decrease interleukin-2dependent T-lymphocyte proliferation [95 • ]. Interestingly, single-arm mutants were at least as potent as intact dodecamers in mediating this effect. SP-D also binds to oligosaccharides associated with dust mite allergen [96 • ], and can inhibit the binding of specific IgE to these allergens, possibly through direct, CRD-dependent binding to lymphocytes [96 • ]. Thus, alterations in the level of SP-D (or the state of oligomerization) might influence the development of immunological responses and contribute to the pathogenesis of asthma and other hypersensitivity disorders. There are other potential interplays between humoral immunity and collectins with regard to antimicrobial host defense. For example, increased glycosylation of IAV coat proteins, an adaptation that is believed to help the virus to evade antibody-mediated neutralization, is associated with increased reactivity with SP-D and other collectins [62]. Thus, the relative potential importance of antibody and collectin-mediated host defenses might be influenced by subtle variations in the structure of the microbial surface. There is little recent information on the developmental regulation of SP-D expression. In general, SP-D increases rapidly late in gestation [97] [98] [99] [100] . The production of SP-D increases during the culture of fetal lung explants, and expression can be increased with glucocorticoids [98, 100, 101] . The exposure of fetal rats to glucocorticoids in vivo leads to precocious expression with increased numbers of SP-D-expressing cells and increased cellular levels of SP-D message [98, 101, 102] . Although SP-D is produced constitutively within the lung, protein accumulation and gene expression are inducible and increases in SP-D expression have been observed in a number of disease states or models (Tables 4 and 5 ). In general, the synthesis and secretion of SP-D increase in association with lung injury and activation of the respiratory epithelium [1] . For example, levels of SP-D mRNA and SP-D accumulation are increased within 24-72 h after intratracheal instillation of LPS [103 • ], and SP-D expression by alveolar and bronchiolar epithelial cells increases after exposure of rats to 95% O 2 for 12 h [104] . Keratinocyte growth factor (KGF) increases SP-D expression and protein production in association with pneumocyte hyperplasia and after injury caused by bleomycin [105] . In addition, the levels of SP-D can increase markedly in response to the overexpression of certain cytokines, such as interleukin-4, or in response to microbial challenge [91, 92] . Studies of the upstream regulatory region of the SP-D gene have demonstrated increased promoter activity in the presence of glucocorticoids, which is consistent with the findings in vivo and in lung organ culture [106] . However, no functional glucocorticoid response elements have been identified, and the effects of dexamethasone seem to be secondary and involve the effects of other transregulatory molecules. The activity of the human SP-D promoter is dependent on a conserved activator protein-1 (AP-1) element (-109) that binds to members of the fos and jun families of transcriptional factors [107] . In addition, the promoter contains multiple functional binding sites for CCAAT-enhancer-binding protein (C/EBP) transcription factors. Mutagenesis experiments suggest that these are required for basal and stimulated promoter activity, and promoter activity is markedly increased in H441 cells after co-transfection with C/EBPβ cDNA (YC He and E crouch, unpublished data). The importance of the conserved AP-1 element and the presence of multiple binding sites for C/EBP transcription factors is consistent with the observed modulation of SP-D expression in the setting of tissue injury. SP-D promoter activity is not dependent on the binding of thyroid transcription factor 1 (TTF-1) [107] . However, promoter activity is dependent on two interacting forkhead binding sites, upstream and downstream of the AP-1 element; these sites bind to hepatic nuclear factor-3α and apparently other forkhead box proteins in H441 lung adenocarcinoma nuclear extracts [107] . Initial comparison of genomic and cDNA sequence suggested the existence of genetic polymorphisms in the SP-D coding sequence, including one in the N-terminal propeptide domain (Thr11 compared with Met11 in the mature protein) and three additional differences within the collagen domain at positions 102, 160, and 186 [108] . The latter substitutions are conservative to the extent that they are not expected to disrupt the collagen helix. Floros Table 5 Increased SP-D accumulation or expression in animal models Silicosis Rat [118] Hyperoxia Rat [104] Endotoxin (LPS) Rat [103] Challenge with P. aeruginosa Mouse [92] Challenge with IAV Mouse [62] Challenge with Pneumocystis carinii SCID mouse [91] Rat [119] Overexpression of interleukin-4 Mouse [120] SCID, severe combined immunodeficiency. and co-workers have recently confirmed the existence of polymorphisms at positions 11 and 160 of the mature protein [109] . The potential biological significance, if any, is not known. Interestingly, the 50 kDa variant of SP-D showed O-linked glycosylation of Thr11 [32 • ], suggesting that this polymorphism might be associated with altered glycosylation. Interestingly, the 50 kDa variant was recovered as trimeric subunits, raising the possibility that differences in the glycosylation of residue 11, which is immediately N-terminal to Cys15, could influence multimerization and the capacity of SP-D to participate in bridging interactions. There is increasing evidence that SP-D interacts specifically with a wide variety of respiratory pathogens, modulates the leukocyte response to these organisms, and participates in aspects of pulmonary immune and inflammatory regulation (Table 6) . SP-D can influence the activity of phagocytes through CRD-dependent and CRD-independent interactions. At least some of the effects of SP-D result from aggregation with enhanced binding of the agglutinated ligand to their natural 'receptors'. Although the lung is the major site of SP-D expression, it is likely that the protein has more generalized roles in host defense and the acute response to infection and tissue injury. 16
3
Role of endothelin-1 in lung disease
Endothelin-1 (ET-1) is a 21 amino acid peptide with diverse biological activity that has been implicated in numerous diseases. ET-1 is a potent mitogen regulator of smooth muscle tone, and inflammatory mediator that may play a key role in diseases of the airways, pulmonary circulation, and inflammatory lung diseases, both acute and chronic. This review will focus on the biology of ET-1 and its role in lung disease.
from Xenopus laevis [16] . ETA receptors in normal lung are found in greatest abundance on vascular and airway smooth muscle, whereas ETB receptors are most often found on the endothelium. Clearance of ET-1 from the circulation is mediated by the ETB receptor primarily in the lung, but also in the kidney and liver [17] . Activation of both ETA and ETB receptors on smooth muscle cells leads to vasoconstriction whereas ETB receptor activation leads to bronchoconstriction. Activation of ETB receptors located on endothelial cells leads to vasodilation by increasing nitric oxide (NO) production. The mitogenic and inflammatory modulator functions of ET-1 are primarily mediated by ETA receptor activity. Binding of the ligand to its receptor results in coupling of cell-specific G proteins that activate or inhibit adenylate cyclase, stimulate phosphatidyl-inositol-specific phosholipase, open voltage gated calcium and potassium channels, and so on. The varied effects of ET-1 receptor activation thus depend on the G protein and signal transduction pathways active in the cell of interest [18] . A growing number of receptor antagonists exist with variable selectivity for one or both receptor subtypes. Regulation of ET-1 is at the level of transcription, with stimuli including shear stress, hypoxia, cytokines (IL-2, IL-1β, tumor necrosis factor α, IFN-β, etc), lipopolysaccharides, and many growth factors (transforming growth factor-β, platelet-derived growth factor, epidermal growth factor, etc) inducing transcription of ET-1 mRNA and secretion of protein [18] . ET-1 acting in an autocrine fashion may also increase ET-1 expression [19] . ET-1 expression is decreased by NO [20] . Some stimuli may additionally enhance preproET-1 mRNA stability, leading to increased and sustained ET-1 expression. The number of ETA and ETB receptors is also cell specific and regulated by a variety of growth factors [18] . Because ET-1 and receptor expression is influenced by many diverse physical and biochemical mechanisms, the role of ET-1 in pathologic states has been difficult to define, and these are addressed in subsequent parts of this article. In the airway, ET-1 is localized primarily to the bronchial smooth muscle with low expression in the epithelium. Cellular subsets of the epithelium that secrete ET-1 include mucous cells, serous cells, and Clara cells [21] . ET binding sites are found on bronchial smooth muscle, alveolar septae, endothelial cells, and parasympathetic ganglia [22, 23] . ET-1 expression in the airways, as previously noted, is regulated by inflammatory mediators. Eosinophilic airway inflammation, as may be seen in severe asthma, is associated with increased ET-1 levels in the lung [24] . ET-1 secretion may also act in an autocrine or paracrine fashion, via the ETA receptor, leading to increased transepithelial potential difference and ciliary beat frequency, and to exerting mitogenic effects on airway epithelium and smooth muscle cells [25] [26] [27] [28] . All three endothelins cause bronchoconstriction in intact airways, with ET-1 being the most potent. Denuded bronchi constrict equally to all three endothelins, suggesting considerable modulation of ET-1 effects by the epithelium [29] . The vast majority of ET-1 binding sites on bronchial smooth muscle are ETB receptors, and bronchoconstriction in human bronchi is not inhibited by ETA antagonists but augmented by ETB receptor agonists [30] [31] [32] . Since cultured airway epithelium secretes equal amounts of ET-1 and ET-3, which have equivalent affinity for the ETB receptor, bronchoconstriction could be mediated by both endothelins [33] . While ET-1 stimulates release of multiple cytokines important in airway inflammation, it does not enhance secretion of histamine or leukotrienes. ET-1 does increase prostaglandin release [32] . Inhibition of cyclo-oxygenase, however, has no effect on bronchoconstriction suggesting that, despite the release of multiple mediators, ET-1 mediated bronchoconstriction is a direct effect of activation of the ETB receptor [32] . ETA mediated bronchoconstriction may also be important following ETB receptor desensitization or denudation of the airway epithelium, as may occur during airway inflammation and during the late, sustained airway response to inhaled antigens [31, 34, 35] . Interestingly, heterozygous ET-1 knockout mice, with a 50% reduction in ET-1 peptide, have airway hyperresponsiveness but not remodeling, suggesting the decrease in ET-1 modulates bronchoconstriction activity by a functional mechanism, possibly by decreasing basal NO production [36, 37] . Asthma is also an inflammatory airway disease characterized by bronchoconstriction and hyperreactivity with influx of inflammatory cells, mucus production, edema, and airway thickening. ET-1 may have important roles in each of these processes. While ET-1 causes immediate bronchoconstriction [38] , it also increases bronchial reactivity to inhaled antigens [35] as well as influx of inflammatory cells [39, 40] , increased cytokine production [40] , airway edema [41] , and airway remodeling [28, 42, 43] . Airway inflammation also leads to increased ET-1 synthesis, possibly perpetuating the inflammation and bronchoconstriction [44] . ET-1 release from cultured peripheral mononuclear and bronchial epithelial cells from asthmatics is also increased [45, 46] . Inhibition of ETA or combined ETA and ETB receptors additionally leads to decreased airway inflammation in antigen-challenged animals, suggesting that the proinflammatory effects of ET-1 in the airway are mediated by ETA receptors [39, 47] . Children with asthma have increased circulating levels of ET-1 [48] . Adult asthmatics have normal levels between attacks but, during acute attacks, have elevated serum ET-1 levels that correlate inversely with airflow measurements and decrease with treatment [49] . Bronchoalveolar lavage (BAL) ET-1 in asthmatics is similarly increased to concentrations that cause bronchoconstriction and inversely correlates with forced expiratory volume in 1 s (FEV 1 ) [29, 50, 51] . As in cultured epithelial cells, ET-1 and ET-3 are found in equal amounts in BAL fluid from asthmatics [33, 52] . There is also a relative increase in ETB versus ETA receptor expression in asthmatic patients, which may contribute to increased bronchoconstriction [53] . Not all asthmatics, however, have increased ET-1 as patients with nocturnal asthma have decreased BAL ET-1 levels [54] . Treatment of acute asthma exacerbations with steroids, beta-adrenergic agonists or phosphodiesterase inhibitors resulted in decreased BAL ET-1 [52, 55] . Immunostaining and in situ hybridization for ET-1 in biopsy specimens from asthmatics have shown an increase in ET-1 in the bronchial epithelium that correlates with asthma symptoms [46, 56] . Cigarette smoking leads to increased circulating ET-1 [57] but patients with chronic obstructive pulmonary disease, in the absence of pulmonary hypertension and hypoxemia, do not have increased plasma ET-1 [58] [59] [60] . Increases in urinary ET-1 instead correlate with decreases in oxygenation, possibly through hypoxic release of ET-1 from the kidney [61, 62] . Smokers also have impaired ET-1 mediated vasodilation that correlates with bronchial hyperresponsiveness and may contribute to pulmonary hypertension [63, 64] . ET-1 has been implicated in the pathogenesis of bronchiectasis by its ability to promote neutrophil chemotaxis, adherence, and activation [65] [66] [67] [68] [69] . Sputum ET-1 levels are increased in patients with cystic fibrosis [59] , and sputum ET-1 correlated with Pseduomonas infection in noncystic fibrosis related bronchiectasis [70] . ET-1 has also been implicated in the pathogenesis of bronchiolitis obliterans (BO), which is characterized by injury to small conducting airways resulting in formation of proliferative, collagen rich tissue obliterating airway architecture. BO is the leading cause of late mortality from lung transplantation, and ET-1 is increased in lung allografts [71] . The pro-inflammatory and mitogenic properties of ET-1 in the airways has led to speculation that ET-1 may be involved in formation of the lesion [28] . This is further supported by the increase in BAL ET-1 in lung allografts [72, 73] . The in vivo gene transfer of ET-1 to the airway epithelium using the hemagglutinating virus of Japan in rats recently resulted in pathologic changes in the distal airways identical to those seen in human BO specimens [74] . These changes were not due to nonspecific effects of the hemagglutinating virus of Japan itself, but could be attributed to the presence of the ET-1 gene, which was localized to the airway epithelium, hyperplastic lesions, and alveolar cells. Pulmonary hypertension is a rare and progressive disease characterized by increases in normally low pulmonary vascular tone, pulmonary vascular remodeling, and progressive right heart failure. ET-1 has been implicated as a mediator in the changes seen in pulmonary hypertension. In the pulmonary vasculature, ET-1 is found primarily in endothelial cells and to a lesser extent in the vascular smooth muscle cells. The endothelium secretes ET-1 primarily to the basolateral surface of the cell. ET-1 secretion may be increased by a variety of stimuli including cytokines, catecholamines, and physical forces such as shear stress, and decreased by NO, prostaglandins, and oxidant stress [20, [75] [76] [77] [78] . Hypoxia has been reported to increase, have no effect, or decrease ET-1 release from endothelial cells [79] [80] [81] [82] [83] . Activation of the receptors for ET-1 in the pulmonary vasculature leads to both vasodilation and vasoconstriction, and depends on both cell type and receptor. In the whole lung, ETA receptors are the most abundant and are localized to the medial layer of the arteries, decreasing in intensity in the peripheral circulation [84, 85] . ETB receptors are also found in the media of the pulmonary vessels, increasing in intensity in the distal circulation, while intimal ETB receptors are localized in the larger elastic arteries [85] . This distribution of receptors has important implications in understanding ET-1 regulation of vascular tone. Vascular ET-1 receptors may be increased by several factors including angiotensin and hypoxia [80, [85] [86] [87] . ET-1 can act as both a vasodilator and vasoconstrictor in the pulmonary circulation. Generation of NO or opening of ATP-sensitive potassium channels leading to hyperpolarization results in vasodilation mediated by ETB receptors on pulmonary endothelium [88, 89] . In hypertensive, chronically hypoxic lungs with increased ETB receptor expression, augmented vasodilation is due to increased ETB mediated NO release that is inhibited by hypoxic ventilation, while inhibition of NO synthesis leads to increased ET-1 mediated vasoconstriction [85, [90] [91] [92] . Both ETA and ETB receptors, conversely, acting on vascular smooth muscle, mediate ET-1 induced vasoconstriction. In the normal lung, ET-1 causes vasoconstriction primarily by activation of the ETA receptors in the large, conducting vessels of the lung [93, 94] . In the smaller, resistance vessels of the lung, ETB receptors in the media predominate and are responsible for the ET-1 induced vasoconstriction [93] . Interestingly, preconstriction of the pulmonary circulation resulted in a shift from primarily ETA mediated to ETB mediated vasoconstriction [94] . The overall effect of ET-1 on vascular tone depends on both the dose and on the pre-existing tone in the lung. ET-1 administration during acute hypoxic vasoconstriction will result in transient pulmonary vasodilation [89] . This effect is dose dependent, with lower doses leading to vasodilation while higher or repetitive doses cause vasoconstriction following an initial, brief vasodilation [89] . The role of ET-1 in the acute hypoxic vasoconstriction in the lung is not certain. ETA receptor antagonism attenuates hypoxic pulmonary vasoconstriction in several species [95] , and ET-1 may be implicated in the mechanism of acute hypoxic response by inhibition of K-ATP channels [96] . Several lines of evidence have suggested the importance of ET-1 in chronic hypoxic pulmonary hypertension. ET-1 is increased in plasma and lungs of rats following exposure to hypoxia [80, 97] . Treatment with either ETA or combined ETA and ETB receptor antagonists additionally attenuates the development of hypoxic pulmonary hypertension [98, 99] . ET-1 has also been implicated in the vascular remodeling associated with chronic hypoxia through its mitogenic effects on vascular smooth muscle cells [98, 100] . ET-1 has also been implicated in other animal models of pulmonary hypertension. ET-1 is increased in fawn hooded rats that develop severe pulmonary hypertension when raised under conditions of mild hypoxia and in monocrotaline treated rats [101, 102] . The increase in ET-1 in both of these forms of pulmonary hypertension may be contributing to increases in vascular tone as well as in vascular remodeling [103] [104] [105] [106] 114] . Interestingly, transgenic mice overexpressing the human preproET-1 gene, with modestly increased lung ET-1 levels (35-50%), do not develop pulmonary hypertension under normoxic conditions or an exaggerated response to chronic hypoxia [107] . Human pulmonary hypertension is classified as primary, or unexplained, or secondary to other cardiopulmonary diseases or connective tissue diseases (ie scleroderma). Hallmarks of the disease include progressive increases in pulmonary vascular resistance and pulmonary vascular remodeling, with thickening of the medial layer small pulmonary arterioles and formation of the complex plexiform lesion [108] . Circulating ET-1 is increased in humans with pulmonary hypertension, either primary or due to other cardiopulmonary disease [109] . Levels are highest in patients with primary pulmonary hypertension. Since the lung is the major source for clearance of ET-1 from the circulation, increased arterio-venous ratios as seen in primary pulmonary hypertension suggest either decreased clearance or increased production in the lung [17, 109] . ET-1 is also increased in lungs of patients with pulmonary hypertension, with the greatest increase seen in the small resistance arteries and the plexiform lesions [110] , and may correlate with pulmonary vascular resistance [111] . Interestingly, treatment with continuous infusion of prostacyclin resulted in clinical improvement and a decrease in the arterio-venous ratio of ET-1 [112] , possibly by decreasing ET-1 synthesis from endothelial cells [76] . Studies using ET-1 receptor antagonists in the treatment of primary pulmonary hypertension are underway and may offer hope to patients with this disease by inhibiting this pluripotent peptide's effects on vascular tone and remodeling. Several lines of evidence suggest the importance of ET-1 in lung allograft survival and rejection. The peptide has been implicated as an important factor in ischemia-reperfusion injury at the time of transplant as well as in acute and chronic rejection of the allograft. Circulating ET-1 is increased in humans undergoing lung transplant immediately following perfusion of the allograft. Plasma ET-1 increased threefold within minutes, remained high for 12 hours following transplantation, and declined to near normal levels within 24 hours [113] . This increase in ET-1 correlated with the increase in pulmonary vascular resistance occurring about 6 hours post-transplantation, suggesting that the release of ET-1 in the circulation may have mediated this event. ET-1 in BAL fluid from recipients of lung allografts is similarly increased several fold and remains elevated up to 2 years post-transplant [72, 73] . In recipients of single lung transplants, ET-1 was increased 10-fold in BAL fluid from the transplanted lung compared with the native lung, suggesting that the increase in ET-1 was due to the graft and not the underlying disease requiring transplant [72] . ET-1 in BAL fluid did not, however, correlate with episodes of infection or rejection. The cellular source of ET-1 in lung allografts is unknown. The expression of ET-1 in nontransplanted human lungs is low and found primarily in the vascular endothelium [114] . Transbronchial biopsy specimens obtained either for surveillance or for clinical suspicion of infection or rejection following transplantation revealed the presence of ET-1 in the airway epithelium and in alveolar macrophages [115] . ET-1 was occasionally seen in lymphocytes but not in the endothelium or pneumocytes. ET-1 localization was no different in surveillance specimens compared with infected or rejecting lungs, or changed over time from transplantation. This study suggests that the source of the increased BAL ET-1 in transplanted lungs is due to the increased number of alveolar inflammatory cells and de novo expression in the airway epithelium. The biologic importance of the ET-1 from inflammatory cells is supported by the observation that peripheral mononuclear cells from dogs with mild to moderate lung allograft rejection cause vasoconstriction in pulmonary arterial rings, which is attenuated by the ETA blocker BQ123 [116] . Analysis of ET-1 binding activity in failed transplanted human lungs suggested that ET-1 binding activity was not different compared with normal lung in the lung parenchyma, bronchial smooth muscle, or perivascular infiltrates. ET-1 binding was, however, decreased in small muscular arteries (pulmonary arteries and bronchial arteries) in the failed transplants, suggesting a role for ET-1 in impaired vasoregulation of transplanted lungs [117] . Ischemia-reperfusion injury is the leading cause of early post-operative graft failure and death. In its severest manifestation, increased pulmonary vascular resistance, hypoxia, and pulmonary edema lead to cor pulmonale and death [118] . ET-1 has been implicated as a mediator of these events. The increase in pulmonary vascular resistance observed in human recipients of lung allografts follows an increase in circulating ET-1 and falls with decreases in circulating ET-1 [113] . A similar pattern is seen in dogs subjected to allotransplantation [119] . Conscious dogs with left pulmonary allografts demonstrate an increase in both resting pulmonary perfusion pressure and acute pulmonary vasoconstrictor response to hypoxia [120] . Administration of ETA selective or combined ETA and ETB receptor blockers did not change the resting tone. ETB receptor mediated hypoxic pulmonary vasoconstriction appeared, however, to be increased in allograft recipients. In another study, administration of a mixed ETA and ETB receptor antagonist (SB209670) to dogs before reperfusion of the allograft resulted in a marked increase in oxygenation, decreases in pulmonary arterial pressures and improved survival compared with control animals [121] . In a model of ischemia reperfusion, inhibitors of ECE additionally attenuated the increase in circulating ET-1 and the severity of lung injury [122] . ET-1 receptor antagonists did not, however, completely eliminate the ischemia-reperfusion injury, suggesting that changes in other vasoactive mediators, such as an increase in thromboxane, a decrease in prostaglandins, or a decrease in NO, may also contribute to the increased pulmonary vascular resistance. Administration of NO donor (FK409) to both donor and recipient dogs before lung transplantation reduced pulmonary arterial pressure, lung edema, and inflammation, and improved survival. This suggests that reductions in NO following transplantation may be partly responsible for early graft failure [123] . Treatment with NO donor was also associated with a decrease in plasma ET-1 levels. Acute rejection is manifested by diffuse infiltrates, hypoxia, and airflow limitation, and may lead to respiratory insufficiency and death. BAL ET-1 was increased in dogs during episodes of acute rejection that decreased with immunosuppressive treatment [124] . Acute episodes of rejection in humans, however, are not associated with further increases in BAL ET-1 [72] . Chronic rejection of allografts, manifested as BO, is the major cause of morbidity and mortality in long-term lung transplant survivors [71] . The etiology of BO following transplant is unclear but may be related to repeated episodes of acute rejection, chronic low-grade rejection, or organizing pneumonia [125] . As discussed earlier, a chronic increase in ET-1, as seen in lung allografts, may contribute to bronchospasm and proliferative bronchiolitis obliterans due to the bronchoconstrictor and smooth muscle mitogenic effects of ET-1 [28, 126] . This is further supported by the increase of BAL ET-1 in the transplanted lung, which is susceptible to BO, but not the native lung in recipients of single lung transplants [72] . The mitogenic effects of ET-1 may play a role in the development of pulmonary malignancy as well as metastasis to the lung. Many human tumor cell lines, including prostate, breast, gastric, ovary, colon, etc, produce ET-1. The importance of the ET-1 may lie in its mitogenic effects on tumor growth and survival. This has been suggested by blockade of ETA receptors resulting in a decrease in mitogenic effects of ET-1 in a prostate cancer and colorectal cell lines [127, 128] . ET-1 receptors in tumor cells may also be altered with increases in the ETA receptor and downregulation of ETB receptors [129] . Other tumors may have an increase in ETB receptors, however, and blockade of ETB results in a decrease in tumor growth [130, 131] . Tumor cells may, as a result of this altered balance, lose the ability to respond to regulatory signals from their environment. ET-1 may additionally protect against Fas-ligand mediated apoptosis [132] . ET-1 has been detected using immunohistochemistry and in situ hybridization in pulmonary adenocarcinomas and squamous cell tumors and, to a lesser extent, small cell and carcinoid tumors [133] . In situ hybridization also demonstrated a similar pattern of ET-1 mRNA expression in non-neuroendocrine tumors. ET-1 receptors have also been found in a variety of pulmonary tumor cell lines. ETA receptors were found in small cell tumors, adenocarcinomas and large cell tumors, while ETB receptors were expressed primarily in adenocarcinomas and small cell tumors [134] . ECE, which converts big ET-1 to ET-1, the committed step in ET-1 biosynthesis, was also found in human lung tumors but not in adjacent normal lung [135] . These findings, combined with the presence of ET-1 in lung tumors, suggest a possible autocrine loop that sustains and supports the growth of lung tumors. A recent study, however, suggested that, while ETA and ECE-1 were detectable in lung tumors, these genes were downregulated compared with normal bronchial epithelial cell lines [136] . It was proposed that the role of ET-1 in lung tumors is not that of an autocrine factor, but that of a paracrine growth factor to the stroma and vasculature surrounding the tumor allowing angiogenesis. Tumor angiogenesis is necessary for continued growth of the tumor beyond the limits of oxygen diffusion. The growth of vessels into the tumor is also important to metastatic potential of the tumor. ET-1 may play an important role in angiogenesis and tumor growth and survival Available online http://respiratory-research.com/content/2/2/090 commentary review reports primary research through induction of vascular endothelial growth factor expression and sprouting of new vessels into the tumor and surrounding tissue [137, 138] . ET-1 binding activity was found in blood vessels and vascular stroma surrounding lung tumors at the time of resection, most markedly surrounding squamous cell tumors [139] . ET-1 production may be further augmented by the hypoxic environment found within large solid tumors [140] . Since metastasis is dependent on neo-vascularization, ET-1 may also be an important mediator of this phenomenon. ET-1 receptor antagonists may have a useful role in the treatment of neoplastic disease by inhibiting growth as well as metastatic potential of human tumors. Experimental lung injury of many different types results in increased circulating ET-1, BAL ET-1, and lung tissue ET-1 [18] . ET-1 levels in humans are also increased in sepsis, burns, disseminated intravascular coagulation, acute lung injury, and acute respiratory distress syndrome (ARDS) [141] [142] [143] [144] [145] [146] [147] . ET-1 increases also correlate with a poorer outcome with multiple organ failure, increased pulmonary arterial pressure, increased airway pressure and decreased PiO 2 /FiO 2 , while clinical improvement correlates with decreased ET-1 levels [144, 145, 147] . The arterio-venous ratio for ET-1 is increased in patients with ARDS but it is not clear whether this is due to increased secretion of ET-1 in the lungs or decreased clearance [142, 144] . In patients who succumbed to ARDS, there was also a marked increase in tissue ET-1 immunostaining in vascular endothelium, alveolar macrophages, smooth muscle, and airway epithelium compared with lungs of patients who died without ARDS. Interestingly, these same patients also had a decrease in immunostaining for both endothelial nitric oxide synthase and inducible nitric oxide synthase in the lung [148] . ARDS is also characterized by the presence of inflammatory cells in the lung. Since ET-1 may act as an immune modulator, an increase in ET-1 may contribute to lung injury by inducing expression of cytokines including tumor necrosis factor and IL-6 and IL-8 [149] . These cytokines may in turn stimulate the production of many inflammatory mediators, leading to lung injury. ET-1 additionally activated neutrophils, and increased neutrophil migration and trapping in the lung [65] [66] [67] [68] [69] . Another hallmark of ARDS is disruption and dysfunction of the pulmonary vascular endothelium leading to accumulation of lung water. The role of endothelin in formation of pulmonary edema is uncertain. Infusion of ET-1 raises pulmonary vascular pressure, but it is uncertain whether ET-1 by itself increased pulmonary protein or fluid transport in the lung [150] [151] [152] . ET-1 may rather be acting synergistically with other mediators to lead to pulmonary edema [153, 154] . Pulmonary fibrosis is the final outcome for a variety of injurious processes involving the lung parenchyma. The final common pathway in response to injury to the alveolar wall involves recruitment of inflammatory cells, release of inflammatory mediators, and resolution. The reparative phase occasionally becomes disordered, resulting in progressive fibrosis. ET-1 in the lung may be important in the initial events in lung injury by activating neutrophils to aggregate and release elastase and oxygen radicals, increasing neutrophil adherence, activating mast cells, and inducing cytokine production from monocytes [65] [66] [67] [68] [69] 149, 155] . Among the many cytokines induced by ET-1 that are important in mediating pulmonary fibrosis are transforming growth factor-β and tumor necrosis factor α [156, 157] . ET-1 is also profibrotic by stimulating fibroblast replication, migration, contraction, and collagen synthesis and secretion while decreasing collagen degradation [158] [159] [160] [161] [162] . ET-1 additionally enhances the conversion of fibroblasts into contractile myelofibroblasts [43, 163] . ET-1 also increases fibronectin production by bronchial epithelial cells [164] . Finally, ET-1 has mitogenic effects on vascular and airway smooth muscle [126, 28] . ET-1 may thus play an important role in the initial injury and eventual fibrotic reparative process of many inflammatory events in the lung. Several lines of evidence regarding the importance of ET-1 in pulmonary fibrosis are available. Plasma and BAL ET-1 levels are increased in idiopathic pulmonary fibrosis [50, 165] . Lung biopsies from patients with idiopathic pulmonary fibrosis have additionally increased ET-1 immunostaining in airway epithelial cells and type II pneumocytes, which correlates with disease activity [166] . Scleroderma is commonly associated with pulmonary hypertension and pulmonary fibrosis. Plasma and BAL ET-1 is increased in these patients [160, 167, 168] , but it is unclear whether the presence of either pulmonary hypertension or pulmonary fibrosis increases these levels further [167] . BAL fluid from patients with scleroderma increased proliferation of cultured lung fibroblasts, which was inhibited by ETA receptor antagonist. This suggests that the ET-1 in the airspace may be contributing significantly to the fibrotic response [160] . An increase in ET-1 binding has also been reported in lung tissue from patients with scleroderma associated pulmonary fibrosis [169] . Pulmonary inflammatory cells also appear to be primed for ET-1 production because cultured alveolar macrophages from patients with scleroderma and lung involvement secrete increased amounts of ET-1 in response to stimulation with lipopolysaccharide [170] . These observations collectively suggest that augmented ET-1 release may contribute to and perpetuate the inflammatory process. Bleomycin-induced pulmonary fibrosis in animals is associated with increased ET-1 expression in alveolar macrophages and epithelium [171] . The increase in ET-1 proceeds the development of pulmonary fibrosis. The use of ET-1 receptor antagonists has produced mixed results in limiting the development of bleomycin-induced fibrosis. A decrease in fibroblast replication and secretion of extracellular matrix proteins in vitro but not a decrease in lung collagen content in vivo has been shown using ETA or combined ETA and ETB receptor antagonists after bleomycin [172] . Another group did, however, observe a decrease in fibrotic area in lungs of rats following bleomycin that were treated with a mixed ETA and ETB receptor antagonist [173] . While ET-1 seems to correlate with pulmonary fibrosis, it remains uncertain whether the increase in ET-1 is a cause or consequence of the lung disease. Pulmonary fibrosis was recently reported in mice that constitutively overexpress human ET-1 [107] . These mice were known to develop progressive nephrosclerosis in the absence of systemic hypertension [174] . The transgene was localized throughout the lung, with the strongest expression in the bronchial wall. In the lung, the mice developed age-dependent accumulation of collagen and accumulation of CD4+ lymphocytes in the perivascular space. This observation suggests that an increase in lung ET-1 alone may play a causative role in the development of pulmonary fibrosis [107, 175] . Since its discovery 12 years ago, much evidence has accumulated regarding the biologic activity and potential role of ET-1 in a variety of diseases of the respiratory track. As compelling as much of this evidence is, the causal relationship between ET-1 activity and disease is not complete. The increasing use of ECE and endothelin receptor antagonists in experimental and human respiratory disorders will help to clarify the role of this pluripotent peptide in health and disease.
4
Gene expression in epithelial cells in response to pneumovirus infection
Respiratory syncytial virus (RSV) and pneumonia virus of mice (PVM) are viruses of the family Paramyxoviridae, subfamily pneumovirus, which cause clinically important respiratory infections in humans and rodents, respectively. The respiratory epithelial target cells respond to viral infection with specific alterations in gene expression, including production of chemoattractant cytokines, adhesion molecules, elements that are related to the apoptosis response, and others that remain incompletely understood. Here we review our current understanding of these mucosal responses and discuss several genomic approaches, including differential display reverse transcription-polymerase chain reaction (PCR) and gene array strategies, that will permit us to unravel the nature of these responses in a more complete and systematic manner.
RSV and PVM are viruses of the family Paramyxoviridae, subfamily pneumovirus; they are enveloped, singlestranded, nonsegmented RNA viruses that can cause intense viral bronchiolitis in humans and mice, respectively. In its most severe form, the lower respiratory tract infection caused by pneumoviruses is associated with the development of peribronchiolar infiltrates that are accompanied by submucosal edema and bronchorrhea, and ultimately leads to bronchiolar obstruction and compromised oxygen transfer. As the infection is confined to the respiratory epithelium, the responses of these cells are clearly of primary importance in determining the nature and extent of the resulting inflammatory process. Most of our understanding of responses to pneumovirus infection has emerged from studies of RSV infection of human epithelial target cells in vitro; a list of genes and/or gene products produced by epithelial cells in response to RSV infection in vitro is provided in Table 1 . At the cellular level, epithelial cells initially respond to RSV infection by reducing their ciliary beat frequency. Production and release of chemoattractant cytokines (chemokines) can be observed as early as 12 h after infection, leading to the recruitment of specific leukocyte subsets to the lung tissue. RSV-infected epithelial cells become resistant to tumor necrosis factor (TNF)-α-induced apoptosis, but later fuse to form giant-cell syncytia and die by cellular necrosis. We review the molecular bases (to the extent that they are understood) of these specific responses, and discuss several novel strategies that may permit us to study the responses to RSV and PVM infection in a more coherent and systematic manner. Tristram et al [1] observed that explanted respiratory epithelial cells slow their ciliary beat frequency almost immediately after exposure to RSV, with complete ciliostasis seen as early as 6 h after the initial infection. The molecular basis of ciliostasis remains completely unknown. The chemokines and cytokines with production and release that has been associated with RSV infection of human epithelial cells are listed in Table 1 . Much of this work was also recently reviewed elsewhere [2, 3] . We focus here on the three chemokines whose molecular mechanisms and physiologic implications are best understood. The earliest reports on this subject described production of the neutrophil chemoattractant IL-8 from tissue culture supernatants from RSV-infected cells [4] [5] [6] and in nasal secretions from patients with viral rhinitis [7] . IL-8 has since been detected in lower airway secretions from patients with severe RSV bronchiolitis [8] , and the neutrophil influx observed in response to this infection is probably due, at least in part, to the activity of this chemokine. At the cellular level IL-8 production can be observed in response to inactivated RSV virions, whereas IL-8 production in response to active infection was inhibited by ribavarin, amiloride, and antioxidants [9, 10] . Several groups have demonstrated activation of the transcription factor nuclear factor-κB (NF-κB) in response to RSV infection, and NF-κB is recognized for its central role in eliciting the production of IL-8 [9, 11, 12] . The transcription factor NF-IL-6 is also produced in response to RSV infection [13] , and participates in a co-operative manner with NF-κB in the regulation of IL-8 gene expression [11] , although later studies suggest that activator protein-1 may function preferentially in this role [14] . Interestingly, the NF-κB regulator IκBα, which functions by inhibiting NF-κB activation in response to TNF-α, was produced with different kinetics and does not promote a reversal of NF-κB activation in response to RSV infection as it does in response to TNF-α [15] . Most recently, Casola et al [16] demonstrated that the IL-8 promoter contains independent response elements, with nucleotides -162 to -132 representing a unique RSV response element that is distinct from elements necessary for IL-8 production in response to TNF-α. This concept of a stimulus-specific response will probably make an important contribution toward our understanding of how pneumoviruses promote transcription of unique and specific sets of independent gene products. The pleiotropic chemokine regulated upon activation, normal T-cell expressed and secreted (RANTES) has also been detected in supernatants from RSV-infected epithelial cells in culture [17, 18] , as well as in upper and lower airway secretions from patients infected with this virus [7, 8] . RANTES acts as a chemoattractant for eosinophils and monocytes in vitro, although its role in vivo is somewhat less clear. Similar to IL-8, RANTES can be produced in vitro in response to inactivated virions [8] , and involves NF-κB activation, binding, and nuclear translocation [19] . However, Koga et al [20] demonstrated that stabilization of RANTES mRNA, a response to RSV infection mediated in part by nucleotides 11-389 of the RANTES gene, is probably the primary mechanism underlying increased production and secretion of RANTES protein. Further studies will determine whether a similar mechanism is also in place for IL-8 and other RSV-mediated responses. Several groups have recently shown that macrophage inflammatory protein (MIP)-1α is released from RSVinfected cells in culture [7, 21] ; MIP-1α was also detected in upper and lower airway secretions from RSV-infected patients [7, 8] . Interestingly, of the three aforementioned chemokines, MIP-1α is the one that is most closely correlated with the presence of eosinophil degranulation products; this, together with data from our PVM model of pneumovirus infection [22] , has suggested to us that MIP-1α plays a pivotal role in eosinophil recruitment in response to primary pneumovirus infection. Interestingly, production of MIP-1α in cell culture requires active viral replication [8] , which suggests that this response may proceed by a mechanism that is completely distinct from that which elicits production of RANTES and IL-8. However, no reports to date have addressed the molecular mechanism that underlies the RSV-mediated MIP-1α response. A list of cell-surface molecules that have been reported as expressed in response to RSV infection is shown in Table 1 . We focus here on the expression of intercellular adhesion molecule (ICAM)-1 (CD54) and the leukocyte integrin CD18. Increased expression of this cell-surface adhesion protein was observed in both respiratory epithelial cell lines [23, 24] and in human nasal epithelial cells [25] in response to infection with RSV in vitro. Chini et al [26] demonstrated that the expression of ICAM-1 mRNA, similar to IL-8 and RANTES, was dependent on an intact NF-κB site in the gene promoter, and demonstrated a role for the consensus binding site for the factor CCAAT/ enhancer-binding protein. Stark et al [27] demonstrated that ICAM-1 and CD18 expressed in response to RSV serve to enhance neutrophil and eosinophil binding to epithelial cells. CD18 is a polypeptide of the integrin family that functions in mediating cell-cell interactions. Several groups have observed expression of CD18 on epithelial cells in response to RSV infection [27, 28] , with CD18 shown to enhance the degranulation of eosinophils in this specific setting [28] . Of particular interest are the recent findings relating expression of CD18 (along with CD14) to earlier literature on bacterial superinfections in the setting of viral infections. Earlier studies [29, 30] reported enhanced binding of bacteria to respiratory epithelial cells that were infected with RSV, findings that had clinical implications relating to acute bacterial otitis media in infants. Two more recent studies addressed the question of binding sites. Saadi et al [31] determined that two strains of the pathogen Bordetella pertussis bound more efficiently to RSV-infected cells, and that the binding was reduced upon pretreatment of the cells with anti-CD14 or anti-CD18 antibodies. Similarly, Raza et al [32] reported that both CD14 and CD18 on RSV-infected epithelial cells contributed to the binding of nonpilate Neisseria meningitidis. In vivo testing is required before the clinical significance of these intriguing findings can be appreciated. RSV-infected epithelial cells in culture do not show features that are suggestive of apoptosis (ie no evidence of membrane blebbing, fragmentation of chromosomal DNA, or characteristic changes in nuclear morphology). Takeuchi et al [33] showed that, although RSV-infected epithelial cells express a number of apoptosis-associated genes, including interferon regulatory factor-1, IL-1β-converting enzyme and caspase 3, they do not undergo formal apoptosis. As part of our attempts to understand mucosal responses in a more systematic manner (see below), we discovered that RSV-infected epithelial cells express the recently described antiapoptosis gene IEX-1L [34] . In our studies, we found that expression of IEX-1L is a response to active virus; no gene expression was observed in response to irradiated, replication-incompetent virus. Moreover, expression of IEX-1L is not observed in response to adenoviral infection, suggesting that expression of this gene is not a universal response to cellular perturbation, or indeed to all viral infections. Functionally, we also demonstrated that RSV infection protects epithelial cells from TNF-α-induced apoptosis, an effect that is temporally associated with the expression of IEX-1L. Apoptosis is generally considered to be a highly efficient self-defense mechanism employed by host target cells, because it permits the infected host to dispose of viral proteins and nucleic acids on a single-cell basis without inducing an inflammatory response. It is thus not surprising that many viruses have evolved strategies to circumvent this response. Of interest, Krilov et al [35] demonstrated that monocytes and cord blood mononuclear cells are similarly protected from apoptosis when infected with RSV. Although virus-induced protection from apoptosis appears advantageous to the virus alone, another interpretation may be considered. Because respiratory epithelial cells are now recognized as a major source of leukocyte chemoattractants, and because leukocyte recruitment to the lung has been associated with enhanced viral clearance and prolonged survival in pneumovirus infection [22] , the ability to maintain chemoattractant production from viable cells may ultimately benefit the host organism as well. Available online http://respiratory-research.com/content/2/4/225 In tissue culture, RSV-infection is characterized by the formation of giant-cell syncytia. The mechanisms for the formation of these fused masses of cells depend in part on the expression of the RSV-specific fusion (F) protein on the surface of infected host cells, and in part on virusmediated changes in cytoskeletal architecture. It is important to note that RSV-induced changes in cytoskeletal architecture are not restricted to cell lines grown in vitro, as giant-cell syncytia have also been found in pathologic lung specimens obtained from both humans and animals that were infected with RSV. Again, as part of our systematic study of gene expression in response to pneumovirus infection, we found that human respiratory epithelial cells respond to RSV infection with increased expression of the cytoskeletal protein cytokeratin-17 [36] . Cytokeratin-17 is a 46-kDa cytoskeletal protein that belongs to the class I acidic cytokeratin family. In the lung, expression of cytokeratin-17 is normally restricted to basal epithelial cells of the larynx, trachea, and bronchi. In RSV-infected cells, we found expression of Ck-17 predominantly at sites of syncytia formation, and thus provided the first description of a unique component of these pathognomonic structures at the molecular level. Similar to what has been reported for the production of IL-8, expression of Ck-17 is dependent on the activity of the transcription factor NF-κB, and future studies will determine the role of the NF-κB consensus site (-200 to -208 of the cytokeratin-17 promoter) in mediating this response. To date, efforts to study pneumovirus-induced alterations in gene expression have relied heavily on in vitro models of virus-infected cells and cell lines. The intrinsic value of characterizing the genes identified in this artificial system is by definition limited, and the clinical and physiologic sig-nificance of any findings must ultimately be tested in vivo. To some extent, the study of gene products in clinical specimens is possible, but this approach is limited, cumbersome, and dictated by sample availability. It is clear that an appropriate animal model of inflammatory bronchiolitis is required to characterize the alterations in gene expression discovered using the available in vitro models. Although RSV has been used for the study of specific allergic reactions to viral antigens, it is not a natural pathogen of mice, and intranasal inoculation of virus at high titer results in, at best, a minimal primary infection with a correspondingly minimal inflammatory response. In order to study gene expression in response to primary pneumovirus infection in vivo, we developed a novel mouse model of inflammatory bronchiolitis using the natural rodent pneumovirus pathogen and closest phylogenetic relative of RSV [37] -PVM. We presented our initial findings on PVM infection in mice in three recent publications [22, 38, 39] . A summary of these findings is presented in Table 2 and Fig. 1 . To begin, we described the cellular and biochemical pathology observed in response to PVM infection in mice [38] . We found that infection could be established with as few as 30 plaque-forming units (pfu) of PVM in the inoculum, with infection resulting in significant morbidity and mortality, and viral recoveries in the order of 10 8 pfu/g lung tissue. We also noted inflammatory bronchiolitis as among the immediate responses to this infection, with bronchoalveolar lavage fluid containing virtually 100% neutrophils and eosinophils obtained as early as 3 days after inoculation. Furthermore, we found that infection was accompanied by the production of the proinflammatory chemokine MIP-1α, which was previously shown by Cook et al [40] to be an important component of the inflammatory response to the orthomyxovirus influenza virus. We also described the role played by MIP-1α in the pathogenesis of PVM-induced bronchiolitis [22] . Specifi- cally, we explored the responses of gene-deleted mice to infection with PVM, and found no inflammatory response in mice deficient in MIP-1α expression (MIP-1α -/-) and only minimal virus-induced inflammation in mice that lacked the major MIP-1α receptor on granulocytes chemokine receptor (CCR)1 (CCR1 -/-). Although the inflammatory response is often considered to be unnecessary and indeed detrimental, we demonstrated that the absence of granulocytic inflammation was associated with enhanced recovery of infectious virions, as well as with accelerated mortality. These results suggest that the MIP-1α/CCR1-mediated acute inflammatory response protects mice by delaying the lethal sequelae of viral infection. Our most recent report on this subject [39] presents a direct comparison between the responses of mice to challenge with PVM and RSV. Although RSV is not a natural pathogen of mice, it has been used extensively in mouse models of human infection because a limited, or 'semipermissive' infection can be established via intranasal inocula-tion of virus at very high titers. In this regard, we found (as have others) that RSV infection did not result in any measurable degree of morbidity, and that viral recovery was significantly lower than that found in the inoculum; these results suggested that there was no significant viral replication in mouse lung tissue. We further demonstrated that the inflammatory response to RSV challenge was minimal, as few leukocytes were recruited to the lungs (Fig. 1) . Taken together, our results suggest that infection of mice with PVM provides a superior model for the study of acute inflammatory bronchiolitis in response to pneumovirus infection in vivo. The advantages of this model include the following: clinical parameters -morbidity and mortalitythat can be measured clearly and specifically; clear evidence of viral replication in lung tissue, with incremental recoveries that, at peak, are in excess of 10 8 pfu/g in response to as few as 30 pfu in the inoculum; and a dramatic granulocytic response that is modulated at least in part by the proinflammatory chemokine MIP-1α and its receptor CCR1. Traditionally, analysis of gene expression through measurement of steady-state levels of individual mRNAs could be conducted only one gene at a time using northern blotting, dot blots, or quantitative reverse transcription-PCR. Differential display, serial analysis of gene expression, and total gene analysis offer great promise, because they are multiplex technologies that provide simultaneous analysis of multiple mRNAs isolated under conditions of interest via PCR amplification techniques. DNA hybridization arrays are theoretically the most efficient of the gene expression analysis techniques. Although many skeptics have described these genome-based approaches as expensive, nonhypothesis-driven 'fishing expeditions', we view them as broad-based screening techniques that will enable us to identify patterns of gene expression that can then be subjected to careful characterization and analysis. Differential display is a semiquantitative, reverse transcription-PCR-based technique that is used to compare mRNAs from two or more conditions of interest. Both increased and decreased expression of specific amplicons will be evident -an obvious advantage to this approach. Total RNA can be isolated from virus infected versus uninfected cells or mouse lungs both before and during infection, and differential display is performed using degenerate T11(XY) anchoring primers and random upstream oligomers, as described elsewhere [34, 36] . The resulting PCR products are separated by electrophoresis, and the gel is dried and exposed to film. An example of our results comparing cDNA amplicons from RNA extracted from RSV-infected epithelial cells daily for 4 days is shown in Fig. 2 . Differentially expressed bands are cut from the gel, eluted and reamplified using the same primers that generated the original signal, and northern blots generated from RNA extracted from pneumovirus-infected cells or tissue over time and probed with the differentially expressed amplicons serve to confirm differential expression of the identified sequence. The DNA sequences of the newly identified differentially expressed amplicons are compared with sequences present in the GenBank database. Viral sequences are expected to be upregulated over time and can be identified immediately, because the entire genomes of both PVM and RSV are present in GenBank. In cases in which the amplicon represents a newly discovered gene, potential openreading frames are compared with sequences that are present in the Swiss protein database; motifs that share homologies with known proteins represent important clues to the identity of the differentially expressed gene. With the help of differential display, we have identified and characterized several genes that are upregulated in RSV-infected respiratory epithelial cells. Two specific examples of genes that were found to be induced during RSV infection, and later characterized as playing independent roles in the pathophysiology of RSV infection, are the antiapoptosis gene IEX-1L [34] and the gene that encodes the cytoskeletal protein cytokeratin-17 [36] . Unlike DNA viruses, which are known to encode virus-specific antiapoptosis genes, RSV -an RNA virus with a small (approximately 15.2 kb) viral genome -was shown to alter host cell expression of the apoptosis inhibitor IEX-1L. After demonstrating that IEX-1L mRNA was present at sevenfold higher concentrations in RSV-infected respiratory epithelial cells when compared with uninfected cells, we concluded that this cellular response protected against TNF-α-induced programmed cell death during viral infection. Further efforts to determine which of the 11 RSV proteins participate in the trans-activation of the IEX-1L gene (either directly or indirectly) are ongoing. A second example of a gene that is specifically upregulated in RSV-infected respiratory epithelium, as identified by differential display, is that which encodes cytokeratin-17 [36] . Upon characterizing the molecular events that are important for cytokeratin-17 induction, we demonstrated a link to an NF-κB signaling pathway. Above, we discussed the importance of this transcription factor in the regulation of proinflammatory cytokine gene expression, and because of this involvement we were not surprised to discover its role in virus-induced cytokeratin-17 gene regulation. Perhaps the most interesting observation made during these experiments was the in situ localization of cytokeratin-17 protein to areas of cytopathic syncytia formation, suggesting a role for this cytoskeletal protein in their formation. Of note, we observed a dramatic decrease in RSV replication and in syncytia formation when we blocked cytokeratin-17 expression, suggesting that blocking syncytia formation, at least in part, impairs the direct cell-cell spread and productive replication of virus. Although there are several companies that market these systems and components, the cytokine gene macroarray systems recently developed by R&D Systems (Sigma Genosys ® ; Minneapolis, MN, USA) and Clontech (Atlas ® ; Palo Alto, CA, USA) represent some of the newer opportunities available that have a focus on gene products that are known to be involved in inflammation. These arrays consist of different cDNAs printed as PCR products onto charged nylon membranes. An example of our experience with the Sigma Genosys array is shown in Fig. 3 . For this example, total RNA was extracted from RSV-infected HEp-2 cells and uninfected controls at day 3 after infection. Three micrograms of total RNA was used in a cDNA synthesis reaction, using a proprietary mixture of 378 primer pairs and trace amounts of 32 P-radiolabeled dCTP. The resulting radiolabeled products were hybridized to the macroarrays overnight at 65°C, and then washed and exposed to film. The arrow highlights one of the most obviously upregulated sequences from this experiment, which was identified as the gene encoding human MIP-1α. The physiologic importance of MIP-1α upregulation during human RSV infection and during rodent PVM bronchiolitis has already been described. Microarrays can be differentiated from macroarrays in several ways. Among these differences, the microarray matrix is a glass or plastic slide, probes are labeled with fluorescent dye rather than via radioisotopes, and, most significantly, microarrays generally include a larger number and a higher density of imbedded sequences than do macroarrays. Although this may seem to be highly appealing at first, the massive amounts of data generated by microarray technology poses new challenges with respect to data normalization, management, and development of mathematical models to assist in data interpretation. The pneumoviruses RSV and PVM enter respiratory epithelial cells via a receptor-mediated event. During hostcell attachment and internalization, the target cell begins to alter its gene expression, which, among other events, involves the transcriptional upregulation of cytokine and chemokine genes. As RSV replication progresses, additional changes in cellular gene expression can be observed, including induction of the potent antiapoptosis gene IEX-1L and increased expression of the otherwise quiescent gene that encodes cytokeratin-17. What we know regarding the physiologic importance of these genes and their gene products has been described, but there is more to be learned. As the available technologies evolve, we can continue to capitalize on the use of Display of amplicons generated from RNA extracted from RSV-infected cells at daily intervals following infection (days 0-4) using a single anchoring primer, T11GC (downstream primer 8) and (A-H) eight random 10mers. Two differentially expressed sequences are highlighted by arrows (the black arrow shows an upregulated amplicon, and the white arrow highlights a downregulated amplicon). Several other differentially expressed signals are also seen. genomic approaches as large-scale screening tools to identify genes that play important roles in the pathophysiology of pneumovirus infection. These elegant and simple tools will provide us with the means for thorough and systematic exploration of gene expression, both in the estab- Cytokine macroarray probed with radiolabelled cDNA generated from total RNA extracted from epithelial cells 48 h after RSV infection (upper panel) or 48 h after exposure to conditioned medium (lower panel). Signal intensity of each gene under each condition is compared. The arrow highlights the signal for human MIP-1α present at 12-fold higher concentration in infected epithelial cells compared with the uninfected controls.
5
Sequence requirements for RNA strand transfer during nidovirus discontinuous subgenomic RNA synthesis
Nidovirus subgenomic mRNAs contain a leader sequence derived from the 5′ end of the genome fused to different sequences (‘bodies’) derived from the 3′ end. Their generation involves a unique mechanism of discontinuous subgenomic RNA synthesis that resembles copy-choice RNA recombination. During this process, the nascent RNA strand is transferred from one site in the template to another, during either plus or minus strand synthesis, to yield subgenomic RNA molecules. Central to this process are transcription-regulating sequences (TRSs), which are present at both template sites and ensure the fidelity of strand transfer. Here we present results of a comprehensive co-variation mutagenesis study of equine arteritis virus TRSs, demonstrating that discontinuous RNA synthesis depends not only on base pairing between sense leader TRS and antisense body TRS, but also on the primary sequence of the body TRS. While the leader TRS merely plays a targeting role for strand transfer, the body TRS fulfils multiple functions. The sequences of mRNA leader–body junctions of TRS mutants strongly suggested that the discontinuous step occurs during minus strand synthesis.
The genetic information of RNA viruses is organized very ef®ciently. Practically every nucleotide of their genome is utilized, either as protein-coding sequence or as cis-acting signals for translation, RNA synthesis or RNA encapsidation. As part of their genome expression strategy, several groups of positive-strand RNA (+RNA) viruses produce subgenomic (sg) mRNAs (reviewed by Miller and Koev, 2000) . The replication of their genomic RNA, which is also the mRNA for the viral replicase, is supplemented with the generation of sg transcripts to express structural and auxiliary proteins, which are encoded downstream of the replicase gene in the genome. Sg mRNAs of +RNA viruses are always 3¢-co-terminal with the genomic RNA, but different mechanisms are used for their synthesis. Some viruses, such as brome mosaic virus, initiate sg mRNA synthesis internally on the full-length minus strand RNA template (Miller et al., 1985) . Others, exempli®ed by red clover necrotic mosaic virus (RCNMV), may rely on premature termination of minus strand synthesis from the genomic RNA template, followed by the synthesis of sg plus strands from the truncated minus strand template (Sit et al., 1998) . Members of the order Nidovirales, which includes coronaviruses and arteriviruses, have evolved a third and unique mechanism, which employs discontinuous RNA synthesis for the generation of an extensive set of sg RNAs (reviewed by Brian and Spaan, 1997; Lai and Cavanagh, 1997; Snijder and Meulenberg, 1998) . Nidovirus sg mRNAs differ fundamentally from other viral sg RNAs in that they are not only 3¢-coterminal, but also 5¢-co-terminal with the genome ( Figure 1A) . A 5¢ common leader sequence of 65±221 nucleotides, derived from the 5¢ end of the genomic RNA, is attached to the 3¢ part of each sg RNA (thè mRNA body'). Various models have been put forward to explain the cotranscriptional fusion of non-contiguous parts of the nidovirus genome during sg RNA synthesis ( Figure 1B and C). Central to each of these models are short transcription-regulating sequences (TRSs), which are present both at the 3¢ end of the leader and at the 5¢ end of the sg RNA body regions in the genomic RNA. The TRS is copied into the mRNA and connects its leader and body part (Spaan et al., 1983; Lai et al., 1984) . Synthesis of sg mRNAs initially was proposed to be primed by free leader transcripts, which would base-pair to the complementary TRS regions in the full-length minus strand, and would be extended subsequently to make sg plus strands ( Figure 1B ; Baric et al., 1983 Baric et al., , 1985 . This model, however, was based on the report that sg minus strands were not present in coronavirus-infected cells (Lai et al., 1982) . The subsequent discovery of such molecules (Sethna et al., 1989) resulted in reconsideration of the initial`leader-primed transcription' model. Sawicki and Sawicki (1995) have proposed an alternative model ( Figure 1C ), in which the discontinuous step occurs during minus instead of plus strand RNA synthesis. In this model, minus strand synthesis would be attenuated after copying a body TRS from the plus strand template. Next, the nascent minus strand, with the TRS complement at its 3¢ end, would be transferred to the leader TRS and attach by means of TRS±TRS base pairing. RNA synthesis would be reinitiated to complete the sg minus strand by adding the complement of the genomic leader sequence. Subsequently, the sg minus strand would be used as template for sg mRNA synthesis, and the presence of the leader complement at its 3¢ end might allow the use of the same RNA signals that direct genome synthesis from the fulllength minus strand. Sequence requirements for RNA strand transfer during nidovirus discontinuous subgenomic RNA synthesis The EMBO Journal Vol. 20 No. 24 pp. 7220±7228, 2001 Using site-directed mutagenesis of TRSs of the arterivirus equine arteritis virus (EAV), we have shown previously that base pairing between the sense leader TRS and antisense body TRSs is crucial for sg mRNA synthesis (van Marle et al., 1999a) . However, base pairing is only one step of the nascent strand transfer process and is essential in both models outlined in Figure 1 . The EAV genomic RNA contains several sequences that match the leader TRS precisely, but nevertheless are not used for sg RNA synthesis (den Boon et al., 1996; Pasternak et al., 2000) . This suggests that leader±body TRS similarity alone is, though necessary, not suf®cient for the strand transfer to occur. To gain further insight into the cis-acting signals regulating sg RNA synthesis, we performed a comprehensive site-directed mutagenesis study of the EAV leader and body TRSs. Every nucleotide of the TRS (5¢-UCAACU-3¢) was substituted with each of the three alternative nucleotides. Our analysis revealed a number of striking similarities with the process of copy-choice RNA recombination, as it occurs in RNA viruses. Whereas the leader TRS plays only a targeting role in translocation of the nascent strand, body TRS nucleotides appear to ful®l diverse position-speci®c and base-speci®c functions. In addition, the sequence of the leader±body junctions of the sg mRNAs produced by these mutants provided strong evidence for the discontinuous minus strand extension model. EAV genome replication is not signi®cantly affected by leader TRS and body TRS mutations To dissect EAV RNA synthesis, we routinely use a fulllength cDNA clone (van Dinten et al., 1997) , from which infectious EAV RNA is in vitro transcribed. Following transfection of the RNA into baby hamster kidney (BHK-21) cells, intracellular RNA is isolated and analysed by northern blot hybridization and RT±PCR (van Marle et al., 1999a) . Due to differences in transfection ef®ciency, the total amount of virus-speci®c RNA (genomic RNA and sg mRNA) isolated from transfected cell cultures is somewhat variable. Thus, the accurate quantitation of sg mRNA synthesis by TRS mutants requires an internal standard for transfection ef®ciency. The amount of viral genomic RNA can be this standard, but only if its ampli®cation is not dramatically affected by the TRS mutations. To prove that this is the case, we used the previously described mutants L4, B4 and LB4 (van Marle et al., 1999a) , in which ®ve nucleotides of the TRS (5¢-UCAAC-3¢) were replaced by the sequence 5¢-AGUUG-3¢, either in the leader TRS (L4), RNA7 body TRS (B4) or both TRSs (LB4). The three mutants were tested in three independent experiments. Intracellular RNA was isolated at 14 h posttransfection, early enough to prevent spread of the wildtype control virus to non-transfected cells (®rst cycle analysis). Transfection ef®ciencies were determined by immuno¯uorescence assays (see Materials and methods) and varied between 10 and 23% (data not shown). Prior to RNA analysis, the amount of isolated intracellular RNA was corrected for the transfection ef®ciency of the sample, so that each lane in Figure 2 represents EAV-speci®c RNA from an approximately equal number of EAV-positive cells. Phosphoimager quantitation revealed that genomic RNA replication of mutants L4, B4 and LB4 varied by not more than 30% (Table I) . These differences could re¯ect, for example, a slight in¯uence of RNA secondary structure changes in the TRS regions on genomic RNA synthesis. Remarkably, however, the genomic RNA level of the leader±body TRS double mutant LB4 was not affected by more than 10%. In view of the results obtained with these pentanucleotide TRS mutants, we assumed that the amount of genomic RNA could indeed be used as an internal standard during the analysis of mutants containing only single nucleotide replacements in leader TRS and/or RNA7 body TRS. The regions of the genome specifying the leader (L) sequence, the replicase gene (ORFs 1a and 1b) and the structural genes are indicated. The nested set of EAV mRNAs (genome and sg mRNAs 2±7) is depicted below. The black boxes in the genomic RNA indicate the position of leader and major body TRSs. (B and C) Alternative models for nidovirus discontinuous sg RNA synthesis. The discontinuous step may occur during either plus strand (B) or minus strand (C) RNA synthesis. In the latter case, sg mRNAs would be synthesized from an sg minus strand template. For details see text. Northern analysis of EAV-speci®c RNA isolated from cells transfected with RNA transcribed either from the wild-type EAV infectious cDNA clone or from TRS pentanucleotide mutants (UCAAC to AGUUG). The results of two independent experiments are shown. The RNA±RNA interaction between the leader and body TRSs is not the only factor that regulates EAV sg RNA synthesis There are numerous examples of regulatory RNA±RNA interactions in both eukaryotic and prokaryotic cells, as well as in RNA viruses. Essential processes such as translation, replication and encapsidation of RNA virus genomes frequently depend on RNA±RNA interactions and higher order RNA structures. Regulation of sg RNA synthesis of +RNA viruses by RNA±RNA interactions is also not without precedent. In tomato bushy stunt virus, an RNA element located 1000 nucleotide upstream of the sg RNA2 promoter base-pairs with the promoter and is necessary for sg RNA production (Zhang et al., 1999) . Similarly, base pairing interactions between complementary sequences in the 5¢ end of the potato virus X genomic RNA and sequences upstream of two major sg RNA promoters are required for ef®cient sg RNA synthesis (Kim and Hemenway, 1999) . In RCNMV, an intermolecular RNA±RNA interaction is required for sg RNA synthesis (Sit et al., 1998) . Recently, we have established the pivotal role of an interaction between sense and antisense RNA sequences in the life cycle of EAV (van Marle et al., 1999a) . In that study, the role of TRS nucleotides C 2 and C 5 was tested by substituting them with G. It was concluded that base pairing between the sense leader TRS and the antisense body TRS plays a crucial role in nidovirus sg RNA synthesis. We now took a more systematic approach and performed an extensive site-directed co-variation mutagenesis study of the entire leader TRS and RNA7 body TRS, which directs the synthesis of the most abundant EAV sg RNA. Every nucleotide of the TRS (5¢-UCA-ACU-3¢) was replaced with each of the other possible nucleotides. As in the study of van Marle et al. (1999a) , every mutation was introduced into leader TRS, RNA7 body TRS and both TRSs, resulting in 54 mutant constructs. Each mutant was given a unique name: e.g. BU 1 A refers to a mutant in which a U has been changed to A at position 1 of the body TRS; LU 1 A refers to the same substitution in the leader TRS; and DU 1 A means that these two substitutions were combined in one double mutant construct. The amount of sg RNA7 was quantitated by phosphoimager scanning of hybridized gels and was corrected for the amount of genomic RNA in the same lane (as outlined above). Figure 3 shows the relative sg RNA7 level of the 54 mutants, compared with the RNA7 level of the wild-type control. For a selection of 11 interesting mutants (see below), the analysis was repeated three times (Figure 4 ), without observing signi®cant variations in sg RNA synthesis. The comprehensive analysis of the effects of TRS mutations considerably expanded our understanding of van Dinten et al., 1997) was taken along as a positive control. For every mutant, the level of sg RNA7 synthesis was calculated as [(sg/g)/(sg/g) wt ] 3 100%: it was corrected for the level of genomic RNA (used as an internal standard; see text) and subsequently was related to the level of sg RNA7 produced by the wild-type control in the same experiment, which was also corrected for the corresponding genomic RNA level. The relative sg RNA7 level of the wild-type control was set at 100%. A.O. Pasternak et al. discontinuous sg RNA synthesis. Remarkably, the effects of single (leader or body) TRS mutations were mostly base speci®c, i.e. different nucleotide substitutions at the same position affected sg RNA7 synthesis to different extents. For example, at position 1, the BU 1 A mutant retained 44% of the wild-type RNA7 synthesis level, whereas both the BU 1 C and BU 1 G mutants lost RNA7 synthesis almost completely. Conversely, when U 1 of the leader TRS was changed to A or G, RNA7 synthesis was completely abolished, whereas 13% of the wild-type level was still maintained by LU 1 C. For position 2, only the BC 2 U mutant retained 30% of the wild-type RNA7 synthesis level, while all the other position 2 single mutants have lost 90% or more of wild-type RNA7 synthesis. Another example is position 6: BU 6 C left only 5% of wild-type RNA7 synthesis, whereas BU 6 A produced much higher RNA7 levels. This implied that for some positions (1, 2 and 6), certain mismatches in the duplex between plus leader TRS and minus body TRS, such as U±U (BU 1 A and BU 6 A) or C±A (LU 1 C and BC 2 U), are allowed to a limited extent. In contrast, no mismatches were allowed for position 5, where all single nucleotide substitutions abolished RNA7 synthesis almost completely. Surprisingly, both body TRS U to C substitutions at positions 1 and 6 (BU 1 C and BU 6 C) resulted in low levels of RNA7, despite the fact that these mutations allow the formation of a G±U base pair between the plus leader TRS, providing the U nucleotide, and the minus body TRS, providing the G. On the other hand, for positions 3 and 4, G±U base pairing was shown to be functional, because mutants LA 3 G and LA 4 G, which can form G±U base pairs between the G in the plus leader TRS and U in the minus body TRS, were the only position 3 and 4 single mutants that produced reasonable levels of RNA7. Taken together, these ®ndings suggest that other factors, besides leader± body base pairing, also play a role in sg RNA synthesis and that the primary sequence (or secondary structure) of TRSs may dictate strong base preferences at certain positions. Our analysis of the degree of complementation by the double mutants provided strong support for this assumption. Differentiating between effects at the level of primary TRS sequence and the level of leader±body duplex formation For some TRS nucleotides (2, 5 and 6, except in the case of DU 6 C), the RNA7 level of double mutants was clearly higher than that of the corresponding single mutants. This means that base pairing between these leader and body TRS nucleotides is involved in sg RNA synthesis. However, none of these double mutants reached the wild-type sg RNA7 level. In the other double mutants (all position 1, 3 and 4 mutants, and DU 6 C), in clear contradiction to the predictions of the`base pairing model', RNA7 synthesis was not signi®cantly restored. Moreover, a comparison of the values for the B and D mutants in Figure 3 showed that, for almost all of these mutants (e.g. the position 1 mutants), the amount of sg RNA7 produced by the double mutant appeared to be limited by the level allowed by the body TRS mutation. Sometimes the RNA7 level of the double mutant was even less than that of the leader mutant (DU 1 C, DA 3 G, DA 4 G or DU 6 C). Clearly, for these substitutions, restoration of the possibilities for leader±body duplex formation did not restore sg RNA synthesis. Apparently this is because the effect of body TRS mutations at the level of primary sequence or secondary structure can be`dominant' over the duplex-restoring effects of the double mutations. Body TRS mutants thus fell into two distinct types, determined by the position and chemistry of the substitution. In mutants of the ®rst type, sg RNA synthesis was impaired mainly because of the disruption of the leader± body TRS duplex. This effect could be compensated for by introduction of the corresponding mutation in the leader TRS and, in the double mutant, sg RNA synthesis was restored compared with the corresponding single mutants. In mutants of the second type, sg RNA synthesis was down-regulated as a consequence of both TRS duplex disruption and disruption of the primary sequence (or secondary structure) of the body TRS. Obviously, the latter effect could not be compensated for by mutating the leader TRS, and the corresponding double mutants did not show restoration of sg RNA synthesis. In contrast to our ®ndings with the body TRS mutants, we did not obtain leader TRS mutations that appeared to determine the level of sg RNA7 synthesis of the corresponding double mutant (Figure 3) . Thus, effects of mutations in the leader TRS were not`dominant' over the duplex-restoring effects of the double mutations, suggesting that they only affected duplex formation. This indicated that the leader TRS probably does not have an additional, sequence-speci®c function in sg RNA synthesis in addition to its participation in TRS±TRS base pairing. The fact that single leader TRS mutations at all six Nidovirus discontinuous subgenomic RNA synthesis positions severely repressed RNA7 synthesis indicated that base pairing of every TRS nucleotide contributes to sg RNA production. In this respect, it was signi®cant that the two leader TRS mutants with the highest RNA7 levels, LA 3 G and LA 4 G, can form G±U base pairs to maintain the duplex. The observation that leader TRS mutations could bè rescued' by introducing complementary mutations in the body TRS, but that many body TRS mutations could not bè rescued' by corresponding changes in the leader TRS, is clearly illustrated by the U 1 A mutants. Due to the restoration of TRS base pairing possibilities, the RNA7 synthesis of double mutant DU 1 A was signi®cantly increased compared with that of LU 1 A, but not above the level of BU 1 A. Thus, restoration of the leader±body duplex in DU 1 A exerted a clear effect on sg RNA7 production compared with LU 1 A, but had no effect on sg RNA synthesis compared with BU 1 A. This exempli®ed the dominant nature of a mutation in the primary sequence of a body TRS. In contrast, for instance, the BC 2 U mutation probably affected duplex formation only, because RNA7 synthesis was restored almost to wildtype levels in the DC 2 U double mutant. These results indicate that there are strong base preference constraints for some body TRS positions. To interpret these base preferences accurately, it is necessary to limit the analysis to the double mutants only, because in these mutants the down-regulation of sg RNA synthesis was only due to the sequence changes in the body TRS, and not to the disruption of the leader±body TRS duplex. There were strict preferences for positions 1, 3 and 4 of the body TRS: at position 1, only the U to A substitution allowed for a signi®cant RNA7 level (~40% of wild-type); and at positions 3 and 4, only the A to U mutants retained 15±20% of the wild-type level. For positions 2 and 5, the sequence constraints were less stringent (all substitutions allowed for >20% of wild-type level), but still only DC 2 A and DC 2 U reached >50%. At position 6 of the body TRS, only U to C was not allowed, whereas the other two double mutants still produced 50% or more of RNA7. In other words, the functional EAV RNA7 body TRS (based on the analysis of our single nucleotide substitutions) can be described as U 1 (C/u/a) 2 A 3 A 4 C 5 (U/a/g) 6 , with wild-type nucleotides shown in upper case and nucleotides that allowed for at least 50% of the wild-type RNA7 level shown in lower case. Remarkably, TRS nucleotides A 3 , A 4 and C 5 are conserved in the TRSs of all other arteriviruses (Snijder and Meulenberg, 1998) . Also the fact that DC 2 U retained 80% of RNA7 synthesis corresponded nicely to the presence of a U at this position in other arteriviruses. Until recently (Almazan et al., 2000; Thiel et al., 2001) , infectious cDNA clones were lacking for coronaviruses. Consequently, most studies on coronavirus sg RNA synthesis were carried out using defective interfering (DI) RNAs. These replicons carried body TRSs from which moderate levels of sg mRNAs could be produced in the presence of helper virus. Using this system, Joo and Makino (1992) and van der Most et al. (1994) performed body TRS mutagenesis studies for the murine coronavirus (MHV). Joo and Makino systematically mutagenized the core of the MHV body TRS. In contrast to our results, they found that in only two of 21 body TRS mutants was sg RNA synthesis from the DI RNA genome abolished, whereas all others supported normal levels of sg RNA production. Thus, it is possible that the MHV TRS which was used in that study is more tolerant to single-nucleotide mismatches than the EAV sg RNA7 TRS. In a similar study, van der Most et al. (1994) observed that U to C substitutions at positions 1 and 3 of the MHV body TRS, which maintained the duplex by changing a U±A base pair into a U±G base pair, reduced sg RNA levels more strongly than substitutions that disrupted the duplex (van der Most et al., 1994) . This implies that, as in the case of EAV, leader±body TRS duplex formation is not the only factor that determines coronavirus sg RNA synthesis. However, because of the limitations of the DI RNA system, the leader TRS could not be mutagenized in these studies, and body TRS-speci®c effects could not be distinguished from effects at the level of leader±body duplex formation. The discontinuous step in nidovirus sg RNA synthesis occurs during minus strand RNA synthesis Due to recent studies of arterivirus and coronavirus sg RNA synthesis (van Marle et al., 1999a; Baric and Yount, 2000; Sawicki et al., 2001) , the discontinuous minus strand extension model ( Figure 1C ) has been gaining more and more ground. This model predicts that the TRSderived sequence that forms the leader±body junction in the sg mRNA is a copy of the body TRS, and not of the leader TRS. The leader-primed transcription model predicts the opposite ( Figure 1B) . Therefore, determining the origin of the leader±body junction of sg mRNAs would help to distinguish between the two models. However, in the wild-type situation, EAV leader and body TRSs are identical and consequently one cannot determine the origin of the sg mRNA leader±body junction. This problem could be overcome by tracing the mutations introduced in leader or RNA7 body TRS mutants, most of which retained part of their ability to produce mRNA7. In a previous study (van Marle et al., 1999a) , we found that nucleotides 2 and 5 of the mRNA7 leader±body junction sequence were derived exclusively from the body TRS, and not from the leader TRS. This was shown by direct sequencing of RT±PCR products obtained from the residual mRNA7 produced by mutants BC 2 G, LC 2 G, BC 5 G and LC 5 G ( van Marle et al., 1999a) . Using the same approach, we analysed mRNA7 from mutants BC 2 A and BC 2 U, and these transcripts also contained the mutated nucleotide derived from the body TRS (data not shown). Assuming that only one crossover event occurs during leader±body joining, we could thus map this crossover between positions ±1 and +2 of the sg RNA junction sequence. This left the intriguing question of whether the crossover site could be mapped even more precisely. In other words, was nucleotide +1 of the junction sequence derived from the body TRS or the leader TRS? Using the position 1 mutants described above, we could answer this question ( Figure 5) . The most striking result was that mRNA7 of mutants BU 1 A, BU 1 G and LU 1 C contained exclusively the body TRS-derived nucleotide at position +1. Thus, for these mutants, the crossover site could be mapped precisely between TRS nucleotide positions ±1 and +1, meaning that the complete leader± body junction sequence in an EAV sg mRNA can be body TRS derived. On the other hand, sg RNAs from mutants LU 1 A, BU 1 C and LU 1 G contained mixed populations of leader TRS-and body TRS-derived nucleotides at position +1 ( Figure 5 ): A and U for LU 1 A, C and U for BU 1 C, and G and U for LU 1 G. Remarkably, this pattern correlated with the relative amounts of sg mRNA7 produced by these mutants (Figure 3 ). Mutants that produced populations of sg RNAs that were mixed with respect to the origin of the nucleotide at position +1 of the leader±body junction had lost RNA7 synthesis almost completely. On the other hand, mutants that contained exclusively the body nucleotide at position +1 retained higher levels of RNA7 synthesis. This observation may be explained as follows: in the wild-type situation, the large majority of the crossovers probably occur between positions ±1 and +1, leading to a body TRS-derived nucleotide at position +1 in the sg RNA; however, a low number of crossovers take place between nucleotides +1 and +2, resulting in a leader TRS-derived nucleotide at position +1. Mutants in which almost all sg RNA synthesis is blocked by a substitution at position +1 may somehow be de®cient in the crossover between ±1 and +1, but may have retained the ability for crossovers between +1 and +2, which were detected by sequence analysis. Conversely, in position +1 mutants that retain reasonable sg RNA synthesis, most crossovers occur between positions ±1 and +1, and they obscure the minority of crossovers between +1 and +2 in the sequencing electropherogram. Alternatively, position +1 TRS mutations that strongly interfere with sg RNA synthesis may force a shift of the crossover site in the remaining molecules. We believe that our present ®ndings strongly support the discontinuous minus strand extension model. Indeed, the fact that a complete body TRS can be copied into the sg RNA is very dif®cult to reconcile with the alternative model, in which sg RNA synthesis from the genomic minus strand template is primed by free plus strand leader transcripts that contain the leader TRS at their 3¢ end ( Figure 1B) . To explain the presence of a complete copy of the body TRS in the sg mRNA in this model, one would have to assume that a 3¢±5¢ exonuclease activity trims back the free leader transcript prior to its extension into an sg mRNA (Baker and Lai, 1990) . Note that there would not be a single base pair left to hold these`trimmed' leader molecules on the template. Such an enzymatic activity, which is unprecedented in +RNA viruses, exists in yeast retrotransposon Ty5 (Ke et al., 1999) , in which reverse transcription is primed by an internal region in a tRNA. However, in this system, it is not a part of the duplex that is removed, but the single-stranded 3¢ tail of the tRNA, which cannot base-pair with the Ty5 RNA. Removal of the TRS at the 3¢ end of the nidovirus leader, which has already base paired with the template, would be very energetically unfavourable for the RdRp. Instead of starting elongation using the intact and properly positioned leader as a primer, it would have to disrupt the newly formed duplex, degrade part of the leader RNA and then reinitiate polymerization, without any base pairing between primer and template. It has been shown that in¯uenza virus transcription does not require a sequence match between the (cellular) RNA primer and the (viral) template (Plotch et al., 1981) . However, if in the nidovirus system the`trimmed' leader RNA could also be ®xed on the template solely by RNA±protein interactions, the targeting of the nascent strand by TRS base pairing would be extremely puzzling. Sequence data of sg RNA leader±body junctions from other arteriviruses are also dif®cult to reconcile with the leader-primed transcription model. For the porcine and simian arteriviruses (Meulenberg et al., 1993; Godeny et al., 1998) , the leader±body junctions of some sg RNAs mapped two nucleotides upstream of the body TRS, which again would not leave a single nucleotide to hold the putative free leader on the template after the hypothetical back trimming'. On the other hand, these ®ndings and our data can be explained readily by the discontinuous minus strand extension model ( Figure 1C ). The six-nucleotide Fig. 5 . Sequence analysis of mRNA7 leader±body junctions from position 1 TRS mutants. Sequences were determined directly from sg mRNA7-speci®c RT±PCR products. For the U 1 A and U 1 C mutants, the sequence shown corresponds to the plus strand of sg RNA7. For sequencing-related technical reasons, the minus strand sequence was determined for the U 1 G mutants; a mirror image of the electropherogram is shown with the corresponding plus strand sequence listed at the top of the panel. For every mutant, a sequence alignment of the leader (red) and body (blue) TRSs and surrounding sequences is shown (TRSs are boxed). The mRNA7 leader±body junctions detected by our sequence analysis are shown in yellow. duplex formed between the body TRS complement at the 3¢ end of the leaderless sg minus strand and the leader TRS in the genomic RNA template should suf®ce to position the nascent minus strand properly for subsequent elongation to add the complement of the leader sequence. In most cases, the nascent minus strand contains the entire body TRS complement at its 3¢ end at the moment of strand transfer, leading to a body TRS-derived leader±body junction sequence in the sg mRNA molecule. In a small number of transcripts, however, minus strand synthesis appears to be interrupted before nucleotide +1 of the body TRS is copied and, after strand transfer, resumes by incorporating the complement of the +1 nucleotide of the leader TRS. As stated above, we postulate that the detection of this phenomenon is determined by the level of crossovers between the ±1 and +1 position that is allowed by the mutations introduced at the +1 position of body TRS or leader TRS. We cannot, however, formally exclude that a`back trimming' activity degrades the 3¢-terminal nucleotide of the minus strand before or after strand transfer. However, note that in the discontinuous minus strand extension model ( Figure 1C ), such an activity would not disturb the proper positioning of the nascent minus strand on the leader template, because the TRS± TRS duplex would be shortened by one nucleotide only. Nidovirus discontinuous minus strand extension resembles similarity-assisted, copy-choice RNA recombination Due to their discontinuous sg RNA synthesis, nidoviruses occupy a special`niche' in the +RNA virus world. Their mode of sg RNA production is clearly different from that of other +RNA viruses and resembles another welldocumented +RNA virus feature: RNA recombination (for recent reviews see Nagy and Simon, 1997; Aaziz and Tepfer, 1999; Worobey and Holmes, 1999) . Most of the experimental evidence supports an RdRp template switch (Kirkegaard and Baltimore, 1986) as the main mechanism of RNA recombination. Mechanistically, such a template switch involves the transfer of a nascent strand from one RNA template (donor) to the other (acceptor). Also, nidovirus discontinuous sg RNA synthesis involves transfer of a nascent RNA strand, the sg RNA, but now from one site to another in the same template. Based on the data currently available, we refer to the discontinuous minus strand extension model as our working model for nidovirus sg RNA synthesis. If one applies the`recombination terms' to this model (Chang et al., 1996; Brian and Spaan, 1997; van Marle et al., 1999a) , the donor strand would be the body part of the genomic RNA template, the acceptor strand would be the leader part of the genomic RNA template and the nascent strand would be the discontinuously synthesized minus strand. Nagy and Simon (1997) have de®ned three main classes of RNA recombination: similarity-essential, similarity-non-essential and similarity-assisted recombination. The latter is de®ned as a mechanism in which strand transfer is determined by both sequence similarity between the parental RNAs and additional RNA determinants, present in only one of the parental RNAs. The results of our present study strongly suggest that nidovirus discontinuous sg RNA synthesis can be considered a special case of high-frequency similarity-assisted RNA recombination. While the only obvious function of the leader TRS is to ensure the ®delity of the strand transfer by base pairing with the 3¢ end of the nascent strand, the body TRS in the donor template indeed has additional, sequence-speci®c functions. One of these functions apparently is to pause (or terminate) nascent strand synthesis and thereby provide the opportunity for strand transfer. In addition, body TRS-derived nucleotides may play a role in the reinitiation of nascent strand synthesis on the acceptor template. Given the compact nature of the EAV TRS, it is quite possible that some nucleotides ful®l multiple tasks. RNA secondary structure of the body TRS may regulate sg RNA synthesis The sequence-speci®c function of the body TRS, revealed in this study, may be exerted at the level of either primary sequence or secondary structure. For a number of +RNA viruses, RNA secondary structure motifs located in the (proximity of) sg RNA promoters are vital for sg RNA synthesis. In alfalfa mosaic virus (Haasnoot et al., 2000) , turnip crinkle virus (TCV) (Wang et al., 1999) and barley yellow dwarf virus (Koev et al., 1999) , stem±loop structures in sg RNA promoter regions of the template strand are required for sg RNA synthesis. The sg RNA1 promoter of the latter virus is especially interesting, since it contains two stem±loop domains. For one of them, secondary structure, but not the primary sequence, is important for sg RNA synthesis, whereas the other domain acts through primary sequence, and not secondary structure (Koev et al., 1999) . Similarly, RNA secondary structure may play only a minor role in the sequence-speci®c recognition of the BMV sg RNA promoter by the RdRp Siegel et al., 1997) . We have suggested previously that RNA secondary structure of body TRS regions contributes to their attenuating potential and thereby determines the relative portion of the nascent minus strands that is transferred to the leader TRS in the template (Pasternak et al., 2000) . At present, it is unknown whether EAV body TRSs are part of an RNA structural motif that is essential for body TRS function, or whether they are recognized by a protein factor in a sequence-speci®c manner. However, the latter seems less likely than the former, since even LB4 (Figure 2 ), in which ®ve TRS nucleotides were substituted, still produced some sg RNA7, although~30-fold less than the wild-type control. The fact that some sequences in the EAV genome match the leader TRS perfectly, but are not used for sg mRNA synthesis, also argues against the recognition of a speci®c sequence (Pasternak et al., 2000) . More probably, mutagenesis of the RNA7 body TRS disturbed an RNA structure that is necessary for its function. This could, for example, explain the fact that the BU 6 C substitution reduced the amount of RNA7 by 20-fold (and could not be rescued by the same mutation in the leader TRS), whereas the wild-type RNA6 body TRS contains a C at the same position. If a protein factor were involved in sequence-speci®c TRS recognition, then one would expect it to recognize all TRSs similarly. If RNA structure is important for recognition by such a protein, then the BU 6 C substitution probably disturbs a structural motif of the RNA7 TRS, which is not present in the RNA6 TRS. On the other hand, conservation of part of the TRS in other arteriviruses suggests a sequence-speci®c recognition. Further studies are required to distinguish between these possibilities. In the TCV satellite RNA recombination system, the hairpin structure in the acceptor strand, as well as the donor±acceptor homology region, are necessary for the template switch . The hairpin has been postulated to bind the RdRp, whereas the homology region targets the nascent strand to the crossover site. The TCV RdRp probably recognizes the secondary and/or tertiary structure of the hairpin, while individual nucleotides play a less important role . In EAV, the leader TRS in the acceptor template is predicted to reside in the loop of an extensive hairpin, and its base pairing interaction with the body TRS complement at the 3¢ end of the nascent minus strand would resemble certain antisense RNA-regulated control mechanisms that are based on interactions between single-stranded tails and hairpin loops (van Marle et al., 1999a, and references therein) . It is possible that the EAV RdRp, or its accessory proteins, also binds to the stem of the long hairpin that presents the leader TRS. In any case, the leader TRS itself does not seem to be recognized by a protein in a sequence-speci®c manner. The body TRS is a better candidate to serve as a protein recognition site. This protein would then mediate the pausing of the nascent strand synthesis and/or nascent strand transfer. This would resemble the DNA-dependent RNA polymerase I termination system, in which speci®c DNA-binding terminator proteins bind to termination sequences (Reeder and Lang, 1997) , or a function of the HIV nucleocapsid protein, which promotes the minus strand strong-stop DNA transfer (Guo et al., 1997) . The EAV replicase component nsp1, which recently was shown to possess an sg RNA synthesis-speci®c activity (Tijms et al., 2001) , may be a good candidate for such a regulatory role. Residues predicted to form a zinc ®nger structure in nsp1 were shown to be necessary for sg RNA synthesis. Interestingly, zinc ®nger structures in the HIV nucleocapsid protein facilitate strand transfer (Guo et al., 2000) . Finally, it should be noted that the RNA structure of the nascent strand may also in¯uence pausing, strand transfer or reinitiation, as illustrated by the fact that stable hairpin structures in the nascent strand promote termination of transcription by Escherichia coli RNA polymerase (Wilson and von Hippel, 1995) . Site-directed mutagenesis, RNA transfections and immuno¯uorescence analysis Site-directed mutagenesis of EAV leader and body TRSs was carried out as described by van Marle et al. (1999a) , and all mutant constructs were sequenced. Following in vitro transcription from infectious cDNA clones, full-length EAV RNA was introduced into BHK-21 cells by electroporation, as described by van Dinten et al. (1997) . Immuno¯uorescence assays with EAV-speci®c antisera were performed at 14 h posttransfection as described by van der Meer et al. (1998) . To visualize the nuclei for cell counting, nuclear DNA was stained with 5 mg/ml Hoechst B2883 (Sigma). Cells were counted using the Scion Image software (Scion Corporation) and the percentage of transfected cells was calculated on the basis of the number of cells positive for the EAV replicase component nsp3 (Pedersen et al., 1999) . For RNA analyses, cells were lysed at 14 h post-transfection. Intracellular RNA isolation was performed using the acidic phenol method as described by Pasternak et al. (2000) . Total intracellular RNA was resolved in denaturing agarose±formaldehyde gels. Hybridization of dried gels with the radioactively labelled oligonucleotide probe E154, which is complementary to the 3¢ end of the EAV genome and recognizes all viral mRNA molecules (genomic and subgenomic), and phosphoimager quantitation of individual bands were performed as described by Pasternak et al. (2000) . To determine the leader±body junction sequence of sg mRNA7, mRNA7-speci®c RT±PCRs were carried out as described by van Marle et al. (1999b) using an antisense (RT and PCR) primer from the RNA7 body region and a sense PCR primer matching a part of the leader sequence. RT±PCR products were sequenced directly as described by Pasternak et al. (2000) using the leader-derived primer, an ABI PRISMÔ sequencing kit (Perkin Elmer) and an ABI PRISMÔ 310 Genetic Analyser (Perkin Elmer).
6
Debate: Transfusing to normal haemoglobin levels will not improve outcome
Recent evidence suggests that critically ill patients are able to tolerate lower levels of haemoglobin than was previously believed. It is our goal to show that transfusing to a level of 100 g/l does not improve mortality and other clinically important outcomes in a critical care setting. Although many questions remain, many laboratory and clinical studies, including a recent randomized controlled trial (RCT), have established that transfusing to normal haemoglobin concentrations does not improve organ failure and mortality in the critically ill patient. In addition, a restrictive transfusion strategy will reduce exposure to allogeneic transfusions, result in more efficient use of red blood cells (RBCs), save blood overall, and decrease health care costs.
Anaemia is a common condition in critically ill patients, and RBC transfusions are often used in the treatment and management of this patient population. In fact, one study [1] reported that 25% of all critically ill patients received RBC transfusions. Many laboratory studies [2] [3] [4] [5] [6] [7] [8] have examined the physiological responses (ie compensatory mechanisms) of the body to anaemia, which include the following [9] : increased cardiac output, decreased blood viscosity, capillary changes, increased oxygen extraction, and other tissue adaptations to meet oxygen requirements. Although critically ill patients are affected by a number of factors that predispose them to the adverse consequences of anaemia, persistence of this condition is of particular concern because it may cause the compensatory mechanisms in these patients to become impaired, risking oxygen deprivation in vital organs [9] . However, critically ill patients may also be at increased risk from the adverse effects of RBC transfusions, such as pulmonary oedema from volume overload, immune suppression resulting in increased risk of infection, and microcirculatory injury from poorly deformable RBCs. It is the aim of the present commentary to justify the statement 'Transfusing to normal haemoglobin concentration will not improve outcome.' If we define normal haemoglobin as being greater than 115 g/l for women and greater than 125 g/l for men, then there is no evidence in the literature to justify maintaining such high concentrations by the use of RBC transfusions in any anaemic patient. There may, however, be some debate about adopting a transfusion threshold of 100 g/l, which is well below 'normal'. transfusion threshold would, obviously, reduce the number of allogeneic RBCs transfused. It is our goal to impress upon the reader that transfusing to a level equal to or greater than 100 g/l does not improve mortality and other clinically important outcomes in a critical care setting. We first explore the reasons why a reduction in the total number of allogeneic blood transfusions would be beneficial. Second, we examine the current evidence for using a lower transfusion strategy, specifically that employed in the Transfusion Requirements In Critical Care (TRICC) trial. RBC transfusions have inherent risks that may be categorized as follows [11] [12] [13] [14] [15] : transfusion-transmitted infections; immune-related reactions (acute or delayed haemolytic reactions, febrile, allergic, anaphylactic reactions and graft-versus-host disease); and nonimmunerelated reactions (fluid overload, hypothermia, electrolyte toxicity and iron overload). Major improvements in donor screening procedures and laboratory testing have dramatically improved the safety of the blood supply [16] . Currently, the risk of transmitting an infectious agent through blood transfusion ranges from 1:100,000 for hepatitis B virus to 1:1,000,000 for HIV (Canadian Blood Services, personal communication, 2000). The most important threats to the blood supply remain new and unknown pathogens. More recently, concern has focused on the potential transmission of prions through RBCs. Also, infectious agents with long latency periods pose particular risks to young individuals who require RBCs, such as multiple trauma victims. The risk : benefit ratio for a 24-year-old trauma victim with a 50-year life expectancy differs markedly from that for a person aged 80 years who is undergoing coronary artery bypass surgery. In summary, because there is a risk of transmitting diseases through the blood supply, we should always strive to use RBCs according to the best available evidence in order to ensure that we do more good than harm to our patients. It is a long-standing observation [17] [18] [19] [20] [21] that blood transfusions are associated with immune suppression. This clinical phenomenon was first observed in renal transplant patients who had received blood transfusions while on dialysis before the transplant [22] , and has been observed repeatedly in transplant centres around the world [23, 24] . Recently, Opelz et al [25] reported a multicentre clinical trial in which all renal allograft recipients received modern immunosuppressive regimens. Those patients who were allocated to receive three allogeneic RBC units before renal transplant had a 1-year graft survival rate of 90%, as compared with 82% for patients who were not transfused (P = 0.02). These data suggest that there are long-term immunosuppressive effects following transfusion of nonleukocyte-reduced allogeneic RBCs. A large number of studies [26] [27] [28] [29] [30] [31] [32] [33] [34] have also suggested that allogeneic transfusions accelerate cancer growth, perhaps due to altered immune surveillance. These altered immune responses after allogeneic RBC transfusions may also predispose critically ill transfusion recipients to nosocomial infections [35] [36] [37] [38] [39] [40] and increased rates of multiplesystem organ failure [41] , which may ultimately result in higher mortality rates. However, most studies that examined the association between cancer recurrence and postoperative infection after transfusion [42, 43] only provided weak causal inferences because of poor study design and the lack of independence between allogeneic RBC transfusions and the potential complication. A recent meta-analysis [44] combined the results from seven RCTs, and was unable to detect clinically important decreases in mortality and postoperative infections. We added the results of a new RCT by van de Watering et al [45] to the above meta-analysis. The relative risk for allcause mortality was 1.05 (95% confidence interval 0.88-1.25), and was 1.10 (95% confidence interval 0.85-1.43) for postoperative infections. However, this meta-analysis excluded two positive RCTs [40, 46] because of the significant statistical heterogeneity introduced by these two studies. If all available RCTs are combined, ignoring heterogeneity, then the relative risk difference for postoperative infections across all studies is 1.60 (95% confidence interval 1.00-2.56; P = 0.05). Thus, the available evidence suggests that universal prestorage leukoreduction could have clinical effects that range from none to decreasing rates of infection by as much as 50% in high-risk patients. In summary, despite convincing laboratory evidence and some supportive clinical studies, the clinical significance of the immunosuppressive effects of allogeneic RBC transfusions have not been clearly established [47] . More importantly, the impact of a leukoreduction programme has not been studied in a large population of patients who are expected to have significant exposure to allogeneic RBCs. The majority of complications from allogeneic RBC transfusion, however, are no more frequent in the intensive care setting than in other patient populations, with the possible exception of pulmonary oedema, hypothermia and electrolyte disturbance. Hypothermia and electrolyte disturbances occur most frequently with massive transfusions. In critically ill patients, the optimal effective circulatory volume may be difficult to determine, and as a consequence pulmonary oedema may be a much more frequent complication of RBC transfusion than in other patient populations. This may explain the significantly higher rate of pulmonary oedema in patients transfused using a threshold of 100 g/l (5.3% in the restrictive transfusion group versus 10.7% in the liberal transfusion group; P < 0.01), as reported in the TRICC trial [10] . As an alternative explanation, the more frequent use of RBCs might have resulted in more frequent episodes of transfusion-related acute lung injury in the liberal strategy group (7.7% in the restrictive strategy versus 11.4% in the liberal strategy; P = 0.06), as reported in the TRICC trial. Clinical evidence is also insufficient to definitively establish a correlation between the age of RBCs being transfused and patient mortality; however, laboratory evidence has shown many storage-related changes that may result in impairment of blood flow and oxygen delivery at the microcirculatory level. Marik et al [48] demonstrated an association between a fall in gastric intramucosal pH and transfusion of RBCs stored for longer than 15 days. In addition, there is ample laboratory evidence that prolonged RBC storage adversely affects RBCs, potentially results in the generation of cytokines, and alters host immune function. In another study, Fitzgerald et al [49] , using an animal model of transfusion, consistently observed a lack of efficacy of transfused, stored rat blood to improve tissue oxygen consumption as compared with fresh cells or other blood substitutes. Three retrospective clinical studies tested the association between the age of transfused blood and duration of stay in the intensive care unit (ICU) [50] and mortality [51, 52] . Martin et al [50] observed a statistically significant association between the transfusion of aged blood (>14 days old) and increased duration of ICU stay (P = 0.003) in 698 critically ill patients. In patients who received a transfusion, aged RBCs was the only predictor of duration of stay (P < 0.0001). In survivors, only median age of blood was predictive of duration of stay (P < 0.0001). Purdy et al [51] demonstrated a negative correlation (r = -0.73) between the proportion of RBC units of a given age transfused to survivors and increasing age of RBCs in patients admitted to the ICU with a diagnosis of severe sepsis (n = 31). Those investigators also noted that these latter units were more likely to be older. A recently reported study by Vamvakas and Carven [52] evaluated the effect of duration of RBC storage on postoperative pneumonia in 416 consecutive patients undergoing coronary artery bypass grafting. Those investigators noted an adjusted increase of 1% in the risk of postoperative pneumonia per day of average increase in the duration of RBC storage (P < 0.005) in transfused patients. Each of these three studies also noted that patients who received a large number of RBC units had a higher mortality. Although these risks are relatively small when viewed collectively, they become significant when one considers that 25% of all critically ill patients in Canada are transfused during their ICU stay [1] . Until recently, physicians have depended on clinical judgement when deciding at what haemoglobin level to transfuse a critically ill patient. As a result, significant variation has been shown to exist in transfusion practice among Canadian critical care physicians [53] , which is due largely to a lack of published data on the subject. An arbitrary haemoglobin level of 100 g/l has historically been used as a threshold to transfuse critically ill patients. Six observational studies investigated the importance of anaemia on transfusion practices in various settings. Of these, three large cohort studies, which were performed in different patient populations (intensive care [1] , coronary artery bypass surgery [54] and hip fractures [55]), reached different conclusions. RBC transfusions in particular improved outcome in critically ill patients with cardiovascular disease, but increased the risk of myocardial infarction in coronary artery bypass surgery patients. Transfusion had no impact on short-term or long-term mortality in hip-fracture patients. Three smaller studies [56] [57] [58] evaluated the relationship between anaemia and adverse outcomes in vascular disease patients. Although increased numbers of ischaemic events were observed in anaemic patients, the validity of these studies is uncertain, given that the decision to transfuse a patient was often correlated with illness burden of the patient. It is also possible that comorbidity was not adequately adjusted for in those studies. Transfusion thresholds were compared in 10 randomized clinical trials [10, [59] [60] [61] [62] [63] [64] [65] [66] [67] . Although the clinical settings varied, each trial randomized patients to be transfused on the basis of a 'conservative' or a 'liberal' strategy. The definitions of conservative and liberal strategies varied, and actually overlapped between studies. Of these 10 trials, only three included more than 100 patients and only one trial evaluated the impact of transfusion on symptoms. In the first trial of patients undergoing elective coronary artery bypass surgery [65] , the differences between perioperative haemoglobin levels were small, event rates were very low, and there were no differences in any outcome. In the second trial [67] , patients undergoing knee arthroplasty were randomly assigned to receive autologous blood transfusion immediately after surgery or to receive autologous blood if haemoglobin level fell below 9 g/dl [67] . Again, no differences in outcome were observed. The third trial of hip fracture patients undergoing surgical repair [64] found no differences in outcomes; however, five deaths were recorded at 60 days after surgery in the symptomatic group, and two deaths occurred in the 10 g/dl group. The numbers of patients in these trials were too small to evaluate the effect of lower transfusion triggers on clinically important outcomes such as mortality, morbidity and functional status. In 1999, Hebert et al [10] reported the results of the TRICC trial. Patients (n = 838) were randomized either to a restrictive strategy (haemoglobin concentration maintained between 70 and 90 g/l, with a trigger set at 70 g/l) or to a liberal strategy (haemoglobin concentration maintained between 100 and 120 g/l, with a trigger at 100 g/l). To date, the TRICC trial is the only large study that has investigated these parameters. The groups were comparable at baseline. The average daily haemoglobin concentration ranged from 85 ± 7.2 g/l in the restrictive group to 107 ± 7.3 g/l in the liberal group (P < 0.01). The average number of transfusions was reduced by 52% in the restrictive group (2.6 ± 4.1 versus 5.6 ± 5.3 RBCs/patient; P < 0.01). Cardiac events, primarily pulmonary oedema and myocardial infarction, were more frequent in the liberal strategy (P < 0.01; Table 1 ). On examination of composite outcomes, the number of patients with multiorgan failure was found to be substantially increased in both groups, with the results being marginally better in the restrictive strategy group (20.6% versus 26.0%; P = 0.07; Table 2 ). Overall, the restrictive transfusion group showed a lower 30-day mortality (18.7% versus 23.3%; P = 0.11; Fig. 1 ). Kaplan-Meier survival curves, however, were significantly different in the subgroup of patients with an Acute Physiology and Chronic Health Evaluation II score of 20 or less (P = 0.02; Fig. 2 ). In addition, although 60-day mortality (22.8% versus 26.5%; P = 0.23) and ICU mortality (13.9% versus 16.2%; P = 0.29) were not statistically significant, they did show a consistent trend in terms of absolute values that favoured the restrictive strategy. The key observation from the TRICC trial is not that the restrictive strategy is better, but rather that it is at worst equivalent to the liberal strategy and at best superior to the liberal strategy. At this juncture, preclinical and clinical evidence support the adoption of a more restrictive transfusion strategy in most critically ill patients. However, there remain divergent views regarding the risks and benefits of treating anaemia in patients with cardiovascular disease. Laboratory-based studies [68, 69] suggest that patients with cardiovascular disease may require higher haemoglobin concentrations to maintain oxygen delivery in partially occluded or diseased coronary arteries. Studies to demonstrate how anaemia affects contractile function of the left ventricle have rarely shown important effects above haemoglobin concentrations of 70 g/l. Indeed, it is more important to address the underlying pathophysiological causes of the acute coronary syndrome with proven therapy such as aspirin and β-blockers, rather than treating mild-to-moderate anaemia as an initial step. If the effects of RBC transfusion were either limited or increased then there would be no debate; however, the use of allogeneic RBCs has been shown to be associated with immunomodulation [12, 47] and/or alteration in the delivery of oxygen in the microcirculation [70, 71] , resulting in increased rates of infections and organ failure. Few clinical studies have attempted to elucidate the risk : benefit ratio of anaemia and transfusion in cardiac patients. Two small RCTs [62, 72] examined transfusion practice in patients undergoing coronary artery bypass grafting, and concluded that a conservative approach to the administration of RBCs may be safe. However, two recent cohort studies suggested that anaemia may increase the risk of mortality in critical illness [73] and following surgery in patients with cardiovascular disease [74] . There were 418 and 420 patients in the restrictive and liberal transfusion groups, respectively. *Difference calculated by subtracting mean values of restrictive group from those of liberal group. † Three patients were lost to 60-day mortality rate; therefore n = 835. ‡ Nonsurvivors are considered to have all organs failing on date of death. § Changes in MOD score from baseline, while also incorporating adjustment for death. Data from Hébert et al [10] . In a study of Jehovah's Witnesses (a group that refuses RBC transfusion on religious grounds) undergoing surgical procedures [74] , it was noted that mortality was significantly increased in patients with cardiac disease after a decrease in haemoglobin levels from 100-110 g/l to 60-69 g/l. In that study, patients with no cardiac disease and similar changes in haemoglobin levels showed no increase in mortality, which is in accordance with the results of the TRICC trial [10] . In the study by Hébert et al. [73] of 4470 critically ill patients, a correlation between Critical Care Vol 5 No 2 Alvarez et al [10] . Kaplan-Meier estimates of survival in the 30 days after admission to the ICU in the restrictive and liberal transfusion strategy groups (all patients). Data from Hébert et al [10] . Kaplan-Meier estimates of survival in the 30 days after admission to the ICU in the restrictive and liberal transfusion strategy groups (patients with APACHE II score ≤20). Data from Hébert et al [10] . anaemia and mortality rates was observed. Those investigators also found that the risk of anaemia appeared to decrease with RBC transfusion in patients with cardiac disease. In patients with cardiac disease, mortality increased when haemoglobin concentrations were below 95 g/l, as compared with anaemic patients with other diagnoses (55% versus 42%; P = 0.09). In the subgroup of patients with cardiac disease, increasing haemoglobin values in anaemic patients was associated with improved survival (odds ratio 0.80 for each 10 g/l increase; P = 0.012). One possible explanation for the discrepancy between the TRICC trial and this observational study may be that the attending physicians who recruited patients into the study did not enter those patients who were considered to have severe cardiac disease. Hébert et al. [73] sought to examine further whether a restrictive transfusion strategy was at least as effective as a liberal strategy in critically ill patients with cardiac disease. In the subgroup of patients with cardiovascular disease from the TRICC trial, those investigators suggested that most haemodynamically stable critically ill patients with cardiovascular disease may be transfused when haemoglobin concentrations fall below 70 g/l, and that the hemoglobin concentration should be maintained between 70 and 90 g/l. Average daily haemoglobin concentrations were 85 ± 6.2 g/l in the restrictive transfusion group and 103 ± 6.7 g/l in the liberal transfusion group (P < 0.01). In the 357 patients with cardiovascular disease, the 30-day mortality rate was 23% in the restrictive transfusion group versus 23% in the liberal group (95% confidence interval of the difference -8.4% to 9.1%; P = 1.00). Other mortality rates, including 60-day (26% versus 27%; P = 0.90), ICU (19% versus 16%; P = 0.49) and hospital mortality (27% versus 28%; P = 0.81), were not significantly different between groups. Kaplan-Meier survival curves comparing time to death demonstrated similar trends in the two groups ( Fig. 3 ; P = 0.98). The multiple organ dysfunction (MOD) scores, during the entire study period, were also not significantly different between groups (8.6 ± 4.9 versus 9.0 ± 4.4; P = 0.40), but the change in MOD score from baseline values was significantly lower in the restrictive group than in the liberal group (0.2 ± 4.2 versus 1.3 ± 4.4; P = 0.02). Combined measures of morbidity and mortality, or composite outcomes, were also examined. When all patients who died were given a score of 24, the total MOD score between groups was not different (P = 0.39), or were the changes in MOD scores significantly different from baseline (2.7 ± 6.9 versus 4.0 ± 7.3; P = 0.08). Among the specific subset of cardiac patients with ischaemic heart disease (n = 257), there were no discernible differences in 30-day and 60-day as well as ICU mortality rates. However, a nonsignificant (P = 0.3) decrease in overall survival rate in the restrictive group was noted in those patients with confirmed ischaemic heart disease, severe peripheral vascular disease or severe comorbid cardiac disease (Fig. 4) . In conclusion, a restrictive RBC transfusion strategy generally appears to be safe in most critically ill patients with cardiovascular disease, with the possible exception of patients experiencing acute myocardial infarction or unstable angina. Survival over 30 days in patients with ischemic heart disease in the restrictive and liberal allogeneic RBC transfusion strategy groups. This graph illustrates Kaplan-Meier survival curves for all patients with ischemic heart disease in both study groups. There is no difference in mortality in patients in the restrictive group (dashed line) as compared to the liberal group (solid line) (P = 0.30). The need to reduce the amount of allogeneic blood transfusions in order to reduce the associated risks has been firmly established. RBCs are associated with clinically important immune suppression, and stored RBCs have been shown to cause adverse microcirculatory effects that result in increased organ failure. The question for some time has been whether critically ill patients are able to tolerate lower levels of haemoglobin without deleterious effects, thus reducing the amount of exposure to allogeneic transfusions. In the only large RCT, Hébert et al [10] established that there was no difference in mortality rates between restrictive and liberal transfusion strategies in noncardiac, critically ill patients. Although those investigators were able to show convincing trends that the liberal strategy may in fact be deleterious in terms of absolute values, statistical significance was not achieved. However, the fact that no difference between the two strategies was achieved is of great importance, because this means that the total number of transfusions can be reduced by approximately half without any impact on mortality. In addition, these findings are easily put into clinical practice. Although many questions remain, the TRICC trial and many laboratory and clinical studies have established that transfusing to normal haemoglobin concentrations does not improve organ failure and mortality in the critically ill patient. As such, a restrictive transfusion strategy will reduce exposure to allogeneic transfusions, result in more efficient use of RBCs, save blood overall, and decrease health care costs.
7
The 21st International Symposium on Intensive Care and Emergency Medicine, Brussels, Belgium, 20-23 March 2001
The 21st International Symposium on Intensive Care and Emergency Medicine was dominated by the results of recent clinical trials in sepsis and acute respiratory distress syndrome (ARDS). The promise of extracorporeal liver replacement therapy and noninvasive ventilation were other areas of interest. Ethical issues also received attention. Overall, the 'state of the art' lectures, pro/con debates, seminars and tutorials were of a high standard. The meeting was marked by a sense of renewed enthusiasm that positive progress is occurring in intensive care medicine.
This year's symposium was dominated by the results of recent clinical trials. After 10 years of 'magic bullet' trials in sepsis, a number of successful therapeutic options are now emerging. In addition, recent advances in our understanding of the soup of mediators observed in sepsis offer yet more tantalizing targets for new therapies. In contrast, the eagerly awaited results from Italy of the prone positioning trial in ARDS were disheartening. The epidemiology of both sepsis and ARDS, and their impact on clinical studies and the future provision of critical care were also hot topics. The era of extracorporeal liver replacement therapy is upon us, with considerable early promise and the probability of wide availability. Finally, as always, ethics remained an area of interest. This report summarizes and discusses the presentations on the above topics. Angus (Pittsburgh, PA, USA) presented his group's work on the epidemiology of sepsis in the USA (accepted for publication in Critical Care Medicine). They developed a method for identifying hospitalized patients with sepsis based on ICD9 criteria, the most widely recorded coding system used in US hospitals. Prospective testing of the method found it to be both sensitive and reliable. They then applied it to a representative selection of US hospitals. Their results indicated that about 50% of intensive care unit (ICU) patients have systemic inflammatory response syndrome, and that approximately 20% of these progress to severe sepsis. Mortality for severe sepsis was greater than 30%. Demographically, those at the extremes of age represent the most at-risk groups, in whom the mortality is also the highest. These data provides yet another reminder that the increasing demands on health care resources caused by the ageing population is predicted to exceed intensive care provision within the next The 21st International Symposium on Intensive Care and Emergency Medicine, Brussels, Belgium, 20-23 March 2001 10-20 years. Finally, those investigators found a striking demographic peak in patients aged 20-30 years, which they attributed largely to human immunodeficiency virus. The long-standing debate between the two schools of sepsis theory -microcirculatory dys-autoregulation versus cellular dysfunction -shows signs of resolution. New techniques for studying tissue oxygen tension, presented by Ince (Amsterdam, The Netherlands), provide more evidence that microcirculatory dys-autoregulation results in significant shunting. This occurs predominantly in the submucosal and serosal portions of organs, and is an early event. These studies show that the macroscopic restoration of global oxygen delivery fails to improve oxygen consumption as the mucosa becomes hyperoxic, whereas the submucosa and serosa remain hypoxic. Somewhat counterintuitively, this can be reversed in the face of resistant hypotension with vasodilators, at least in animal models. The cellular dysfunction camp, although still somewhat doubtful as to the importance of these microcirculatory findings, have now clearly established that their championed mechanism of mitochondrial failure is a late but crucial event in the evolution of sepsis. Fink (Pittsburgh, PA, USA) presented evidence that mitochondrial failure in septic cells is triggered by the activation of the enzyme poly-adenosine diphosphate ribose polymerase [1] . This enzyme represents a significant target for novel therapies, which are apparently already in development. The debate regarding the toxicity of oxygen and the formation of free radicals continues despite the absence of demonstrated effectiveness of scavenging therapies, and is a testament to the incomplete understanding of this area. The round-table conference preceding this year's symposium concentrated on distilling current knowledge on the microscopic events in critically ill patients into an explanation of the macroscopic multiorgan failure that is so commonly encountered. The conclusions of the conference appeared to relate mostly to future directions for research, in particular the study of organ-organ interactions. Marshall (Toronto, Canada) proposed the development of an alternative to the much-maligned physiological scoring systems, based on the staging systems widely used in the field of oncology. He proposed that mediator levels, in addition to physiological variables, will soon be used usefully to characterize septic patients. He also suggested that, in the light of the recent successful mediator trials in sepsis, future therapies will be directed in a manner analogous to the control of glucose in diabetic patients. The natural anticoagulants antithrombin III (AT III), tissue factor pathway inhibitor (TFPI) and activated protein C (APC), and the cytokine tumour necrosis factor (TNF)-α are the latest inflammatory mediators to be targeted in large multicentre clinical trials in an attempt to improve the current dismal outcome for patients with severe sepsis. The KyberSept AT III study recruited over 2300 patients from 200 centres, with high Simplified Acute Physiology Scale scores (median 50), and a mortality of nearly 40% [2] . Unfortunately, no overall benefit was shown between AT III and placebo, although results were more encouraging in an analysis of the subgroup of patients who received AT III but no heparin, which is known to inhibit AT III. Interestingly, improvements in quality of life scores were seen in survivors who received AT III in comparison to those who received placebo, suggesting that morbidity may be reduced, although again this was an analysis of a subgroup. Patients in the AT III group who received concomitant heparin had a significantly higher incidence of bleeding events, and outcome worsened as the dose of heparin increased. Explanations for the failure of this study included the inhibitory effects of heparin and the failure to achieve AT III activity levels of greater that 200% from baseline in the treatment population, a level established as required for therapeutic benefit in phase II trials. Phase II clinical trial results using TFPI (TFPI n = 141, placebo n = 69; unpublished data) show a mortality benefit in the sicker sepsis patients who already have coagulation problems. Results of the phase III multicentre study are expected to be presented at the 22nd International Symposium on Intensive Care and Emergency Medicine, in Brussels in 2002. Human trials of various anti-TNF-α formulations have been variable to date, and include North American sepsis trial (NORASEPT) I [3] , International sepsis trial (INTERSEPT) [4] and NORASEPT II [5] . Possible reasons have included a lack of biological activity of the anti-TNF-α formulation studied, inappropriate timing of therapy, redundancy of proinflammatory mediators and hetereogeneity of patient populations. The Monoclonal Anti-TNF, A Randomized controlled Sepsis Trial (MONARCS) study used a different anti-TNF-α formulation (F[ab′]2 fragment of a murine monoclonal antibody to human TNF-α), and stratified patients based on demonstrable abnormalities in immunological pathways (highly elevated interleukin-6 levels -a circulating cytokine that is induced by TNF-α). Unpublished results revealed 28-day mortality rates of 44 and 48% in the anti-TNF-α and placebo groups, respectively, in those patients who had high interleukin-6 levels on recruitment to the study (n = 488 anti-TNF-α, n = 510 placebo). This represented a relative mortality reduction of 14%. Relative mortality reduction in all patients (n = 1305 anti-TNF-α, n = 1329 placebo), independent of baseline interleukin-6 levels, was only 10%. The Recombinant Human Activated Protein C Worldwide Evaluation in Severe Sepsis (PROWESS) study is hot off the press [6] , and presentation of the results at the congress allowed those of us who still carry the unopened New England Journal of Medicine issue in our briefcases to catch up! A total of 164 sites from 11 countries recruited 1690 patients with severe sepsis, before the trial was prematurely stopped following the second safety analysis. Twenty-eight-day all-cause mortality rates for placebo and APC were 31 and 25% respectively, with a relative risk reduction of 19%. Resolution of cardiovascular and respiratory function was more rapid in survivors who received APC, although ICU and hospital stay did not differ. There was a trend towards an increase in serious bleeding events in the APC group (3% APC versus 2% placebo), but these events were primarily due to trauma or instrumentation. Although this is an exciting breakthrough, we all recognize that when APC reaches the market place it will seriously stretch ICU finances, especially because there appear to be other mediators on the horizon that we will be encouraged to use, in combination, to fight the inflammatory 'soup'. Two opposing epidemiological views of ARDS were presented by Lemaire (Créteil, France) and Evans (London, UK). Broad agreement does seem to exist as to the incidence of this condition, which is in the order of 10/100,000, although there is significant variation between countries. It was argued that this variation results from the availability of ventilated beds, with higher incidences apparent in countries with greater provision, emphasizing that this condition can be considered the result of intensive care intervention or, as one speaker put it, 'physician-induced lung injury'. Early results from the Acute Lung Injury Verification of European Epidemiology (ALIVE) study (unpublished data), sponsored by the European Society of Intensive Care Medicine, are at odds with recent trial findings. The ALIVE study, which included over 6000 patients surveyed in 1998, found a 50-60% 28-day mortality, which compares to only 20-30% in the control groups of recent trials. Pneumonia was the commonest cause, responsible for 50% of cases, with sepsis identified as the cause in a further 20-30%. Astonishingly, this study found the ratio of arterial oxygen tension to fractional inspired oxygen at ICU admission was highly predictive of mortality, despite continuing controversy regarding this measurement. A diverse range of views were presented from the Third International Consensus Conference on ARDS (unpublished data), held in Barcelona late last year. The decision as to how to change the defining criteria for this condition remains unresolved. The debates surrounding chest X-ray criteria, the use of the ratio of arterial oxygen tension to fractional inspired oxygen, and the level/utility of pulmonary artery wedge pressure measurements continue. In addition, a debate has arisen as to whether ARDS can be a unilateral process, and whether it can coexist with cardiac failure. There appears to be increasing recognition that ARDS represents only a small subset of patients with acute lung failure (approximately 30%). Surprisingly little is known about the remainder of this larger group. In contrast to the ALIVE study, several centres have reported their 28-day mortality at 40%, which represents an improvement from the 60% of 10 years ago. However, it was argued that a 28-day follow-up period is too short for clinical trials, as the long-term quality of life for patients with ARDS is poor compared with that of critically ill patients without this condition. Results suggest that the recovery of lung function is good overall, but is dependent on severity. Treatment recommendations include the universal adoption of the US National Institutes of Health protective lung ventilation strategy [7] . There was general agreement that recruitment manoeuvres are beneficial, but how and when to employ them remains controversial. Rouby (Paris, France) put forward a new classification for ARDS based on computed tomography findings. He observed that patients can be split into three groups, depending on the appearance of the upper lobes. In group 1 the upper lobes are normal, and positive end-expiratory pressure (PEEP) is of little benefit and results in significant over-distension. Survival in this group is approximately 60%. In group 2 the upper lobes are abnormal, PEEP is of dramatic benefit, but survival is only approximately 25%. In group 3 there are mixed/patchy abnormalities, the effects of PEEP are less predictable, but, as in group 1, survival is approximately 60%. Gattinoni (Milan, Italy) presented the results of the longawaited Italian multicentre randomized controlled trial of prone positioning in ARDS (unpublished data). This trial was terminated after 3 years despite having only recruited 304 patients; enrollment of 750 patients was originally planned, in order to achieve sufficient statistical power. At trial outset, recruitment was encumbered by the lack of familiarity with and scepticism regarding this procedure in many of the centres. However, by the end of the study many participants were unwilling to enter patients into the trial, because they felt it unethical to deny them this intervention. The trial protocol resulted in patients in the treatment group being prone for an average of only 7 out of 24 h for a 10-day period. Overall there was no difference in mortality between the control and treatment groups at day 10, time of ICU discharge, or at 6 months. Interestingly, analysis of subgroups revealed a significant difference in the outcome at 10 days for patients with the most severe disease, although this disappeared by ICU discharge. In retrospect, the design of this ambitious trial was flawed by its failure to establish the optimal utilization of this manoeuvre. The opening session reported that we are making progress in supporting the failing liver (Wendon, London, UK). Current optimism should probably be limited to extracorporeal methods, because the molecular adsorbent recirculating system (essentially extracorporeal albumin dialysis) has been shown to have beneficial clinical effects as well as improved survival in two small randomized controlled trials [8, 9] . The equipment is familiar to all those who use dialytic therapies, and we will undoubtedly hear more about this system in the next few years. The slide of a patient reading the newspaper through a transparent helmet, while receiving noninvasive ventilation (NIV) resembled pictures of a NASA astronaut! However, it was reported to be well tolerated for prolonged periods, and significantly reduces the complications associated with NIV (pressure areas, tolerance of mask). The recent Consensus Conference [10] examined weaning aspects of NIV and emphasized the reduced weaning time and avoidance of reintubation, but called for more randomized controlled trials. Finally, although continuous positive airway pressure has been shown to be beneficial in pulmonary oedema, caution is still advised with the use of bilevel positive airway pressure because of the reporting of myocardial infarction in several studies. However, the groups studied were unmatched and starting points were different, so conclusions should not be drawn until randomized controlled trial results are available in this area. This was a well-attended session, which, according to Levy (Providence, RI, USA), was in complete contrast to the interest shown in the USA for the subject. Although there were few new data in the session, the emphasis on a strategy for lawsuits was welcome. Suggestions included statements from scientific societies at a national and international level, open reporting in medical files of decisions to withdraw or withhold treatment, and family involvement in decision making that will ultimately involve better media education. The last day of this year's symposium was sadly abandoned by many due to the Belgian rail strike. Despite this, the usual convivial atmosphere, both in and around the congress, was as abundant as ever. Overall, the 'state of the art' lectures, pro/con debates, seminars and tutorials were of the usual high standard, although, yet again, access to many of the symposium's venues was limited by the lack of capacity of the secondary rooms. The 21st International Symposium was marked by a sense of renewed enthusiasm that real positive progress is occurring at the coal face of intensive care.
8
Heme oxygenase-1 and carbon monoxide in pulmonary medicine
Heme oxygenase-1 (HO-1), an inducible stress protein, confers cytoprotection against oxidative stress in vitro and in vivo. In addition to its physiological role in heme degradation, HO-1 may influence a number of cellular processes, including growth, inflammation, and apoptosis. By virtue of anti-inflammatory effects, HO-1 limits tissue damage in response to proinflammatory stimuli and prevents allograft rejection after transplantation. The transcriptional upregulation of HO-1 responds to many agents, such as hypoxia, bacterial lipopolysaccharide, and reactive oxygen/nitrogen species. HO-1 and its constitutively expressed isozyme, heme oxygenase-2, catalyze the rate-limiting step in the conversion of heme to its metabolites, bilirubin IXα, ferrous iron, and carbon monoxide (CO). The mechanisms by which HO-1 provides protection most likely involve its enzymatic reaction products. Remarkably, administration of CO at low concentrations can substitute for HO-1 with respect to anti-inflammatory and anti-apoptotic effects, suggesting a role for CO as a key mediator of HO-1 function. Chronic, low-level, exogenous exposure to CO from cigarette smoking contributes to the importance of CO in pulmonary medicine. The implications of the HO-1/CO system in pulmonary diseases will be discussed in this review, with an emphasis on inflammatory states.
The heme oxygenase-1/carbon monoxide (HO-1/CO) system has recently seen an explosion of research interest due to its newly discovered physiological effects. This metabolic pathway, first characterized by Tenhunen et al. [1, 2] , has only recently revealed its surprising cytoprotective properties [3, 4] . Research in HO-1/CO now embraces the entire field of medicine where reactive oxygen/nitrogen species, inflammation, growth control, and apoptosis represent important pathophysiological mechanisms [3] [4] [5] [6] . Indeed, the number of publications in recent years concerning HO-1 has increased exponentially, while the list of diseases and physiological responses associated with changes in HO-1 continues to expand [5] . Until now, relatively few studies have addressed the role of HO-1/CO in pulmonary medicine. Several investigators have focused on the diagnostic application of the HO-1/CO system, by measuring exhaled CO (E-CO) in various pathological pulmonary conditions, such as asthma or chronic obstructive pulmonary disease (COPD) [7] . In another experimental approach, investigators have examined the expression of HO-1 in lung tissue from healthy or diseased subjects [8, 9] . This review will highlight the actions of HO-1/CO in the context of heme degradation have antioxidant properties [18, 19] . The liberated heme iron undergoes detoxification either by extracellular efflux or by sequestration into ferritin, an intracellular iron-storage molecule with potential cytoprotective function [20] [21] [22] [23] . Of the three known isoforms of HO (HO-1, HO-2, and HO-3), only HO-1 responds to xenobiotic induction [24] [25] [26] [27] . Constitutively expressed in many tissues, HO-2 occurs at high levels in nervous and vascular tissues, and may respond to regulation by glucocorticoids [25, 28, 29] . HO-1 and HO-2 differ in genetic origin, in primary structure, in molecular weight, and in their substrate and kinetic parameters [25, 26] . HO-3 displays a high sequence homology with HO-2 but has little enzymatic activity [27] . This review will focus on the inducible, HO-1, form. In addition to the physiological substrate heme, HO-1 responds to induction by a wide variety of stimuli associated with oxidative stress. Such inducing agents include hypoxia, hyperoxia, cytokines, nitric oxide (NO), heavy metals, ultraviolet-A (320-380 nm) radiation, heat shock, shear stress, hydrogen peroxide, and thiol (-SH)-reactive substances [3] . The multiplicity of toxic inducers suggest that HO-1 may function as a critical cytoprotective molecule [3, 4] . Many studies have suggested that HO-1 acts as an inducible defense against oxidative stress, in models of inflammation, ischemia-reperfusion, hypoxia, and hyperoxia-mediated injury (reviewed in [3] ). The mechanisms by which HO-1 can mediate cytoprotection are still poorly understood. All three products of the HO reaction potentially participate in cellular defense, of which the gaseous molecule CO has recently received the most attention [30, 31] . The administration of CO at low concentrations can compensate for the protective effects of HO-1 in the presence of competitive inhibitors of HO-1 activity [32] [33] [34] . While HO-1 gene transfer confers protection against oxidative stress in a number of systems, clearly not all studies support a beneficial role for HO-1 expression. Cell-culture studies have suggested that the protective effects of HO-1 overexpression fall within a critical range, such that the excess production of HO-1 or HO-2 may be counterprotective due to a transient excess of reactive iron generated during active heme metabolism [35, 36] . Thus, an important caveat of comparative studies on the therapeutic effects of CO administration versus HO-1 gene delivery arises from the fact that the latter approach, in addition to producing CO, may have profound effects on intracellular iron metabolism. HO-1 expression is primarily regulated at the transcriptional level. Genetic analyses have revealed two enhancer sequences (E1, E2) in the murine HO-1 gene located at -4 kb (E1) and -10 kbp (E2) of the transcriptional start site [37, 38] . These enhancers mediate the induction of HO-1 by many agents, including heavy metals, phorbol esters, endotoxin, oxidants, and heme. E1 and E2 contain repeated stress-responsive elements, which consist of overlapping binding sites for transcription factors including activator protein-1 (AP-1), v-Maf oncoprotein, and the cap'n'collar/basic-leucine zipper family of proteins (CNC-bZIP), of which Nrf2 (NF-E2-related factor) may play a critical role in HO-1 transcription [39] . The promoter region of HO-1 also contains potential binding sites for nuclear factor κB (NF-κB), though the functional significance of these are not clear [40] . Both NF-κB and AP-1 have been identified as regulatory elements responsive to oxidative cellular stress [40, 41] . In response to hyperoxic stress, AP-1 factors mediated the induction of HO-1 in cooperation with signal-transducer and activator of transcription (STAT) proteins [41] . Furthermore, a distinct hypoxia-response element (HRE), which mediates the HO-1 response to hypoxia, represents a binding site for the hypoxia-inducible factor-1 (HIF-1) [42] . The toxic properties of CO are well known in the field of pulmonary medicine. This invisible, odorless gas still claims many victims each year by accidental exposure. CO evolves from the combustion of organic materials and is present in smoke and automobile exhaust. The toxic actions of CO relate to its high affinity for hemoglobin (240-fold greater than that of O 2 ). CO replaces O 2 rapidly from hemoglobin, causing tissue hypoxia [43] [44] [45] . At high concentrations, other mechanisms of CO-induced toxicity may include apoptosis, lipid peroxidation, and inhibition of drug metabolism and respiratory enzyme functions [44] . Only recently has it become known that, at very low concentrations, CO participates in many physiological reactions. Where a CO exposure of 10,000 parts per million (ppm) (1% by volume CO in air) is toxic, 100-250 ppm (one hundredth to one fortieth as much) will stimulate the physiological effects without apparent toxicity [4] . The majority of endogenous CO production originates from active heme metabolism (>86%), though a portion may be produced in lipid peroxidation and drug metabolism reactions [46] . Cigarette smoking, still practiced by many lung patients, represents a major source of chronic lowlevel exposure to CO. Inhaled CO initially targets alveolar macrophages and respiratory epithelial cells. The exact mechanisms by which CO acts at the molecular level remain incompletely understood. CO potentially exerts its physiological effects by influencing at least three known pathways (Fig. 2 ). By complexation with the heme moiety of the enzyme, CO activates soluble guanylate cyclase (sGC), stimulating the production of cyclic 3':5'guanosine monophosphate (cGMP) [47] . The sGC/cGMP pathway mediates the effects of CO on vascular relaxation, smooth muscle cell relaxation, bronchodilation, neurotransmission, and the inhibition of platelet aggregation, coagulation, and smooth muscle proliferation [48] [49] [50] [51] . Furthermore, CO may cause vascular relaxation by directly activating calcium-dependent potassium channels [52] [53] [54] . CO potentially influences other intracellular signal transduction pathways. The mitogen-activated protein kinase (MAPK) pathways, which transduce oxidative stress and inflammatory signaling (i.e. response to lipopolysaccharide), may represent an important target Possible mechanism(s) of carbon monoxide action Figure 2 Possible mechanism(s) of carbon monoxide action. Endogenous carbon monoxide (CO) arises principally as a product of heme metabolism, from the action of heme oxygenase enzymes, although a portion may arise from environmental sources such as pharmacological administration or accidental exposure, or other endogenous processes such as drug and lipid metabolism. The vasoregulatory properties of CO, including its effects on cellular proliferation, platelet aggregation, and vasodilation, have been largely ascribed to the stimulation of guanylate cyclase by direct heme binding, leading to the generation of cyclic GMP. The anti-inflammatory properties of CO are associated with the downregulation of proinflammatory cytokine production, dependent on the selective modulation of mitogen-activated protein kinase (MAPK), such as the 38 kilodalton protein (p38MAPK). In addition to these two mechanisms, CO may potentially interact with any hemoprotein target, though the functional consequences of these interactions with respect to cellular signaling remain poorly understood. Anti-Platelet Aggregation Anti-Proliferation ? Inhibition of pro-inflammatory cytokine production Modulation of hemoprotein function of CO action [32, 34, 55, 56 ]. An anti-apoptotic effect of CO and its relation to MAPK has recently been described. The overexpression of HO-1 or the exogenous administration of CO prevented tumor necrosis factor α (TNF-α)induced apoptosis in murine fibroblasts [57] . In endothelial cells, the anti-apoptotic effect of CO depended on the modulation of the p38 (38 kilodalton protein) MAPK pathway [34] . The role of the remaining heme metabolites, (i.e. Fe and biliverdin IXα) in the modulation of apoptosis is currently being investigated and is beyond the scope of this review. Recent studies have reported a potent anti-inflammatory effect of CO, involving the inhibition of proinflammatory cytokine production after endotoxin stimulation, dependent on the modulation of p38 MAPK [32] . The clinical relevance of p38 MAPK lies in the possibility of modulating this pathway in various clinical conditions to downregulate the inflammatory response [58] . Oxidative stress arising from an imbalance between oxidants and antioxidants plays a central role in the pathogenesis of airway disease [59] . In lung tissue, HO-1 expression may occur in respiratory epithelial cells, fibroblasts, endothelial cells, and to a large extent in alveolar macrophages [41, 60, 61] . HO-1 induction in these tissues, in vitro and in vivo, responds to common causes of oxidative stress to the airways, including hyperoxia, hypoxia, endotoxemia, heavy metal exposure, bleomycin, diesel exhaust particles, and allergen exposure [4, 41, 61] . Induction of HO-1 or administration of CO can protect cells from these stressful stimuli [10, 41] . In one of the experiments that best illustrate the protective role of CO in vivo, rats were exposed to hyperoxia (>98% O 2 ) in the absence or presence of CO at low concentration (250 ppm). The CO-treated rats showed increased survival and a diminished inflammatory response to the hyperoxia [11] . As demonstrated in a model of endotoxin-induced inflammation, the protection afforded by CO most likely resulted from the downregulated synthesis of proinflammatory cytokines (i.e. TNF-α, IL-1β) and the upregulation of the anti-inflammatory cytokine interleukin-10 (IL-10) [32] . Furthermore, increases in exhaled CO (E-CO) have been reported in a number of pathological pulmonary conditions, such as unstable asthma, COPD, and infectious lung disease; these increases may reflect increased endogenous HO-1 activity [7] . Elevated carboxyhemoglobin (Hb-CO) levels have also been reported in these same diseases in nonsmoking subjects, where both the E-CO and Hb-CO levels decrease to normal levels in response to therapy [62] . E-CO in humans originates primarily from both systemic heme metabolism, which produces CO in various tissues, and localized (lung) heme metabolism, as a result of the combined action of inducible HO-1 and constitutive HO-2 enzymatic activity. Endogenously produced or inspired CO is eliminated exclusively by respiration [63] . Elevation of E-CO may also reflect an increase in exogenous sources such as smoking or air pollution. In addition to changes in environmental factors, elevations of E-CO in lung diseases may reflect an increase in blood Hb-CO levels in response to systemic inflammation, as well as an increase in pulmonary HO-1 expression in response to local inflammation [9, 62, 64] . The diagnostic value of measuring E-CO remains controversial due to many conflicting reports (i.e. some reports indicate differences in E-CO measurements between disease activity and controls, and some reports do not). The possible explanations for these discrepancies include large differences in patient populations and in the methods used for measuring E-CO, and undefined corrections for background levels of CO. Furthermore, remarkable differences arise between studies in the magnitude of the E-CO levels in the control groups as well as in treated or untreated asthma patients. When active or passive smoking occurs, or in the presence of high background levels of CO, the measurement of E-CO is not particularly useful for monitoring airway inflammation. In patients who smoke, E-CO can be used only to confirm the smoking habit [65, 66] . Comparable to the beginning era of measurements of exhaled NO, a standardization in techniques and agreement on background correction should be reached for E-CO measurements, to allow proper conclusions to be drawn in this area of investigation. Asthma, a form of allergic lung disease, features an accumulation of inflammatory cells and mucus in the airways, associated with bronchoconstriction and a generalized airflow limitation. Inflammation, a key component of asthma, involves multiple cells and mediators where an imbalance in oxidants/antioxidants contributes to cell damage. Several pathways associated with oxidative stress may participate in asthma. For example, the redox-sensitive transcription factors NF-κB and AP-1 control the expression of proinflammatory mediators [59, [67] [68] [69] . In light of the potential protective effects of HO-1/CO on inflammatory processes, the study of HO-1 in asthma has gained popularity. In a mouse model of asthma, HO-1 expression increased in lung tissue in response to ovalbumin aerosol challenge, indicating a role for HO-1 in asthma [70] . In a similar model of aeroallergen-induced asthma in ovalbumin-sensitized mice, exposure to a CO atmosphere resulted in a marked attenuation of eosinophil content in bronchoalveolar lavage fluid (BALF) and downregulation of the proinflammatory cytokine IL-5 [10] . This experiment showed that exogenous CO can inhibit asthmatic responses to allergens in mice. Recent human studies have revealed higher HO-1 expression in the alveolar macrophages and higher E-CO in untreated asthmatic patients than in healthy nonsmoking controls [71, 72] . Patients with exacerbations of asthma and patients who were withdrawn from inhaled steroids showed higher E-CO levels than steroid-treated asthmatics or healthy controls [73] . Higher levels of E-CO may also occur in children with persistent asthma than in healthy controls [74] . E-CO levels may correlate with functional parameters such as peak expiratory flow rate. A low rate in asthma exacerbations correlated with high E-CO, whereas normalization of the rate with oral glucocorticoid treatment resulted in a reduction of E-CO [75] . Furthermore, increased E-CO was associated with greater expression of HO-1 in airway alveolar macrophages obtained by induced sputum in untreated asthmatic patients than in controls. These asthma patients also showed higher bilirubin levels in the induced sputum, indicating higher HO activity [71] . Furthermore, patients with asthma show an increased Hb-CO level at the time of exacerbation, with values decreasing to control levels after oral glucocorticoid treatment [62] . In human asthmatics, E-CO and airway eosinophil counts decreased in response to a one-month treatment with inhaled corticosteroids [73] . In direct contrast to such studies promoting E-CO as a useful noninvasive tool for monitoring airway inflammation, other studies reported no difference in E-CO levels of asthma patients versus healthy controls, or between patients with stable and unstable asthma. In one such report, no further change in E-CO occurred in asthma patients after a one-month treatment of inhaled corticosteroids, despite observed decreases in airway eosinophil content and bronchial responsiveness to metacholine [76] . A recent study accentuates this finding in asthma excerbations, where no decrease in E-CO of children with asthma could be detected after oral prednisolone treatment [77] . In human allergic responses, results on elevation of E-CO are also inconclusive. A clear elevation of E-CO after allergen exposure occurred in patients with asthma during the late response, and during the early response immediately after the inhalation [78] . However, another report showed that no elevation of E-CO occurred in allergen-induced asthma within 48 hours after allergen challenge [79] . Finally, increases in E-CO were measured in allergic rhinitis, correlating with seasonal changes in exposure to allergen (pollen) [80] . Airway inflammation plays an important role in the development of COPD, characterized by the presence of macrophages, neutrophils, and inflammatory mediators such as proteinases, oxidants, and cytokines. Further-more, the inflammatory consequences of chronic microbiological infections may contribute to the progression of the disease. The current paradigm for the pathogenesis of COPD involves imbalances in protease/antiprotease activities and antioxidant/pro-oxidant status. Proteases with tissue-degrading capacity, (i.e. elastases and matrix metalloproteinases), when insufficiently inhibited by antiproteases, can induce tissue damage leading to emphysema. Oxidants that supersede cellular antioxidant defenses can furthermore inactivate antiproteases, cause direct injury to lung tissue, and interfere with the repair of the extracellular matrix. Smoking plays an important role in both hypotheses. Cigarette smoke will act primarily on alveolar macrophages and epithelial cells, which react to this oxidative stress by producing proinflammatory cytokines and chemokines and releasing growth factors. Nevertheless, smoking cannot be the only factor in the development of COPD, since only 15-20% of smokers develop the disease [81, 82] . Exposure to reactive oxygen species (from cigarette smoke or chronic infections) and an imbalance in oxidant/antioxidant status are the main risk factors for the development of COPD. To defend against oxidative stress, cells and tissues contain endogenous antioxidant defense systems, which include millimolar concentrations of the tripeptide glutathione (GSH). A close relation exists between GSH concentration and HO-1, whereby depletion of GSH augments the transcriptional regulation of HO-1 by oxidants, suggesting that the HO-1/CO system acts as a secondary defense against oxidative stress [83] [84] [85] [86] . Accumulating clinical evidence suggests that HO-1/ CO may also play an important part in COPD. Alveolar macrophages, which produce a strong HO-1 response to stimuli, may represent the main source of CO production in the airways [60, 64] . Patients with COPD have displayed higher E-CO than healthy nonsmoking controls [87] . Furthermore, much higher levels of HO-1 have been observed in the airways of smokers than in nonsmokers [64] . Among subjects who formerly smoked, patients with COPD have lower HO-1 expression in alveolar macrophages than healthy subjects [88] . A microsatellite polymorphism that is linked with the development of COPD may occur in the promoter region of HO-1, resulting in a lower production of HO-1 in people who have the polymorphism. Thus, a genetically dependent downregulation of HO-1 expression may arise in subpopulations, possibly linked to increased susceptibility to oxidative stress [89] [90] [91] . Future studies on both genetic predisposition and possible therapeutic modalities will reveal the involvement of the HO-1/CO system in COPD. Cystic fibrosis (CF) involves a deposition of hyperviscous mucus in the airways associated with pulmonary dysfunc-tion and pancreatic insufficiency, which may be accompanied by chronic microbiological infections. E-CO readings were higher in untreated versus oral-steroidtreated CF patients [92] . Furthermore, E-CO increased in patients during exacerbations of CF, correlating to deterioration of the forced expiratory volume in one second (FEV 1 ), with normalization of the E-CO levels after treatment [93] . E-CO levels may correlate with exhaled ethane, a product of lipid peroxidation that serves as an indirect marker of oxidative stress. Both E-CO and exhaled ethane were higher in steroid-treated and untreated CF patients than in healthy controls [94] . E-CO was higher in children with CF than in control patients. In addition to the inflammatory and oxidative stress responses to continuous infectious pressure in these patients, E-CO may possibly respond to hypoxia. E-CO increased further in CF children following an exercise test, and correlated with the degree of oxyhemoglobin desaturation, a finding suggestive of an increased HO-1 expression in CF patients during hypoxic states induced by exercise [95] . In patients with pneumonia, higher Hb-CO levels can be measured at the onset of illness, with values decreasing to control levels after antibiotic treatment [62] . E-CO levels were reported to be higher in lower-respiratory-tract infections and bronchiectasis, with normalization after antibiotic treatment [96, 97] . Furthermore, E-CO levels in upper-respiratory-tract infections were higher than in healthy controls [74, 80] . The relationship between higher measured E-CO in these infectious states and higher Hb-CO levels cannot be concluded from these studies. The role of HO-1 in the development of interstitial lung disease remains undetermined. Comparative immunohistochemical analysis has revealed that lung tissue of control subjects, patients with sarcoidosis, usual interstitial pneumonia, and desquamative interstitial pneumonia, all showed a high expression of HO-1 in the alveolar macrophages but a weak expression in the fibrotic areas [98] . The antiproliferative properties of HO-1 suggest a possible beneficial role in limiting fibrosis; however, this hypothesis is complicated by a newly discovered relation between IL-10 and HO-1. IL-10 produced by bronchial epithelial cells promotes the growth and proliferation of lung fibroblasts [99] . HO-1 expression and CO treatment have been shown to increase the production of IL-10 in macrophages following proinflammatory stimuli [32] . Conversely, IL-10 induces HO-1 production, which is apparently required for the anti-inflammatory action of IL-10 [100] . A recent report clearly shows the suppression of bleomycin-induced pulmonary fibrosis by adenovirus-mediated HO-1 gene transfer and overexpression in C57BL/6 mice, involving the inhibition of apoptotic cell death [101] . Overall, more research is needed to elucidate the mechanisms of HO-1 in interstitial lung disease and its possible therapeutic implications. HO-1 action may be of great importance in solid tumors, an environment that fosters hypoxia, oxidative stress, and neovascularization. HO-1 may have both pro-and antagonistic effects on tumor growth and survival. HO-1 and CO cause growth arrest in cell-culture systems and thus may represent a potential therapeutic modality in modulating tumor growth [16] . The overexpression of HO-1 or administration of CO in mesothelioma and adenocarcinoma mouse models resulted in improved survival (>90%) as well as reduction in tumor size (>50%) [17] . Furthermore, HO-1 expression in oral squamous cell carcinomas can be useful in identifying patients at low risk of lymph node metastasis. High expression of HO-1 was detected in groups without lymph node metastasis in this report [102] . In contrast to growth arrest, HO-1 may protect solid tumors from oxidative stress and hypoxia, possibly by promoting neovascularization. In one study, zinc protoporphyrin, a competitive inhibitor of HO-1 enzyme activity, suppressed tumor growth [103] . CO may represent a critical mediator of the body's adaptive response to hypoxia, a common feature in pulmonary vascular disease [104] . Since CO can modulate vascular tone by inducing cGMP and large, calcium-dependent potassium channels, HO-1 and CO probably play important roles in pulmonary vascular diseases [54] . A NOmediated HO-1 induction occurred in the hepatopulmonary syndrome during cirrhosis, associated with enhancement of vascular relaxation [105] . In portopulmonary hypertension, elevated levels of cGMP and inducible nitric oxide synthase (iNOS) expression in the vascular endothelium, and HO-1 expression in macrophages and bronchial epithelium have been described [106] . In transgenic mice models, ho-1 -/and ho-1 +/+ mice did not differ in their development of pulmonary hypertension following chronic hypoxia treatment, despite the development of right ventricular dilation and right myocardial infarction in ho-1 -/mice [107] . The preinduction of HO-1 protein with chemical inducers, however, prevented the development of pulmonary hypertension in the rat lung as a consequence of chronic hypoxia treatment [108] . Transgenic mice overexpressing HO-1 in the lung were resistant to hypoxia-induced inflammation and hypertension [109] . Further research is needed to elucidate the potential role of HO-1 and CO in primary human lung vascular diseases such as primary pulmonary hypertension. Supplemental oxygen therapy is often used clinically in the treatment of respiratory failure. Exposure to high oxygen tension (hyperoxia) may cause acute and chronic lung injury, by inducing an extensive inflammatory response in the lung that degrades the alveolar-capillary barrier, leading to impaired gas exchange and pulmonary edema [110, 111] . Hyperoxia-induced lung injury causes symptoms in rodents that resemble human acute respiratory distress syndrome [112] . Hyperoxia induced HO-1 expression in adult rats but apparently not in neonatal rats, in which the expression and activities of HO-1 and HO-2 are developmentally upregulated during the prenatal and early postnatal period [113] . Both HO-1 and HO-2 potentially influence pulmonary adaptation to high O 2 levels. In one example, the adenoviral-mediated gene transfer of HO-1 into rat lungs protected against the development of lung apoptosis and inflammation during hyperoxia [114] . In vitro studies showed that the overexpression of HO-1 in lung epithelial cells or rat fetal lung cells caused growth arrest and conferred resistance against hyperoxia-induced cell death [15, 16] . An oxygen-tolerant variant of hamster fibroblasts that moderately overexpressed HO-1 in comparison with the parent line resisted oxygen toxicity in vitro. The treatment of this oxygen-tolerant strain with HO-1 antisense oligonucleotides reduced the resistance to hyperoxia. In contrast, additional, vector-mediated, HO-1 expression did not further increase oxygen tolerance in this model [115] . In vivo studies with gene-deleted mouse strains have provided much information on the roles of HO-1 and HO-2 in oxygen tolerance. Dennery et al. demonstrated that heme oxygenase-2 knockout mice (ho-2 -/-) were more sensitive to the lethal effects of hyperoxia than wild-type mice [116] . In addition to the absence of HO-2 expression, however, the mice displayed a compensatory increase in HO-1 protein expression, and higher total lung HO activity. Thus, in this model, the combination of HO-2 deletion and HO-1 overexpression resulted in a hyperoxiasensitive phenotype. Recent studies of Dennery et al. have shown that HO-1-deleted (ho-1 -/-) mice were more resistant to the lethal effects of hyperoxia than the corresponding wild type [117] . The hyperoxia resistance observed in the ho-1 -/strain could be reversed by the reintroduction of HO-1 by adenoviral-mediated gene transfer [117] . In contrast, mouse embryo fibroblasts derived from ho-1 -/mice showed increased sensitivity to the toxic effects of hemin and H 2 O 2 and generated more intracellular reactive oxygen species in response to these agents [118] . Both ho-1 -/-and ho-2 -/strains were anemic, yet displayed abnormal accumulations of tissue iron. Specifically, ho-1 -/accumulated nonheme iron in the kidney and liver and had decreased total iron content in the lung, while ho-2 -/mice accumulated total lung iron in the absence of a compensatory increase in ferritin levels [116, 119] . The mechanism(s) by which HO-1 or HO-2 deletions result in accumulation of tissue iron remain unclear. These studies, taken together, have indicated that animals deficient in either HO-1 and HO-2 display altered sensitivity to oxidative stress conditions. Aberrations in the distribution of intra-and extra-cellular iron, may underlie in part, the differential sensitivity observed [116, 117] . Otterbein et al. have shown that exogenous CO, through anti-inflammatory action, may protect the lung in a rat model of hyperoxia-induced lung injury. The presence of CO (250 ppm) prolonged the survival of rats in a hyperoxic (>95% O 2 ) environment, and inhibited the appearance of markers of hyperoxia-induced lung injury (i.e. hemorrhage, fibrin deposition, edema, airway protein accumulation, and BALF neutrophil influx) [11] . Furthermore, in a mouse model, CO inhibited the expression of proinflammatory cytokines (TNF-α, IL-1β, and IL-6) in mice induced by the hyperoxia treatment. Using genedeleted mice, Otterbein and colleagues also observed that the protection afforded by CO in this model, similar to a lipopolysaccharide-induced model of lung injury, depended on the p38 MAPK pathway (Otterbein et al., unpublished observation, as reviewed in [3] ). In direct contrast to these studies, the group of Piantadosi and colleagues reported no significant difference in the hyperoxia tolerance of rats at CO doses between 50 and 500 ppm [120] . In their model, CO did not alter the accumulation of fluid in the airway. Furthermore, CO, when applied in combination with hyperoxia, increased the activity of myeloperoxidase, a marker of airway neutrophil influx. This study also suggested that inhalation of CO (50-500 ppm) did not alter the expression of HO-1 or other antioxidant enzymes such as Manganese superoxide dismutase (MnSOD) in vivo [120] . Furthermore, Piantadosi and colleagues were able to induce oxygen tolerance in rats and HO-1 expression with hemoglobin treatment, but this tolerance also occurred in the presence of HO inhibitors, thereby not supporting a role for HO activity in oxygen tolerance [121] . Although no consensus has been reached as to the protective role of CO inhalation and/or HO-1 induction in hyperoxic lung injury, human studies will be required to show if CO will supersede NO in providing a significant therapeutic benefit in the context of severe lung diseases [122] . While antioxidant therapies have been examined, until now no human studies exist on the role of HO-1 and CO in acute respiratory distress syndrome (ARDS) and bronchopulmonary dysplasia [123] . Lung transplantation is the ultimate and often last therapeutic option for several end-stage lung diseases. After lung transplantation, there remains an ongoing hazardous situation in which both acute and chronic graft failure, as well as complications of the toxic immunosuppressive regimen used (i.e. severe bacterial, fungal, and viral infections; renal failure; and Epstein-Barr-virus-related lymphomas), determine the outcome [124] . The development of chronic graft failure, obliterative bronchiolitis (OB), determines the overall outcome after lung transplantation. OB, which may develop during the first months after transplantation, is the main cause of morbidity and death following the first half-year after transplantation, despite therapeutic intervention. Once OB has developed, retransplantation remains the only therapeutic option available [124, 125] . Little is known about the pathophysiological background of OB. The possible determinants of developing OB include ongoing immunological allograft response, HLADR mismatch, cytomegalovirus infection, acute rejection episodes, organ-ischemia time, and recipient age [125] . OB patients displayed elevated neutrophil counts in the BALF, and evidence of increased oxidant activity, such as increased methionine oxidation in BALF protein and decreases in the ratio of GSH to oxidized glutathione (GSSG) in epithelial lining fluid. [126, 127] . So far, only very limited research data are available on the possible role for HO-1 in allograft rejection after lung transplantation. Higher HO-1 expression has been detected in alveolar macrophages from lung tissue in lung transplant recipients with either acute or chronic graft failure than in stable recipients [128] . The protective role of HO-1 against allograft rejection has been shown in other transplantation models, in which solid organ transplantation typically benefits from HO-1 modulation. A higher expression of protective genes such as HO-1 has been observed in episodes of acute renal allograft rejection [129] . Furthermore, the induction of HO-1 alleviates graft-versus-host disease [130] . Adenoviral-HO-1 gene therapy resulted in remarkable protection against rejection in rat liver transplants [131] . The upregulation of HO-1 protected pancreatic islet cells from Fas-mediated apoptosis in a dose-dependent fashion, supporting an anti-apoptotic function of HO-1 [132, 133] . HO-1 may confer protection in the early phase after transplantation by inducing Th2-dependent cytokines such as IL-4 and IL-10, while suppressing interferon-γ and IL-2 production, as demonstrated in a rat liver allograft model [134] . Beneficial effects of HO-1 modulation have also been described in xenotransplantation models, in which HO-1 gene expression appears functionally associated with xenograft survival [135] . In a mouse-to-rat heart trans-plant model, the effects of HO-1 upregulation could be mimicked by CO administration, suggesting that HOderived CO suppressed the graft rejection [136] . The authors proposed that CO suppressed graft rejection by inhibition of platelet aggregation, a process that facilitates vascular thrombosis and myocardial infarction. HO-1 may also contribute to ischemic preconditioning, a process of acquired cellular protection against ischemia/ reperfusion injury, as observed in guinea pig transplanted lungs [137] . HO-1 overexpression provided potent protection against cold ischemia/reperfusion injury in a rat model through an anti-apoptotic pathway [138, 139] . The induction of HO-1 in rats undergoing liver transplantation with cobalt-protoporphyrin or adenoviral-HO-1 gene therapy resulted in protection against ischemia/ reperfusion injury and improved survival after transplantation, possibly by suppression of Th1-cytokine production and decreased apoptosis after reperfusion [140, 141] . Until now, no reports have addressed E-CO measurements in lung transplantation, where it is possible that differences in E-CO will be found in patients with acute and chronic allograft rejection. The evolution of CO in exhaled breath may serve as a general marker and diagnostic indicator of inflammatory disease states of the lung, though more research will be required to verify its reliability. Increases in exhaled CO presumably reflect changes in systemic and airway heme metabolic activity from the action of HO enzymes. Evidence from numerous in vitro and animal studies indicates that HO-1 provides a protective function in many, if not all, diseases that involve inflammation and oxidative stress. Thus, the exploitation of HO-1 for therapeutic gain could be achieved through the modulation of HO-1 enzyme activity or its up-and downstream regulatory factors, either by gene transfer, pharmacological inducers, or direct application of CO by gas administration or chemical delivery [142] [143] [144] [145] . The CO-releasing molecules (transition metal carbonyls) developed by Motterlini et al. [144] show promise in the pharmacological delivery of CO for therapeutic applications in vascular and immune regulation. The CO-releasing molecules have been shown to limit hypertension in vivo and promote vasorelaxation in isolated heart and aortic rings [144] . Ultimately, the challenge remains in applying the therapeutic potentials of HO-1 to the treatment of human diseases. In vivo models of transplantation have shown that HO-1 gene therapy protects against allograft rejection [129, 134] . Given the toxic therapy that every transplant patient receives, especially after lung transplantation, the field of transplantation medicine may bring the first frontier for human applications of HO-1 gene therapy or exogenous CO administration. The potential use of inhalation CO as a clinical therapeutic in inflammatory lung diseases has also appeared on the horizon. In one promising study, an inhalation dose of 1500 ppm CO at the rate of 20 times per day for a week produced no cardiovascular side effects [146] . Cigarette smoking and CO inhalation at identical intervals produced comparable Hb-CO levels of approximately 5%. The question of whether or not CO can be used as an inhalation therapy will soon be replaced by questions of "how much, how long, and how often?" The fear of administering CO must be weighed against the severe toxicity of the immunosuppressive agents in current use, and the often negative outcome of solid organ transplantation.
9
Technical Description of RODS: A Real-time Public Health Surveillance System
This report describes the design and implementation of the Real-time Outbreak and Disease Surveillance (RODS) system, a computer-based public health surveillance system for early detection of disease outbreaks. Hospitals send RODS data from clinical encounters over virtual private networks and leased lines using the Health Level 7 (HL7) message protocol. The data are sent in real time. RODS automatically classifies the registration chief complaint from the visit into one of seven syndrome categories using Bayesian classifiers. It stores the data in a relational database, aggregates the data for analysis using data warehousing techniques, applies univariate and multivariate statistical detection algorithms to the data, and alerts users of when the algorithms identify anomalous patterns in the syndrome counts. RODS also has a Web-based user interface that supports temporal and spatial analyses. RODS processes sales of over-the-counter health care products in a similar manner but receives such data in batch mode on a daily basis. RODS was used during the 2002 Winter Olympics and currently operates in two states—Pennsylvania and Utah. It has been and continues to be a resource for implementing, evaluating, and applying new methods of public health surveillance.
Unfortunately, conventional public health disease surveillance-which relies on physician and laboratory reporting and manual analysis of surveillance data-is ill equipped for timely detection of such threats. 3 The reportable disease system relies on health care professionals to recognize, diagnose, and report cases and suspected outbreaks to public health officials 4, 5 ; however, it is unlikely that without an event or alert to raise his or her index of suspicion, a physician will attribute the early symptoms and signs of disease in a bioattack victim appropriately and report the case. 6 A key limitation of the current system is that the lone physician is blind to the cases his or her colleagues in a nearby hospital are seeing-knowledge that might lead the physician to consider uncommon diseases more strongly in his or her diagnostic reasoning. Mandatory laboratory reporting 4 is also illequipped for early detection, because it takes time before tests are ordered and specimens are obtained, transported, processed, and resulted. Sufficiently early detection of a biological attack may be accomplished through surveillance schemes that can detect infected individuals earlier in the disease process. For completeness, we note that biosensors are being developed (and deployed) that detect organisms in the air and that this type of detection, if feasible, occurs fundamentally much earlier, because the delay introduced by the incubation period of the disease is eliminated from the surveillance system. 7 However, such approaches face unsolved technical problems in the analysis of contaminated specimens (the norm in air sampling). Biosensors also need to be in the right place-on every person's lapel or every street corner and hallway-to provide complete surveillance coverage. Surveillance methods that can detect disease at an earlier stage are an important research direction for public health surveillance. These methods are generally referred to as syndromic surveillance because they have the goal of recognition of outbreaks based on the symptoms and signs of infection and even its effects on human behavior prior to first contact with the health care system. 8 Because the data used by syndromic surveillance systems cannot be used to establish a specific diagnosis in any particular individual, syndromic surveillance systems must be designed to detect signature patterns of disease in a population to achieve sufficient specificity. For example, it would be absurd to use only the symptom of fever to attempt to establish a working diagnosis of inhalational anthrax in an individual, but it would be very reasonable to establish a working diagnosis of anthrax release in a community if we were to observe a pattern of 1,000 individuals with fever distributed in a linear streak across an urban region consistent with the prevailing wind direction two days earlier. It would be beyond reasonable and, in fact, imperative to establish a working diagnosis of public health emergency if presented with such information. One recent example of a form of syndromic surveillance is drop-in surveillance-the stationing of public health workers in emergency departments (EDs) and special clinics during high-profile events such as the Super Bowl to capture data on patients presenting with symptoms potentially indicative of bioterrorism. The major disadvantage of this approach is the cost of round-the-clock staffing for manual data collection. A less expensive approach-and the one taken in the Realtime Outbreak and Disease Surveillance (RODS) system-is detection based on data collected routinely for other purposes. Examples of such data include absenteeism data, sales of over-the-counter (OTC) health care products, and chief complaints from EDs. 9 The expenses of manual data collection are avoided; however, the data obtained typically are noisy approximations of what could be obtained by direct interviewing of the patient (in the case of individual level data). Both approaches may play complementary roles with current methods of public health surveillance 10-12 by assisting the physician and public health official with a continuously updated picture of the ''health status'' of a population. 13, 14 A focus of our research has been syndromic surveillance from free-text chief complaints routinely collected by triage nurses in EDs and acute care clinics during patient registration. We have deployed this type of surveillance at the 2002 Winter Olympics and in the States of Pennsylvania and Utah. We described a previous version of the RODS system, 12 but the system has undergone considerable subsequent development both architecturally and functionally. This report provides a detailed description of the current version of RODS, an example of a computer-based public health surveillance system that adheres to the National Electronic Disease Surveillance System (NEDSS) specifications of the Centers for Disease Control and Prevention (CDC). 15, 16 Background The role of public health surveillance is to collect, analyze, and interpret data about biological agents, diseases, risk factors, and other health events and to provide timely dissemination of collected information to decision makers. 17 Conventionally, public health surveillance relies on manual operations and off-line analysis. Existing syndromic surveillance systems include the CDC's drop-in surveillance systems, 8 Early Notification of Community-based Epidemics (ESSENCE), 10,18 the Lightweight Epidemiology Advanced Detection and Emergency Response System (LEADERS), 19 the Rapid Syndrome Validation Project (RSVP), 20 and the eight systems discussed by Lober et al. 11 Lober et al. summarized desirable characteristics of syndromic surveillance systems and analyzed the extent to which systems that were in existence in 2001 had those characteristics. 11 A limitation of most systems (e.g., ESSENCE, 10 Children's Hospital in Boston, 11 University of Washington 11 ) was batch transfer of data, which may delay detection by as long as the time interval (periodicity) between batch transfers. For example, a surveillance system with daily batch transfer may delay by one day the detection of an outbreak. Some systems required manual data input (e.g., CDC's dropin surveillance systems, RSVP, 20 and LEADERS 19 ), which is labor-intensive and, in the worst case, requires round-theclock staffing. Manual data input is not a feasible mid-or long-term solution even if the approach is to add items to existing encounter forms (where the items still may be ignored by busy clinicians). A third limitation for existing surveillance systems is that the systems may not exploit existing standards or communication protocols like Heath Level 7 (HL7) even when they are available. The data type most commonly used among surveillance systems is symptoms or diagnoses of patients from ED and/or physician office visits. Other types of data identified in that study include emergency call center and nurse advice lines. Other types of data being used include sales of over-thecounter health care products, prescriptions, telephone call volumes to health care providers and drug stores, and absenteeism. We have conducted studies demonstrating that the free-text chief complaint data that we use correlate with outbreaks. 21, 22 Design Objectives The overall design objective for RODS is similar to that of an early warning system for missile defense; namely, to collect whatever data are required to achieve early detection from as wide an area as necessary and to analyze the data in a way that they can be used effectively by decision makers. It is required that this analysis be done in close to real time. This design objective is complex and difficult to operationalize because of the large number of organisms and the even larger number of possible routes of dissemination all requiring potentially different types of data for their detection, different algorithms, and different time urgencies. For this reason, our focus since beginning the project in 1999 has been on the specific problem of detecting a large-scale outbreak due to an outdoor (outside buildings) aerosol release of anthrax. Additional design objectives were adherence to NEDSS standards to ensure future interoperability with other types of public health surveillance systems, scalability, and that the system could not rely on manual data entry, except when it was done in a focused way in response to the system's own analysis of passively collected data. This report describes RODS 1.5, which was completely rewritten as a Java 2 Enterprise Edition (J2EE) application since the previous publication describing it. RODS 1.5 is multidata type enabled, which means that any time series data can be incorporated into the databases and user interfaces. The deployed RODS system currently displays and analyzes health care delivery site registrations and separately monitors sales of OTC health care products. Overview RODS uses clinical data that are already being collected by health care providers and systems during the registration process. When a patient arrives at an ED (or an InstaCare in Utah), the registration clerk or triage nurse elicits the patient's reason for visit (i.e., the chief complaint), age, gender, home zip code, and other data and enter the data in a registration computer. The registration computer then generates an HL7 ADT (admission, discharge, and transfer) message and transmits it to the health system's HL7 message router (also called an integration engine). There usually is only one message router per health system even if there are many hospitals and facilities. These processes are all routine existing business activities and do not need to be created de novo for public health surveillance. Figure 1 shows the flow of clinical data to and within RODS. The hospital's HL7 message router, upon receipt of an HL7 message from a registration computer, deletes identifiable information from the message and then transmits it to RODS over a secure virtual private network (VPN), or a leased line, or both (during the 2002 Winter Olympics we utilized both types of connections to each facility for fault tolerance). The RODS HL7 listener maintains the connection with the health system's message router and parses the HL7 message as described in more detail below. It then passes the chief complaint portion of the message to a Bayesian text classifier that assigns each free-text chief complaint to one of seven syndromic categories (or to an eighth category, other). The database stores the category data, which then are used by applications such as detection algorithms and user interfaces. Data about sales of OTC health care products are processed separately by the National Retail Data Monitor, which is discussed in detail in another article in this issue of JAMIA. 23 The processing was kept separate intentionally because, in the future, the servers for the National Retail Data Monitor may operate in different physical locations than RODS. The RODS user interfaces can and do display sales of OTC health care products as will be discussed, but other user interfaces can be connected to the National Retail Data Monitor as well. Prior to September 2001, RODS received data only from hospitals associated with the UPMC Health System, and efforts to recruit other hospitals met with resistance. After the terrorist attacks (including anthrax) in the Fall of 2001, other hospitals agreed to participate. Although data in this project are de-identified, certain information such as the number of ED visits by zip code were considered proprietary information by some health systems. Health Insurance Portability and Accountability Act (HIPAA) concerns also were very prominent in the discussions. Data-sharing agreements were executed with every participating health system that addressed these concerns. As an additional precaution, all RODS project members meet annually with University of Pittsburgh council to review obligations and are required to sign an agreement every year stating that they understand the terms of the data-sharing agreements and agree to abide by the terms. RODS began as a research project at the University of Pittsburgh in 1999 and has functioned with IRB approvals since that time. Health care facilities send admission, discharge, and transfer (ADT) HL7 messages to RODS for patient visits in EDs and walk-in clinics. A minimal data set is sent, as shown in Figure 2 , which qualifies as a HIPAA Limited Data Set. 24 Currently the data elements are age (without date of birth), gender, home zip code, and free-text chief complaint. The HL7 listener receives HL7 messages from the message routers located in each health system. The HL7 listener then passes the received HL7 message to the HL7 parser bean, an Enterprise JavaBean (EJB) in the RODS business logic tier. The HL7 parser bean uses regular expressions to parse the fields in an HL7 message. The HL7 parser bean then stores the parsed elements into a database through a managed database connection pool. Although nearly all health systems utilize the HL7 messaging standard, the location of individual data elements in an HL7 message may differ from health system to health system. For example, some care providers' systems record free-text chief complaint in the DG1 segment instead of the PV2 segment of an HL7 message. To resolve this mapping problem, a configuration file written in eXtensible Markup Language (XML), a standard protocol often used to define hierarchical data elements, defines where each of the data elements can be found in the HL7 message. When an HL7 listener starts up, it reads the hospital-dependent configuration file and passes the configuration information to the parser bean. We also use this configuration file to define the database table and field in which the HL7 parser bean should store each data element. This approach is useful because it allows the HL7 data to be stored to an external database. We anticipate that health departments with existing NEDSS or other public health surveillance databases may wish to use just this component of RODS for real-time collection of clinical data. For hospitals that do not have HL7 message routers (two of approximately 60 in our experience to date), RODS accepts ED registration data files through either a secure Web-based data upload interface or a secure file transfer protocol. In general, these types of data transfers are technically trivial and for that reason are used by many groups but do not have the reliability of a HL7 connection (and have very undesirable time latencies). RODS checks the integrity of the data in the HL7 messages that it receives. This processing is necessary because hospital data flows may have undesirable characteristics such as duplicates. RODS identifies and deletes duplicates by using a database trigger that creates a composite primary key before inserting the data. RODS also filters out scheduling messages, which are identified by the fact that they have future admitted date and time. RODS monitors all data feeds to ensure continuous connections with health systems. If RODS does not receive data for six hours, it sends an alert to the RODS administrator and the sending health system's administrator. Because the commercial message routers that hospitals use queue up HL7 messages when encountering networking or system problems, data integrity is preserved. RODS uses an Oracle8i database to store ED registration data. (Oracle, Redwood Shores, CA). To ensure fast response for an online query (e.g., the daily counts of respiratory syndrome in a county for the past six months), we developed a cache For connectivity with the HL7 message routers, we utilize hardware-based routers. The VPN router is a Cisco PIX 501 and the leased-line routers are a pair of Cisco 2600s (Cisco Systems, Inc., San Jose, CA). All of the RODS processes can be run on a single computer, but in our current implementation-serving Pennsylvania F i g u r e 2. Sample HL7 admission, discharge, and transfer (ADT) message from an emergency department. The circled fields are age, gender, home zip code, admitted date and time, and free-text chief complaint, respectively. and Utah as an application service provider-we use five dedicated servers: firewall, database, Web server, a geographic information system (GIS) server, and computation. The processes are written in Java code and can run on most platforms, but here we describe the specific platforms we use to indicate approximate sizing and processing requirements. We developed RODS applications using the Java 2 Enterprise Edition Software Toolkit (J2EE SDK) from Sun Microsystems for cross-platform Java application development and deployment. 26 We followed contemporary application programming practices-a multitiered application consisting of a client tier (custom applications such as HL7 listeners and detection algorithms), business logic tier, database tier, and Web tier. Business logic such as the HL7 parser bean was implemented as Enterprise JavaBeans (EJBs). NEDSS specifies EJB as the standard for application logic. RODS uses Jboss, an opensource J2EE application server, to run all EJBs. 10 The Web tier comprises the graphical user-interface to RODS and uses Java Server Pages (JSP), Java Servlets, and ArcIMS. The database tier was implemented in Oracle 8i. RODS uses a naive Bayesian classifier called Complaint Coder (CoCo) to classify free-text chief complaints into one of the following syndromic categories: constitutional, respiratory, gastrointestinal, neurological, botulinic, rash, hemorrhagic, and other. CoCo computes the probability of each category, conditioned on each word in a free-text chief complaint and assigns a patient to the category with the highest probability. 27 The probability distributions used by CoCo are learned from a manually created training set. CoCo can be retrained with local data, and it can be trained to detect a different set of syndromes than we currently use. CoCo runs as a local process on the RODS database server. CoCo was developed at the University of Pittsburgh and is available for free download at <http://health.pitt.edu/rods/sw>. Over the course of the project, RODS has used two detection algorithms. These algorithms have not been formally field tested because the emphasis of the project to date has been on developing the data collection infrastructure more than field testing of algorithms. The Recursive-Least-Square (RLS) adaptive filter 28 currently runs every four hours, and alerts are sent to public health officials in Utah and Pennsylvania. RLS, a dynamic autoregressive linear model, computes an expected count for each syndrome category for seven counties in Utah and 16 counties in Pennsylvania as well as for the combined counts for each state. We use RLS because it has a minimal reliance on historical data for setting model parameters and a high sensitivity to rapid increases in a time series e.g., a sudden increase in daily counts. RLS triggers an alert when the current actual count exceeds the 95% confidence interval for the predicted count. During the 2002 Olympics we also used the What's Strange About Recent Events (WSARE 1.0) algorithm. 29 WSARE performs a heuristic search over combinations of temporal and spatial features to detect anomalous densities of cases in space and time. Such features include all aspects of recent patient records, including syndromal categories, age, gender, and geographical information about patients. The criteria used in the past for sending a WSARE 1.0 alert was that there has been an increase in the number of patients with specific characteristics relative to the counts on the same day of the week during recent weeks and the p-value after careful adjustment for multiple testing for the increase was #0.05. Version 3.0 of WSARE, which will incorporate a Bayesian model for computing expected counts rather than using unadjusted historical counts currently, is under development. When an algorithm triggers an alert based on the above criteria, RODS sends e-mail and/or page alerts to its users. RODS uses an XML-based configuration file to define users' e-mail and pager addresses. The e-mail version of the alert includes a URL link to a graph of the time series that triggered the alarm with two comparison time series: total visits for the same time period and normalized counts. RODS has a password-protected, encrypted Web site at which users can review health care registration and sales of OTC health care products on epidemic plots and maps. When a user logs in, RODS will check the user's profile and will display data only for his or her health department's jurisdiction. The interface comprises three screens-Main, Epiplot, and Mapplot. The main screen alternates views automatically among each of the available data sources (currently health care registrations and OTC products in Pennsylvania and Utah and OTC sales only for other states). The view alternates every two minutes as shown in Figure 3 . The clinic visits view shows daily total visits and seven daily syndromes for the past week. The OTC data view shows daily sales for five product categories and the total, also for the past week. Users also can set the view to a specific county in a state. If the normalize control box is checked, the counts in the time series being displayed will be divided by (normalized by) the total daily sales of OTC health care products or ED visits for the region. The Epiplot screen provides a general epidemic plotting capability. The user can simultaneously view a mixture of different syndromes and OTC product categories for any geographic region (state, county, or zip code), and for any time interval. The user also can retrieve case details as shown in Figure 4 . The Get Cases button queries the database for the admission date, age, zip code, and chief complaint (verbatim, not classified into syndrome category) of all patients in the time interval and typically is used to examine an anomalous density (spike) of cases. The Download Data button will download data as a compressed comma separated file for further analyses. The Mapplot screen is an interface to ArcIMS, an Internetenabled GIS product developed by Environmental Systems Research Institute, Inc. Mapplot colors zip code regions to indicate the proportion of patients presenting with a particular syndrome. The GIS server also can overlay state boundaries, county boundaries, water bodies, hospital locations, landmarks, streets, and highways on the public health data as shown in Figure 5 . Similar to Epiplot, Mapplot also can display case details for a user-selected zip code. RODS has been in operation for four years and, like most production systems, has acquired many fault-tolerant features. For example, at the software level, HL7 listeners continue to receive messages and temporarily store the messages when the database is off-line. A data manager program runs every ten minutes and, on finding such a cache, it loads the unstored messages to the database when the database is back on-line. In addition, the data manager program monitors and restarts HL7 listeners as necessary. The database uses ''archive log'' mode to log every transaction to ensure that the database can recover from a system failure. The hardware architecture also is fault tolerant. All servers have dual power supplies and dual network cards. All hard drives use Redundant Arrays of Inexpensive Disk configurations. In addition to dual power supplies, all machines are connected to an uninterrupted power supply that is capable of sending an e-mail alert to the RODS administrator when the main power is down. An important component of RODS that currently is used only at the UPMC Health System in Pittsburgh is the Health System Resident Component (HSRC). The HSRC is located within the firewall of a health system and connects directly to the HL7 message router. The HSRC currently receives a diverse set of clinical data from the HL7 message router including culture results, radiology reports, and dictated F i g u r e 3. Health care registrations view in the Main screen of RODS. The Main screen alternates views every 2 minutes among data types available in the public health jurisdiction. The figure shows eight plots of health care registration data-total visits, botulinic, constitutional, gastrointestinal (GI), hemorrhagic, neurological, rash, and respiratory. After 2 minutes, over-the-counter data will be displayed. The Main screen can be used as a ''situation room'' display. emergency room notes. Its purpose is to provide additional public health surveillance functions that would not be possible if it were located outside of the firewall due to restrictions on the release of identifiable clinical data. The HSRC uses patient identifiers to link laboratory and radiology information to perform case detection. In the past, we have used HSRC to monitor for patients with both a gram-positive rod in a preliminary microbiology culture report and ''mediastinal widening'' in a radiology report. The HSRC is a case detector in a distributed outbreak detection system that is capable of achieving much higher specificity of patient diagnostic categorization through access to more information. HSRC also removes identifiable information before transmitting data to the RODS system, a function provided by the health system's message router in other hospitals that connect to RODS. The HSRC at UPMC Health System functions as an electronic laboratory reporting system, although the state and local health departments are not yet ready to receive real-time messaging from the system. Currently, it sends email alerts to the director of the laboratory and hospital infection control group about positive cultures for organisms that are required to be reported to public health in the state of Pennsylvania. 30 It also sends messages to hospital infection control when it detects organisms that cause nosocomial infections. These organisms include Clostridium difficile, methicillin-resistant Staphylococcus aureus, and vancomycin-resistant Enterococcus. We have been able in HSRC to prototype one additional feature, which is a ''look-back'' function that facilitates very rapid outbreak investigations by providing access to electronic medical records to public health investigators as shown in Figure 6 . This feature requires a token that can be passed to a hospital information system that can uniquely identify a patient, and the reason we have prototyped this feature in the HSRC and not in RODS is simply that HSRC runs within the firewall so an unencrypted token can be used. The lookback is accomplished as follows: when a public health user identifies an anonymous patient record of interest (e.g., one of 20 patients with diarrhea today from one zip code), HSRC calls the UPMC Health System Web-based electronic medical record system and passes it the patient identifier. UPMC Health System then requests the user to log in using the UPMC-issued password before providing access to the record directly from its own secure Web site. This approach is not intended to be implemented in HSRC, but rather in the RODS system outside of the firewall of a health system. It is intended to use encrypted identifiers that the health system would decrypt to retrieve the correct record. The HSRC could provide the encryption-decryption service or it could be provided by another data system in the hospital. We estimate that the prevalence of health systems that have Web-based results review in the United States is 30% to 50% and growing so that this approach could very quickly improve the efficiency of outbreak investigations. For these reasons, we have moved to an application service provider model for dissemination in which we encourage state and local health departments to form coalitions to support shared services. We also have been fortunate to have sufficient grant funding from the Commonwealth of Pennsylvania to be able to support these services on an interim basis while sustainable funding models evolve. Our original design objectives for RODS were real-time collection of data with sufficient geographic coverage and sampling density to provide early syndromic warning of a large-scale aerosol release of anthrax. Although we have not achieved all of our initial design objectives, progress has been substantial. The research identified two types of data-freetext chief complaints and sales of OTC health care prod- ucts-that can be obtained in real time or near real time at sampling levels of 70% or higher for most of the United States. These results were obtained through large-scale deployments of RODS in Pennsylvania and Utah and through building the National Retail Data Monitor described in the accompanying article in this issue of JAMIA. The deployments also provided insights about organizational and technical success factors that would inform an effort to scale the project nationally. The project established the importance of HL7 message routers (also known as integration engines) for public health surveillance. HL7 message routers are a mature, highly prevalent technology in health care. We demonstrated that free-text triage chief complaints can be obtained in real time from most U.S. hospitals through message routers and that these data represent early syndromal information about disease. Many other clinical data of value to public health are transmitted using the HL7 standard (e.g., orders for diagnostic tests, especially microbiology tests, reports of chest radiographs, medications, and test results) and can be integrated into RODS or other surveillance systems capable of receiving HL7 messages. As a result of our efforts to disseminate this technology by giving it away, we have learned that most health departments do not have the technical resources to build and maintain real-time electronic disease surveillance systems. Our application service provider model has been much more success-ful, and we now recommend that states form coalitions to share the costs of such services. The project very early identified the need for a computing component to reside within the firewall of a health system, connected to the hospital's HL7 message router. This component would function as a case detector in a distributed public health surveillance scheme linking laboratory and radiology data to increase the specificity of case detection. It has proven very difficult to disseminate this technology, perhaps due to the complexity of the idea. Nevertheless, the threat of bioterrorism has created a need for such technology, and this approach, or something with equivalent function, must be deployed. Adherence to NEDSS architectural standards was an early design objective that we have met. RODS 1.5 closely follows NEDSS architectural, software, messaging, and data specifications. Our success is a strong validation of those standards. We will gain further understanding of the standards as we attempt to use RODS components including HL7 listeners, natural language parsers, message parsers, databases, user interfaces, notification subsystems, and detection algorithms with other NEDSS compliant systems. An ongoing project will use RODS to collect chief complaints and integrate them into the Utah Department of Health's planned NEDSS system. We have demonstrated the ability to rapidly deploy RODS in a special event with the added advantage that the system F i g u r e 6. Look-back function of RODS. The user has selected one patient to investigate using the screen that is in the background and partly hidden by overlap. RODS has logged the user into the results-review function of an electronic medical record and requested that patient's chart, which is shown on the screen in the foreground. persisted after the event. This experience suggests strongly that RODS or similar systems be considered an alternative to drop-in surveillance. Our future plans are to meet our initial design objective to develop early-warning capability for a large, outdoor release of anthrax, especially ensuring that the data and analysis produced by RODS are reviewed by public health. This goal will require improvements in the interfaces and the detection algorithms to reduce false alarms and to vastly improve the efficiency with which anomalies are evaluated by use of multiple types of data, better interfaces, and implementation of the look-back function. We would like to enlarge as quickly as possible the application service provider to include more states and more types of clinical data so that states will be in a position to prospectively evaluate the detection performance from different types of data on naturally occurring outbreaks. Our long-term goals are to add additional disease scenarios to the design objectives such as detection of in-building anthrax release, vector-borne disease, food-borne disease, and a communicable disease such as severe acute respiratory syndrome (SARS). RODS is a NEDSS-compliant public health surveillance system that focuses on real-time collection and analysis of data routinely collected for other purposes. RODS is deployed in two states and was installed quickly in seven weeks for the 2002 Olympics. Our experience demonstrates the feasibility of such a surveillance system and the challenges involved. Outbreaks, emerging infections, and bioterrorism have become serious threats. It is our hope that the front-line of public health workers, astute citizens, and health care workers will detect outbreaks early enough so that systems such as RODS are not needed. However, timely outbreak detection is too important to be left to human detection alone. The notion that public health can operate optimally without timely electronic information is as unwise as having commercial airline pilots taking off without weather forecasts and radar.
10
Conservation of polyamine regulation by translational frameshifting from yeast to mammals
Regulation of ornithine decarboxylase in vertebrates involves a negative feedback mechanism requiring the protein antizyme. Here we show that a similar mechanism exists in the fission yeast Schizosaccharomyces pombe. The expression of mammalian antizyme genes requires a specific +1 translational frameshift. The efficiency of the frameshift event reflects cellular polyamine levels creating the autoregulatory feedback loop. As shown here, the yeast antizyme gene and several newly identified antizyme genes from different nematodes also require a ribosomal frameshift event for their expression. Twelve nucleotides around the frameshift site are identical between S.pombe and the mammalian counterparts. The core element for this frameshifting is likely to have been present in the last common ancestor of yeast, nematodes and mammals.
The ef®ciency of +1 ribosomal frameshifting at a speci®c codon is used as a sensor to regulate polyamine levels in mammalian cells. The frameshifting occurs in decoding the gene antizyme 1, which has two partially overlapping open reading frames (ORFs). Protein sequencing showed that the reading-frame shift occurs at the last codon of ORF1, causing a proportion of ribosomes to enter ORF2 to synthesize a transframe protein (Matsufuji et al., 1995) . ORF2 encodes the main functional domains (Matsufuji et al., 1990; Miyazaki et al., 1992) of antizyme but has no ribosome initiation site of its own. The antizyme 1 protein binds to ornithine decarboxylase (ODC) (Murakami et al., 1992a; Cof®no, 1993, 1994) , inhibits it (Heller et al., 1976) and targets it for degradation by the 26S proteosome without ubiquitylation (Murakami et al., 1992b (Murakami et al., , 1999 . ODC catalyzes the ®rst and usually ratelimiting step in the synthesis of polyamines, conversion of ornithine to putrescine. Putrescine is a substrate for the synthesis of spermidine and spermine. Because of its inhibition of ODC, antizyme 1 is a negative regulator of the synthesis of polyamines. In addition, antizyme 1 is a negative regulator of the polyamine transporter (Mitchell et al., 1994; Suzuki et al., 1994; Sakata et al., 1997) . As discovered by Matsufuji and colleagues (Gesteland et al., 1992) and Rom and Kahana (1994) , increasing polyamine levels elevate frameshifting in decoding antizyme 1 mRNA and so increase the level of antizyme 1. Since antizyme 1 negatively regulates the synthesis and uptake of polyamines, the frameshifting is the sensor for an autoregulatory circuit. A second mammalian paralog of antizyme, antizyme 2, has very similar properties to antizyme 1, including the regulatory frameshifting, but does not stimulate degradation of ODC under certain conditions where antizyme 1 is active (Ivanov et al., 1998a; Zhu et al., 1999; Y.Murakami, S.Matsufuji, I.P.Ivanov, R.F.Gesteland and J.F.Atkins, in preparation) . Just like antizyme 1, antizyme 2 mRNA is ubiquitously expressed in the body but is 16 times less abundant than mRNA of antizyme 1 (Ivanov et al., 1998a) . In addition to antizyme 1 and 2, mammals have a third paralog of the gene, antizyme 3 (also encoded by two ORFs), which is expressed only during spermatogenesis (Ivanov et al., 2000) . Zebra®sh also have multiple antizyme genes, which differ in their expression patterns and activities (Saito et al., 2000) . Numerous studies have addressed the regulation of fungal ODC in response to exogenously added polyamines. In the cases examined, Physarum polycephalum (Mitchell and Wilson, 1983) , Saccharomyces cerevisiae (Fonzi, 1989; Toth and Cof®no, 1999) and Neurospora crassa (Barnett et al., 1988; Williams et al., 1992) , added polyamines, especially spermidine, result in signi®cant repression of ODC activity. The mechanisms of repression seem to vary from fungus to fungus and are apparently different from the mechanism of polyamine-dependent regulation of ODC in higher eukaryotes. In some cases, the existence of an antizyme-like protein has been suggested but has either been disproved, as in the case of N.crassa (Barnett et al., 1988) , or has never been substantiated, as is the case with S.cerevisiae. As expected from their small cationic nature and ability to neutralize negative charges locally, polyamines play key roles in processes ranging from the functioning of certain ion channels (Williams, 1997) , nucleic acid packaging, DNA replication, apoptosis, transcription and translation. The role of polyamines can be complex as illustrated by the transfer of the butylamine moiety of spermidine to a lysine residue to form hypusine in mammalian translation initiation factor eIF-5A, the only known substrate for this reaction (Tome et al., 1997; Lee et al., 1999) . Spermine negatively regulates the growth of prostatic carcinoma cells at their primary site (Smith et al., 1995) , but at later stages of tumor progression it fails to induce antizyme, which correlates with cells becoming refractory to spermine (Koike et al., 1999) . Lack of antizyme function is also important in the early deregulation of cellular proliferation in oral tumors (Tsuji Conservation of polyamine regulation by translational frameshifting from yeast to mammals The EMBO Journal Vol. 19 No. 8 pp. 1907±1917, 2000 ã European Molecular Biology Organization et al., 1998) and probably others. The levels of polyamines are altered in many tumors, and inhibitors of polyamine synthesis are being tested for antiproliferative and cell death effects. The synthesis of ODC varies during the cell cycle in normal cells (Linden et al., 1985; Fredlund et al., 1995) . It is induced by many growth stimuli and is constitutively elevated in transformed cells (Pegg, 1988; Auvinen et al., 1992) with some phosphorylated ODC being translocated to the surface membrane where it is important for mitotic cytoskeleton rearrangement events (Heiskala et al., 1999) . Antizyme is one example of certain mRNA-contained signals that can elevate speci®c frameshifting >1000-fold above the background level of normal translational errors. In addition to antizyme, frameshifting is also involved in the decoding of some bacterial and yeast genes and especially in many mammalian Retroviruses and Coronaviruses, plant viruses and bacterial insertion sequences (Atkins et al., 1999) . The site of frameshifting in both mammalian antizyme 1 and 2 mRNAs is UCC UGA, where quadruplet translocation occurs at UCCU (underlined) to shift reading to the +1 frame, immediately before the UGA stop codon of the initiating frame (Matsufuji et al., 1995; Ivanov et al., 1998a) . For the frameshifting to occur with an ef®ciency of 20% or more, it is important that the 3¢ base of the quadruplet is the ®rst base of a stop codon. Other important features are a pseudoknot just 3¢ of the shift site and a speci®c sequence 5¢ of the shift site (Matsufuji et al., 1995; Ivanov et al., 1998a) . A pseudoknot 3¢ of the shift site is a common stimulator for eukaryotic ±1 frameshifting, but the synthesis of antizyme is the only known case utilizing +1 frameshifting. Comparative analysis of RNA sequences from different organisms is informative about important features and the different options selected by evolution. Since most of the known examples of programmed frameshifting are in viruses or chromosomal mobile elements, the opportunity for comparison of frameshift cassettes in divergent organisms where the time of divergence can be approximated is limited. A start has been made with the frameshifting required for bacterial release factor 2 expression (Persson and Atkins, 1998) , but antizyme provides the ®rst opportunity for such a comparison in eukaryotes. Antizyme genes in genetically tractable lower eukaryotes would be helpful for understanding the functionally important interactions responsible for autoregulatory programmed frameshifting. Identi®cation of an antizyme gene in Schizosaccharomyces pombe A search for DNA sequences encoding protein sequences homologous to Drosophila melanogaster antizyme (Ivanov et al., 1998b) and Homo sapiens antizyme 1 identi®ed the same S.pombe anonymous cDNA clone (DDBJ/EMBL/GenBank accession No. D89228). The similarity is limited (~10% identity, 24% similarity to both human antizyme 1 and D.melanogaster antizyme); however, it is highest in regions that are most highly conserved among the previously identi®ed antizymes ( Figure 1A ). Closer examination of the cDNA nucleotide sequence provided further evidence that it encodes an S.pombe homolog of antizyme. The initiating AUG codon for the ORF that is similar to higher eukaryotic antizymes (ORF2 of those genes) is not the 5¢-most AUG in this cDNA. In fact, there are eight AUGs closer to the 5¢ end. The ®rst or the second AUGs would initiate translation of an ORF (ORF1) that overlaps the longer downstream ORF (ORF2) such that a +1 translational frameshifting event in the overlap would generate a protein product analogous to the products of antizyme genes from higher eukaryotes. Furthermore, the last 12 nucleotides of ORF1 (UGG-UGC-UCC-UGA) are identical to the last 12 nucleotides of mammalian antizyme 1 ORF1s, including the frameshift site. Eleven of these 12 nucleotides are identical to the corresponding regions of all previously identi®ed antizyme genes ( Figure 1B ). Previous experiments with the mammalian frameshift sequence tested in S.pombe have shown that this short 12 nucleotide sequence, by itself, is suf®cient to stimulate measurable levels (up to 0.5%) of +1 frameshifting (Ivanov et al., 1998c) . To con®rm the ORF con®guration of the putative S.pombe antizyme gene, a region corresponding to the two overlapping ORFs plus~80 nucleotides of the 5¢ UTR and 370 nucleotides of the 3¢ UTR, was ampli®ed from both S.pombe genomic DNA and a cDNA library. The sequence of the ampli®ed DNA con®rmed that there are indeed two overlapping ORFs with the deduced con®guration. This sequence (DDBJ/EMBL/GenBank accession No. AF217277) differs from the previously sequenced cDNA clone by three nucleotides (two in the coding region and one in the 3¢ UTR); one changes an alanine codon to proline, another is a silent mutation within a proline codon. Since the sequences from the cDNA library and genomic DNA are identical, we conclude that the differences with clone No. D89228 are most likely due to strain variation. This gene contains no introns within the ampli®ed region. The S.pombe protein was tested for antizyme activity using a gene fusion with glutathione S-transferase (GST). In this construct, ORF1 and ORF2 of antizyme are fused in-frame by deleting the T nucleotide that encodes U of the stop codon of ORF1. This GST±antizyme fusion gene was expressed in Escherichia coli and the protein was puri®ed by af®nity chromatography. ODC inhibitory activity was tested by incubating the recombinant antizyme protein with an S.pombe crude extract and then assaying the mixture for ODC activity. The results ( Figure 2) show that the recombinant protein can inhibit S.pombe ODC. GST alone (1 mg) does not inhibit S.pombe ODC (data not shown). In light of these results, the S.pombe gene will be called S.pombe ODC antizyme (SPA). Interestingly, the S.pombe ODC was also inhibited by mouse antizyme 1 and antizyme 2 (both expressed as GST fusions); however, the yeast fusion protein did not inhibit mouse ODC (data not shown). Deletion and overexpression of SPA Although the effects of overexpression of antizyme on cellular physiology have been tested previously in mammalian cells, the physiological changes associated with complete absence of antizyme activity have not yet been investigated because of the complication of multiple antizymes. The single S.pombe antizyme provides the chance to explore a knockout. SPA deletion strains were I.P. Ivanov et al. generated by replacing the two ORFs of the gene with the ORFs of either URA4 or LEU2 (see Materials and methods). Complete deletion of SPA (both ORFs) did not affect the viability of S.pombe cells in rich (YE) or minimal (MM) media. Temperature had no differential effect on mutant and wild-type cell growth. Similarly, the growth rates, mating ef®ciencies and overall morphology of the knockout strains are apparently indistinguishable from those of wild-type cells (results not shown). In wild-type S.pombe cells the most abundant polyamine is spermidine followed by putrescine ( Figure 3 ). Spermine and cadaverine are found in much smaller amounts. This distribution of polyamine content is very similar to that in other fungi for which polyamine concentrations have been measured (for references, see review by Tabor and Tabor, 1985) . The effect of SPA deletion on cellular polyamine contents was examined in both exponentially growing and stationary phase cells ( Figure 3 ). The cellular concentrations of putrescine, spermidine and cadaverine (but not spermine) were higher in the knockout strains than in wild-type cells. The greatest effect was seen on putrescine and cadaverine content, with smaller effects on spermidine, presumably because eukaryotic ODC activity directly catalyzes decarboxylation of both ornithine and lysine to produce putrescine and cadaverine, respectively (Pegg and McGill, 1979) , but subsequent regulatory events affect homeostasis of spermidine and spermine. The effect of inactivating antizyme on the polyamine contents in exponentially growing cells is modest (<2-fold in all cases). The effect becomes very pronounced in cells in stationary phase with up to 40-and 10-fold increases of putrescine and cadaverine contents, respectively, in the knockout strains. To test overexpression of SPA, two versions of the gene were cloned into pREP3 expression vector behind a strong, thiamine-repressible promoter (nmt1). One had the wild- type SPA sequence while in the second, ORF1 and ORF2 are fused in-frame. SPA wild type and an SPA deletion strain were transformed with each of the overexpression constructs. Derepression of the nmt1 promoter is a gradual process since it requires dilution of the intracellular pool of thiamine (the repressor) through cell division. After 2.5 days of exponential growth under derepressed conditions, yeast strains transformed with either SPA overexpression construct show signi®cant increases in doubling time ( Figure 4A ). The growth inhibition is greater with the construct expressing the in-frame version of SPA and after prolonged incubation (5±7 days); these cells cease growth and accumulate in G 1 as determined bȳ ow cytometry (data not shown). The fact that the inframe overexpression construct, which differs by a single nucleotide from the wild-type construct, confers a more severe phenotype is consistent with the hypothesis that translational frameshifting is required for expression of SPA. The growth phenotype associated with SPA overexpression is only partially relieved by adding 100 mM putrescine to the media (1 mM had no further effect) (data not shown). To see whether the slower growth is correlated with aberrant polyamine levels the polyamine contents of the deletion strain carrying in-frame SPA overexpression vector were measured under derepressed and repressed conditions, in both cases after 2 days of exponential growth ( Figure 4B ). As expected, overexpression of SPA results in signi®cant reduction in the intracellular levels of all four polyamines. After longer (4±5 days) incubation under derepressed conditions, no putrescine and cadaverine can be detected (data not shown). Translational frameshifting during expression of SPA Previously, we developed an assay for measuring antizyme translational frameshifting in both S.cerevisiae (Matsufuji et al., 1996) and S.pombe (Ivanov et al., 1998c) . Brie¯y, the nucleotide sequence to be assayed is inserted between GST and lacZ, such that ORF1 of the assayed sequence is fused in-frame to GST, while ORF2 is fused in-frame to lacZ. b-galactosidase activity provides a measure of frameshifting ef®ciency. To determine whether translational frameshifting occurs in the overlap of ORF1 and ORF2 of SPA, a region of SPA including all but the ®rst codon of ORF1 plus 180 nucleotides downstream of the ORF1 stop codon was tested. +1 frameshifting occurred at 2.2% compared with a construct in which ORF1 and ORF2 are fused in-frame. This result is consistent with +1 frameshifting being crucial for expression of SPA. Previous experiments have shown that the frameshift cassette of mammalian antizyme 1 can direct ef®cient +1 frameshifting when tested in S.pombe. The reverse experiment was conducted here. The SPA gene was translated in vitro in rabbit reticulocyte lysate and its resulting frameshift ef®ciency measured. With no addition of polyamines, frameshifting ef®ciency is~1.5%. Addition of spermidine to the translation mixture to a ®nal concentration of 1 mM results in a 3.7-fold increase in frameshifting to~5.5%, a level even higher than that observed in the endogenous system in vivo (autoradiogram not shown). The observed ef®ciency of frameshifting with the SPA frameshifting cassette in vivo in S.pombe is signi®cantly more than that expected from its limited nucleotide similarity to the antizyme frameshift sites of higher eukaryotes. This prompted a search for additional stimulatory elements within the SPA frameshift cassette. The following experiments were done in a strain carrying deletion of SPA (high polyamines) because it gives higher frameshifting and higher b-galactosidase activity in general; however, we obtained similar ratios for mutant to wild-type frameshifting ef®ciency in a strain with the intact SPA gene. Deleting 5¢ sequences up to the third to last sense codon of ORF1 has little or no effect on frameshifting ef®ciency. Deleting all but the last sense codon (UCC) of ORF1 leads to a 4-to 5-fold reduction in frameshifting ef®ciency ( Figure 5A ). This implies that the conservation of the six nucleotides 5¢ of the UCC-UGA frameshift site is due to their importance for stimulating +1 frameshifting. It also suggests that no additional ORF1 sequences of SPA stimulate the +1 recoding event. The 180 nucleotide 3¢ region was searched for possible structure by computer RNA folding algorithms plus visual inspection. The algorithms predicted several minimal structures in that region. 3¢ deletion constructs (constructs del.3,3¢±81,3¢) tested the importance of any putative structure on the frameshifting ef®ciency. The results ( Figure 5B and C) show that all of these deletions lead to a signi®cant (~10-fold) reduction in +1 frameshifting, indicating the presence of a major 3¢ stimulatory element in the 180 nucleotide region immediately following the frameshift site of SPA. However, the results indicate that none of the putative RNA structures in this region are suf®cient for the activity of this element. Several additional 3¢ deletions delineated the boundaries of this stimulatory element from the frameshift site to 150 and 180 nucleotides downstream (since construct del.150,3¢ stimulates 5.5-fold more +1 frameshifting than del.129,3¢, 150 nucleotides downstream probably contain most of the 3¢ stimulator). In the experiments described above, two of the characteristics of the autoregulatory circuit of mammalian antizyme 1 were con®rmed: SPA inhibition of ODC and the +1 translational frameshifting. The key question left is whether the recoding event is responsive to polyamine levels in cells. As shown above, overexpression of SPA leads to signi®cant reduction of polyamine levels in S.pombe. An SPA + strain was co-transformed with an SPA wild-type overexpressing plasmid (cells overexpressing wild-type SPA grow slowly but continuously) and a construct that monitors the +1 frameshifting from an SPA frameshift sequence. The +1 frameshifting was compared with that in SPA non-overexpressing cells (in both cases frameshifting was measured relative to in-frame control). The results ( Figure 6 ) show a signi®cant reduction (6.5-fold) in frameshifting ef®ciency in SPA-overproducing cells that correlates with a decrease of polyamine content (4.5-fold for putrescine and 3.9-fold for spermidine). This indicates that polyamines modulate the frameshifting ef®ciency of SPA. An alternative but less likely possibility is that SPA overexpression reduces frameshifting because high levels of SPA transcript titrate some factor limiting for frameshifting. The SPA frameshift signals direct 2-fold more frameshifting in Dspa::LEU2 cells (4.4%) than in SPA + cells (in both cases the measurement is done during stationary phase); however, the relatively high standard deviations for both measurements make it dif®cult to draw ®rm conclusions from this particular result. A search of Caenorhabditis elegans expressed sequence tag (EST) sequences with mammalian antizyme 1 sequence identi®ed 20 clones. These sequences could be deconvoluted into a contiguous cDNA sequence. Primers designed on the basis of this sequence were used to PCR amplify and subclone this cDNA from a C.elegans cDNA library. The sequence of the subcloned cDNA was con®rmed (DDBJ/EMBL/GenBank accession No. AF217278); the subsequently released genomic sequence of this C.elegans gene (DDBJ/EMBL/GenBank accession No. AF040659) con®rms our cDNA data. The amino acid sequence deduced from the cDNA sequence revealed that the longer ORF has similarity to previously reported antizyme sequences (overall 27% identity, 39% similarity to human antizyme 1; 19% identity, 34% similarity to Drosophila antizyme). These similarities are higher than that of SPA to these two antizyme genes and again are concentrated in the regions most highly conserved among previously identi®ed antizymes ( Figure 1A ). Just like mammalian antizymes, the longer ORF (ORF2) lacks an appropriate in-frame initiation codon, and expression could be provided by initiation in a short upstream overlapping ORF (ORF1) leading to +1 ribosomal frameshifting in the overlap. The putative C.elegans antizyme frameshift site (the nucleotides proximal to the end of ORF1) has 18 of 26 nucleotides identical to the consensus sequence for antizyme frameshift sites ( Figure 1B) . Frameshifting for expression of C.elegans antizyme was investigated in heterologous systems. Two constructs containing the entire antizyme cDNA, one with the wildtype sequence and one with a single nucleotide deletion that fuses ORF1 to ORF2 in-frame (in-frame control), were transcribed in vitro and the RNA was translated in rabbit reticulocyte lysate. The products were examined by SDS±PAGE (Figure 7) . The main product from both constructs has an apparent M r of 21 kDa, slightly greater than the predicted M r of 17.7 kDa [aberrant, slower than expected, mobility is observed with antizyme proteins from other species (Ivanov et al., 1998a) ]. From the ratio of wild-type to in-frame product, we estimate that the ef®ciency of frameshifting of C.elegans antizyme in reticulocyte lysate is~0.8%, which is somewhat lower than SPA frameshifting in the same system. Addition of spermidine to the translation reactions almost doubles the ef®ciency of frameshifting to~1.5% (the exact numbers are not easy to determine because of dif®culty in de®ning background values). The frameshifting properties of C.elegans antizyme mRNA were also tested in vivo in S.pombe cells. A sequence including all but the ®rst codon of ORF1 plus 180 nucleotides downstream was inserted between GST and lacZ of the PIU-LAC plasmid. Comparison of the b-galactosidase activity of cells (Dspa::LEU2 strain) transformed with the wild-type construct and the in-frame control constructs indicated 3.5% +1 frameshifting. From the frameshifting observed in the heterologous systems, as well as the sequence considerations discussed above, we conclude that expression of this C.elegans gene requires ribosomal frameshifting. Searching the EST database with the newly discovered C.elegans antizyme identi®ed antizyme orthologs in four other nematode species. In two cases (Necator americanus and Haemonchus contortus), the cDNA sequences in the database were suf®cient to make contigs of the complete coding regions. In the other two cases [Onchocerca volvulus (DDBJ/EMBL/GenBank accession No. AF217279) and Pristioncus paci®cus (DDBJ/EMBL/ GenBank accession No. AF217280)] the complete cDNA sequences were obtained by PCR amplifying and sequencing the full genes from cDNA libraries. As with the previously identi®ed eukaryotic antizyme genes, the ORF con®guration of the newly found nematode orthologs implies the necessity for +1 frameshifting for synthesis of full-length protein. The C.elegans antizyme mRNA frameshift site UUU-UGA is unique, differing from the UCC-UGA of previously known antizyme mRNAs. The C.elegans antizyme gene shares this feature with N.americanus and H.contortus but not with P.paci®cus and O.volvulus antizymes. The phylogenetic tree of nematode antizyme protein sequences matches exactly the phylogenetic relationship (Blaxter, 1998) of the nematodes expressing them, indicating that these gene sequences are the result of divergent evolution within the nematode lineage (data not shown). These results also show that the UUU-UGA frameshift site evolved after the last common ancestor of P.paci®cus and C.elegans but before the divergence of C.elegans, N.americanus and H.contortus (probably 450± 500 million years ago). The ability of UUU-UGA sequence to direct +1 frameshifting was further tested in a mammalian system in the context of the mammalian antizyme mRNA (i.e. in the presence of the 3¢ RNA pseudoknot and 5¢ stimulator). A BMV-coat-protein±antizyme 1 gene fusion construct, which has a TCC-TGA to TTT-TGA substitution, was transcribed and then translated in a rabbit reticulocyte lysate. Eleven percent frameshift ef®ciency was seen in the absence of exogenously added polyamines, 2.2 times the ef®ciency seen with the UCC-UGA transcript. The frameshift ef®ciency becomes 18% when 0.6 mM spermidine is added, which is 1.3 times that with the wild type (Matsufuji et al., 1995) . Similar results were obtained in cultured mammalian (Cos7) cells transfected with TTT-TGA mutant construct, the frameshift being higher than that of wild-type construct in both high-and lowpolyamine conditions (our unpublished results). These results demonstrate that the putative C.elegans frameshift site (UUU-UGA) is, if anything, shiftier than UCC-UGA in the antizyme 1 context and is subject to polyamine stimulation. The results presented show that the yeast S.pombe has a homolog of mammalian antizyme. This is the ®rst documented example of antizyme-type regulation of ODC in a lower eukaryote. Deleting SPA from the yeast genome has no detectable effect on viability or any other overt phenotypic effect but, as expected, it results in altered accumulation of polyamines in the cell. Interestingly, the effect is most pronounced in cells in stationary phase, where the knockout cells accumulate up to 40 times more putrescine than wild-type counterparts. This compares with a <2-fold increase of putrescine in exponentially growing cells. A likely explanation for this observation is that the usual rate of ornithine decarboxylation in exponentially growing cells is close to capacity given`normal' concentrations of substrate, enzyme and product. At the same time, all newly synthesized polyamines are continuously diluted through Fig. 6 . Effect of polyamine depletion on SPA +1 frameshifting. Polyamine depletion is achieved by overexpression of the wild-type version of SPA. The same cultures were assayed both for frameshifting and polyamine content. Numbers above columns indicate fold reduction of frameshifting and polyamine content compared with cells that do not overexpress SPA. Antizyme genes in S.pombe and C.elegans cell growth and division at a rate that is almost identical to the rate of maximum capacity synthesis. Cells in stationary phase can no longer dilute newly synthesized polyamines, and more importantly lack an effective antizymeindependent mechanism of shutting off ODC. This suggests that SPA is the primary regulator of ODC activity in S.pombe, not only during cell growth (short term regulation) but also in non-dividing cells (longer term regulation). Overexpression of SPA (5±7 days derepression) leads to complete depletion of intracellular putrescine. This result implies that in S.pombe ornithine decarboxylation is the only source of putrescine synthesis (the pathway from arginine via agmatine is not utilized). The complete depletion of cadaverine in SPA overexpressing cells suggests that ODC is the only enzyme in S.pombe that can decarboxylate lysine, which is also the case in rat tissues (Pegg and McGill, 1979) . It is somewhat perplexing that addition of putrescine to the media leads to only partial relief of the growth phenotype associated with SPA overexpression. There are two likely explanations. (i) Perhaps S.pombe imports putrescine poorly. (ii) Alternatively, like the mammalian system, maybe SPA inhibits not only ODC but also the polyamine transporter. Further experiments will help to distinguish between these two models. It is unclear how widespread the antizyme gene is within the fungal kingdom. We have identi®ed and cloned antizyme homologs from two other ®ssion yeasts (Schizosaccharomyces octosporus and Schizosaccharomyces japonicus) and from two distantly related fungi (Botryotinia fuckeliana and Emericella nidulans) (our unpublished results). The antizyme frameshift site of the latter two fungi has evolved in a unique way different from all other known antizymes, but nevertheless even these two distantly related fungi have conserved the autoregulatory +1 frameshifting. The fact that the yeast S.pombe has an antizyme gene suggests the possibility that the higher eukaryotic metazoans may all have an antizyme gene. The only previously reported antizyme activity in unicellular organisms is from E.coli, but recent analyses suggest that E.coli does not have a true antizyme (Ivanov et al., 1998d) . This makes SPA the ®rst bona ®de antizyme in a unicellular organism. The remarkable similarity of the core sequence important for antizyme frameshifting from S.pombe to humans could be due to convergent or divergent evolution. The near identity of this sequence in worms, Drosophila, Xenopus, zebra®sh and humans argues against convergent evolution, as if antizyme frameshifting arose in a common ancestor perhaps more than one billion years ago. Three cis-acting RNA elements are known to stimulate mammalian antizyme 1 frameshifting. One is a 50 nucleotide sequence immediately 5¢ of the shift site (Matsufuji et al., 1995; our unpublished results) . A second stimulator is the UGA stop codon of ORF1 and the third is an RNA pseudoknot starting 3 nucleotides 3¢ of the UGA stop codon. Among frameshift sites of the previously identi®ed antizymes from mammals all the way to Drosophila, there is substantial similarity in the sequences immediately 5¢ of the shift site. Sixteen of the last 18 nucleotides of ORF1 are completely conserved in these genes. Schizosaccharomyces pombe and C.elegans antizymes have 9 of 9 and 6 of 9 (14 out of 19 in O.volvulus) nucleotides identical to the consensus, respectively. For the 5¢ sequences, generally, the more distantly related two antizymes are, the more the similarity is con®ned to the 3¢ end of that region. Our SPA ORF1 deletion data show that mutation of nucleotides that are part of the 5¢ consensus sequence leads to reduced frameshifting ef®ciency. This is another indication that conservation of nucleotide sequence in this region is because of its importance for stimulating ef®cient +1 frameshifting. It is quite striking that in all antizyme gene sequences identi®ed so far, including a number of unpublished ones, ORF1 ends with a UGA stop codon. This is particularly surprising since any of the other two stop codons can substitute for UGA to stimulate antizyme 1 frameshifting, although slightly less ef®ciently, in vitro (Matsufuji et al., 1995) and in vivo (our unpublished results). The 3¢ pseudoknot that stimulates frameshifting in antizyme 1 is highly conserved in all known vertebrate antizymes, including mammalian antizyme 2 ( Figure 1B) . None of the invertebrate antizyme mRNAs identi®ed so far, including those presented here, has a sequence in the equivalent region that can be simply folded to a comparable RNA structure. However, sequences immediately 3¢ of the frameshift site are conserved between invertebrates and vertebrates. The conservation of this region between Drosophila and the vertebrate counterparts has already been noted (Ivanov et al., 1998b) . The C.elegans antizyme gene contains the sequence YGYCCCYCA (Y = pyrimidine) in this region, which is identical to the consensus. The antizyme genes from the other four nematodes also have a similar sequence ( Figure 1B) . The signi®cance of this similarity is not clear [in fact, sequences in this region appear to play no role in antizyme 1 in vitro frameshifting outside of the RNA pseudoknot context (Matsufuji et al., 1995) ]. Only two examples are known where RNA elements 3¢ of the frameshift site stimulate +1 frameshifting. One is the RNA pseudoknot of mammalian antizyme 1 and the second is a short RNA sequence immediately following the frameshift site of Ty3 (Farabaugh et al., 1993) . Additional examples would be very helpful in deciphering the role such elements play in the mechanism of +1 frameshifting. It is currently not known how many and which of the invertebrate antizyme genes contain 3¢ frameshift stimulators. The results presented here show that an S.pombe 3¢ stimulator enhances frameshifting 10-fold. This stimulator appears completely different from the 3¢ RNA pseudoknot in vertebrates. Our deletion experiments indicate that none of the predicted RNA structures contained within the minimally required 3¢ region [up to 150±180 nucleotides downstream of the frameshift site ( Figure 5C )] are suf®cient to confer the stimulatory effect. The SPA 3¢ stimulator may act directly through sequence or may have an unusual RNA structure involving non-Watson±Crick base pairing. More detailed mutagenesis combined with phylogenetic analysis would be required to discern the nature of the 3¢ stimulator of SPA. The nematode antizymes were analyzed for the presence of possible 5¢ or 3¢ stimulators¯anking the core frameshift site. Computer RNA folding programs did not identify any potentially interesting structure. More importantly, phylogenetic analysis with the ®ve identi®ed nematode antizymes failed to identify any conservation of primary RNA sequence (or for that matter potential secondary structure) outside of the core region that is shared between two or more members. This could indicate that no such extra cis-acting stimulators exist in nematode antizymes or that they are located in a very different place within the mRNA, for example the 3¢ untranslated region (the latter suggestion is not supported by our sequence analysis). A common mechanism for frameshifting is re-pairing of the peptidyl tRNA in the new reading frame. However, an alternative mechanism whereby the peptidyl tRNA merely occludes the ®rst base of the next codon, has been documented for yeast Ty3 frameshifting (Farabaugh et al., 1993) . Results of experiments with some mutants of the mammalian antizyme 1 shift site pointed to an occlusion mechanism (Matsufuji et al., 1995) . However, the mechanism with the wild-type, UCC-UGA, shift site is not clear. For C.elegans antizyme the UUU-UGA sequence would be an obvious candidate for a re-pairing since Phe-tRNA could pair perfectly with UUU in both frames. But with UCC-UGA the Ser-tRNA ®rst reading UCC could at best pair two out of three with CCU. This important problem warrants further investigation. The frameshift ef®ciency of SPA frameshift site is lower than that observed with mammalian antizyme 1 even when both are tested in the same organism (S.pombe) [for the frameshift ef®ciency of antizyme 1 cassette in S.pombe, see Ivanov et al. (1998c) ]. It is possible that the observed ef®ciencies for S.pombe antizyme are arti®cially low because the constructs do not include all the cis-acting stimulatory elements. On the other hand there is no reason why a lower level of frameshifting does not correctly re¯ect the evolved balance with the other characteristics of the complex system such as relative protein stabilities. Like other core cellular processes, the antizyme polyamine regulatory scheme is conserved from yeast S.pombe to human. It is not obvious why this very special mechanism is so exquisitely preserved over vast evolutionary time. Perhaps there is another whole aspect to the system that our experiments do not yet detect. From this viewpoint it would seem very important to exploit the genetics systems of S.pombe and C.elegans to understand more thoroughly the physiological effects of perturbing the antizyme system. The SPA gene was ampli®ed using the following primers: 5¢-CAAAACAAGTTTTCATTATTGGTTTTTTTTAAATCAATCCCC (sense) and 5¢-CGTAAATCCAATCTAAATTTAATCTTCAACTAA-ATCATGAAAAGCCTC (antisense). The S.pombe cDNA library used as a template in the ampli®cation was kindly provided by R.Rowley (University of Utah). The C.elegans antizyme gene was ampli®ed using the following primers: 5¢-CCCAGGAATTCCTCGAGTATTTTGA-GTATAATTTTAC (sense) and 5¢-CGGCCGCTCGAGTTAGACCTT-GTAGCTCATGATG (antisense). This same ampli®ed DNA was used to make the constructs for in vitro transcription and translation of C.elegans antizyme by cloning it into pTZ18U plasmid using the SacI and HindIII sites incorporated in the two primers. The in-frame construct was made using a two-step PCR. The cDNA sequences of O.volvulus and P.paci®cus antizyme genes were obtained by performing 5¢ and 3¢ RACE PCR with cDNA libraries, which were kindly provided by Ralf Sommer (P.paci®cus) and Susan Haynes (O.volvulus). The SPA overexpression constructs were made by amplifying the gene with the primers 5¢-GCATCCGAATTCCCAAATCCAAGCATCATACGCC (sense) and 5¢-GCATCCGGATCCGCCAGTGTTCTTACTTTGAGA-TGC (antisense), and then inserting BamHI-digested product between the MscI and BamHI sites of pREP3 plasmid. The in-frame construct was made by two-step PCR and subsequently all in-frame SPA constructs described below were made by one-step PCR using this plasmid's DNA as a template. To make the constructs for frameshift assays in S.pombe, DNA fragments with a given nucleotide length (as described in the main text), were ampli®ed from both the SPA and C.elegans antizyme constructs described above. These fragments were then cloned between the KpnI and BstEII sites of PIU-LAC plasmid (Ivanov et al., 1998c) . The PCR primers included an`AC' spacer between the 5¢ cloning site (BstEII) and the antizyme sequences in order to correct the reading frame. The in vivo frameshifting assays in S.pombe (strains ura4-D18 leu1-32 ade6-M216 h ± and Dspa::LEU2 ura4-D18 leu1-32 ade6-M216 h ± ) were done as described (Ivanov et al., 1998c) . The plasmid for GST±SPA expression was made by PCR amplifying SPA (all but the ®rst codon of ORF1 through the downstream ORF2) from an in-frame template and cloning the product into the EcoRI and XhoI restriction sites of pGEX-5X-3 plasmid. The antizyme frameshift site in the BMV-coatantizyme fusion construct (C3NE) (Matsufuji et al., 1995) was mutated with a two-step PCR. To generate the two knockout strains, Dspa::URA4 and Dspa::LEU2, both ORFs of SPA were replaced exactly with the ORF of either URA4 or LEU2. To accomplish this, two pairs of primers ampli®ed URA4 and LEU2 such that 50±60 nucleotides, which normallȳ ank the two ORFs of SPA,¯ank the ORFs of the two genes. The ampli®ed DNA products were gel puri®ed and 2 mg of each were used to electroporate into ura4-D18 leu1-32 ade6-M216 h ± cells. URA + and LEU + transformants were selected by growth on URA ± and LEU ± media, respectively. PCR screen and partial sequencing, with primers¯anking the regions used for the homologous recombination, con®rmed the SPA disruptions. All DNA clones were sequenced with automated sequencing machines (ABI 100). Schizosaccharomyces pombe ODC active crude extracts were prepared as follows: S.pombe (strain 1519, leu1-32, h ± ) provided by R.Rowley was grown to OD 600 0.7 in 50 ml of minimal media + LEU. Ten milligrams of lysing enzymes (Sigma) were added, followed by continued incubation for 30 min at 30°C. Cells were harvested and washed once with cold homogenization buffer [25 mM Tris±HCl pH 7, 0.25 M sucrose, 1 mM dithiothreitol (DTT), 20 mM pyridoxal-5-phosphate, 2 mM EDTA] then resuspended in 0.75 ml of homogenization buffer. Cells were broken open and the lysate was clari®ed by centrifugation at 10 000 r.p.m. for 15 min at 4°C. Extracts were dialyzed overnight in dialysis buffer (25 mM Tris± HCl pH 7.4, 1 mM DTT, 20 mM pyridoxal-5-phosphate, 0.1 mM EDTA). A volume of 25 ml of extract was used for each ODC assay. ODC activity was assayed by measuring the release of 14 CO 2 from L-[1-14 C]ornithine (Amersham) as described (Nishiyama et al., 1988) . Each reaction took 1 h. Pre-incubation of S.pombe extract with 0.1 mM di¯uoromethyl ornithine (DFMO) for 15 min led to >99% inhibition of 14 CO 2 release. The cells were collected by centrifugation, washed twice with 1 ml of phosphate-buffered saline (PBS) and then the pellet was frozen at ±80°C until use. The pellet was resuspended in 0.1 ml of PBS. An aliquot of the suspension was mixed with an equal volume of 8% perchloric acid, vortexed for 1 min, kept on ice for 5 min and centrifuged at 15 000 r.p.m., 4°C for 5 min. Ten microliters of the supernatant were subjected to polyamine analysis using¯uorometry on high-performance liquid chromatography as described previously (Murakami et al., 1989) . Protein concentrations were determined with the BCA protein assay kit (Pierce). The experiments with the BMV-coat-antizyme fusion constructs were performed as described previously (Matsufuji et al., 1995) . All other plasmid DNA templates were prepared using QIAGEN Miniprep Kit and then digested with HindIII. Transcripts for SPA in vitro translation were made from PCR templates that had a T7 promoter incorporated into the PCR primers. Linearized DNA (1 mg) was used as a template for in vitro transcription with Ambion MEGAshortscript TM T7 Kit. The DNasetreated RNAs were recovered and resuspended in 40 ml of RNase-free water. One microliter of each speci®ed transcript suspension was used in each in vitro translation reaction [0.5 ml of 1 mM amino acid mix ±Met, 7 ml of reticulocyte lysate (Promega), 0.5 ml of [ 35 S]Met (Amersham)] to a total volume of 10 ml. The reactions were stopped by adding 1 ml of RNase (10 mg/ml). The frameshift ef®ciencies were quanti®ed as described (Ivanov et al., 1998a) .
11
Heterogeneous nuclear ribonucleoprotein A1 regulates RNA synthesis of a cytoplasmic virus
Heterogeneous nuclear ribonucleoprotein (hnRNP A1) is involved in pre-mRNA splicing in the nucleus and translational regulation in the cytoplasm. In the present study, we demonstrate that hnRNP A1 also participates in the transcription and replication of a cytoplasmic RNA virus, mouse hepatitis virus (MHV). Overexpression of hnRNP A1 accelerated the kinetics of viral RNA synthesis, whereas the expression in the cytoplasm of a dominant-negative hnRNP A1 mutant that lacks the nuclear transport domain significantly delayed it. The hnRNP A1 mutant caused a global inhibition of viral mRNA transcription and genomic replication, and also a preferential inhibition of the replication of defective-interfering RNAs. Similar to the wild-type hnRNP A1, the hnRNP A1 mutant complexed with an MHV polymerase gene product, the nucleocapsid protein and the viral RNA. However, in contrast to the wild-type hnRNP A1, the mutant protein failed to bind a 250 kDa cellular protein, suggesting that the recruitment of cellular proteins by hnRNP A1 is important for MHV RNA synthesis. Our findings establish the importance of cellular factors in viral RNA-dependent RNA synthesis.
Introduction hnRNP A1 is an RNA-binding protein that contains two RNA-binding domains (RBDs) and a glycine-rich domain responsible for protein±protein interaction. It is involved in pre-mRNA splicing and transport of cellular RNAs (reviewed by Dreyfuss et al., 1993) . It is predominantly located in the nucleus, but also shuttles between the nucleus and the cytoplasm (Pin Äol-Roma and Dreyfuss, 1992) . The signal that mediates shuttling has been identi®ed as a 38 amino acid sequence, termed M9, located near the C-terminus of hnRNP A1 between amino acids 268 and 305 (Michael et al., 1995; Siomi and Dreyfuss, 1995; Weighardt et al., 1995) . Yeast two-hybrid screening with M9 as bait resulted in the discovery of a novel transportin-mediated pathway for nuclear import of hnRNP A1 (Pollard et al., 1996; Fridell et al., 1997; Siomi et al., 1997) . The function of the cytoplasmic hnRNP A1 has not been well de®ned. Studies have shown that cytoplasmic and nuclear hnRNP A1 exhibit different RNA-binding pro®les. Cytoplasmic hnRNP A1 is capable of high-af®nity binding to AU-rich elements that modulate mRNA turnover and translation (Hamilton et al., 1993 (Hamilton et al., , 1997 Henics et al., 1994) . It has also been shown to promote ribosome binding to mRNAs by a cap-mediated mechanism, and prevent spurious initiation at aberrant translation start sites (Svitkin et al., 1996) . MHV belongs to the Coronaviridae family of positivesense, single-stranded RNA viruses. MHV replication and transcription occur exclusively in the cytoplasm of infected cells via the viral RNA-dependent RNA polymerase (RdRp) (reviewed by Lai and Cavanagh, 1997) . Initially, the 5¢-most gene 1 of the viral genome is translated into the viral RdRp, which then replicates the viral genomic RNAs into negative-strand RNAs. Subsequently, the negative-strand RNAs are used as templates to transcribe mRNAs, which include a genomic-sized RNA and a nested set of subgenomic mRNA transcripts, all with an identical 5¢ non-translated leader sequence of 72±77 nucleotides and 3¢ co-terminal polyadenylated ends. The subgenomic mRNA transcription of MHV utilizes a unique discontinuous mechanism in which the leader sequence, often derived from a different molecule, is fused to RNAs at the intergenic (IG) sites (i.e. transcription initiation site) to generate subgenomic mRNAs (Jeong and Makino, 1994; Liao and Lai, 1994; Zhang et al., 1994) . The exact mechanism of how these mRNAs are made is still controversial. However, it has been shown that the process of discontinuous RNA transcription is regulated by several viral RNA elements, including the cis-and trans-acting leader RNA Zhang et al., 1994) , IG sequence (Makino et al., 1991) and 3¢-end untranslated sequence (Lin et al., 1996) . There is considerable biochemical evidence suggesting possible direct or indirect interactions between the various RNA regulatory elements. hnRNP A1 binds MHV negative (±)-strand leader and IG sequences (Furuya and Lai, 1993; Li et al., 1997) . Site-directed mutagenesis of the IG sequences demonstrated that the extent of binding of hnRNP A1 to the IG sequences correlated with the ef®ciency of transcription from the IG site (Zhang and Lai, 1995; Li et al., 1997) . Immunostaining of hnRNP A1 showed that hnRNP A1 relocated to the cytoplasm of MHV-infected cells, where viral RNA synthesis occurs (Li et al., 1997) . hnRNP A1 also mediates the formation of a ribonucleoprotein complex containing the MHV (±)-strand leader and IG sequences . These results suggest that hnRNP A1 may serve as a protein mediator for distant RNA regions to interact with each other. Heterogeneous nuclear ribonucleoprotein A1 regulates RNA synthesis of a cytoplasmic virus The EMBO Journal Vol. 19 No. 17 pp. 4701±4711, 2000 ã European Molecular Biology Organization Many cellular proteins, including calreticulin (Singh et al., 1994) , polypyrimidine tract-binding protein (PTB) (Hellen et al., 1994; Wu-Baer et al., 1996) , La protein (Pardigon and Strauss, 1996), Sam68 (McBride et al., 1996) , poly(rC)-binding protein (Parsley et al., 1997) and nucleolin (Waggoner and Sarnow, 1998) , have been implicated to be involved in viral RNA transcription or replication. In addition to MHV, hnRNP A1 has also been reported to interact with human cytomegalovirus immediate-early gene 2 protein, which plays an important role in the regulation of virus replication (Wang et al., 1997) . Furthermore, a yeast protein related to human core RNA splicing factors, Lsm1p, has been shown to be required for the ef®cient replication of brome mosaic virus RNA (Diez et al., 2000) . Recently, Reddy and colleagues demonstrated an inhibition of HIV replication by dominant-negative mutants of Sam68 (Reddy et al., 1999) . However, none of these cellular proteins has been shown experimentally to participate directly in RNA-dependent RNA synthesis. In order to demonstrate the involvement of hnRNP A1 in MHV RNA replication and transcription, we established several DBT cell lines stably expressing either the wildtype (wt) hnRNP A1 or a C-terminus-truncated mutant lacking the M9 sequence and part of the glycine-rich domain. We showed that the mutant hnRNP A1, which was localized predominantly in the cytoplasm, exhibited dominant-negative effects on viral genomic RNA replication and subgenomic mRNA transcription. In contrast, overexpression of the wt hnRNP A1 accelerated the synthesis of all viral RNAs. Our results provide strong evidence that hnRNP A1 is directly or indirectly involved in MHV RNA synthesis in the cytoplasm and that the C-terminal part of the protein is important for its function. This ®nding thus reveals a novel function for hnRNP A1 in the cytoplasm. Characterization of stable cell lines expressing the wt and a C-terminus-truncated hnRNP A1 To explore a potential role for hnRNP A1 in MHV RNA synthesis, we established murine DBT cell lines stably expressing the Flag-tagged wt hnRNP A1 (DBT-A1) or a mutant hnRNP A1, which has a 75 amino acid deletion from the C-terminus (DBT-A1DC) ( Figure 1A ). This mutant lacks part of the glycine-rich domain and the M9 sequence responsible for shuttling hnRNP A1 between the nucleus and the cytoplasm. Immunoblot of the whole-cell lysates with an anti-Flag antibody detected a 34 kDa protein in DBT-A1 cells and a 27 kDa protein in three independent clones of DBT-A1DC cells ( Figure 1B ), whereas no protein was cross-reactive to the anti-Flag antibody in the control cell line stably transfected with the pcDNA3.1 vector (DBT-VEC). The amounts of the Flagtagged wt and truncated hnRNP A1 were comparable in these cell lines. A chicken polyclonal antibody against hnRNP A1 detected two endogenous hnRNP A1 isoforms or hnRNP A1-related proteins in the whole-cell lysates of all of the cell lines. The bottom band (34 kDa) overlaps the Flag-tagged wt hnRNP A1 in DBT-A1 cells. There was only a slight increase in the overall amount of hnRNP A1 in DBT-A1 cells as compared with DBT-VEC cells, indicating that the exogenous hnRNP A1 constituted a small fraction of the total hnRNP A1 in the cells. In DBT-A1DC cells, an additional band of smaller size (27 kDa) corresponding to the mutant hnRNP A1 was detected. The overall expression levels of the exogenous hnRNP A1 and hnRNP A1DC were~3-fold lower than that of the endogenous hnRNP A1 in whole-cell lysates ( Figure 1B ). Similar to the endogenous hnRNP A1 protein (Pin Äol-Roma and Dreyfuss, 1992) , the Flag-tagged wt hnRNP A1 was localized almost exclusively in the nucleus ( Figure 1C ). The mutant hnRNP A1, however, was localized predominantly in the cytoplasm ( Figure 1C) , consistent with the previous ®nding that the M9 nuclear localization signal is necessary to localize hnRNP A1 to the nucleus Weighardt et al., 1995) . Thus, hnRNP A1DC was much more abundant than the endogenous hnRNP A1 in the cytoplasm. The expression levels of the wt or mutant hnRNP A1 varied among individual cells based on immuno¯uorescent staining ( Figure 1C ). The growth rate ( Figure 1D ) and cell morphology (data not shown) were similar among the different cell lines. The effects of overexpression of the wt and mutant hnRNP A1 on syncytium formation and virus production We ®rst assessed the effects of hnRNP A1 overexpression on the morphological changes induced by MHV-A59 infection using several different clones of DBT cell lines. Virus infection was performed at a multiplicity of infection (m.o.i.) of 0.5 to detect the subtle morphological differences among the different cell lines. Syncytia appeared at~7 h post-infection (p.i.) in DBT-VEC cells and~1 h earlier in DBT-A1 cells. At both 8 and 14 h p.i., syncytia were signi®cantly larger and more spread out in DBT-A1 cells than those in DBT-VEC cells ( Figure 2A ). Similar differences were observed with two additional clones of DBT-A1 cells (data not shown). In contrast, no syncytium was observed in three different clones of DBT-A1DC cells, even at 14 h p.i. At 24 h p.i., almost all DBT-A1 cells detached from the plate, but~10±20% of DBT-VEC cells still remained on the plate (data not shown). Remarkably, there was no sign of syncytium formation in DBT-A1DC cells until 24 h after virus infection, when the overall morphology of the cells was similar to that of DBT-VEC cells at 7 h p.i. (data not shown). All of the DBT-A1DC cells were eventually killed at~48 h p.i., suggesting that the inhibition of viral replication was not a result of the disruption of the MHV receptor. Correspondingly, virus production from these cell lines was signi®cantly different. Between 6 and 14 h p.i., virus production from DBT-A1DC cells was 100-to 1000-fold less than that from DBT-VEC and DBT-A1 cells ( Figure 2B ). DBT-A1 cells produced twice as many viruses as those from DBT-VEC cells during that time period. Relocalization of hnRNP A1 during MHV infection MHV RNA synthesis occurs exclusively in the cytoplasm of infected cells. In order for hnRNP A1 to participate directly in viral transcription, it has to be recruited to the site of RNA synthesis. Although hnRNP A1 shuttles between the nucleus and the cytoplasm in normal cells (Pin Äol-Roma and Dreyfuss, 1992) , the level of cytoplasmic hnRNP A1 is very low. We have demonstrated previously that hnRNP A1 relocates from the nucleus to the cytoplasm of MHV-infected cells (Li et al., 1997) . To determine whether the overexpressed hnRNP A1 may participate in MHV RNA synthesis, we performed immunostaining experiments using an anti-Flag antibody to localize Flag-tagged hnRNP A1. In DBT-A1 cells, a signi®cant increase in the cytoplasmic level of hnRNP A1 and a corresponding decrease of nuclear hnRNP A1 were observed in virus-infected cell syncytia at 7 h p.i. ( Figure 3B ); these cells express the MHV nucleocapsid (N) protein in the cytoplasm ( Figure 3A ). By comparison, in the uninfected cells, which did not have N protein staining, hnRNP A1 was predominantly localized to the nucleus (arrow in Figure 3B ). In DBT-A1DC cells, very few cells were stained positive for the MHV N protein at 7 h p.i. ( Figure 3C ). Signi®cantly, the viral N protein was detected only in the cells that were stained weakly or not at all for Flag-hnRNP A1 ( Figure 3D ), suggesting that the expression of a high level of hnRNP A1DC interfered with viral replication. The effects of wt and mutant hnRNP A1 on MHV protein production We further investigated the effects of the wt and mutant hnRNP A1 on the production of MHV structural and nonstructural proteins. Cytoplasmic protein was extracted from infected cell lines at different time points after infection for immunoblot analysis to detect an open reading frame (ORF) 1a product, p22 (Lu et al., 1998) and the N protein. p22 expression in DBT-VEC cells was clearly detected at 6 h p.i. and peaked at~16 h p.i. ( Figure 4A ). In DBT-A1 cells, p22 appeared at 5 h p.i. and peaked at~8 h p.i. In DBT-A1DC cells, no p22 protein was detected until 16 h p.i. Similar patterns of differences were observed for the N protein in these three cell lines. Actin levels in different cell lines remained relatively constant throughout the infection, except that, in DBT-A1 cells, actin was not detected at 16 and 24 h p.i. due to the loss of the dead cells ( Figure 4A ). These results clearly demonstrated that overexpression of the wt hnRNP A1 accelerated viral protein production, whereas expression of the mutant hnRNP A1 delayed it. We also performed immuno¯uorescent staining of the N protein at 7 h p.i. to further con®rm the western blot results. As represented by images shown in Figure 4B , there were more DBT-A1 cells stained positive for the N protein than DBT-VEC cells. Very few cells were found to express the N protein in DBT-A1DC cells. The p22 and N proteins appeared as doublets in some of the lanes of Figure 4A , but the results varied from experiment to experiment. The N protein is known to be phosphorylated (Stohlman and Lai, 1979) . Whether p22 is post-translationally modi®ed is not known. Figure 5A ). DBT-A1 cells showed a signi®cantly higher level of [ 3 H]uridine incorporation, which peaked at~8 h p.i. DBT-A1DC cells did not show any detectable level of incorporation of the radioactivity. These results suggest that hnRNP A1 regulates MHV RNA synthesis. We further assessed the production of genomic and subgenomic MHV RNAs in these cell lines by northern blot analysis. The genomic and the six subgenomic RNA species were detected at 8 h p.i. in both DBT-VEC and DBT-A1 cells; there were signi®cantly higher steady-state levels of all of the RNA species in DBT-A1 cells ( Figure 5B ). In contrast, no viral RNA was detected in DBT-A1DC cells at that time point. At 16 h p.i., MHV RNA levels in DBT-VEC and DBT-A1 cells decreased generally because of the loss of the dead cells, while the smaller subgenomic RNAs became detectable in DBT-A1DC cells. By 24 h p.i., most viral RNA species became detectable in DBT-A1DC cells ( Figure 5B , lane 10), while most of the DBT-A1 cells were dead (lane 9). These results con®rmed that the synthesis of all of the viral RNA species is accelerated by overexpression of the wt hnRNP A1 and delayed by a dominant-negative mutant of hnRNP A1. In this analysis, we also detected an additional RNA species (arrow in Figure 5B ), which was determined to be a defective-interfering (DI) RNA by northern blot analysis using a probe representing the 5¢-untranslated region (without the leader), which is present only in genomic and DI RNAs (data not shown). Interestingly, this DI RNA was inhibited to a greater extent than other RNA species in DBT-A1DC cells. This result suggests that the replication of DI RNAs is more sensitive to the dominant-negative inhibition by cytoplasmic hnRNP A1. To demonstrate further that MHV RNA transcription machinery is defective in cells expressing the mutant hnRNP A1, we studied transcription of an MHV DI RNA, 25CAT, which contains a transcription promoter (derived from the IG sequence for mRNA 7, IG7) and a chloramphenicol acetyltransferase (CAT) reporter gene . CAT activity can be expressed from At 1 h p.i., serum-free medium was replaced by virus growth medium containing 1% NCS and 5 mg/ml actinomycin D. [ 3 H]uridine (100 mCi/ml) was added to the infected cells at 2, 3, 4, 5, 6, 7, 8, 9, 16 and 24 h p.i. After 1 h labeling, cytoplasmic extracts were prepared and precipitated with 5% TCA. The TCA-precipitable counts were measured in a scintillation counter. (B) Northern blot analysis of MHV genomic and subgenomic RNA synthesis in DBT cells. Cytoplasmic RNA was extracted from MHV-A59-infected cells at 8, 16 and 24 h p.i. for northern blot analysis. The naturally occurring DI RNA of MHV-A59 is indicated by an arrow. this DI RNA only if a subgenomic mRNA containing CAT sequences is produced . The 25CAT RNA was transfected into MHV-A59-infected cells 1 h after infection. At 8 h p.i., CAT activity in DBT-A1 cells was signi®cantly higher than that in DBT-VEC cells ( Figure 6A ). On the other hand, CAT activity was very low in DBT-A1DC cells. At 24 h p.i., CAT activity in DBT-A1 cells became slightly lower than that in DBT-VEC cells because of the loss of the dead DBT-A1 cells. The CAT activity in DBT-A1DC was still signi®cantly lower than that in DBT-VEC or DBT-A1 cells. These results established that mRNA transcription from the DI RNA was also inhibited by hnRNP A1DC. The results shown above ( Figure 5B ) also suggest that DI RNA replication is more sensitive to the inhibitory effects of the hnRNP A1 mutant. To con®rm this result, we further studied replication of another DI RNA during serial virus passages. DBT cells were infected with MHV-A59 and transfected with DIssE RNA derived from JHM virus (Makino and Lai, 1989) ; the virus released (P0) was passaged twice in DBT cells to generate P1 and P2 viruses. DBT cells were infected with these viruses, and cytoplasmic RNA was extracted for northern blot analysis using glyoxalated RNA for a better resolution of smaller RNAs. For DBT-A1DC cells, RNA was extracted at 36 h p.i. since viral RNA synthesis was delayed in this cell line. Cells infected with P0 viruses did not yield detectable amounts of DIssE, but contained the naturally occurring A59 DI RNA, whose replication was inhibited more strongly than the synthesis of MHV genomic and subgenomic RNAs in DBT-A1DC cells ( Figures 5B, lanes 8±10 and 6B, lanes 1± 3). However, this A59 DI RNA was not detectable in cells infected with P1 and P2 viruses ( Figure 6B , lanes 4±9). In contrast, DIssE appeared in cells infected with P1 viruses and further increased in cells infected with P2 viruses, indicating that the replication of the smaller DIssE may have an inhibitory effect on the replication of the larger A59 DI RNA (Jeong and Makino, 1992) . Similar to the A59 DI RNA, the replication of DIssE RNA was much more strongly inhibited than that of MHV genomic and subgenomic RNAs in DBT-A1DC cells ( Figure 6B , lanes 6 and 9). Our results thus suggest that MHV DI RNA replication is more dependent on the function of cytoplasmic hnRNP A1. The mechanism of dominant-negative inhibition by the C-terminal deletion mutant of hnRNP A1 To understand the underlying mechanism of the inhibition of MHV RNA transcription by the C-terminal-deletion mutant of hnRNP A1, we ®rst examined the RNA-and protein-binding properties of this mutant protein. Electrophoretic mobility shift assay demonstrated that hnRNP A1DC retained the ability to bind the MHV (±)strand leader RNA and to form multimers with itself, similar to the wt hnRNP A1 (data not shown); this is consistent with the fact that both of its RBDs are intact ( Figure 1A) . Furthermore, UV-crosslinking experiments showed that increasing amounts of puri®ed glutathione S-transferase (GST)±hnRNP A1DC ef®ciently competed with the endogenous hnRNP A1 for the binding of the MHV (±)-strand leader RNA ( Figure 7A ), indicating that the binding of hnRNP A1DC to RNA was not affected. These results suggest that the RNA-binding properties of hnRNP A1DC were intact. We next examined the protein-binding properties of hnRNP A1DC. Since hnRNP A1 has been shown to interact with the N protein, which also participates in MHV RNA synthesis (Compton et al., 1987; Wang and Zhang, 1999) , we ®rst determined whether the dominantnegative mutant of hnRNP A1 retained the ability to interact with the N protein in vitro. GST pull-down assay using various truncation mutants of hnRNP A1 showed that the N protein bound the N-terminal domain (aa 1±163) of hnRNP A1 ( Figure 7B) ; thus, the binding of hnRNP A1DC [equivalent to hnRNP A1(1±245)] to the N protein was not affected. We next examined the in vivo interaction of the wt and mutant hnRNP A1 with an MHV ORF 1a product, p22, which has been shown to co-localize with the de novo synthesized viral RNA (S.T.Shi and The viruses were passaged twice in wt DBT cells to obtain P1 and P2 viruses. Cytoplasmic RNA was extracted from the DBT cells infected with P0, P1 and P2 viruses and treated with glyoxal before electrophoresis and northern blot analysis using a 32 P-labeled (±)-strand mRNA 7 as a probe. The A59 DI RNA and DIssE RNA are indicated by arrows. M.M.C.Lai, unpublished results) and associate with the viral replicase complex (Gibson Bost et al., 2000) . Cytoplasmic extracts from MHV-A59-infected cells were immunoprecipitated with anti-Flag antibody-conjugated beads, followed by western blotting with a rabbit polyclonal antibody against p22. At 8 h p.i., p22 was co-precipitated with the Flag-tagged hnRNP A1 from DBT-A1 cells, whereas no precipitation of p22 was observed in DBT-VEC cells ( Figure 7C ). For DBT-A1DC cells, co-immunoprecipitation was performed at 24 h p.i., when abundant MHV proteins were synthesized. p22 was shown to co-precipitate with hnRNP A1DC, indicating that hnRNP A1DC still formed a complex with the viral polymerase gene product. These results suggest that the ability of hnRNP A1DC to interact with the N and polymerase proteins was not altered. We next investigated whether the mutant hnRNP A1 is de®cient in the interaction with any other cellular proteins in this RNA±protein complex. We labeled proteins in MHV-infected cells or mock-infected cells at different time points after infection and immunoprecipitated with the anti-Flag antibody. Signi®cantly, a cellular protein of 250 kDa was shown to be associated only with the wt hnRNP A1, but not the mutant hnRNP A1 ( Figure 7D ), suggesting that hnRNP A1 binds to this protein through its C-terminal domain. We propose that this cellular protein is another important component of the MHV RNA transcription/replication complex. There is an accumulating body of evidence signifying the importance of cellular factors in RNA synthesis of RNA viruses (reviewed by Lai, 1998) . Previous studies have shown that hnRNP A1 binds to the cis-acting sequences of MHV template RNA and that this interaction correlates with the transcription ef®ciency of viral RNA in vivo (Zhang and Lai, 1995; Li et al., 1997) . In addition, hnRNP A1 is also implicated in viral RNA replication by the recent ®nding that hnRNP A1 interacts with the 3¢-ends of both positive-and negative-strand MHV RNA (P.Huang and M.M.C.Lai, unpublished results). However, hnRNP A1 modulates cytoplasmic viral RNA synthesis the functional importance of hnRNP A1 in viral RNA synthesis has so far not been directly demonstrated. In the present study, we established that MHV RNA transcription and replication were enhanced by overexpression of the wt hnRNP A1 protein, but inhibited by expression of a dominant-negative hnRNP A1 mutant in DBT cell lines. Our results suggest that hnRNP A1 is a host protein involved in the formation of a cytoplasmic transcription/ replication complex for viral RNA synthesis. This represents a novel function for hnRNP A1 in the cytoplasm. Our results indicate that the inhibitory effects on MHV replication exhibited by the dominant-negative mutant of hnRNP A1 were relatively more prominent than the enhancement effects by overexpression of the wt hnRNP A1. This is consistent with the subcellular localization patterns of the wt and mutant hnRNP A1 proteins. The overexpressed exogenous wt hnRNP A1 in DBT-A1 cells was predominantly localized in the nucleus, similar to the endogenous hnRNP A1 ( Figure 1C ). The C-terminal-deletion mutant, however, was localized mainly in the cytoplasm. Thus, the level of hnRNP A1DC was much higher than the endogenous wt hnRNP A1 in the cytoplasm of DBT-A1DC cells, where MHV replication occurs. This result explains why hnRNP A1DC could have a strong dominant-negative inhibitory effect, despite the fact that it was expressed at a lower level than the endogenous hnRNP A1 ( Figure 1B) . The effects of the expression of the wt and mutant hnRNP A1 on virus production ( Figure 2B ), viral protein synthesis ( Figure 4A ) and viral RNA synthesis ( Figure 5A ) correlated with each other. Furthermore, hnRNP A1DC caused not only a global inhibition of genomic RNA replication and subgenomic mRNA transcription, but also a preferential inhibition of at least two DI RNA species. These results suggest that the inhibition of MHV replication by the hnRNP A1 mutant was most likely a direct effect on viral RNA synthesis rather than an indirect effect on other aspects of cellular or viral functions. Since hnRNP A1 binds directly to the cis-acting MHV RNA sequences critical for MHV RNA transcription (Li et al., 1997) and replication (P. Huang and M.M.C.Lai, unpublished results) , it is most likely that hnRNP A1 may participate in the formation of the transcription/replication complex. Indeed, our data show that hnRNP A1 interacts directly or indirectly with the N protein and a gene 1 product, p22, both of which are probably associated with the viral transcription/replication complex (Compton et al., 1987; Wang and Zhang, 1999; Gibson Bost et al., 2000) . hnRNP A1 may participate directly in viral RNA synthesis in a similar role to that of transcription factors in DNAdependent RNA synthesis, e.g. by maintaining favorable RNA conformation for RNA synthesis. Alternatively, hnRNP A1 may modulate MHV RNA transcription or replication by participating in the processing, transport and controlling the stability of viral RNAs. It has been reported that RNA processing of retroviruses, human T-cell leukemia virus type 2 (Black et al., 1995) and HIV-1 (Black et al., 1996) , is altered by the binding of hnRNP A1 to the viral RNA regulatory elements. It is also possible that hnRNP A1 may participate in MHV RNA synthesis indirectly by affecting the production of other host cell proteins, which may, in turn, regulate MHV RNA synthesis. Since hnRNP A1 is a dose-dependent altern-ative splicing factor (Caceres et al., 1994) , even small changes in the intracellular level of hnRNP A1 can alter the splicing of other cellular proteins. Regardless of the mechanism, our study established the importance of cellular factors in viral RNA-dependent RNA synthesis. The transcription from 25CAT RNA was strongly inhibited by the dominant-negative mutant of hnRNP A1, as shown by CAT assays ( Figure 6A ). In addition, the replication of the naturally occurring A59 DI RNA and the arti®cial DIssE RNA was completely abolished ( Figure 6B) . Surprisingly, the replication of MHV DI RNAs suffered a stronger inhibition by the dominantnegative mutant of hnRNP A1 than the synthesis of MHV genomic and subgenomic RNAs, suggesting that DI RNA replication may be more dependent on hnRNP A1. Although DI RNAs contain all of the cis-acting replication signals that are essential for their replication in normal cells (Kim and Makino, 1995) , the small size of DI RNA may cause it to require more hnRNP A1 to maintain a critical RNA structure. It has been shown that different DI RNAs require different cis-acting signals for RNA replication (Kim and Makino, 1995) . Our results demonstrate that the C-terminal domain of hnRNP A1, including the M9 sequence and the glycinerich region, is important for MHV RNA transcription and replication, but the mechanism of the dominant-negative effects of hnRNP A1DC is still not clear. hnRNP A1DC retains the RNA-binding and self-association ability and is capable of binding the viral proteins N and p22, which are associated with the transcription/replication complex. It is possible that hnRNP A1DC is not productive due to its inability to interact with other viral or cellular proteins that are involved in MHV RNA synthesis. We have found a protein of~250 kDa that binds only the wt, but not the mutant hnRNP A1 ( Figure 7D ). It remains to be shown whether this cellular protein is involved in MHV RNA synthesis. In our preliminary study, we found that MHV could replicate in an erythroleukemia cell line, CB3, which was reported to lack detectable hnRNP A1 expression as a result of a retrovirus integration in one allele and loss of the other allele (Ben-David et al., 1992) . Since hnRNP A1 protein is involved in a variety of important cellular functions, including RNA splicing, transport, turnover and translation, it is conceivable that other redundant gene products may substitute for the function of hnRNP A1 in CB3 cells. Indeed, UV-crosslinking assays using CB3 cell extracts detected two proteins comparable to hnRNP A1 in size that could interact with the MHV negative-strand leader RNA (data not shown). These proteins may represent hnRNP A1-related proteins, since many of such hnRNPs exist in the cells (Buvoli et al., 1988; Burd et al., 1989) . Therefore, multiple cellular proteins may have the capacity to be involved in MHV RNA synthesis. Based on previous ®ndings (Kim and Makino, 1995; Zhang and Lai, 1995; Li et al., 1997) and the results from this study, we propose a model for the regulation of transcription/replication of MHV RNA by hnRNP A1. We hypothesize that hnRNP A1 is one of the components of the MHV RNA transcription or replication complex, and the crosstalk between hnRNP A1 and another viral or cellular RNA-binding protein (designated X in Figure 8 ) is essential for MHV replication and transcription. The X protein binds to the C-terminus of hnRNP A1 and cooperates with hnRNP A1 to recruit more proteins to form the transcription or replication complex. The C-terminaldeletion mutant of hnRNP A1 loses the ability to interact with the X protein and to bring it into the initiation complex, resulting in an inhibition of MHV RNA transcription and replication. The residual replication and transcription activities of MHV RNA in the absence of functional hnRNP A1 may be due to a limited af®nity of the X protein to a cis-acting signal that is only present in MHV genomic RNA (site B). On the other hand, DI RNAs may lack this cis-acting signal. When the crosstalk between the X protein and hnRNP A1 is abolished by the dominant-negative mutant of hnRNP A1, the X protein can no longer participate in the formation of the initiation complex, resulting in a complete loss of DI RNA replication. In summary, our data provide direct experimental evidence that hnRNP A1 is involved directly or indirectly in MHV RNA synthesis, probably by participating in the formation of an RNA transcription/replication complex. This ®nding reveals a novel cytoplasmic function for hnRNP A1. Cells and viruses DBT cells, a mouse astrocytoma cell line (Hirano et al., 1974) , were cultured in Eagle's minimal essential medium (MEM) supplemented with 7% newborn calf serum (NCS) and 10% tryptone phosphate broth. MHV strain A59 (Robb and Bond, 1979) was propagated in DBT cells and maintained in virus growth medium containing 1% NCS. Plasmid construction and establishment of DBT stable cell lines The cDNA of the murine hnRNP A1 gene was ampli®ed by RT±PCR using RNA extracted from DBT cells and a set of primers representing the 5¢-and 3¢-ends of hnRNP A1-coding region, and cloned into pcDNA3.1 (Invitrogen, Carlsbad, CA). The 8 amino acid Flag tag was attached to the N-terminus of hnRNP A1 by including the Flag tag in the forward PCR primer. The truncated hnRNP A1DC was similarly constructed using a PCR-ampli®ed fragment that represents hnRNP A1 (aa 1±245). For the establishment of permanent DBT cell lines, pcDNA3.1 alone or the plasmid containing the Flag-tagged hnRNP A1 or hnRNP A1DC was transfected into 60% con¯uent DBT cells using DOTAP according to the manufacturer's instructions (Boehringer Mannheim, Indianapolis, IN) . After 4 h, the transfected cells were selected in DBT cell medium containing 0.5 mg/ml Geneticin (G418) (Omega Scienti®c, Tarzana, CA) for 10 days. Single colonies were then collected and cultured individually for 10 additional days before screening for the expression of Flag-tagged proteins. The polyclonal rabbit antibody against p22 was a gift from Dr Susan C.Baker at Loyola University, IL. The chicken polyclonal antibody against hnRNP A1 was produced by Aves Labs, Inc. (Tigard, OR) by immunizing chickens with the puri®ed mouse hnRNP A1 protein expressed in bacteria. The polyclonal anti-Flag antibody was purchased from Af®nity Bioreagents (Golden, CO). The goat polyclonal antibody against actin was obtained from Santa Cruz Biotechnology (Santa Cruz, CA). The mouse monoclonal antibody against the N protein has been described previously (Fleming et al., 1983) . Examination of growth rate of permanent DBT cells Equal numbers (1 3 10 5 ) of DBT-VEC, DBT-A1 and DBT-A1DC cells were plated in 10-cm culture plates and maintained in culture medium for 4 days. Cells were trypsinized, stained with Trypan Blue (Gibco-BRL, Grand Island, NY) and counted at 24-h intervals with a hemacytometer (Hausser Scienti®c, Horsham, PA). Plaque assay DBT cells in 10-cm plates were infected with MHV-A59 at an m.o.i. of 2. After 1 h for virus adsorption, the cells were washed three times with serum-free MEM, which was then replaced with virus growth medium containing 1% serum. At 1, 6, 8, 10, 14 and 24 h p.i., 1 ml of medium was taken from each plate for plaque assay. [ 3 H]uridine labeling of MHV RNA Cells plated in 6-well plates were infected with MHV-A59 at an m.o.i. of 2. At 1 h p.i., 5 mg/ml actinomycin D was added to the virus growth medium to inhibit cellular RNA synthesis. To label newly synthesized MHV RNA, 100 mCi/ml of [ 3 H]uridine (NEN, Boston, MA) were added to the medium at hourly intervals. After 1 h of labeling, the cells were washed twice in ice-cold PBS and scraped off the plates in 1 ml of PBS. The cells were then collected by centrifugation and incubated in 200 ml of NTE buffer (150 mM NaCl, 50 mM Tris pH 7.5, 1 mM EDTA) containing 0.5% NP-40, 0.5 mM dithiothreitol (DTT) and 400 U/ml of RNasin on ice for 15 min. After centrifugation, 5 ml of the cytoplasmic extract were spotted on a piece of 3 mm paper and incubated with 5% trichloroacetic acid (TCA). The radioactivity remaining on the 3 mm paper was measured in a scintillation counter. Northern blot analysis DBT cells were infected with MHV-A59 at an m.o.i. of 2. At 8, 16 and 24 h p.i., cytoplasmic extract was prepared as described above and subjected to phenol/chloroform extraction and ethanol precipitation to purify cytoplasmic RNA. Approximately 10 mg of RNA were separated by electrophoresis on a 1.2% formaldehyde-containing agarose gel and transferred to a nitrocellulose membrane. For a better resolution of the DIssE RNA ( Figure 6B ), RNA was glyoxalated before being electrophoresed on a 1% agarose gel. An in vitro transcribed, 32 P-labeled negative-strand mRNA 7 of MHV-JHM was used as a probe to detect MHV genomic and subgenomic RNAs. For detecting DI RNA species, RNA blots were probed with an RNA representing a sequence complementary to the sequence of the 5¢-untranslated region of MHV-JHM RNA, but excluding the leader sequence. Western blot analysis DBT cells in 6-well plates were infected with MHV-A59 and cytoplasmic extracts were prepared as described previously (Li et al., 1997) at various hnRNP A1 modulates cytoplasmic viral RNA synthesis time points p.i. The extracts were electrophoresed on a 12% polyacrylamide gel and transferred to a nitrocellulose membrane for western blotting. Immuno¯uorescence staining Cells were washed in phosphate-buffered saline (PBS) and ®xed in 4% formaldehyde for 20 min at room temperature, followed by 5 min in ±20°C acetone. Primary antibodies were diluted in 5% bovine serum albumin and incubated with cells for 1 h at room temperature. After three washes in PBS,¯uorescein-conjugated secondary antibodies were added to cells at 1:200 dilution for 1 h at room temperature. FITC-or TRITCconjugated secondary antibodies were used to generate green or red uorescence. Cells were then washed in PBS and mounted in Vectashield (Vector Laboratories, Burlingame, CA). UV-crosslinking assay UV-crosslinking assay was performed as described previously (Huang and Lai, 1999) . In brief, DBT cell extracts (30 mg protein), 200 mg/ml tRNA and 10 4 c.p.m. of an in vitro transcribed, 32 P-labeled negativestrand MHV 5¢-end RNA (182 bp) were incubated for 10 min at 30°C. Increasing amounts of puri®ed GST (0, 0.5, 1.5 and 5 ng) or recombinant GST±hnRNP A1 fusion protein (0, 1, 3 and 10 ng) were included in the reaction to compete with the endogenous hnRNP A1 for binding. The reaction mixture was placed on ice and UV-irradiated in a UV Stratalinker 2400 (Stratagene) for 10 min, followed by digestion with 400 mg/ml RNase A for 15 min at 37°C. The protein±RNA complexes were then separated on a 10% SDS±polyacrylamide gel and visualized by autoradiography. GST pull-down assay GST pull-down was performed as described previously (Tu et al., 1999) . In brief, GST±hnRNP A1 fusion proteins on glutathione beads (Pierce, Rockford, IL) were incubated with the in vitro translated, 35 S-labeled N protein in 0.3 ml of GST-binding buffer containing 0.1% NP-40 for 2 h at 4°C. The beads were washed ®ve times with the GST-binding buffer containing 0.3% NP-40. Proteins bound to beads were eluted by boiling in Laemmli buffer for 5 min and separated on a 10% polyacrylamide gel. [ 35 S]methionine labeling and immunoprecipitation DBT cells were infected with MHV-A59 at an m.o.i. of 2. The cells were incubated with methionine-free medium for 30 min before labeling and were labeled in 100 mCi/ml [ 35 S]methionine starting at 1.5, 7 or 24 h p.i. After labeling for 2 h at each time point, the cells were harvested for protein extraction as described previously (Li et al., 1997) . The protein extracts were immunoprecipitated with anti-Flag antibody-conjugated beads (Sigma, St Louis, MO) in Tm 10 buffer (50 mM Tris±HCl pH 7.9, 0.1 M KCl, 12.5 mM MgCl 2 , 1 mM EDTA, 10% glycerol, 1 mM DTT, 0.1% NP-40, 1 mM phenylmethylsulfonyl¯uoride) at 4°C for 2 h. The immunoprecipitates were washed and separated on a 4±15% gradient SDS±polyacrylamide gel and visualized by autoradiography. Plasmid 25CAT was linearized by XbaI and in vitro transcribed with T7 RNA polymerase to produce the DI RNA . The DI RNA was transfected into MHV-A59-infected DBT cells using DOTAP as described previously (Huang and Lai, 1999) . In brief,~80% con¯uent DBT cells were infected by MHV-A59 at an m.o.i. of 10. At 1 h p.i., the cells were transfected with 5 mg of in vitro transcribed DI RNA and incubated at 37°C for the desired lengths of time. To amplify the DI RNA, viruses (P0) were passaged twice in wt DBT cells to generate P1 and P2 viruses. Cells were harvested at 8 or 24 h p.i. and lysed by freezing and thawing for three times. After centrifugation at 12 000 r.p.m. for 10 min, the supernatant was used in a CAT assay as described previously (Lin et al., 1996) .
12
A Method to Identify p62's UBA Domain Interacting Proteins
The UBA domain is a conserved sequence motif among polyubiquitin binding proteins. For the first time, we demonstrate a systematic, high throughput approach to identification of UBA domain-interacting proteins from a proteome-wide perspective. Using the rabbit reticulocyte lysate in vitro expression cloning system, we have successfully identified eleven proteins that interact with p62’s UBA domain, and the majority of the eleven proteins are associated with neurodegenerative disorders, such as Alzheimer’s disease. Therefore, p62 may play a novel regulatory role through its UBA domain. Our approach provides an easy route to the characterization of UBA domain interacting proteins and its application will unfold the important roles that the UBA domain plays.
p62 is a novel cellular protein which was initially identified in humans as a phosphotyrosine independent ligand of the src homology 2 (SH2) domain of p56 lck (1, 2) . p56 lck is a member of the c-src family of cytoplasmic tyrosine kinases that is found predominantly in cells of lymphoid origin (3, 4) . In addition to the interaction with p56 lck , p62 also associates with the Ser/Thr kinase (1, 2) , atypical protein kinase C (5, 6) , and ubiquitin (7) . In addition to the SH2 domain, p62 possesses several structural motifs, including a ubiquitin associated (UBA) domain that is capable of binding ubiquitin nonconvalently (8, 9) . Ubiquitin (Ub) is a small polypeptide of 76 amino acids that can be convalently attached to other proteins at specific lysine residues through chains composed of one (mono) or several ubiquitin moieties (poly). In addition to its classical role in protein degradation, ubiquitin is emerging as a signal for protein transport and processing (10) (11) (12) . Conjugation of ubiquitin to substrate proteins requires three enzymes: a ubiquitin activating enzyme E1, a ubiquitin-conjugating enzyme E2, and a ubiquitin ligase E3. Initially, E1 activates ubiquitin by forming a high energy thioester intermediate with the C-terminal glycine using ATP. The activated ubiquitin is sequentially transferred to E2, then to E3 which catalyzes isopeptide bond formation between the activated C-terminal glycine of ubiquitin and ε-amino group of a lysine residue of the substrate. Following the linkage of the first ubiquitin chain, additional molecules of ubiquitin are attached to lysine side chains of the previously conjugated moiety to form branched polyubiquitin chains. The fate of ubiquitinated substrates depends on the number of ubiquitin moieties conjugated, as well as, the lysine linkage of Ub-Ub conjugation. The conjugation of ubiquitin to eukaryotic intracellular proteins is one way in which those proteins are targeted to the proteasome for subsequent rapid degradation. This mechanism is particularly important for short-lived regulatory proteins such as cyclins, cyclin-dependent protein kinase-inhibitors, p53, the nuclear factor kappa B precursor, and IκB (13) . The ubiquitinproteasome system consists of two steps: 1) the target protein is conjugated with polyubiquitin molecules, which mark the substrate for degradation; 2) the target protein is transferred to the 26S proteasome, unfolded and degraded. The UBA domain is a conserved sequence motif among proteins that can bind polyubiquitin. It is comprised of ~45 amino acids (13) . The amino acids 386-434 of p62, which bind polyubiquitin, has been shown to possess homology to other recently described UBA domains (9) . Interestingly, proteins with UBA domains are more likely to bind polyubiquitin chains over monoubiquitin, such as the yeast UBA protein Rad23, a highly conserved protein involved in nucleotide excision repair (13) . Recently, it has been shown that yeast cells lacking two UBA proteins (Dsk2 and Rad23) are deficient in protein degradation and that the UBA motif is essential for their function in proteolysis (14) . In addition to the important role in recycling of amino acids from damaged or misfolded proteins, ubiquitin-protein conjugation also has functions unrelated to proteasomal targeting. For example, polyubiquitination is required for the internalization of several yeast and mammalian cell surface proteins into the endocytic pathway (15, 16) . Interestingly, p62 appears to sequester ubiquitinated substrates into a cytoplasmic structure referred to as a sequestosome, into which excess ubiquitinated proteins are segregated (17) . In addition, p62 is an immediate early response gene product for a variety of signals (18) . Thus, p62 appears to play a novel regulatory role for polyubiquitinated proteins and may have an essential function in cell proliferation and differentiation. We have developed a method that will enable identification of protein(s) that interact with p62's UBA domain. Human adult brain library 10×96 well plates with 100 cDNAs per well and Gold TNT SP6 To search for novel proteins that bind to the UBA domain of p62, we performed in vitro expression cloning (IVEC) using the ProteoLink IVEC system. The human adult brain library 96 well plates with 100 cDNAs per well was transcribed and translated employing the Gold TNT SP6 Express 96 plate and [ 35 S] methionine. The TNT Quick-coupled transcription-translation system contained a rabbit reticulocyte lysate pre-mixed with most of the reaction components necessary to carry out transcription/translation in the lysate, including all of the amino acids except methionine. [ 35 S] Methionine was used to label newly synthesized proteins. The reactions were set up according to the manufacturer's instructions. Rabbit reticulocyte lysate has been shown to be capable of carrying out ubiquitination of proteins that were translated in such an in vitro translation system (19, 20) . The reactions mixtures also contained ubiquitin so that the newly synthesized proteins could be ubiquitinated. The reactions were incubated at 30°C for 2 hours. The resulting proteins were assayed to determine their binding ability with p62's UBA domain. Potential positive "hits" were further subdivided and reassayed to link individual clones to the protein of interest (Fig. 1 ). Each translated pool was resuspended in binding buffer (25 mM Tris pH 7.5, 125 mM NaCl, 0.1% NP-40) and used as a source of protein in p62 UBA pull down assays. Proteins that specifically interact with the UBA domain of p62 were isolated by interaction with agarose-immobilised p62-UBA peptide (amino acid 387-436 of p62) (5 µg) for 2 hours at 4ºC, then washed three times in washing buffer (25 mM Tris pH 7.6, 100 mM NaCl, 1% NP-40). Bound proteins were released by addition of SDS-sample buffer and separated by SDS-PAGE. The SDS-PAGE gels were fixed in 50% methanol, 10% acetic acid for 30 min, stained in 0.2% Commassie Brilliant Blue R-250, 45% methanol, 10% acetic acid for 15 min, destained in 10% acetic acid, 50% methanol overnight, and enhanced in autoradiography enhancer En 3 HANCE for 1 hr and exposed to X-Ray film. By combining 4 pools as one mixed pool, 96 protein pools were divided into 24 mixed protein pools for use in p62 UBA pull down assays. Positive mixed protein pools were selected and individual pools were retested for its ability to bind p62's UBA domain. The individual cDNA pool from which the positive protein pool was generated was transformed into JM109 competent cells and plated on LB ampicillin plate. Individual colonies were chosen to grow overnight in 1 ml of LB media plus ampicillin. Plasmid DNA was purified from the cell culture and used for TNT Quick coupled in vitro transcription/ translation. The individual protein synthesized from each plasmid DNA chosen was screened for its ability to bind p62's UBA domain. To confirm the interaction with p62's UBA domain, the final resulting individual proteins were used in the coupled TNT/p62 UBA pull down assays. The cDNA inserts were sequenced in the Genomics Core Facility at Auburn University and the sequences were compared with known sequences in NCBI database by BLAST analysis. Human embryonic kidney 293 (HEK 293) cells were cultured in high glucose Dulbecco's modified Eagle's medium (DMEM) containing 10% heat-inactivated fetal calf serum and transfected with myc-tagged HSP70 plasmid using the Mammalian Cell Transfection Kit. Cells were harvested and lysed in 1 ml of SDS lysis buffer (50 mM Tris-HCl pH 7.5, 150 mM NaCl, 10 mM NaF, 0.5% TX-100, 1 mM Na 3 VO 4 , 2 µg/ml aprotinin, 2 µg/ml leupeptin, 1 mM PMSF, 1% SDS) for 30 min on ice, followed by centrifugation at 14000 rpm for 15 min at 4°C to remove the insoluble fraction. The protein concentration of the supernatant was determined using the Bio-Rad DC protein assay reagent with bovine serum albumin (BSA) as standard. Equal amount of protein (750 µg) was immunoprecipitated with anti-myc and collected with agarose-coupled secondary antibody. To the agarose beads containing the immunoprecipitated HSP70, 50 µl of reaction buffer (50 mM Tris-HCl pH 7.5, 2.5 mM MgCl 2 , 2 mM DTT, 2 mM ATP) was added containing 100 ng E1, 200 ng E2 (UbcH7), and 100 µg of E3 (Flag-tagged TRAF6) along with 5 µg GST-WT-Ub, GST-K29R Ub, GST-K48R Ub, GST-K63R Ub, or K63 Ub. Control samples without HSP70, E1, E2, E3, or GST-WT-Ub were also included. Reactions were carried out by continuous shaking at 37°C for 2 hours and then washed three times with reaction buffer. The proteins were released by boiling for 2 min in SDS-PAGE sample buffer, separated on 7.5% SDS-PAGE and Western blotted for anti-ubiquitin. To search for novel proteins that bind to the UBA domain of p62, we performed in vitro expression cloning (IVEC) using the ProteoLink IVEC system from Promega (Cat. No. L6500). The human adult brain library 96 well plates with 100 cDNAs per well were transcribed and translated employing the Gold TNT SP6 Express 96 plate in the presence of [ 35 S] methionine and ubiquitin (25 µg/µl, Sigma). By combining 4 protein pools as one mixed pool, 96 protein pools were divided into 24 mixed pools ( Fig. 2A, 2B ). Each lane contained more than 100 proteins (theoretically 400) with different molecular weight. Therefore, each lane appeared as a smear, indicating that the in vitro transcription/translation system from Promega worked successfully. In order to examine whether proteins synthesized in the IVEC system are also ubiquitinated, Western blot analysis was performed by blotting the newly synthesized proteins (in the presence of cold methionine instead of 35 S methionine) with ubiquitin monoclonal antibody. In the mixed protein pools, each of the 24 lanes appeared as a smear, indicating that proteins synthesized by the IVEC system are also ubiquitinated (Fig. 3A) . Furthermore, the rabbit reticulocyte lysate in the IVEC system can utilize different lysine linkages of ubiquitin (i.e., Ub K29, Ub K48, and Ub K63) for ubiquitination (Fig. 3B) . In order to investigate whether the agarose-immobilised p62 UBA peptide has binding specificity, a mixed protein pool synthesized by IVEC system was tested in a pull down assay in the presence of agarose beads alone or in the presence of p62 UBA agarose beads (Fig. 3C ). Our results revealed that proteins that bound to p62's UBA domain could not be pulled down by agarose beads alone, indicating that the agarose-immobilised p62 UBA peptide had binding specificity. In order to identify proteins that bind to p62's UBA domain, p62 UBA pull down assays were performed. Out of the 24 mixed protein pools, several pools contained [ 35 S] methionine-labeled bands in the primary p62 UBA pull down assays (Fig. 4) . We chose 6 pools (pool # 2, 4, 8, 14, 20, 21) because of their stronger signal to specifically identify which individual protein pool in the mixed pools has the ability to bind to p62's UBA domain. Therefore, a secondary screen was conducted on the 6 positive individual mixed pools (representing 24 individual protein pools) which bound with p62's UBA domain (Fig. 5) . Mixed protein pool # 2 generated a positive protein with molecular weight of 51 KDa (Fig. 4) , and only individual protein pool "c" out of the four protein pools (a, b, c, d) that comprised protein pool #2 had a protein with the same molecular weight (Fig. 5) . Depending on the size of the protein pulled down in the secondary screen compared to the primary screen (Fig. 4) , individual protein pools "c", "h", "i", "o", "t", and "v" were identified (Fig. 5) . To specifically identify which protein in the individual protein pool has the ability to bind p62's UBA domain, the cDNAs from the positive individual protein pools were then transformed into JM109 competent cells and plated out on LB ampicillin plates. Individual colonies were chosen to grow overnight in 1 ml of LB media plus ampicillin. Plasmid DNAs were purified from the cell culture and used for TNT Quick coupled in vitro transcription/translation. The individual protein synthesized from each plasmid DNA was retested for its ability to bind p62's UBA domain. By synthesizing individual protein from individual plasmid using the Gold TNT Quick coupled in vitro transcription/translation system and subjecting them to p62 UBA pull-down assays, 11 positive clones were isolated from the 6 positive individual pools. It is not surprising that 5 more clones showed binding ability with p62's UBA domain since there are 100 cDNAs in each positive individual pool and some of them could have lower binding ability and therefore showed weak signal in the mixed protein pool. It is also possible that they are not as efficiently synthesized in the mixed TNT reaction as in the individual TNT reaction in which only one cDNA was used as template. The 11 positive plasmids were sequenced and compared with known cDNA sequences in NCBI database using BLAST analysis with results shown in Table 1 . Interestingly, the proteins identified in the screen fall into three distinct categories. One set are proteins that are associated with Alzheimer's disease, including myelin basic protein, 14-3-3 protein, syntaxin binding protein munc18, transketolase, heat shock protein HSP70, reelin, and calcium/calmodulin kinase II (Table 1 ). Significant decrease in the amount of myelin basic protein has been reported in the white matter of Alzheimer's disease patients, accompanied by increased quantities of βamyloid peptides (21) . The presence of β-amyloid peptides containing senile plaques and neurofibrillary tangles are the two major pathological features in the brain of patients with Alzheimer's disease (22) . Interestingly, 14-3-3 proteins have also been demonstrated to be components of neurofibrillary tangles of Alzheimer's disease brains (23) . Syntaxin binding protein munc18 can powerfully regulate amyloid precursor protein metabolism and β-amyloid secretion through direct and indirect interactions with X11 proteins (24) . The activity of transketolase has been reported to be reduced in dementia of Alzheimer's type brain (25) . Heat shock protein HSP70 expression is significantly increased in the temporal cortex of patients with Alzheimer's disease (26) . Besides HSP70, other heat shock proteins are also linked with Alzheimer's disease. For example, increased synthesis of HSP27 has been suggested to play a role in preventing neuronal injury in AD (27) , and alpha-crystallin heat shock protein has a close relationship with neurofibrillary tangles of AD brains (28) . Reelin is a large secreted protein that controls cortical layering by signaling through the very low density lipoprotein receptor and apolipoprotein E receptor 2, thereby inducing tyrosine phosphorylation of the adaptor protein Disabled-1 (Dab1) and suppressing tau phosphorylation (29) . Neurofibrillary tangles comprised of highly phosphorylated tau proteins are a key component of Alzheimer's disease (30) . Enhanced activity of calcium/calmodulin kinase II has been suggested to contribute to phosphorylation of tau protein and lead to neurofibrillary tangle deposition and neuronal death in Alzheimer's disease (31) . Although the relationship between p62 and neurofibrillary tangles or neuritic plaques is unclear, both neurofibrillary tangles and dystrophic neuritis of neuritic plaques are associated with ubiquitin (32) , suggesting that dysfunction in ubiquitin-mediated proteolysis and the resulting accumulation of ubiquitinconjugated proteins may contribute to the origination of dystrophic neuritis and neurofibrillary tangles. Furthermore, p62 has been recently reported to accumulate early in neurofibrillary tangles in Alzheimer's disease (33) , suggesting that p62 may play an important role in Alzheimer's disease by interacting with those proteins through its UBA domain. A second set of proteins identified in the screen that bind to p62's UBA domain are associated with brain development, including homeobox protein Meis2 and unc51 like kinase II (Table 1) . Although Meis proteins are not extensively studied in humans, these proteins have been shown to be required for hindbrain development in the zebrafish (34) . Unc51 like kinase II has been demonstrated to play a role in axonal elongation (35, 36) , which is needed for the formation of complicated neuronal networks. The third set of proteins that exhibit ability to bind p62's UBA domain are proteins that are linked with other neurodegenerative diseases, including FK506 binding proteins and nuclear receptor corepressor I (Table 1) . FK506 (tacrolimus) is a potent immunosuppressive drug used in the treatment of patients after organ transplantation and in selected autoimmune disorders (37) . FK506 is activated upon binding to members of the immunophilin family of proteins, which were designated as FK506 binding proteins (38) . Immunophilins are chaperone proteins and FK506 binding proteins have been suggested as therapeutics for neurological disorders (39, 40) . Nuclear receptor corepressor I has been suggested to play a role in Huntington's disease because it is able to interact with huntingtin (41) . The proteins identified here suggest that p62's UBA domain has the ability to interact with multiple proteins that play important roles in neurodegenerative diseases. Further screening from the whole genome-wide perspective will be necessary to define the important role that p62's UBA domain plays. It has been reported that polyubiquitin chains assembled through lysine 48 of ubiquitin act as a signal for substrate proteolysis by the 26S proteasome (42) (43) (44) . In order to understand whether the proteins identified in our screen bind to the p62's UBA domain through lysine 48 (K48), polyubiquitin K48 chains were added to the p62 UBA pull down assay (Fig. 6) . Inclusion of polyubiquitin K48 chains in the assay should compete for the binding of substrate to the p62's UBA domain and reduce the interaction of those proteins with the p62's UBA domain if those proteins are assembled through K48 chains. An alternative interpretation for polyubiquitin K48 chain competition is that the ubiquitin chains are competing for the same binding site as the binding partners which are either ubiquitinated or non-ubiquitinated. We randomly chose five proteins out of the 11 binding partners for the competition pull down (Fig. 6) . Out of the five proteins, four proteins (# 2, 3, 4, and 5) showed reduced binding ability with p62's UBA domain when polyubiquitin K48 chains were included (Fig. 6A, 6B ). However, K48 chains failed to compete with HSP70, suggesting that p62's UBA domain binds to HSP70 through a ubiquitin lysine linkage other than K48. Interestingly, it has been reported that heat shock protein 70 cognate (HSP70) is ubiquitinated by CHIP (carboxyl terminus of Hsc70-interacting protein) via ubiquitin chain synthesis that uses either K29 or K63 (45) . In order to examine which lysine linkage utilized by HSP70 binds to p62's UBA domain, in vitro ubiquitination assay was performed by incubating lysates from HEK cells expressing HSP70 with E1, E2, and E3 in reaction buffer (50 mM Tris-HCl pH 7.5, 2.5 mM MgCl 2 , 2 mM DTT, 2 mM ATP). As control, the ubiquitination of HSP70 utilizing the rabbit reticulocyte lysate was also investigated by Western blot analysis. Our results revealed that HSP70 was ubiquitinated in the IVEC system (Fig. 7A, 7B) , and the rabbit reticulocyte lysate contained enzymes such as TRAF6 (E3) and UbcH7 (E2) for in vitro ubiquitination (Fig. 7C ). TRAF6 was chosen as an E3 in this in vitro ubiquitination assay due to its RING domain, a common feature of E3 ligases, and the observation that p62 is a scaffold for TRAF6 interaction (46) . Therefore, in vitro ubiquitination assays using the E1-E2-E3 system were performed in the presence of either ubiquitin wild type or ubiquitin mutants (K29R, K48R, and K63R). If one lysine mutant blocks the ubiquitination of HSP70, it would suggest that the ubiquitination of HSP70 utilizes that specific lysine linkage. Our results revealed that HSP70 utilizes K63 linkage to assemble polyubiquitin chains to bind to p62's UBA domain since only the K63R ubiquitin mutant blocked the ubiquitination of HSP70 (Fig. 8A) . A similar result was also observed when reactions were conducted with wild type ubiquitin or mutant ubiquitin with all lysines mutated to arginines except K63 and the ubiquitination of HSP70 occurred only in the reaction that has either intact K63 ubiquitin or wild type ubiquitin (Fig. 8B ). This finding is consistent with previous reports (45) , demonstrating that HSP70 is K63-polyubiquitinated. Furthermore, the in vivo interaction of HSP70 and p62 was confirmed by transfecting myc-tagged HSP70 into HEK 293 cells in the presence of the proteasome inhibitor MG132 and subjecting cell lysates to p62 immunoprecipitation and Western blot with anti-myc antibody (Fig. 8C) . The interaction between HSP70 and p62 in vivo took place only when MG132 was included, suggesting that the interaction in vivo is dependent upon the ubiquitination of HSP70. The specific type of polyubiquitin chain recognized by p62's UBA domain is not yet known and studies are underway lab to determine p62's interaction with specific polyubiquitin chains, however, our preliminary studies suggest that p62's UBA domain may recognize K63 linked polyubiquitin chains. Protein was synthesized employing TNT Quick Coupled in vitro transcription/translation system in the presence of ubiquitin, resolved on 10% SDS-PAGE gels, transferred to nitrocellulose membrane and western blotted with ubiquitin monoclonal antibody. B: HSP70 Protein was synthesized employing TNT Quick Coupled in vitro transcription/translation system in the presence of ubiquitin and 35 S-methionine, resolved on 10% SDS-PAGE and exposed to X-ray film. C: Western blot of rabbit reticulocyte lysate with TRAF6 (E3) and UbcH7 (E2). In summary, for the first time, we demonstrate a systematic approach to identify UBA domain binding proteins from a proteome wide perspective. This approach could be readily adapted to high throughput screening. Using the rabbit reticulocyte lysate in vitro expression cloning system, we have successfully identified eleven proteins in the human adult brain that interact with the UBA domain of p62, and the majority of the eleven proteins are associated with neurodegenerative disorders, such as Alzheimer's disease. This is a very interesting finding since 9600 cDNAs have been screened and only 11 of them showed binding specificity with p62's UBA domain. Studies are underway to unfold the functional roles of p62 in the ubiquitin system. Our approach provides an easy route to the characterization of UBA domain binding proteins at the level of the whole proteome, its application will unfold the important roles that p62's UBA domain plays. This method could be easily adapted to identify proteins that interact with other UBA domains as well.
13
Vaccinia virus infection disrupts microtubule organization and centrosome function
We examined the role of the microtubule cytoskeleton during vaccinia virus infection. We found that newly assembled virus particles accumulate in the vicinity of the microtubule-organizing centre in a microtubule- and dynein–dynactin complex-dependent fashion. Microtubules are required for efficient intracellular mature virus (IMV) formation and are essential for intracellular enveloped virus (IEV) assembly. As infection proceeds, the microtubule cytoskeleton becomes dramatically reorganized in a fashion reminiscent of overexpression of microtubule-associated proteins (MAPs). Consistent with this, we report that the vaccinia proteins A10L and L4R have MAP-like properties and mediate direct binding of viral cores to microtubules in vitro. In addition, vaccinia infection also results in severe reduction of proteins at the centrosome and loss of centrosomal microtubule nucleation efficiency. This represents the first example of viral-induced disruption of centrosome function. Further studies with vaccinia will provide insights into the role of microtubules during viral pathogenesis and regulation of centrosome function.
Intracellular bacterial and viral pathogens have evolved numerous mechanisms to appropriate and exploit different systems of the host during their life cycles in order to facilitate their spread during entry and exit from the host (Cudmore et al., 1997; Finlay and Cossart, 1997; Dramsi and Cossart, 1998) . In the case of viruses, perhaps the best studied example is the exploitation of the actin cytoskeleton by vaccinia virus during its exit from infected cells (Cudmore et al., 1997) . Vaccinia virus is a large DNA virus with a genome of~191 kb encoding 260 open reading frames (ORFs) that is a close relative of variola virus, the causative agent of smallpox (Johnson et al., 1993; Massung et al., 1993) . Vaccinia virus morphogenesis is a complex process which occurs in the cytoplasm of infected cells and results in the formation of the intracellular mature virus (IMV) and the intracellular enveloped virus (IEV). IMV consist of a viral core of DNA and protein enveloped in a membrane cisterna derived from the intermediate compartment (Sodeik et al., 1993) . The IMV core contains ®ve major proteins, A3L, A4L, A10L, F17R and L4R (Vanslyke and Hruby, 1994; Jensen et al., 1996a) , while 12 proteins, A12L, A13L, A14L, A14.5L, A17L, A27L, D8L, G4L, G7L, H3L, I5L and L1R, are associated with the membranes around the virus particle (Jensen et al., 1996a; Betakova et al., 2000) . Depending on the virus strain and cell type, a proportion of IMV can become enwrapped by a membrane cisterna derived from the trans-Golgi apparatus to give rise to IEV particles (Schmelz et al., 1994) . To date, six IEV-speci®c proteins, A33R (Roper et al., 1996) , A34R (Duncan and Smith, 1992) , A36R (Parkinson and Smith, 1994) , A56R (Payne and Norrby, 1976; Shida, 1986) , B5R (Engelstad et al., 1992; Isaacs et al., 1992) and F13L (Hirt et al., 1986) , have been identi®ed. Studies using recombinant viruses have shown that A33R, A34R, B5R and F13L play an important role in IEV assembly (Blasco and Moss, 1991; Engelstad and Smith, 1993; Wolffe et al., 1993 Wolffe et al., , 1997 Roper et al., 1998; Sanderson et al., 1998a; Ro Èttger et al., 1999) . Vaccinia virus is thought to leave the cell by fusion of the outer IEV membrane with the plasma membrane, to give rise to the extracellular enveloped virus (EEV) (Morgan, 1976; Payne, 1980; Blasco and Moss, 1991) or the cell-associated enveloped viruses (CEV) which remain associated with the outer surface of the plasma membrane (Blasco and Moss, 1992) . During the complex vaccinia infection process, the actin cytoskeleton is dramatically reorganized and numerous actin comet-like tails are induced by IEV particles (Cudmore et al., 1995; Ro Èttger et al., 1999) . Using actin polymerization as the driving force, IEV particles are propelled on actin tails until they contact the plasma membrane and extend outwards, thereby facilitating infection of neighbouring cells (Cudmore et al., 1995) . In addition, vaccinia infection results in stimulation of cell motility, loss of contact inhibition and changes in cell adhesion (Sanderson and Smith, 1998; Sanderson et al., 1998b) . Vaccinia virus-induced cell motility can be subdivided further into cell migration and extension of neurite-like projections, the latter of which is dependent on microtubules (Sanderson et al., 1998b) . The dependence of neurite-like projection formation on microtubules suggests that the microtubule cytoskeleton may also play a role during the life cycle of vaccinia virus. Indeed, recently, the vaccinia A27L protein and microtubules have been shown to be required for ef®cient IMV dispersion (Sanderson et al., 2000) . Furthermore, in the absence of vaccinia actin-based motility, cell to cell spread still occurs although it is less ef®cient (Wolffe et al., 1997 Sanderson et al., 1998a) , suggesting that additional transport mechanisms must exist. Given these observations, we wondered whether the microtubule cytoskeleton has a function during the life Vaccinia virus infection disrupts microtubule organization and centrosome function The EMBO Journal Vol. 19 No. 15 pp. 3932±3944, 2000 cycle of vaccinia virus. We now report that the microtubule cytoskeleton and the dynein±dynactin complex play an important role during the early stages of vaccinia infection. However, later during the infection cycle, loss of centrosome function and accumulation of viral-encoded microtubule-associated proteins (MAPs) result in a dramatic rearrangement of the microtubule cytoskeleton. Vaccinia localization in the vicinity of the MTOC depends on microtubules and the dynein±dynactin complex Indirect immuno¯uorescence labelling shows that by 6 h post-infection the majority of vaccinia virus particles are concentrated in the area coinciding with the centre of the microtubule aster ( Figure 1A and C). To examine whether this localization is indeed microtubule dependent, we infected cells pre-treated with nocodazole to depolymerize microtubules. In the absence of microtubules, virus particles were distributed throughout the cytoplasm ( Figure 1B and D) . The accumulation of virus particles in the area around the centre of the microtubule aster suggested that a microtubule minus end-directed motor may be involved in establishing the position of the virus in this location. To examine this possibility, we infected cells overexpressing p50/dynamitin which acts as a dominantnegative for dynein±dynactin function (Echeverri et al., 1996) . We found in cells overexpressing p50/dynamitin that virus particles did not accumulate at the centre of the microtubule aster but rather throughout the cytoplasm, as occurs in the absence of microtubules (compare Figure 2B with Figure 1D ). As vaccinia morphogenesis involves wrapping by host membranes, it was possible that the effects of nocodazole and p50/dynamitin on virus localization were in fact due to disruption of the intermediate compartment and Golgi apparatus by these reagents (Burkhardt et al., 1997) . However, two independent experiments showed that this is not the case. First, in cells infected in the absence of microtubules, the Golgi apparatus as well as vaccinia virus particles are dispersed throughout the cytoplasm but do not co-localize ( Figure 3F and O). Secondly, vaccinia particles remain in the vicinity of the microtubule-organizing centre (MTOC) when the Golgi but not the microtubules was disrupted by treatment with brefeldin A ( Figure 3G and P). Similar results were obtained using other markers: A17L for vaccinia, galactosyltransferase for the Golgi or ERGIC53 for the intermediate compartment (data not shown). Taken together, our data indicate that the microtubule cytoskeleton is required for the localization of newly assembled virus particles in the vicinity of the MTOC during vaccinia infection. Formation of functional IEV, but not IMV, is microtubule dependent Given the requirement for microtubules in vaccinia localization, we subsequently examined whether this localization has a role in morphogenesis of the two different intracellular forms of vaccinia virus, IMV and IEV. From electron microscopic examination of cells infected in the presence of nocodazole, it became clear that IMV particles which are morphologically indistinguishable from controls are formed ( Figure 4 ). Although IMV particles are assembled in the absence of microtubules, we wondered whether their number is reduced and whether those that are formed are infectious, since the integrity of the intermediate compartment depends on microtubules (Burkhardt et al., 1997) . To address this question, three independent virus stocks were prepared in the presence or absence of nocodazole. To simplify the interpretation of the data, we used the recombinant vaccinia virus mutant DF13L, which is unable to form IEV (Blasco and Moss, 1991) . The ®nal concentration of virus particles produced, Vaccinia uses and abuses the microtubule cytoskeleton as determined by the method of Joklik (1962) , was 30.2 6 5.2 3 10 10 particles/ml in the presence of microtubules and 9.0 6 6.7 3 10 10 particles/ml in the absence of microtubules. Although there is a 3-fold decrease in the number of virus particles formed in the absence of microtubules, the particles that are formed are infectious (data not shown). While infectious IMV are formed in the absence of microtubules, we found no evidence for IEV formation, based on electron microscope examination of cells infected in the presence of nocodazole ( Figure 4 ). We did, however, observe IMV particles partially wrapped in trans-Golgi membranes most probably in the process of abortive IEV formation ( Figure 4D ). Given these data, we examined by indirect immuno¯uorescence whether low amounts of IEV particles are formed in the absence of microtubules. However, we could ®nd no evidence for colocalization of the IEV protein markers A36R, A34R or A33R with vaccinia particles formed in the presence of nocodazole ( Figure 5F ). We also found no evidence for IEV formation, based on their ability to nucleate actin tails ( Figure 5O ). As IEV particle assembly involves wrapping by the Golgi apparatus (Schmelz et al., 1994) , we examined the effects of only disrupting this membrane compartment using brefeldin A. We could ®nd no evidence for IEV formation, based on co-localization of IEV protein markers with virus particles and actin tails in cells infected in the presence of brefeldin A ( Figure 5G±I and P±R). Indeed, in brefeldin A-treated cells, the IEV membrane proteins required for assembly were observed in the endoplasmic reticulum and not the trans-Golgi ( Figure 5H ). In summary, our data indicate that the microtubule cytoskeleton is required for ef®cient IMV assembly and is essential for IEV formation. In the course of our experiments, it became obvious that the Golgi apparatus becomes progressively dispersed during infection co-concominantly with disruption of the microtubule network ( Figure 6 ). Further analysis showed that during infection the normal morphology of the microtubule cytoskeleton is replaced by morphologically aberrant microtubule forms, which vary among each other but have in common the absence of a discrete MTOC ( Figure 7 ). These aberrant forms can be broadly classi®ed into three types: (i) cells with a disorganized microtubule network where microtubules seem randomly oriented ( Figure 7E ); (ii) cells in which microtubules form rings around the nucleus and throughout the cytoplasm ( Figure 7H ); or (iii) cells with long projections consisting of microtubule bundles ( Figure 7K ). We quanti®ed the appearance of the different morphological forms in ®ve independent infection experiments, in which 200 cells were counted for each time point for each experiment ( Figure 7C , F, I and L). Small compact cells, representing 20.7 6 2.6, 21.8 6 12.4 and 29.7 6 15.2% for 5, 8 and 24 h post-infection, respectively, in which the microtubule cytoskeleton morphology was not evident were not included in the analysis. Already by 5 h post-infection, when virus particle assembly has occurred, the normal aster microtubule con®guration has been disrupted and replaced in the majority of cells by microtubules without obvious organization from the MTOC ( Figure 7F ). Furthermore,~10% of cells have microtubule rings and 5% of cells have long projections by this time point ( Figure 7I and L). As the infection proceeds, microtubules become progressively more disrupted and bundled ( Figure 7I and L). From our observations, there seems to be no obvious connection between the disruption and changes in the actin and the microtubule cytoskeletons ( Figure 7) . Moreover, the same reorganization of the microtubule network occurs in cells infected with the vaccinia deletion mutants DF13L and DA36R which do not make actin tails (data not shown). The effects of vaccinia virus infection on the reorganization of the microtubule cytoskeleton were also observed in all cell lines we examined (BHK-21, C 2 C 12 , PtK2, RK 13 and Swiss 3T3) to varying degrees (data not shown). Our data show that vaccinia infection results in severe disruption of the normal morphology of the microtubule cytoskeleton. The formation of microtubule bundles and the loss of organization from the MTOC in vaccinia-infected cells is strongly reminiscent of the phenotype observed in cells overexpressing a MAP (Weisshaar et al., 1992; Togel et al., 1998) . As overexpression of MAPs stabilizes microtubules, we examined whether the microtubule cytoskeleton in infected cells was more resistant to depolymerization by nocodazole or cold treatment ( Figure 8 ). This was indeed the case, suggesting that the virus genome may encode viral proteins with MAP-like properties. To identify viral proteins which exhibit microtubule-binding properties, we performed microtubule co-sedimentation assays using extracts prepared from uninfected and vaccinia-infected cells (Figure 9 ). Initial experiments, however, revealed that intact virus particles in the extracts were prone to pellet even in the absence of microtubules, making identi®cation of viral MAPs impossible. To avoid this problem, we prepared extracts from cells infected in the presence of rifampicin, a drug that inhibits vaccinia virus particle assembly but does not affect viral protein expression (Moss et al., 1969; Tan and McAuslan, 1970) . The morphological effects of vaccinia infection on the microtubule cytoskeleton were the same in the presence or absence of rifampicin (data not shown). Comparison of the proteins present in pellets from microtubule co-sedimentation assays reveals that a number of additional prominent and minor bands are present in extracts prepared from vaccinia-infected but not from uninfected cells (Figure 9 ). Co-sedimentation assays Vaccinia uses and abuses the microtubule cytoskeleton performed in the presence of nocodazole or with coldtreated extracts reveal that the majority of these additional bands disappear in the absence of microtubules. To identify the viral proteins co-sedimenting with microtubules, we performed in-gel protease digestion followed by analysis of the resulting peptides by MALDI mass spectrometry. Using this approach, we identi®ed a number of potential vaccinia-encoded MAPs: A10L (a structural protein), I1L and L4R (which are DNA-binding proteins), all of which are associated with viral cores (Vanslyke and Hruby, 1994; Jensen et al., 1996a; Klemperer et al., 1997) , and A6L which is conserved in all poxvirus genomes but is of unknown function (Figure 9 ). A10L and L4R associate with microtubules in vivo and mediate binding of viral cores to microtubules in vitro Using available antibodies, we examined the localization of A10L, L4R and I1L in infected cells to see whether they associate with microtubules in vivo, in addition to their essential role in the virus core (Vanslyke and Hruby, 1994; Jensen et al., 1996a) . As a negative control, we also examined the localization of the A3L core protein which was identi®ed as the prominent 70 kDa protein pelleting in the absence of microtubules (Figure 9 ). Indirect immunouorescence analysis showed that A10L and L4R are associated with microtubules, in both the presence and absence of rifampicin ( Figure 10 ). As expected, A10L and L4R were also associated with viral particles (data not shown). In contrast, I1L and A3L were never observed in association with microtubules, regardless of the ®xation conditions, but were localized to viral factories and viral particles, respectively (data not shown). Interestingly, A10L and L4R were not associated with all microtubules but were co-localized with a subset of acetylated microtubules ( Figure 10) . The association of A10L and L4R with virus particles and microtubules raises the question of whether there is a role for this microtubule-binding activity during infection. We wondered whether these two proteins mediate the interaction of incoming viral cores with microtubules at the beginning of infection, as cores and not virus particles are released in the cytoplasm at the start of the infection cycle (Ichihashi, 1996; Vanderpasschen et al., 1998; Pedersen et al., 2000) . To examine this possibility, we investigated whether puri®ed viral cores would bind microtubules in vitro. We found that viral cores were able to bind microtubules, while protease-treated cores showed no association ( Figure 11A and B) . Pre-incubation of puri®ed viral cores with antibodies against A10L and L4R speci®cally inhibited the interaction of viral cores with microtubules ( Figure 11C and D); in contrast, IgG or antibodies against A3L had no inhibitory effect ( Figure 11E and F). Taken together, our data suggest that A10L and L4R have MAP-like properties and may play a role in mediating interactions of incoming viral cores with microtubules. The dramatic rearrangement of the microtubule cytoskeleton which occurs during vaccinia infection is unlikely to be attributed exclusively to the action of A10L and L4R since they only associate with a subset of microtubules ( Figure 10) . Furthermore, the loss of microtubule organization precedes detectable association of A10L and L4R with microtubules, which occurs from~8 h post-infection. We therefore wondered whether vaccinia infection disrupts centrosome function, given the loss of microtubule aster con®guration during infection (Figure 7 ). Since microtubules are nucleated by the centrosome in animal cells, we examined whether vaccinia infection affects g-tubulin, which is critically required for this process (Stearns and Kirschner, 1994) . We observed that g-tubulin labelling of the centrosome is greatly reduced from as early as 2 h post-infection ( Figure 12 ). The same result was obtained when we infected PtK1 cells stably expressing green¯uorescent protein (GFP)-labelled g-tubulin (Khodjakov and Rieder, 1999) . In addition, the centrosomal and centriolar components pericentrin, C-Nap 1, Nek 2 and centrin are reduced by immuno¯uorescence in the centrosomes/centrioles of vaccinia-infected cells ( Figure 12) . Furthermore, the reduction of centrosomal markers requires viral protein synthesis as their levels are not affected when cells are infected in the presence of cycloheximide (data not shown). The dramatic reduction of g-tubulin from the centrosome implies that vaccinia infection perturbs centrosome Vaccinia uses and abuses the microtubule cytoskeleton function. To test this hypothesis, we examined whether the centrosome in vaccinia-infected cells could re-nucleate microtubules, following their depolymerization by nocodazole. We found that by 2 h post-infection, when we already see a reduction in g-tubulin, microtubule nucleation from the centrosome was very inef®cient, as compared with uninfected controls, indicating that vaccinia has disrupted`normal' centrosome function (Figure 13 ). At later times post-infection, microtubule re-nucleation ef®ciency from the centrosome was even lower (data not shown). However, following nocodazole washout, microtubules eventually are repolymerized throughout the cytoplasm of infected cells but do not display any organization from the MTOC, as do controls (compare Figure 13I and K). The size of virus particles is such that they are unlikely to move within and between cells by diffusion alone, suggesting that their movements will require interactions with the host cytoskeleton. Previous data have shown that vaccinia virus both disrupts and hijacks the actin cytoskeleton to facilitate movement of the intracellular enveloped form of vaccinia virus (Cudmore et al., 1995; Ro Èttger et al., 1999) and of the infected cell itself (Sanderson and Smith, 1998; Sanderson et al., 1998b) . The data described here now show that vaccinia also uses and subsequently disrupts the microtubule cytoskeleton during its infection cycle. It is clear from our experiments and the previous observations of Ulaeto et al. (1995) that microtubules are required to maintain the integrity of the Golgi apparatus which is in turn required for IMV wrapping to form IEV (Schmelz et al., 1994) . In contrast, IMV are assembled in the absence of microtubules, albeit at reduced levels. While microtubules are not required for IMV assembly, they are required together with the dynein±dynactin complex for virion accumulation in the vicinity of the microtubule aster. One can envisage that minus enddirected microtubule-dependent movements of IMV particles from their site of assembly in the viral factory towards the MTOC, by the dynein±dynactin complex, would enhance the possibility of wrapping with the Golgi apparatus and subsequent IEV formation. Recently it has been shown that the IMV protein A27L and microtubules are required for ef®cient IMV dispersion from the viral factories (Sanderson et al., 2000) . In the absence of A27L, mature IMV particles accumulate at the periphery of the virus factory but do not subsequently wrap to form IEV, presumably because they are unable to move on microtubules (Sanderson et al., 2000) . The microtubule-and dynein±dynactin-dependent accumulation of vaccinia in the vicinity of the MTOC is analogous to the microtubule-dependent movements required for herpes simplex virus 1 (HSV-1) and adenovirus to reach their site of replication in the nucleus (Sodeik et al., 1997; Suomalainen et al., 1999; Leopold et al., 2000) . In the case of HSV-1, the UL34 protein, 9 . Vaccinia encodes proteins that co-sediment with microtubules. Analysis of pellets from in vitro microtubule co-sedimentation assays performed with protein extracts from vaccinia-infected (inf.) and uninfected (uninf.) cells. Twice the amount of pellet has been loaded in control assays performed in the absence of microtubules (nocodazole or 4°C). Proteins co-sedimenting with microtubules that were only present in extracts from infected cells are indicated by an asterisk. The identity of proteins determined by in-gel proteolysis MALDI mass spectrometry is indicated (arrowheads). which is associated with the incoming nucleocapsids, interacts with the intermediate chain of cytoplasmic dynein (IC-1a) (Ye et al., 2000) . It has also been reported that incoming nucleocapsids of pseudorabies virus, an alphaherpes virus closely related to HSV-1, are associated with and dependent on microtubules for their movement to the nucleus (Kaelin et al., 2000) . This interaction may be mediated by the UL25 protein, a minor but essential component of the capsid, which co-localizes with microtubules and accumulates at the MTOC (Kaelin et al., 2000) . The accumulation of UL25 at the MTOC is consistent with a possible interaction with the dynein± dynactin motor complex which is known to be localized at the MTOC (Echeverri et al., 1996) . It would not be surprising, based on observations with HSV-1 and pseudorabies, if microtubules and dynein±dynactin were also involved in establishing the infection cycle of cytomegalovirus (CMV), Epstein±Barr virus and varicella-zoster virus, all of which are herpes viruses. The other clear example of microtubule-dependent virus movements during the establishment of infection is that of incoming human foamy virus (HFV) which is dependent on microtubules and presumably a minus end-directed microtubule motor to get to its nuclear replication site (Saib et al., 1997) . In the absence of protein expression, HFV Gag proteins, which are associated with the viral genome, accumulate at the centrosome in a microtubuledependent fashion prior to nuclear import (Saib et al., 1997) . The centrosomal accumulation of Gag proteins of HFV, however, appears to be unique for this class of retroviruses as no similar localization has been reported for human immunode®ciency virus (HIV) or other retroviruses. On the other hand, the Gag protein of murine leukaemia virus and HIV has been shown to interact with KIF4, a microtubule plus end-directed kinesin motor, both in vitro and in vivo (Kim et al., 1998; Tang et al., 1999) , suggesting that additional roles may exist for microtubules and motors during the outward movement of virus particles. Indeed, vaccinia virus particles are able to reach the cell periphery in the absence of actin-based motility (see images in Wolffe et al., 1997; Sanderson et al., 1998a Sanderson et al., , 2000 Ro Èttger et al., 1999) , suggesting that viral particles can also move out on microtubules (Sanderson et al., 2000) . Microtubule-dependent motordriven movements of virus particles represent an ef®cient mechanism to achieve a peri-nuclear localization, required to facilitate entry into the nucleus during establishment of infection. They also provide an excellent way for newly assembled virus particles to reach the cell periphery, facilitating the continued spread of infection. Our data show that although vaccinia virus uses the microtubule cytoskeleton to achieve a peri-nuclear localization, microtubule and Golgi organization becomes disrupted later during the infection process. Interestingly, HSV-1 and CMV have also been reported to disrupt the microtubule cytoskeleton and Golgi organization in their infection cycles (Avitabile et al., 1995; Fish et al., 1996) . While disruption of the microtubule network might at ®rst sight not appear to be bene®cial to the virus, it may not actually hinder viral spread but could enhance it. First, extensive virus assembly and spread to the cell periphery have already occurred by the time the microtubule cytoskeleton and Golgi organization are disrupted. Secondly, disruption of microtubule organization may overcome potential microtubule motor anchoring effects at the MTOC, thus allowing viral spread to the periphery to occur more easily. Lastly, the formation of long projections of up to 200 mm supported by extensive microtubule bundles provides a means to achieve long range spread of virus particles (Sanderson et al., 1998b) . It is clear that disruption and reorganization of the microtubule cytoskeleton by vaccinia virus is mediated by the combined effects of viral proteins with MAP-like properties and loss of microtubule-organizing function from the MTOC. The same may also be true for HSV-1, although disruption of centrosome function remains to be established, as late in infection microtubules are organized in bundles around the nucleus and do not show MTOCorchestrated organization (Avitabile et al., 1995) . The identi®cation of viral proteins with MAP-like properties is not unique to vaccinia virus. The VP22 tegument protein from HSV-1 co-localizes with microtubules in infected cells and induces microtubule bundles when expressed in uninfected cells (Elliott and O'Hare, 1998) . Other examples of viral MAPs based on their in vivo localization Fig. 11 . Vaccinia cores bind directly to microtubules in vitro. Puri®ed viral cores labelled by DAPI (green) bind to rhodamine-labelled microtubules (red) in the absence of ®xation (A). Binding to microtubules is not observed if cores are pre-treated with protease (B) or pre-incubated with antibodies against the A10L (C) or L4R (D) proteins. In contrast, pre-incubation of puri®ed viral cores with control IgG (E) or antibody against the A3L protein (F) does not inhibit their interaction with microtubules. Scale bar = 5 mm. Vaccinia uses and abuses the microtubule cytoskeleton or in vitro association with microtubules are the N protein from murine coronavirus (Kalicharran and Dales, 1996) , the movement protein from tobamovirus (Heinlein et al., 1995) , the aphid transmission factor from cauli¯ower mosaic virus (Blanc et al., 1996) , the UL25 protein from pseudorabies virus (Kaelin et al., 2000) , the VP4 spike protein from rotavirus (Nejmeddine et al., 2000) and the M protein of vesicular stomatitis virus (VSV) (Melki et al., 1994) . The identi®cation of A10L and L4R, two viral core proteins, as MAP-like proteins was, however, unexpected given their previously characterized role in viral morphogenesis (Vanslyke and Hruby, 1994) . The interaction of A10L and L4R with microtubules in vivo, together with the in vitro microtubule-binding data, suggest a potential mechanism for the association of viral cores with microtubules. One could envisage that viral cores which are released into the cytoplasm at the beginning of infection (Ichihashi, 1996; Vanderpasschen et al., 1998; Pedersen et al., 2000) bind directly to microtubules in a manner analogous to adenovirus or HSV-1 nucleocapsids. Further work is required to determine whether incoming cores do in fact move towards the MTOC by the dynein± dynactin complex and/or use the complex for anchoring on microtubules. The loss of centrosome function must enhance disruption of the microtubule cytoskeleton during infection. Indeed, the loss of microtubule organization from the MTOC precedes detectable association of A10L and L4R with microtubules, which occurs from~8 h post-infection. Vaccinia-induced loss of centrosomal proteins is inhibited by cycloheximide, indicating that viral protein expression is required for disruption of the centrosome microtubule nucleation activity. To our knowledge, vaccinia virus infection represents the ®rst example of virus-induced disruption of centrosome function, although we would predict that HSV-1 may have a similar effect. The mechanism by which vaccinia virus disrupts the centrosome requires further study; nevertheless, it is clear that understanding the molecular basis of this disruption will provide important insights into the regulation and stability of centrosome function which currently is the subject of intense research (Ohta et al., 1993; Lane and Nigg, 1997; Karsenti, 1999) . HeLa cells (ATCC CCL2) were infected with the wild-type vaccinia virus strain Western Reserve (WR) or with the vaccinia deletion mutants DF13L (vRB12) (Blasco and Moss, 1991) or DA36R (Parkinson and Smith, 1994) at a multiplicity of infection of 1 p.f.u. (plaque-forming unit) per cell, as described previously (Ro Èttger et al., 1999) . Nocodazole dissolved in dimethyl sulfoxide (DMSO) and brefeldin A dissolved in ethanol were added to the culture medium to ®nal concentration of 10 mM and 5 mg/ml, respectively unless otherwise stated. In non-treated controls, an equal volume of DMSO or ethanol was added. Cells transfected with a myc-tagged p50/dynamitin expression construct (Echeverri et al., 1996) were infected 24 h later with WR and subsequently ®xed 6 h postinfection. All experiments described have been repeated 3±10 times. The following antibodies were kindly provided: anti-a-tubulin by Dr E.Karsenti, anti-centrin (20H5) by Professor J.L.Salisbury (Sanders and Salisbury, 1994; Paoletti et al., 1996) , anti-Nek2 and anti-C-Nap1 by Professor E.Nigg (Fry et al., 1998a,b) , anti-myc and anti-gp27 by Dr T.Nilsson (Fu Èllekrug et al., 1999) and antibodies against the corresponding vaccinia proteins: A3L, A10L and L4R by Professor D.Hruby (Vanslyke and Hruby, 1994) , I1L by Professor P.Traktman (Klemperer et al., 1997) , A27L (C3) by Dr M.Esteban (Rodriguez et al., 1985) , A33R, A34R and A36R (Ro Èttger et al., 1999) . In addition, the following antibodies were obtained from commercial sources: anti-a-tubulin (N356) (Amersham International, UK), anti-acetylated a-tubulin (6-11B-1) (Sigma, USA), anti-g-tubulin (GTU-88; Sigma), anti-pericentrin and anti-TGN46 (BAbCO, USA), and rabbit IgG (Sigma). Actin was visualized with¯uorescently labelled phalloidin derivatives (Molecular Probes, USA). Cells were ®xed in ±20°C methanol or in 5% paraformaldehyde in BRB80 (80 mM PIPES pH 6.8, 1 mM MgCl 2 , 1 mM EGTA) followed by 0.1% Triton X-100 permeabilization. Fixed cells were processed for immuno¯uorescence, viewed and images recorded as described previously (Ro Èttger et al., 1999) . HeLa cells were pre-incubated with 25 mM nocodazole in the medium for 1 h to depolymerize microtubules, prior to infection with vaccinia DF13L at 1 p.f.u./cell. Nocodazole was kept in the medium throughout the infection, while an equal volume of DMSO was added to the controls. At 24 h post-infection, the cells were scraped from the¯asks into the medium and sedimented by centrifugation (300 g, 7 min, 4°C). The cell membranes were disrupted and the nuclei were removed by centrifugation. The resulting post-nuclear supernatant was centrifuged through a 36% sucrose cushion (76 000 g, 30 min, 4°C). The virus pellet was resuspended in 10 mM Tris pH 9; the virus was collected by centrifugation (76 000 g, 30 min, 4°C), resuspended in 10 mM Tris pH 9 and stored at ±80°C. The concentration of the virus (elemental bodies) was determined by OD 260 measurement (Joklik, 1962) . Fig. 13 . Vaccinia infection reduces centrosome microtubule nucleation ef®ciency. In uninfected cells, microtubules (A, E and I) nucleate from centrosomes (B, F and J) after nocodazole washout for the times indicated. In contrast, 2 h after infection with vaccinia, microtubules (C, G and K) are nucleated inef®ciently from centrosomes (D, H and L). All images were collected with identical camera settings, to allow comparison of uorescence intensity between centrosomes. Inserts (B, D, F, H, J and L) are adjusted as in Figure 12 to facilitate visualization of the weak g-tubulin centrosomal labelling. Arrowheads indicate the position of the centrosome. Scale bar = 10 mm. In vitro microtubule binding assays Puri®ed EEV particles were prepared as described previously (Ro Èttger et al., 1999) and subsequently were used to prepare virus cores following the method of Cudmore et al. (1996) . Rhodamine-labelled microtubules were prepared according to Hyman et al. (1991) . Vaccinia virus cores were incubated with rhodamine-labelled microtubules in BRB80 buffer containing 10 mM taxol for 5 min at room temperature. 4¢,6-diamidino-2phenylindole (DAPI) was added subsequently to a ®nal concentration of 0.1 mg/ml to label the virus cores. Finally, the mixture was diluted 1:1± 1:10 with antifade solution (0.1 mg/ml catalase, 0.1 mg/ml glucose oxidase, 10 mM glucose) and viewed without ®xation. Proteinase K or trypsin treatment of core particles prior to incubation with microtubules was performed as described previously (Roos et al., 1996) . Anti-A3L, A10L, L4R or control IgG antibodies were incubated with puri®ed cores for 1 h at room temperature prior to incubation with microtubules. Cell extracts and microtubule co-sedimentation assay Extracts from HeLa cells infected for 24 h or uninfected controls maintained in the presence of 0.1 mg/ml rifampicin were prepared as described previously (Ro Èttger et al., 1999) . The extract was clari®ed by centrifugation at 150 000 g for 20 min at 4°C and cytochalasin D added to a ®nal concentration of 1 mg/ml to depolymerize actin ®laments. Endogenous tubulin in the extract was polymerized in a two-step procedure. First, the extract supernatant was supplemented with protease inhibitors, 2 mM MgGTP and 5 mM taxol and incubated for 5 min at room temperature; subsequently, an additional 15 mM taxol was added to the mix and the reaction incubated at 33°C for 30 min. For controls, no taxol was added at any stage and microtubule polymerization was inhibited either by the addition of nocodazole to a ®nal concentration of 40 mM or by maintaining the extract at 4°C throughout the experiment. Following microtubule assembly, each 400 mg extract reaction was diluted 5-fold in BRB80 buffer (containing protease inhibitors and 20 mM taxol) and centrifuged through a 10% sucrose cushion containing protease inhibitors and 20 mM taxol at 165 000 g for 20 min at 25°C. The microtubule pellet was solubilized in SDS±PAGE sample buffer and analysed by SDS± PAGE. In-gel proteolytic cleavage was performed automatically in the`Progest' as described (Houthaeve et al., 1997) [Genomic Solutions Cambridge (http://www.genomicsolutions.com)] and the peptides obtained were analysed on a Bruker REFLEX MALDI mass spectrometer (Bruker Analytik, Germany) (Jensen et al., 1996b) . Proteins were identi®ed by peptide mass ®ngerprinting (Jensen et al., 1997) using the program PeptideSearch (http://www.narrador.embl-heidelberg.de/Services/ PeptideSearch/PeptideSearchIntro.html. At 1 h post-infection, nocodazole was added to the culture medium to a ®nal concentration of 25 mM to depolymerize microtubules. At 2 h postinfection, the cells were washed 3±4 times in warm medium to remove nocodazole. Washed cells were incubated in medium without nocodazole for the indicated time at 37°C to re-initiate microtubule polymerization; they were then washed brie¯y in warm phosphate-buffered saline (PBS) and immediately ®xed. In parallel, samples were also removed at the same time point, brie¯y rinsed in ice-cold PBS, ®xed and processed for immuno¯uorescence to con®rm complete microtubule depolymerization before initiation of microtubule assembly. Uninfected control HeLa cells were treated and processed in an identical fashion. The same numbers of images were integrated using identical camera settings to allow direct comparison between infected and uninfected samples from the same experiment.
14
Multi-faceted, multi-versatile microarray: simultaneous detection of many viruses and their expression profiles
There are hundreds of viruses that infect different human organs and cause diseases. Some fatal emerging viral infections have become serious public health issues worldwide. Early diagnosis and subsequent treatment are therefore essential for fighting viral infections. Current diagnostic techniques frequently employ polymerase chain reaction (PCR)-based methods to quickly detect the pathogenic viruses and establish the etiology of the disease or illness. However, the fast PCR method suffers from many drawbacks such as a high false-positive rate and the ability to detect only one or a few gene targets at a time. Microarray technology solves the problems of the PCR limitations and can be effectively applied to all fields of molecular medicine. Recently, a report in Retrovirology described a multi-virus DNA array that contains more than 250 open reading frames from eight human viruses including human immunodeficiency virus type 1. This array can be used to detect multiple viral co-infections in cells and in vivo. Another benefit of this kind of multi-virus array is in studying promoter activity and viral gene expression and correlating such readouts with the progression of disease and reactivation of latent infections. Thus, the virus DNA-chip development reported in Retrovirology is an important advance in diagnostic application which could be a potent clinical tool for characterizing viral co-infections in AIDS as well as other patients.
There are hundreds of viruses that infect different human organs and cause diseases. Some fatal emerging viral infections have become serious public health issues worldwide. Early diagnosis and subsequent treatment are therefore essential for fighting viral infections. Current diagnostic techniques frequently employ polymerase chain reaction (PCR)-based methods to quickly detect the pathogenic viruses and establish the etiology of the disease or illness. However, the fast PCR method suffers from many drawbacks such as a high false-positive rate and the ability to detect only one or a few gene targets at a time. Microarray technology solves the problems of the PCR limitations and can be effectively applied to all fields of molecular medicine. Recently, a report in Retrovirology described a multi-virus DNA array that contains more than 250 open reading frames from eight human viruses including human immunodeficiency virus type 1. This array can be used to detect multiple viral co-infections in cells and in vivo. Another benefit of this kind of multi-virus array is in studying promoter activity and viral gene expression and correlating such readouts with the progression of disease and reactivation of latent infections. Thus, the virus DNA-chip development reported in Retrovirology is an important advance in diagnostic application which could be a potent clinical tool for characterizing viral co-infections in AIDS as well as other patients. Microarray technology has been proven to be a powerful tool with great potential for biological and medical uses. In this technique, recombinant DNA fragments or synthesized oligonucleotides affixed on the surface of glass slides or nylon membranes are used for detecting complementary nucleic acid sequences (frequently representing a few hundred to >10,000 genes/expressed sequence tags) as well as for genotyping microorganisms and for profiling the gene-expression patterns in cells from higher organisms [1]. A new report by Ghedin, et al. [2] in Retrovirology describes the successful use of a multi-virus array (termed multivi-rus-chip) to detect multiple viral co-infections in cultured cells as well as to study viral gene expression and promoter activities (Figure 1 ). Ghedin's multivirus-chip contains genes from eight human viruses including human immunodeficiency virus type 1 (HIV-1). Conceptually, this chip can be used to detect viral co-infections in AIDS patients who are frequently rendered susceptible to additional opportunistic infections. In developing their multivirus-chip, Ghedin, et al. tested more than 250 ORFs from HIV-1, human T cell leukemia virus types 1 (HTLV-1) and 2 (HTLV-2), hepatitis C virus (HCV), Epstein-Barr virus (EBV), human herpesvirus 6A (HHV6A) and 6B (HHV6B), and Kaposi's sarcoma-associated herpesvirus (KSHV) which were PCR-amplified and spotted on glass slides. They then hybridized their slides with Cy3-or Cy5labeled genomic DNA or cDNAs derived from various virus-infected cells. Their multivirus-chip was found to be highly specific and sensitive for detecting different viral genomic sequences in cell lines. Moreover, the chip could also detect the effect of various drugs on viral gene expression. In such instance, cell lines latently infected with HIV-1 and KSHV were used to generate profiles of viral gene expression in the presence of cyclin-dependent Schematic drawing of the multivirus-chip that possesses multiple functions kinase inhibitor (CKI), Roscovitine, which was applied to cells to suppress the reactivation of latently infected viruses. Ghedin, et al. [2] also studied the role of cellular chromatin structure on viral gene expression using their multivirus-chip. They employed the chromatin immunoprecipitation technique (ChIP) [3] to isolate cellular DNA fragments that were bound to phosphorylated histone H3 (P-H3). These DNA fragments were hybridized to the viral ORFs contained on the multivirus-chip to investigate the role of phospho-H3 on viral gene expression. They showed that whether transcriptionally active or silent the chromatin state played a role in regulating the expression of KSHV genes under the different cellular context. Current routine clinical diagnostics employ PCR, Southern blotting, Northern blotting, DNA sequencing and microarray hybridization to detect and characterize genes of interest in biomedicine. PCR is generally regarded as the most sensitive diagnostic method. However, Iyer, et al. [4] have shown that the sensitivity of cDNA-chip hybridization is comparable to that of TaqMan-driven quantitative PCR assay, and that the microarray hybridization technique is less likely to be complicated by high false positive rates due to carry-over contaminations. Furthermore, using microarrays, the viral gene transcripts in infected cells can be easily detected by hybridization without any prior amplification steps, and the microarray technique requires much less experimental material when compared to Southern or Northern blotting and can provide high sensitivity in the setting of large throughput. In view of the above, the multivirus-chip described in Retrovirology [2] holds several advantages over other more commonly used techniques (e.g. PCR, DNA sequencing) for the diagnosis of viral infections. First, this chip provides a more accurate diagnosis of viral infection by simultaneously evaluating the transcription of all viral genes, and can use such cumulative data to correlate infection with clinical disease manifestations. Second, the high throughput and flexible synthesis nature of DNA microarray construction can allow scientists to tailor-make and rapidly alter arrays to match evolving emergence of new pathogens. The SARS genome chip made by the US NIAID, NIH is a good example [5] of how diagnostic arrays can be developed quickly and be used in a timely manner. Finally, the most novel application described by Ghedin, et al. is their use of microarrays to correlate the cellular "histone code" [6] with the promoter activity of KSHV. Usually the transcription of a gene located on chromosomal DNA is influenced not only by the cis-acting ele-ments (or DNA-binding motifs), but also by the structure of chromatin. The latter can be vary depending on the post-translational modifications of histone proteins. Methylation, acetylation, and/or phosphorylation of certain amino acid residues at the amino terminal "tails" of histone H3 and/or H4 can indeed influence chromatin structure. Thus accumulating evidence has shown that chromatin-associated proteins and their modifications play vital roles in many physiological processes such as growth, differentiation, and development in mammals, plants and fungi [6, 7] . Many studies have used DNA array technology to investigate viral gene expression or to genotype viral isolates; however, none has used this technique to study the influence of cellular chromatin structure on viral gene expression [1]. Ghedin, et al. [2] demonstrated that only DNA fragments derived from ChIP of latent BCBL-1 cell genomic DNA captured using phospho-H3 antibody bound specifically to the KSHV ORF on the multivirus-chip. This result suggests that latent KSHV genome in BCBL-1 cells is packed into a nucleosomal structure and that histone H3 proteins near the viral promoter can be phosphorylated at serines to make the DNA at the promoter region less tightly packed with histones and more easily accessible to transcription factors. In conclusion, the multivirus-chip improvements developed by Ghedin, et al. [2] provide versatile clinical and basic uses. In the near future, such chips are likely to be used to detect viral co-infections in many different clinical settings.
15
Herpes simplex virus type 1 and normal protein permeability in the lungs of critically ill patients: a case for low pathogenicity?
INTRODUCTION: The pathogenicity of late respiratory infections with herpes simplex virus type 1 (HSV-1) in the critically ill is unclear. METHODS: In four critically ill patients with persistent pulmonary infiltrates of unknown origin and isolation of HSV-1 from tracheal aspirate or bronchoalveolar lavage fluid, at 7 (1–11) days after start of mechanical ventilatory support, a pulmonary leak index (PLI) for (67)Gallium ((67)Ga)-transferrin (upper limit of normal 14.1 × 10(-3)/min) was measured. RESULTS: The PLI ranged between 7.5 and 14.0 × 10(-3)/min in the study patients. Two patients received a course of acyclovir and all survived. CONCLUSIONS: The normal capillary permeability observed in the lungs argues against pathogenicity of HSV-1 in the critically ill, and favors that isolation of the virus reflects reactivation in the course of serious illness and immunodepresssion, rather than primary or superimposed infection in the lungs.
In some critically ill patients herpes simplex virus (HSV)-1 is isolated from the upper or lower respiratory tract [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] . Immunodepressed patients may be susceptible to transmission and acquisition of viral diseases; alternatively, viral reactivation may occur and may contribute relatively little to morbidity and mortality. Indeed, reactivation of human herpesvirus-6 is common in critically ill patients and does not worsen outcome [16, 17] . In immunocompetent patients, however, isolation of HSV-1 may be associated with viral pneumonia, even if reactivation rather than primary infection is responsible [6, 8, 18] . HSV-1 has been associated with acute respiratory distress syndrome (ARDS) and ventilator-associated pneumonia in the critically ill [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] , as either a primary or a superimposed infection. However, there are few reports of the virus eliciting an infectious host response, as demonstrated by a rise in serum antibodies, by bronchoscopic airway disease, by 'typical' findings on computed tomography of the lungs, or by the presence of giant cells or nuclear inclusion bodies on cytology or biopsy of the lower respiratory tract [3, 5, 9, 10, 18] . Indeed, Tuxen and coworkers [4] observed that prophylactic antiviral therapy in ARDS prevented respiratory HSV-1 emergence but it had no impact on duration of mechanical ventilation or on patient outcome. The pathogenicity of the virus therefore remains unknown, and the rare association in the critically ill of HSV-1 isolation with mortality may represent reactivation of the virus in immunodepressed patients with multiple organ failure and poor outcome [1, 2, 11, 14, 15] , rather than a symptomatic primary infection or superinfection contributing to death. Assessing pulmonary capillary protein permeability noninvasively at the bedside to yield the pulmonary leak index (PLI) could help in determining the extent of tissue injury, as was previously described [18] [19] [20] . This radionuclide technique involves gallium-67-labelled transferrin ( 67 Gatransferrin) and technetium-99m-labelled red blood cells ( 99m Tc-RBCs). In bacterial pneumonia, for instance, the PLI is elevated and the increase above normal directly relates to the severity of pneumonia, expressed as the lung injury score (LIS) [19] . In patients with acute lung injury (ALI) or ARDS during the course of bacterial pneumonia, the PLI is uniformly and greatly elevated above normal (up to 14.1 × 10 -3 /min) when LIS is greater than 2.5; the PLI is also elevated in 80% of patients with mild injury and a LIS between 1.5 and 2.5 [19] . Hence, the technique is a direct measure of permeability and an indirect measure of capillary injury in the lungs. The PLI is also elevated in interstitial lung disease [21] . In order to help differentiate between symptomatic and asymptomatic viral shedding and spread, which could inform the decision regarding whether to institute antiviral therapy and help in determining the pathogenicity of the virus, we measured the PLI in four consecutive critically ill patients with persistent pulmonary infiltrates of unknown origin on ventilatory support, in whom a HSV-1 had been isolated. We studied a small series of consecutive patients in whom respiratory secretions, sent for viral culture because of persistent pulmonary infiltrates of unknown origin, were found to be positive for HSV-1 (Table 1) . Tracheal aspirates or bronchoalveolar lavage fluid were transported directly to the microbiology laboratory or placed in viral transport medium (Copan Diagnostics Inc., Corona, CA, USA). For isolation of HSV-1, specimens were inoculated using standard procedures in triplicate flat bottom tubes on human embryonal lung fibroblasts and incubated at 37°C. Cultures were studied three times weekly for 10 days to identify the presence of a cytopathic effect. If a cytopathic effect, indicating the presence of HSV-1, was apparent or otherwise at days 2 and 7, the cells were fixed in methanol:acetone (1:1) and typed by immunofluorescence with labelled specific HSV-1 and HSV-2 antibodies (Syva Mikrotac HSV-1/HSV-2 typing kit, Palo Alto, CA, USA). In the four patients studied, the results were available within 3 days after samples had been inoculated in culture medium. On the day of specimen collection for viral culture, demographic, chest radiographic and respiratory data were recorded, as were clinical features. In three out of four patients on mechanical ventilation after intubation, the total respiratory compliance was calculated from ventilator settings as follows (ml/cmH 2 O): tidal volume/(plateau -end-expiratory pressure). From the radiographic score (ranging from 0 to 4 depending on the number of quadrants with radiographic opacities), the ratio of arterial oxygen tension to fractional inspired oxygen, the level of positive end-expiratory pressure and the compliance, the LIS was calculated [22] . (LIS ranges between 0 and 4, with values up to 2.5 denoting ALI and those above 2.5 ARDS.) None of the patients had visible oropharyngeal vesicles. To characterize further the persistent pulmonary infiltrates, the PLI was measured using a modification to a method described previously [19, 20] . Because this is a routine procedure, informed consent was waived. Autologous RBCs were labeled with 99m Tc (11 MBq, physical half-life 6 hours; Mallincrodt Diagnostica, Petten, The Netherlands), using a modified in vitro method. Ten minutes after injection of the labelled RBCs, transferrin was labelled in vivo, following intravenous injection of 67 Ga-citrate (6 MBq, physical half-life 78 hours; Mallincrodt Diagnostica). Patients were in the supine position, and two scintillation detection probes were positioned over the right and left lung apices. The probe system (manufactured by Eurorad C.T.T., Strasbourg, France) consists of two small cesium iodide scintillators (15 × 15 × 15 mm 3 ), each in a 2-mm tungsten and 1-mm aluminium housing cover (35 mm in diameter and 40 mm in height). The front end of each probe has an aluminium flange attached (3 mm in thickness and 70 mm in diameter) to facilitate easy fixation to the patient's chest with tape. Each probe weighs approximately 255 g. The probe signals are led into a dual amplifier, from which the output is fed into a multichannel analyzer system connected to a personal computer. Because the probes have separate channels, there is no electronic crossover. Starting at the time of the intravenous injection of 67 Ga, radioactivity was measured each minute for 1 hour. For each measurement interval, the entire spectrum of photon energies was stored on disk. During processing, the 99m Tc and 67 Ga -and plotted against time. The PLI was calculated, using linear regression analysis, from the slope of increase of the radioactivity ratio divided by the intercept, in order to correct for physical factors in radioactivity detection. By taking pulmonary blood volume and thus presumably surface area into account, the radioactivity ratio represents the ratio of extravascular to intravascular 67 Ga radioactivity. The PLI represents the transport rate of 67 Ga-transferrin from the intravascular to the extravascular spaces in the lungs, and it is therefore a measure of pulmonary capillary permeability to transferrin [19, 20] . The mean PLI from the two lungs was taken. The upper limit of normal PLI is 14.1 × 10 -3 /min. Where appropriate, numbers are summarized as median (range). Patient data are presented in Table 1 . The patients had stayed for some time in the hospital or intensive care unit before HSV-1 was isolated, and they had been admitted primarily because of respiratory insufficiency during the course of pneumonia. Patient 4 was admitted into the coronary care unit a few days before intensive care unit admission for cardiogenic pulmonary oedema. All patients had been dependent on mechanical ventilatory support for some time before sampling. They had received adequate antibiotic therapy for pneumonia and had ALI at the time of sampling, which was of otherwise unknown origin. Table 1 shows that patients had radiographic abnormalities but without an increased PLI. Central venous pressure was not elevated, which suggests that the persistent pulmonary infiltrates were not caused by overhydration. In patients 1 and 3 a high-resolution computed tomography scan of the lungs with contrast was obtained; the findings were nonspecific, however, with alveolar consolidations and pleural fluid, even in the presence of interstitial abnormalities with a ground glass appearance in patient 3. In patient 1 a bronchoscopy was performed and there were no mucosal lesions. There was a normal distribution of lymphocyte subtypes in the lavage fluid. A transbronchial biopsy revealed interstitial inflammation with many macrophage deposits, and immunohistochemical staining for HSV-1 was negative. No multinucleated cells or cell inclusions were observed, either in bronchoalveolar lavage fluid from patient 1 or in tracheal aspirates from the other patients. In patients 1-3 concomitant isolation of bacteria by culture was regarded as bacterial colonization. Antibody testing was not done in patients 2-4 but was found to be positive for anti-HSV-1 IgG in patient 1, which is indicative of prior HSV-1 infection. The antiviral agent aciclovir (10 mg/kg three times daily) was started when cultures became positive in two patients, at the discretion of the treating physician. Aciclovir was withheld in the other two patients because it was presumed that the pulmonary infiltrates were not caused by HSV-1, on the basis of a normal PLI among other findings. In patient 1, who had a normal PLI, a course of steroids was initiated on the day after the PLI was measured, and was continued despite positivity for HSV-1, reported 5 days later. All patients survived until discharge from the intensive care unit. The 67 Ga-transferrin PLI is a sensitive and specific measure of pulmonary capillary permeability, which is utilized for noninvasive assessment of severity of a broad range of pulmonary conditions [19] [20] [21] . The PLI roughly parallels clinical severity (i.e. the LIS) [19, 20] . Although it involves the use of relatively routine equipment, the diagnostic method has not gained broad application, partly because of its laborious nature [20] . It has the advantage that bedside measurements are possible in mechanically ventilated critically ill patients, who cannot easily be transported. Pulmonary inflammation, of whatever cause, increases the PLI up to four times normal values in the most severe forms of lung injury, including ARDS. In less severe injury, such as impending ARDS and interstitial lung disease, the PLI is also elevated, albeit to a lesser extent, as reported by us and other groups [20, 21] . The patients had in common a prior infectious episode, followed by a relatively prolonged period of respiratory insufficiency. They had persistent and nonspecific pulmonary infiltrates of unknown origin, after treatment of their primary disease, which prompted viral culture. The normal PLI observed suggests the involvement of a relatively harmless reactivation of HSV-1, rather than the presence of a primary and damaging infection. Indeed, critically ill patients with sepsis may have late immunodepression, with lymphocytic apoptosis, lymphocytopenia and T-cell anergy, promoting viral reactivation [23, 24] . Apparently, the virus must have been latent in the nerve endings of the mucous membranes of the upper respiratory tract in these patients [2, 15] . Herpesviruses (HSV-1) have frequently been isolated in vivo from respiratory secretions of patients with ARDS [3, 4] and detected in surveillance cultures from the respiratory tract of patients following burns, trauma, transplantation, major surgery and others. However, these viruses are detected in only 3% of lung biopsies from patients with prolonged and unresolving ARDS [3, 7, [9] [10] [11] [12] [13] 15] . The literature is thus widely divergent on the precise role of the virus in pulmonary disease in the critically ill and its contribution to patient morbidity and mortality [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] . We believe that the tracheal aspirates were representative of lower respiratory tract secretions, in the absence of herpes orolabialis and oral epithelial cells in smears for Gram stain of the secretions. Concurrent colonization with other pathogens has previously been described [5, 13] . Because there was no overlap in the duration of stay of the patients, transmission of the virus from one patient to another can be excluded. This further suggests that respiratory HSV-1 infections in the critically ill may result from relatively harmless endogenous reactivation. Although the normal PLI argues against pulmonary parenchymal pathogenicity, tracheobronchitis caused by the virus [18, 25] cannot be ruled out, even in the absence of orolabial lesions, because bronchoscopy was not performed in three of the four patients, even though it was unremarkable in patient 1. The persistent pulmonary infiltrates in our patients may thus relate to slow radiographic resolution of prior bacterial or aspiration pneumonia, rather then superimposed infection. Moreover, computed tomographic images of the lung may be largely nonspecific [26] , and so the precise diagnostic criteria for HSV-1 pneumonia remain unclear. When properly standardized, for instance with respect to cell numbers in bronchoalveolar fluid or tracheal aspirates, quantitative cultures, viral RNA and DNA by polymerase chain reaction, could be helpful together with the PLI in further studies to quantitate viral load and the ratio of replication to shedding, and therefore the pathogenicity of the virus in the lower respiratory tract. In conclusion, the anecdotal data presented here suggest that isolation of HSV-1 from respiratory secretions in the critically ill patient with a persistent pulmonary infiltrate may warrant evaluation of tissue injury potentially caused by the virus to judge its pathogenicity. This could be done using a radionuclide PLI measurement, and would help to inform decisions regarding antiviral therapy, which may have adverse effects. In some patients a normal PLI may argue against viral pathogenicity, and withholding of aciclovir in such patients may be safe.
16
Logistics of community smallpox control through contact tracing and ring vaccination: a stochastic network model
BACKGROUND: Previous smallpox ring vaccination models based on contact tracing over a network suggest that ring vaccination would be effective, but have not explicitly included response logistics and limited numbers of vaccinators. METHODS: We developed a continuous-time stochastic simulation of smallpox transmission, including network structure, post-exposure vaccination, vaccination of contacts of contacts, limited response capacity, heterogeneity in symptoms and infectiousness, vaccination prior to the discontinuation of routine vaccination, more rapid diagnosis due to public awareness, surveillance of asymptomatic contacts, and isolation of cases. RESULTS: We found that even in cases of very rapidly spreading smallpox, ring vaccination (when coupled with surveillance) is sufficient in most cases to eliminate smallpox quickly, assuming that 95% of household contacts are traced, 80% of workplace or social contacts are traced, and no casual contacts are traced, and that in most cases the ability to trace 1–5 individuals per day per index case is sufficient. If smallpox is assumed to be transmitted very quickly to contacts, it may at times escape containment by ring vaccination, but could be controlled in these circumstances by mass vaccination. CONCLUSIONS: Small introductions of smallpox are likely to be easily contained by ring vaccination, provided contact tracing is feasible. Uncertainties in the nature of bioterrorist smallpox (infectiousness, vaccine efficacy) support continued planning for ring vaccination as well as mass vaccination. If initiated, ring vaccination should be conducted without delays in vaccination, should include contacts of contacts (whenever there is sufficient capacity) and should be accompanied by increased public awareness and surveillance.
Concerns about intentional releases of smallpox have prompted extensive preparations to improve our ability to detect and respond to an outbreak of smallpox [1, 3, 4, 2] . Many factors contribute to the public health challenge of understanding and preparing for smallpox, including the age and quality of epidemiological data on native smallpox and the smallpox vaccine, the difficulty of extrapolating that data to our current populations, the possible terrorist use of altered smallpox, our ignorance of terrorist methods of release, and the relatively high risk of adverse events caused by the smallpox vaccine. The Centers for Disease Control and Prevention (CDC) established ring vaccination (selective epidemiological control [5] ), a strategy in which contacts of cases are identified and vaccinated, as the preferred control measure in the event of a smallpox outbreak (interim plan). The successful use of ring vaccination during the smallpox eradication campaign and its logical emphasis of case-contacts for immediate vaccination support its use (though the attribution of the success of the eradication program to ring vaccination has been challenged [6] ). Health Officers should initiate ring vaccination upon identification of the first cases of smallpox. However, there are legitimate concerns regarding the ability of public health practitioners to mount a quick, comprehensive and successful ring vaccination program, particularly in the face of a moderatesized or large smallpox outbreak. To guide preparation efforts and inform incident decision-making, we attempt to identify outbreak characteristics and response capacities that significantly impact the ability of ring vaccination to control a smallpox outbreak and to determine whether ring vaccination is useful in the presence of a mass vaccination campaign. Our analysis uses a newly developed mathematical model: a continuous-time, event-driven network simulation model of smallpox ring vaccination. Mathematical models can advance our understanding of how a smallpox outbreak might progress. Several mathematical and computer models address the question of smallpox transmission [7] [8] [9] [10] [11] [12] [13] . The first model to appear [8] concluded that ring vaccination would be effective, but did not treat response logistics in detail; the model was linear and did not treat the depletion of susceptibles as the epidemic progressed (appropriate, however, for assessing control early in an epidemic, when the number infected is small compared to the number of susceptibles, e.g. [14] ). The innovative model by Kaplan et al. [9] emphasized the importance of resource limitation and the logistics of smallpox response, but assumed that full infectiousness began before the onset of symptoms (and the subsequent identification and removal), and did not separately monitor close epidemiological contacts of patients (which are at greatest risk, but also easiest to find and vac-cinate); the conclusions were highly critical of ring vaccination. The model by Halloran et al. [11] , a stochastic, discrete-time network model omitted the explicit inclusion of response logistics while otherwise used parameter values similar to those in Kaplan et al. [9] ; the inclusion of residual immunity from individuals vaccinated prior to the discontinuation of routine vaccination, however, led to a more favorable view of ring vaccination. The model by Bozzette et al. [12] assumed that ring vaccination would reduce the number of transmissions and focused on health care workers (but did not explicitly include the network structure of the population nor the response logistics of ring vaccination). The model by Eichner [15] did not explicitly include the network structure of the population nor the logistics of ring vaccination, but did use parameters based on data from an outbreak in Nigeria, and did distinguish close and casual contacts, case isolation, and surveillance of contacts; it concluded that case isolation and contact tracing could prevent the spread of smallpox. Finally, the individual-based model by Epstein et al. [16] presented scenarios illustrating certain alternatives to pure mass vaccination and ring vaccination of contacts of cases in preventing smallpox transmission in small populations of 800 individuals; this model includes no homogeneity assumptions, but did not analyze tracing of contacts of contacts. Because none of the available models includes both network structure (with explicit contact tracing) and response logistics limited by the number of available disease control investigators [9] , we included these features in a continuous-time event-driven network simulation model of smallpox ring vaccination. Specifically, the model we developed includes the following features: exposed individuals and vaccinate them in time, resulting in a "race to trace" [9] . Mild, ambulatory cases of smallpox may spread disease because such cases may be harder to recognize. Vaccination of individuals prior to the discontinuation of routine vaccination may provide some, possibly considerable, protection against infection [11, 23, 24] , although it may also result in more mild cases which may be harder to detect. Public awareness may lead to more rapid detection of cases. We use this model to determine what factors promote or hinder the success of ring vaccination during a smallpox outbreak, and whether ring vaccination is useful in the presence of a mass vaccination campaign. In particular, the goal of this paper is to examine the control of smallpox by contact tracing and ring vaccination using a network model which includes response logistics [9] . Natural history of smallpox We briefly review relevant features of the natural history and epidemiology of smallpox [17, [25] [26] [27] 8, 28] . Following infection by the variola virus, individuals exhibited an incubation period of approximately 7-19 days with 10-14 being most typical. Sudden onset of fever and malaise, often with accompanying headache and backache, began the initial (or pre-eruptive) phase of smallpox. After 2-3 (or perhaps 4) days, individuals with the most common form, ordinary type smallpox, developed the characteristic focal rash, preceded in many cases by oropharyngeal lesions. In fatal cases of ordinary smallpox, death often occurred between the tenth and sixteenth day of symptoms; among survivors, most scabs had separated by day 22-27 of illness [26] . The course of smallpox varied widely between individuals, and several different clinical classifications were developed [29] [30] [31] 17, 26] . Consideration of the clinical features and severity of smallpox is important from the standpoint of mathematical transmission modeling because (1) the clinical features affect the ease of diagnosis (and thus of case identification), (2) more severe forms of smallpox may result in more transmission, (3) vaccinated individuals may develop less severe disease. We utilize a modified or simplified version of the classification system developed by Rao [32, 31, 26] ; for the mathematical model, we will classify smallpox into five categories: early hemorrhagic, flat and late hemorrhagic, ordinary, modified, and mild. However, the clinical features and severity of smallpox in different populations may have been affected by underlying host factors, differences in viral strains, or differences in the infectious dose owing to different prevailing modes of transmission, and thus robust and precise quantitative estimates of the effects of (pre-or post-exposure) vaccination on the resulting smallpox severity, or of the infectivity differences between individuals exhibiting different forms of smallpox, are not available. The significance of such differences will be revealed through sensitivity analysis. Further details are given in Appendix 1 [see Additional file: 1]. Vaccination with vaccinia virus provided substantial protection against infection. Dixon assessed the risk of infection for an individual successfully vaccinated 3 years prior to exposure to be 0.1% the infection risk of an unvaccinated individual [17] . However, smallpox vaccination did not always take when applied, and moreover, in many instances, individuals who experienced a repeated vaccination failure developed severe smallpox upon exposure. The probability of a successful take depended on the vaccination method used; we assume that the take rate is between 95% and 100% [22, 28] . In addition to protection against infection, vaccination could in many cases modify the course of infection and reduce the severity. Vaccine protection waned over time, but individuals vaccinated 20 years prior to exposure were believed to still have half the infection probability that an unvaccinated person had [17] , and to have some protection against the most severe manifestations of smallpox. Dixon [17] believed that vaccine protection had at least three components, which decayed at different rates; for the purpose of this paper, we will assume that the severity of smallpox in previously any (recently or otherwise) vaccinated individuals follows the same distribution as for the vaccinated subjects seen in the case series observed by Rao in Madras [26] , except that anywhere from 0 to 5% of vaccinated subjects develop smallpox too mild to diagnose without special surveillance or awareness. Observe that the vaccinated cases studied by Rao were vaccinated (at some point in their lives) before exposure, rather than after exposure to smallpox. Smallpox was largely a disease of close contacts [17, 26, 33] , spread primarily through face to face contact with an infected person (or occasionally through contaminated clothing). Individuals in the incubation period of smallpox were not infectious, and long term carriers did not exist. Patients were believed to be infectious following the development of oropharyngeal lesions, which could precede the rash by 24 hours [26] . However, patients were believed to be most infectious during the first week of the rash [26] ; Dixon (1962) believed that patients could be infectious from the onset of acute viremia, but most evidence suggested that little transmission occurred prior to the development of the rash [26, 33] . The more severe the case, the more infectious they appeared to be [34] ; mild cases were believed to have very little infectiousness. While scabs contained infectious material and patients were considered to be infectious until the last scab fell off, in practice patients were not highly infectious during the scabbing phase. Importantly, patients who had been vaccinated were found to cause fewer secondary cases [34] . Very severe cases, such as hemorrhagic or flat smallpox, occasionally resulted in considerable transmission, owing to diagnostic difficulties; mild cases, in which the patient remained ambulant during the course of the disease, could cause considerable spread as well [35, 36] . Within a household or family dwelling, the secondary attack rate of unvaccinated susceptibles depended on the time and place, occasionally below 50% [29] , but often approaching 100% [37] . Drier conditions were often believed to favor transmission [17, 27] , so that lower rates of transmission derived from tropical regions may not be applicable to the temperate zone [38] . The number of secondary cases resulting from a given importation into Europe varied widely [39] , with most importations yielding few cases, but with the occasional large outbreak being seen. Mathematically, we represent the course of smallpox according to Figure 1 . We distinguish eight epidemiologically relevant states: (1) just following exposure, during which time vaccination could afford complete protection against disease, (2) a period of several days during which vaccination will not prevent disease, but may still reduce the severity of disease, (3) still prior to the development of symptoms, but too late for vaccination, (4) the beginning of the pre-eruptive period, during which the patient exhibits fever, malaise, and possibly other symptoms, but is not yet infectious, (5) a short period prior to the appearance of the rash, during which the appearance of oropharyngeal lesions will permit variola transmission, (6) the first week of the rash, during which time the patient is most infectious, (7) and (8), succeeding stages of the rash, during which time the patient is less infectious. For each of these states, we assume that conditional on surviving, the waiting time until the next stage is chosen from a uniform distribution as indicated in Appendix 2 [see Additional file: 2], except that the incubation period (the time from infection until Stage 4) is derived from estimates of the incubation distribution of smallpox based on importation cases in Europe [26] (see Appendix 2 [see additional file 2] for details). We chose to sample from a uniform distribution as a simple way to ensure a minimum waiting time in each state; many alternatives to this choice are possible. We simulate the transmission of smallpox on a "smallworlds" network (highly clustered, but with short characteristic path lengths) [40] . Specifically, we assume that each person is located in a single household, and that the transmission rates were greatest in the household. We also assume that a fraction of the population are grouped into workplace or social groups, in which transmission may also occur, but with a lower rate per unit time than for household contacts. Finally, we assume that with a still smaller probability, any individual may transmit infection to any other individual in the population (casual contacts). In general, in a network-structured model, the number of secondary cases caused by an index case in a completely susceptible population is not a useful index of epidemic potential [41, 42] (for a simple example, see [43] ), since (for instance) an individual could infect everyone in his or her household, and not cause a widespread epidemic unless between-household transmission were sufficiently frequent. Rather than constructing the appropriate generalized basic reproduction number for our model (leading to highly cumbersome expressions), we chose an alternative (ad hoc) index of epidemic potential. For any given scenario of interest, we simulated the introduction of 10 index cases at random into a population of size 10000, and operationally defined "containment" to occur whenever the final size of the epidemic was less than 500 cases within 250 days (we showed, in the discussion of Figure Smallpox stages used in the simulation model Figure 1 Smallpox stages used in the simulation model. Flat and ordinary smallpox rashes are indicated with more dots than modified and "mild" smallpox, suggesting potentially greater infectiousness. Hemorrhagic smallpox is indicated by horizontal line shading. Further details are provided in Table 6 . 5A below, that in nearly all cases, the 250-day window differs very little from a 1000-day window). Because we simulate a disease with a finite duration on a finite and nonrenewing population, epidemic extinction always occurs in finite time. We assume that even in the absence of specific case investigations, the presence of smallpox symptoms will prompt patients to be diagnosed; we assume, however, a higher diagnosis rate for all forms of ordinary smallpox than for the severe flat and hemorrhagic forms, or for the mildest form. We assume that once an individual is diagnosed, their household and workplace contacts are investigated and detected with some probability; we assume that a high fraction (such as 95%) of household contacts are assumed to be traceable (see below). We assume that the fraction of workplace/social contacts that are traceable is less than the fraction of household contacts that are traceable; we assume that no casual contacts are traceable. High contact-finding rates may be plausible; we examined San Francisco Department of Public Health records of contact investigations for meningococcal disease (like smallpox, a potentially fatal disease for which rapid intervention may prevent mortality and morbidity). Records were available from December 2001 to April 2002; 13 such investigations during this period resulted in identification of 62 household contacts, all of which were contacted; out of 38 workplace/social contacts identified, 32 were contacted (84%). In our model, we assume that identified asymptomatic contacts are vaccinated, quarantined, and monitored for symptom development, while symptomatic patients are isolated and treated as necessary [9] ; thus, the modeled interventions include more than ring vaccination alone. Finally, we include the possibility that all contacts (of both symptomatic and asymptomatic) traced and the same procedure applied, so that all contacts of contacts would be investigated. We assume that uninfected or asymptomatic individuals who are visited or traced individuals will be diagnosed more rapidly than if they had not been traced; in fact, such individuals would be isolated and would not be able to continue a chain of transmission. We follow previous models [9] in assuming a limited vaccination capability of K r per day for ring vaccination. We assumed one of two strategies for contact tracing: (1) tracing only of direct contacts of diagnosed cases, and (2) tracing of contacts of contacts of diagnosed cases as well as direct contacts. The contact structure of the network is illustrated in Figure 2 . Observe that individuals b and c are household contacts of individual a, so that if individual a were a smallpox case, an attempt would be made to find and vaccinate individuals b and c as household contacts of a case. If individuals a and b were both cases, then two attempts could be made to find individual c. We have modeled the effect of multiple contact-finding attempts conservatively in the sense that if the first attempt to find an individual as a household contact (of a case or of a contact) is determined to fail, no further attempts will be made. This maintains the failure rate of contact tracing (looked at from the standpoint of finding individuals) even in large households. Similar considerations apply to workplace/social groups. Figure 2 Network structure shown for households (joined by thick lines) of size 3 and workplace/social groups of size 4 (joined by thin lines); a small portion of the network is shown. Individual a has two household contacts (b and c), and three workplace/social contacts (d, e, and f). If individual a were a smallpox case, the household contacts would be at highest risk for acquiring smallpox, followed by workplace/social contacts; all individuals in the population are at a low risk of casual transmission from individual a. Case investigation of individual a would identify the direct contacts b-f with probabilities that depend on whether the contact is household or workplace/social; if such individuals are identified, they will be vaccinated. If contacts of contacts are being traced, the investigation will subsequently identify individuals g-p. We analyzed the model in three ways. First, we selected a Latin Hypercube sample [44] [45] [46] of parameters chosen uniformly from the parameter ranges given in Appendix 2 [see additional file 2] , and simulated the transmission and control of smallpox to determine which parameters were most important for contact tracing and ring vaccination to be effective. Second, we used the same Latin Hypercube Sample of input parameters, but assumed that all disease control efforts were inactive. We used these parameters to simulate smallpox transmission, but then iteratively selected transmission parameters so that (1) between 1% and 10% of new infections resulted from casual (random) transmission, and (2) each index case resulted in between two and five secondary cases (thought to be plausible for historic smallpox; [8] suggest three secondary cases). For each of the resulting smallpox parameter sets using 100 stochastic simulations per set, we determined the daily ring vaccination/case tracking capacity needed to contain all simulated smallpox epidemics (i.e., keep the total number of cases below 500 within 250 days). Third and finally, we chose parameter values to yield an moderately large smallpox epidemic (with each index case causing approximately six secondary cases), and present illustrative scenarios for ring vaccination. These scenarios are intended to complement the simulations which were calibrated to historic smallpox, since the characteristics of smallpox that may be used in a deliberate release are not known. It is important to realize that in our model, the case finding time determines the fraction of contacts that will become infected, and that our model parameters have been chosen to yield quite rapid transmission to close contacts; in reality, much transmission of natural smallpox occurred through "sickbed" routes which would not occur in a modern setting [47] , so that in this sense our model errs considerably on the side of caution and pessimism. To determine which of the input parameters were most important in determining the total number of smallpox cases, we selected a Latin Hypercube sample of size 1000 from the input parameter ranges indicated in Appendix 2 [see additional file 2] and simulated the mean number of cases within 250 days in a population of 10000. We then computed the partial rank correlation coefficient [46] (PRCC; see Appendix 2 [see additional file 2]) between each input parameter and the number of smallpox cases; when the PRCC is close to zero, the value of the parameter has little relation to the simulation output; when the PRCC is close to +1.0 or -1.0, the value of the parameter is highly important in determining the simulation output. Neglecting the number of index cases (which is directly related to the number of new cases), those parameters whose PRCC exceeds 0.1 are shown in Table 2 . Most of these parameters identified as important are related to the density of available contacts (mean household size, prior vaccination fraction, and protection due to prior vaccination) or the transmission rate and infectivity (including the length of the pre-eruptive infectious period (stage 5 in Figure 2 )). Note, however, that the speed of ring vaccination (household tracing delay) and faster diagnosis due to awareness of the outbreak are important parameters. Additionally, the infectivity of mild cases appears as an important parameter as well. To explore factors which contribute to the success of ring vaccination, we chose smallpox scenarios which resulted in severe and fast-moving epidemics in the absence of disease control; these simulated epidemics are considerably more severe than is believed likely under present circumstances. We used these parameters to simulate smallpox epidemics beginning with 10 cases, for a variety of levels of ring vaccination capacity per day (contact tracing capacity per day), as shown in Figure 3A . In this Figure, we assume that the population size is 10000, and that the epidemic began with 10 infected individuals. The mean household size is assumed to be 4, the mean size of the workplace/social contact group is 8, and contacts of contacts are traced. We assume that each day, the number of contacts that can be traced and vaccinated as a result of case investigation is 0, 10, 20, 30 and 40 per day; the probability of finding a workplace/social contact is assumed to be 80%. The Table 1 . Because we assumed nonzero diagnosis probabilities during the prodromal period for all individuals in Figure 3A , we repeated the simulation assuming no diagnosis in the prodromal period unless individuals were under specific surveillance. The results were nearly identical: assuming 30 contact tracings (ring vaccinations) per day, we found 26% of the scenarios in Figure 3A exhibited decontainment, and 28% assuming no diagnosis during the prodromal period; assuming 40 contact tracings per day, we found 1 out of 100 scenarios showed loss of containment in Figure 3A and when we repeated the scenario of Figure 3A assuming no diagnosis during the prodromal period. In Figure 3B , we illustrate control of an epidemic for which all parameters are identical to Figure 3A , except that the workplace/social group size is 12 (instead of 8, as in Figure 3A ), and the probability of finding workplace/ social contacts is 60% (instead of 80%, as in Figure 3A ). In this case, the larger size of the workplace/social groups and the lower contact finding probability makes it necessary to have a higher ring vaccination capacity to attain a high probability of containing the epidemic, and on average it takes longer for eradication to finally occur. Finally, in Figure 3C , we show control of an epidemic in a population of 100,000, beginning with 1000 initial infectives, keeping all other parameters the same as in Figure 3A . Each curve corresponds to the indicated number of possible ring vaccinations per day. This figure shows that assuming sufficient capacity, ring vaccination is in principle capable of containing even epidemics beginning with very many infected individuals. However, mass vaccination in such cases is justified because of the far larger number of individuals at risk and the inability to perform such extensive contact tracing. In Figure 3D , we compare the effect of tracing contacts of contacts (as described in Appendix 2 [see additional file 2]) at different levels of ring vaccination capacity. Thin Figure 5A , 5B φ Prob. of finding household contact 0.95 Delay, tracing household contacts 1-5 days Expanding severe smallpox epidemic Figure 3 3A -Expanding severe smallpox epidemic beginning with 10 initial cases, assuming 0, 10, 20, 30, and 40 possible ring vaccinations per day. The household size is 4 and the workplace/social group size is 8; we assume 95% of household contacts are traceable (with a mean delay of 1 day) and 80% of workplace/social contacts are traceable (with a mean delay of 2 days). We also assume that 25% of the population have 50% protection from infection resulting from vaccination prior to the discontinuation of routine vaccination. We assume that infection will be transmitted to close contacts with a mean time of 0.2 days, and that each person while infective causes on average 0.15 casual (untraceable) infections per day. We assume that individuals are 20% as infectious in the day just before the appearance of the rash as they will be during the first week of the rash, and that individuals are 20% as infectious as this (4% as infectious as during the first week of the rash) during the prodromal period. We assume that diagnosis rates will increase by a factor of 50% after smallpox becomes known to the community; we assume that each individual contacted during an investigation has a additional diagnosis or removal rate of 0.75 per day following the onset of symptoms (reflecting enhanced surveillance or contact isolation). Important parameters are summarized in Table 1 ; the full set of parameter choices is outlined in Tables 8-11 in Appendix 2 [see additional file 2] . Diagnosis times are discussed in Appendix 2 [see additional file 2]. 3B -An expanding severe smallpox epidemic under inadequate ring vaccination is shown for parameters identical to Figure 3A , except that workplace/social group sizes are 12 (instead of 8), and the probability of tracing workplace/social contacts is 0.6 (instead of 0.8). 3C -A severe smallpox epidemic is controlled by ring vaccination despite the large number of initial cases. The parameters are identical to Figure 3A , except that 1000 index cases inaugurate the attack in these scenarios (and ring vaccination capacity is much greater, as indicated). While not recommended, ring vaccination may ultimately halt epidemics beginning with many index cases if sufficient vaccination capacity were available, contact finding feasible, and follow-up sufficient. 3D -Tracing contacts of contacts (red) is beneficial when sufficient contact tracing/ring vaccination capacity exists (dotted lines). In these scenarios, all parameters are the same as in Figure 3A ; the number of contact tracings possible per day is either 20 or 40 per day. Contacts of contacts are traced in two scenarios; in the other two, only direct contacts of cases are traced. For low levels of ring vaccination (20 per day), tracing contacts of contacts is harmful; for high levels (40 per day) of ring vaccination, it is beneficial to trace contacts of contacts. When the contact tracing/ring vaccination capacity is too small to adequately cover contacts of the cases themselves, diversion of resources to contacts of contacts is harmful; however, provided that sufficient capacity exists, tracing contacts of contacts helps outrun the chain of transmission. Each line corresponds to the average of 100 realizations. The average number infected on each day is plotted in the Figure. The figure illustrates that when ring vaccination capacity is low, tracing contacts of contacts (as modeled) yields a more severe average epidemic; when ring vaccination capacity is large, tracing contacts of contacts results in a less severe average epidemic; if the contact tracing/ring vaccination capacity is too low to cover adequately the contacts of contacts in addition to the contacts of cases, extension of tracing to the contacts of contacts (the second ring) is harmful; however, if there is sufficient capacity to cover the contacts of contacts, then the tracing of contacts of contacts is beneficial. Finally, in Figure 4 , we illustrate the considerable variability that may be seen from simulation to simulation. This figure shows twenty simulations when contacts of contacts are not traced. Stochastic variability between realizations is considerable, even when all parameters are held constant; this variability is expected to limit the ability to make inferences based on observation of a single realization of the process. Because our baseline hazard for infection of individuals may be larger than would be expected for naturally occurring smallpox, we examined the effect of more realistic values of this hazard. In particular, we chose different levels of ring vaccination capacity (10, and 20) , and of the relative hazard for workplace/social contacts, and then chose values of the baseline hazard for infection varying from 0.5 per day (for a mean time to infection of 2 days) to 2 per day (for a mean time to infection of one half day), Table 3 : Estimated decontainment probability for different levels of ring vaccination capacity (Kr) and relative hazard for infection due to workplace/social contacts (h2), for different levels of the baseline hazard for infection from household contacts λ (based on replications of 100 simulations for each level). For each scenario, 10 index cases were introduced into a population of size 10000. All other parameters were the same as for Figure 3A . As before, we define decontainment to mean that the total number of cases from 10 index cases eventually exceeded 500 by day 250. and introduced 10 index cases into a population of 10000. We then repeated this 100 times, and reported the fraction of scenarios in which the number of infections ultimately exceeded 500 (as before, chosen as a cutoff to indicate the ultimate "escape" of containment of the epidemic). These results, shown in Table 3 , support the idea that ring vaccination can easily control introduced smallpox provided there is sufficient capacity and efficacy of tracing. Because of considerable uncertainty in the model parameters, we chose a collection of parameter values, and for each, estimated the containment probability (operationally defined as fewer than 500 total cases as a result of 10 index cases, within 250 days). We estimated this containment probability by simulating the smallpox epidemic 100 times for the same parameter values, and computing the frequency out of these 100 realizations for which fewer than 500 index cases resulted within 250 days. (Using a 1000 day window produces slightly smaller containment estimates; for 3 out of 1000 parameter set choices, this difference was greater than 0.06; the maximum difference seen was 0.23; the mean absolute difference was 0.0029; in only one case out of 1000 did we see containment in all 100 cases for the 250-day window, but not in all 100 cases for the 1000-day window). One thousand scenarios chosen from a Latin Hypercube sample were analyzed, and as indicated before, we chose the hazard for close contact transmission and the hazard for random transmission to guarantee that between 2 and 5 secondary cases per case occur, and that no more than 5% of cases are attributable to random transmission (we refer to this set as the "calibrated" scenarios further in this text). Having chosen this collection of 1000 parameter sets, we considered two levels of two different control parameters which were applied to each (so that each of the 1000 parameter sets were simulated under four different control conditions). The first of the two control parameters was the probability of workplace/social group contact finding; we chose values of 0.8 and 0.9 for this parameter (the household contact finding probability was 0.95 in all cases). The second of the control parameters was the rate of diagnosis (and effective removal) from the community of cases developing among previously identified and traced contacts who were initially asymptomatic (we refer to this as the monitored diagnosis rate); we assumed first a low level corresponding to a mean diagnosis time of one day from the onset of symptoms, and a high level corresponding to a mean time of 3 hours from the onset of symptoms (high levels of the monitored diagnosis rate correspond effectively to isolation of contacts). Finally, we assumed a doubling of the diagnosis rate after the beginning of widespread community awareness of smallpox. We then computed the containment fraction at different levels of ring vaccination capacity (contact tracing capacity per day). Thus, for each of 1000 scenarios (parameter set choices), we assigned the workplace/social group contact tracing success probability (υ 2 ), the monitored diagnosis rate φ (Appendix 2 [see additional file 2]), and the contact tracing/ring vaccination capacity per day (K r ). We then performed 100 realizations beginning with 10 index cases, and computed the containment fraction (fraction showing fewer than 500 cases in 250 days, beginning with 10 index cases). Thus, for each of the two choices each of υ 2 and φ, and for each value of K r we examined, we obtained 1000 values of the containment fraction. We use the resulting distributions in Figure 5A (averaging over these 1000 containment fractions), and Figure 5B (displaying the minimum value of the 1000 containment fractions). In Figure 5A , we plot the mean containment fraction (averaging the containment fraction over all 1000 scenarios), as ring vaccination capacity varies, for the two levels of workplace/social group contact finding probabilities (0.8 and 0.9), and for the two levels of monitored diagnosis rate among initially asymptomatic contacts (1 day -1 and 8 day -1 ). For low levels of ring vaccination (traceable contacts per day), the epidemic is almost never contained, but for ring vaccination levels near 50-60 per day (5-6 per index case per day), the average containment fraction Figure 4 Stochastic variability is illustrated by plotting the number of infectives over time over multiple replications. In this example, most simulations exhibit rapid containment of smallpox. The mean number of cases (averaging over simulations) is influenced by a small number of simulations exhibiting an uncontained epidemic. The parameters are the same as in Figure 3A , except that contacts of contacts are not traced in these replications. The mean containment probability Figure 5 5A -The mean containment probability increases as the number of ring vaccinations per day is increased. For this figure, the 1000 "calibrated" parameter sets were chosen, and for each parameter set, 100 realizations were simulated and the fraction of these for which the epidemic was contained to fewer than 500 cases was determined. The average of these 1000 containment fractions is plotted on the vertical axis. We assumed a household contact finding probability of 95% and that the diagnosis rates double after community awareness of the epidemic. We considered high levels of workplace/social (w/s) contact finding (0.9), as well as moderate levels (0.8). We also considered two levels of diagnosis of smallpox among investigated (alerted) contacts: high levels (corresponding to a 3 hour mean delay, indicated by "high contact isolation"), and moderate levels (corresponding to a one day delay, and indicated by "less contact isolation"). The figure shows four such conditions, a. high workplace/social contact finding probability and high contact isolation, b. moderate workplace/social contact finding probability and high contact isolation, c. high workplace/social contact finding probability and less contact isolation, and d. moderate workplace/social contact finding probability and less contact isolation. All other parameter values were chosen from the uncertainty analysis (the 1000 "calibrated" parameter sets). In this figure, "contact isolation" refers to the monitored diagnosis rate, i.e. the rate at which previously asymptomatic contacts who subsequently develop disease will be diagnosed (φ, Table 1 , Table 8 ). 5B -The minimum containment probability out of the same 1000 scenarios chosen in Figure 5A . Whereas in Figure 5A , we averaged the simulated containment frequency (out of 100 realizations for each scenario), in this figure we determined which of the 1000 scenarios led to the lowest containment frequency, and we plotted this single worst (out of 1000) containment frequency, at different levels of ring vaccination capacity, for the same four conditions as in Figure 5A : a. high workplace/ social contact finding probability (0.9) and high contact isolation (effective 3 hour delay following symptoms), b. moderate workplace/social contact finding probability (0.8) and high contact isolation, c. high workplace/social contact finding probability (0.9) and less contact isolation (effective one day delay), and d. moderate workplace/social contact finding probability (0.8) and less contact isolation. All parameters are the same as in Figure 5A (the household contact finding probability is 0.95 for all scenarios, and the diagnosis rates are doubled after the onset of community awareness). In this figure, "contact isolation" refers to the monitored diagnosis rate, i.e. the rate at which previously asymptomatic contacts who subsequently develop disease will be diagnosed (φ, c. d. became close to 1. However, this average conceals the fact that for some scenarios (parameter sets chosen from the calibrated uncertainty analysis), control remains difficult or impossible even at high levels of ring vaccination. Therefore, in Figure 5B , we plotted the single lowest containment fraction seen out of the 1000 computed; focusing on the single worst scenarios reveals a different picture, and shows that isolation of asymptomatic contacts and very high probabilities of finding workplace or social contacts would be needed to control smallpox under these most pessimistic parameter choices. Rapid contact tracing in ring vaccination may play an important role in suppressing the epidemic, since the longer it takes to trace a contact, the less likely the vaccine is to be efficacious, and the more opportunities the infected individual may have to transmit disease before they are finally located, isolated, and vaccinated if appropriate. We illustrate this possibility in Figure 6 by examining the same scenario we showed earlier in Figure 3A (e.g. households of size 4, workplace/social groups of size 8, 95% of household contacts traceable, 80% of workplace/social groups traceable, an average time to infection for a household contact of an infective given by 0.2 days). We assume in one case that contacts may be traced quickly (1 day for a household contact, 2 days for a workplace/ social contact), and in the other that the contacts are on average found slowly (5 days for a household contact, 10 days for workplace/social contacts); we assumed 30 ring vaccinations (traceable contacts) possible per day. In this scenario, the epidemic is more severe and containment (as we have been defining it) less likely when contact tracing is slow: in the fast scenario, 238 infections occurred on average and the (estimated) containment probability was 99%; for the slow scenario, on average 3587 infections occurred and the (estimated) containment probability was only 1%. While Figure 6 illustrates the possibility that rapid contact tracing may be of decisive importance in some scenarios (parameter set choices), this is not always the case. For some parameter sets, the probability of tracing contacts (household or workplace/social) may be too low, or the transmission rate too high, for more rapid contact tracing to make any difference. Conversely, for other parameter sets, the smallpox transmission rate may be so low that smallpox is easily contained even with slow contact tracing. While rapid contact tracing is never harmful, overall, how typical are the results of Figure 6 (in which rapid contact tracing was important in ensuring the efficacy of ring vaccination)? To address this question, we simulated the growth of smallpox for the 1000 "calibrated" scenarios we used in Figure 5A and 5B. As before, we assumed ten initial cases, and (as in Figure 6 ) that 30 ring vaccinations were possible per day; then we simulated 100 epidemics assuming one day to find a household contact (and 2 days to find a workplace/social contact). We then simulated 100 epidemics assuming that it takes five days to find a household contact and 10 days to find a workplace/social contact (as in Figure 6 ). For each of these 1000 scenarios, we calculated the fraction of simulations for which the total number of cases is 500 or less within 250 days, i.e. the containment fraction. For nearly all scenarios (parameter set choices), the containment fraction was smaller (sometimes much smaller) when the contact finding time is faster (since faster contact finding, all else being equal, improves smallpox control, as illustrated in Figure 6 ). However, for 64.5% of the scenarios (parameter set choices) examined, the difference was less than 2.5% in absolute terms (smallpox was either contained or not contained depending on other factors, and rapid contact tracing did not make the difference). On the other hand, for 18.7% of the scenarios examined, the absolute difference in the containment probability was 20% or more; thus, a substantial difference in containment probability is occa-Faster contact tracing Figure 6 Faster contact tracing may improve the efficacy of ring vaccination. We assume the same baseline parameters as in Figure 3A (e.g. households of size 4, workplace/social groups of size 8, 95% of household contacts traceable, 80% of workplace/social contacts traceable), and 30 ring vaccinations available per day (with contacts of contacts not traced). The fast scenario corresponds to an average one day delay for household and two days for workplace/social contacts (as in Figure 3A ); the slow scenario corresponds to an average five day delay for household and ten day delay for workplace/ social contacts. This figure shows the average of one hundred realizations starting with ten index cases. Effect of more rapid diagnosis Public awareness of smallpox, leading to more rapid isolation and identification, may play an important role in eliminating the epidemic, as illustrated by the scenarios in Figure 7 . We assumed 20 ring vaccinations possible per day, a capacity too small to contain the epidemic in the absence of increased surveillance or diagnosis; the black line in the figure shows the steeply rising average number of cases for the first 100 days. If, however, surveillance or public awareness of the symptoms of smallpox increases the diagnosis rate by 50% (multiplies the baseline diagnosis rates by 1.5), containment becomes possible (blue line); with a doubling of the diagnosis rate (red line) the peak number of cases is lower still. In these scenarios, increased diagnostic rates markedly improve the ability of ring vaccination to control the epidemic, this suggest that any ring vaccination effort be accompanied by increased public awareness and surveillance. In many cases, however, more rapid diagnosis was not required for ring vaccination to be effective. As before, we simulated smallpox epidemics for each of 1000 calibrated scenarios, performing 100 realizations each beginning with 10 index cases, and computed the fraction of scenarios for which the epidemic was always contained (as defined earlier), assuming no change in diagnosis rates. We assumed 80 ring vaccinators per day, contact finding probabilities of 0.95 for households and 0.8 for workplace/social contacts (as in Figure 3A ). Under these assumptions, for 83.4% of the scenarios, the epidemic was contained within 500 total cases in each of the 100 realizations, even with no change in diagnosis rates. Uncertainty analysis (using the 1000 calibrated scenarios, and based on the fraction of 100 replications showing decontainment) revealed the most important parameters which predict the failure of ring vaccination without more rapid diagnosis were the same as we found in the earlier uncertainty analysis; a higher fraction vaccinated before the epidemic, smaller households or workplace/social groups, less transmissibility, lowered infectivity prior to the rash, more rapid diagnosis, and a higher rate of diagnosis for alerted individuals all contribute to a greater containment probability even without an overall increase in the diagnosis rate. We have been assuming that whenever an individual is contacted during an investigation, the individual will be diagnosed more quickly should they subsequently develop symptoms. When transmission is assumed to be very rapid (smallpox is assumed to be highly contagious), most individuals may already be infected when identified through contact tracing from an infective. Using the scenario we examined in Figure 3A , we see that continued surveillance of contacts is an essential component of effective ring vaccination designed to control rapidly spreading smallpox: if smallpox in a contact is not diagnosed any more quickly than for a non-contact, containment by ring vaccination requires over 98% contact finding probabilities for both household and workplace/ social contacts -even if unlimited numbers of ring vaccinators are available; containment cannot be guaranteed by adding additional ring vaccination capacity if the contact finding rates are too low and/or the follow-up for contacts is insufficient. Smallpox which is transmitted less rapidly to contacts would, however, be containable with a lower contact finding probability (results not shown). Finally, we used the "calibrated" scenarios (parameter set choices) to explore the levels of contact finding probability needed to contain the epidemic (as before, defined to mean 500 or fewer cases ultimately resulting from ten initial cases) ( Table 4 ). In these scenarios, we assumed that all traceable contacts were followed up very More rapid diagnosis Figure 7 More rapid diagnosis due to public awareness or increased surveillance may lead to far more effective epidemic control. We assume the same baseline parameters as in Figure 3A , and averaged 100 realizations of the epidemic beginning with 10 index cases and assumed a ring vaccination capacity of 20 per day (and contacts of contacts not traced). For the black line, the diagnosis rate of cases does not change after the first case is identified (the multiplier is 1.0); for the blue line, the diagnosis rate increases by 50% (multiplier 1.5) after the first case is identified (as in Figure 3A ), resulting in substantially fewer cases; and for the red line, the diagnosis rate is doubled (multiplier 2.0) after the first case is identified, resulting in still fewer cases. quickly (1/a = 1 hour, so that cases arising in previously contacted persons almost never transmit the infection further). We chose different levels of household and workplace/social contact finding probabilities and different levels of ring vaccination capacity, and performed 100 replications of each of the 1000 different scenarios. In Table 4 we report the fraction of scenarios for which all 100 replications exhibited containment. Scenarios in which smallpox is highly contagious require high contact finding probability to ensure the containment of the epidemic. Transmission prior to the rash makes epidemic control more difficult. In Figure 8 , we show an expanding smallpox epidemic assuming differing levels of infectivity prior to the rash (adding increased infectivity prior to the rash, keeping constant the infectivity after the rash). We assume all parameters are the same as in Figure 3A (and that the ring vaccination capacity is 40 per day). Infectivity prior to the rash is modeled as the relative infectivity during the short (1 day) period of oropharyngeal lesions just prior to the rash (compared to the infectivity during the first week of the rash), and as the relative infectivity during the prodromal period (relative to the period just prior to the rash). We consider three scenarios: a relative infectivity during entire period is one (i.e., infectivity during the prodromal period and just prior to the rash is the same as during the first week of the rash), b the relative infectivity just prior to the rash is the same as during the first week of the rash, but during the prodromal period is 4% (as in Figure 3A ) of this value, and c the relative infectivity just prior to the rash is 20% of the infectivity during the first week of the rash, and during the prodromal period is 20% of this value. The figure shows that increased infectivity just prior to the rash leads to a larger epidemic (comparing b and c); in case b (high infectivity just prior to onset of rash), loss of containment occurs 36% of the time (but in none of the 100 realizations shown in case c (low infectivity prior to rash)). Scenario a (full infectivity during entire the prodromal period) showed loss of control in every realization. Increasing the ring vaccination capacity from 40 per day to 80 per day (results not shown) led to containment in all of the realizations with high infectivity just prior to the rash and low infectivity during the prodromal period (case b), but made no difference if the infectivity was as high during the prodromal period as during the rash (case a). While intuitively adding additional infectiousness must increase the number of secondary cases and make control more difficult, these results do illustrate that even a small amount of increased infectiousness prior to the rash (when diagnosis is more difficult) may substantially increase the difficulty of smallpox control. Finally, in Figure 9 , we present scenarios in which each of four other parameters are modified from the baseline values of Figure 3A , assuming 40 contact tracings (ring vaccinations) are possible per day (line a in the figure) . Specifically, we assume that severe smallpox (hemorrhagic and flat) on average takes four times longer to diagnose and isolate than ordinary smallpox (case b), Table 4 : Containment of severe smallpox at different levels of contact finding. The first three columns are assumed levels for the probability of finding a household contact, the probability of finding a workplace/social (W/S) contact, and for the number of contact tracings/ring vaccinations possible per day; the last two columns express (as percentages) the resulting probability of containment given the assumed contact finding probabilities and contact tracing capacities; two containment probabilities are given: the containment probability when only contacts of cases are traced (first column, "Contacts"), and the containment probability when contacts of contacts of cases are traced in addition to the contacts of cases (second column, "Contacts of Contacts"). All other parameters are the same as in Figure 3A . that no one in the population has prior vaccination protection (from before the discontinuation of routine vaccination, case c), that 10% more smallpox is too mild to diagnose (but still contagious, case d) compared to baseline, and finally that the vaccine is completely ineffective (case e). Each of these scenarios will be discussed further below. Scenario b was motivated by the possibility that individuals with severe forms of smallpox may be more difficult to diagnose, and thus remain infectious in the community longer (despite the much greater degree of illness of such patients), or that such patients may be more infectious. In this particular case, quadrupling the mean diagnosis time led to one additional replication out of 100 in which containment was not achieved (2/100, compared to the baseline of 1/100). However, we assumed that community awareness of smallpox leads to the same relative rate of increased diagnosis among severe cases as for ordinary cases, and that the most severe forms are relatively rare. In addition to the scenario shown in the figure, we also replicated the same 1000 "calibrated" simulations, assuming that in each case 40 contact tracings per day are possible and that the diagnosis time for severe cases was four times that of ordinary cases. Finally, we repeated each "calibrated" scenario 100 times assuming long diagnosis times for severe cases, and not making this assumption, and found that the difference in the decontainment fraction was not large (results not shown). Scenario c illustrates that vaccination prior to the discontinuation of routine vaccination does play a role in smallpox control by ring vaccination; there were more decontainment scenarios (5/100) when no prior Transmission prior to the rash Figure 8 Transmission prior to the rash makes epidemic control more difficult. The figure shows a expanding smallpox epidemic assuming differing levels of infectivity prior to the rash. We assume all parameters are the same as in Figure 3A (and that the ring vaccination capacity is 40 per day). Infectivity prior to the rash is modeled as the relative infectivity during the short (1 day) period of oropharyngeal lesions just prior to the rash (compared to the infectivity during the first week of the rash), and as the relative infectivity during the prodromal period (relative to the period just prior to the rash). For scenario a, relative infectivity during the prodromal period and just prior to the rash is the same as during the first week of the rash, for scenario b, the relative infectivity just prior to the rash is the same as during the first week of the rash, but during the prodromal period is 4% (as in Figure 3A ) of this value, and for scenario c, the relative infectivity just prior to the rash is 20% of the infectivity during the first week of the rash, and during the prodromal period is 20% of this value (these two parameters are the same as in Figure 3A ). Additional scenarios, assuming 40 ring vaccinations or con-tact tracings possible per day, and that contacts of contacts are traced; all parameters are identical to those in Figure 3A unless otherwise indicated Figure 9 Additional scenarios, assuming 40 ring vaccinations or contact tracings possible per day, and that contacts of contacts are traced; all parameters are identical to those in Figure 3A unless otherwise indicated. The figure shows the average of 100 replications of five scenarios (Case a repeats the result from Figure 3A for reference); the numbers in parentheses in the legend are the corresponding fraction of the 100 scenarios for which decontainment occurred. For case b, we assumed that flat and hemorrhagic smallpox cases took four times as long on average to diagnose as ordinary cases; for case c., we assumed that no one in the population had prior protection (as opposed to 25% for Figure 3A) ; for case d, we assumed that an additional 10% of individuals (13% instead of 3%) would develop mild smallpox (with 75% developing ordinary smallpox instead of 85% as in Figure 3A ); and for case e, we assumed that the vaccine is completely ineffective and provides no protection against infection. protection exists in the population. The results suggest that prior vaccination aids in the control of smallpox, but that it is not strictly necessary for control (in this scenario, 95% of the replications exhibited containment). In Figure 3A , we assumed 25% of individuals had protection due to vaccination prior to the discontinuation of routine vaccination; in scenario c of Figure 9 , we assumed this fraction was zero. Scenario d demonstrates that if 10% more smallpox infections (in absolute terms, i.e. 13% compared to 3% in Figure 3A ) lead to mild cases among individuals with no prior protection, the epidemic is more difficult to contain (13/100 replications showed loss of containment). Finally, scenario e demonstrates that containment is still possible even when the vaccine is completely ineffective in everyone -because of case isolation and isolation of contacts (and of contacts of contacts). Here, with 40 contact tracings possible per day, 55% of the replications nevertheless exhibited containment even with a vaccine which offered no protection whatever. With 90 contact tracings possible per day, all replications exhibited containment even assuming no vaccine protection. Although less efficient than ring vaccination in the sense that more vaccinations must be delivered to eliminate infection, comprehensive mass vaccination following the introduction of smallpox is sufficient to eliminate the infection. In Figure 10 , we show the probability of achieving containment (defined to be fewer than 500 total cases resulting from 10 index cases) for different levels of ring vaccination (0, 5, 10, and 20 vaccinations per day) and mass vaccination (0, 0.5%, 1%, and 2%; compare with the 10%-20% per day many jurisdictions in the United States are planning to vaccinate). Specifically, for each level of ring vaccination and mass vaccination, we used the same 1000 parameter sets used in Figure 5 , and performed 100 simulated epidemics for each parameter set. On the vertical axis, we plot the fraction of the 1000 scenarios for which each of the 100 simulated epidemics was contained. We further computed the fraction of scenarios for which none of the 100 simulated epidemics was contained; this is indicated by the colored segment in the small pie chart at each symbol. When the mass vaccination rate was 2% per day, the mean number of deaths (averaging over all scenarios and all simulations within each scenario) was 47.7, 33.7, 26.4, and 20.1 for a ring vaccination level of 0, 5, 10, and 20 per day (respectively) out of a population of 10000. Moreover, when we increased the mass vaccination level to 3%, an average of 28.9 deaths occurred when no ring vaccination was used, but this fell to 22.3 deaths when only 5 ring vaccinations per day were available (again assuming a population of 10000, and 10 index cases). With a mass vaccination level of 5% per day, an average of 18.8 deaths occurred without ring vaccination, and 15.8 deaths occurred when only 5 ring vaccinations per day were possible. (At a mass vaccination rate of 3% per day, containment as defined above was achieved in all 100 replications for 95% of the scenarios even without ring vaccination; at a mass vaccination rate of 5% per day, containment was achieved in all replications for all scenarios.) These results show that over a wide range of simulated epidemics, even seemingly small levels of ring vaccination (coupled with follow-up) may have a substantial effect in preventing epidemic spread and reducing deaths from smallpox, even during a mass vaccination campaign. Note that many jurisdictions in the United States are planning mass vaccination campaigns which could reach 10%-20% of the population per day, far greater than the mass vaccination levels examined here; it is interesting to note that mass vaccination cam- Figure 10 Mass and ring vaccination together. Low-level mass vaccination programs are improved substantially by the addition of ring vaccination. The shaded pie segments represent the fraction of 1000 scenarios for which containment (as defined in the text) was never realized; the vertical position of the pie chart represents the fraction of the 1000 "calibrated" scenarios for which containment was always achieved. As the fraction of the population mass vaccinated increases or the ring vaccination capacity increases, the probability of containment increases. paigns may be effective in preventing a widespread epidemic even at much lower levels than are being planned for. Where feasible, such rapid mass vaccination rapidly eliminates smallpox transmission in our model; vaccination of contacts is still beneficial, since we are assuming that earlier vaccination yields a greater probability of preventing or ameliorating infection (results not shown). We constructed a simple network model of smallpox transmission, and addressed the question of what circumstances contribute to the success of a ring vaccination campaign designed to control smallpox. Our analysis focused on the use of contact tracing/ring vaccination to prevent a widespread epidemic following a deliberate release. We conducted a sensitivity analysis based on particular, but reasonable, ranges for the unknown parameters. Our results are consistent with prior vaccination models in identifying prior vaccination and ring vaccination capacity as significant factors in determining the spread of smallpox. Unsurprisingly, we also find that household size and ring vaccination speed are particularly important parameters; these results are intuitively plausible. The contact finding probability did not appear important in this analysis only because a narrow range of values was chosen. We illustrated smallpox control by presenting scenarios based on control of moderately severe smallpox epidemics. We find that swift, aggressive contact tracing and ring vaccination is is usually sufficient to bring the infection under control. Provided that there is sufficient capacity, vaccination of contacts of contacts is beneficial, and results in fewer infected individuals and more rapid elimination of infection; investigating contacts of contacts allows the chain of transmission to be outrun to some extent. When ring vaccination capacity is small, diversion of crucial resources away from contacts is harmful; contacts of contacts should only be traced and vaccinated provided that no resources are diverted away from contacts of cases. The increased surveillance (or isolation) of contacts, together with improved rates of diagnosis due to community awareness, play important roles in smallpox control; we note that in some cases, lowered diagnosis rates among severe cases contributed to a small extent to loss of epidemic control, and suggest that any public awareness campaign include information to help the public be more aware of the full spectrum of the clinical features of smallpox. One limitation of our analysis is that we chose not to explicitly incorporate the specific epidemiology of health care workers (or mortuary workers), who are likely to be exposed to infected individuals during any smallpox epidemic (e.g. [17, 22] ), and who may then infect further members of the community [22] (as was also seen in the recent outbreak of SARS, e.g. [48] ). Transmission to health care workers may be considered to amplify the initial attack or to be simply accounted among the exposures we considered (and thus be approximated by the behavior of our model), since health care workers and their household contacts are in all likelihood traceable contacts, and ring vaccination/contact tracing would identify and halt these chains of transmission as in our model. The disruption of smallpox control and patient care that may occur is not accounted for in our analysis, however, causing our model in this sense to err on the side of optimism. The appropriateness of pre-event vaccination of health care workers or other first responders has been addressed by other analyses [12, 49] , and is beyond the scope of our model. While we analyzed the effect of contact tracing, case and contact isolation, and ring vaccination (together with mass vaccination), in a real smallpox epidemic, in practice, control efforts are unlikely to be limited strictly to vaccinating contacts (and health care workers, as likely contacts) and isolating cases. Indeed, making vaccine available to individuals who believe they live near cases or to others on a voluntary basis occurred in smallpox control efforts in the past [22] . Vaccination of such individuals can only harm the disease control effort if it hinders or delays the diagnosis of cases or the investigation and vaccination of contacts; our results show that even relatively low levels of vaccination of the general population may have a beneficial effect in preventing the epidemic from escaping control. More serious is the possibility that individuals who should be vaccinated or isolated would be missed; this could occur either because individuals or institutions did not cooperate with the disease control effort, or because the individuals simply could not be found. Our analysis suggests that ring vaccination need not be perfect to successfully contain the epidemic, and yet, under conditions where there is a high rate of infection among contacts, or a relatively high rate of casual transmission, high rates of contact finding (in excess of 90%), together with increased surveillance and contact isolation, are needed to contain the epidemic. Finally, the vaccination of individuals at low risk of contracting smallpox will cause harm due to adverse events of the vaccine; in our model, the assumed death rate due to vaccination was small compared to the probability of death from smallpox, and played essentially no role in the analysis. In practice, individuals suspected to be at high-risk for vaccine complications, but at relatively low risk for contracting smallpox, might simply be isolated or closely monitored even during an outbreak; while the presence of individuals in the population at higher risk for vaccine complications would increase the death rate during an outbreak, such individuals are unlikely to impair the containment of the epidemic (the primary focus of this analysis). Our results support ring vaccination against epidemics of smallpox (even assuming high rates of transmission to close contacts), but do note that stochastically, for severe (rapidly transmissible) smallpox, scenarios of loss of control are seen, with resulting widespread epidemics. In scenarios in which the transmission potential of smallpox is smaller, such loss-of-control scenarios occur less frequently (results not shown). Mass vaccination campaigns, when conducted quickly and with very high coverage, do not result in loss of control in our model. Nevertheless, fewer deaths due to smallpox result when ring vaccination is conducted along with mass vaccination. Simulated smallpox epidemics with ring vaccination suggest that aggressive, fast ring vaccination can control epidemics of smallpox. To do so, however, smallpox must be identified quickly and contacts vaccinated promptly. We also identify public awareness of smallpox -leading to prompt identification of cases -as a major factor in smallpox control; in some simulations, it may play a role as significant as ring vaccination itself [15] . However, we also found that uncertainty in (1) transmission from mild cases, (2) the household size, and (3) casual transmission contributed to the overall uncertainty in the epidemic size. Other parameters to which the number of infections were highly sensitive were the prior vaccination fraction, parameters related to infectiousness, and parameters related to transmission prior to the rash. Because our model combines network structure with response logistics, our results support and complement the results of other investigators. Our results support the notion that prior vaccine protection may play an important role in slowing the epidemic [11] , despite the possibility that some vaccinated individuals may develop mild cases which are harder to identify, but which nevertheless transmit disease. Likewise, our results provide support for the view that ring vaccination should play a central part in smallpox control. If initiated, ring vaccination should be conducted without delays in vaccination, should include contacts of contacts (whenever there is sufficient capacity to cover all contacts of cases), and should be accompanied by a vigorous campaign of public awareness which can facilitate more rapid identification and isolation of cases. We assumed that ring vaccination could be fast (little delay between identification of a case and vaccination of the contacts), effective (nearly all household contacts can be found, and most of workplace and social contacts), and available (there is sufficient capacity). To be effective, ring vaccination planning must yield a system capable of meeting these benchmarks; we should not only be able to assess the number of contact vaccinations that will be possible per day, but should have a plan in place to (1) identify contacts by working with individuals, employers, schools, community representatives, and authorities or businesses who may have access to information facilitating contact tracing, (2) rapidly investigate and vaccinate such individuals, perhaps using field teams managed by central dispatch. It is important to realize that for highrisk, transient, or unstably housed populations where reliable contact tracing is impossible, the conclusions of the model we present cannot be applied. It is important to note that while our model suggests that ring vaccination together with contact tracing and isolation is likely to be successful, we found that for some scenarios (where smallpox was more transmissible, or was relatively more transmissible before the rash), epidemic containment required not only ring vaccination, but increased public awareness, the isolation of contacts, and tracing of contacts of contacts. For scenarios in which the smallpox was less transmissible, epidemic containment was possible at lower contact finding probabilities. Thus, while our simulations suggest that contact tracing/ring vaccination need not be perfect to succeed, because of uncertainties in our knowledge of the behavior of bioterrorist smallpox, it is impossible to know in advance how good it will have to be. Thus, that high contact finding rates, mass public awareness leading to early identification of cases, isolation of contacts, and investigation of contacts of contacts should all be conducted with maximum effectiveness to reduce the probability of a widespread epidemic. While the possibility of smallpox uncontrollable by ring vaccination has made mass vaccination preparations wise, and while mass vaccination may be unavoidable in the event of a deliberate release of smallpox, we believe that ring vaccination is essential in any case. This is not only because individuals recently exposed to smallpox may be protected if they are vaccinated promptly, but because each contact identified potentially lies in the immediate future of the transmission chain. From the standpoint of epidemic control, it is far more valuable to vaccinate individuals next in the transmission chain than to vaccinate other persons. Our results support the idea that ring vaccination/case isolation may in many, if not most cases, eliminate smallpox even without mass vaccination, but also support planning for mass vaccination (so that the vastly more costly and difficult policy of mass vaccination will be available in the event of an explosive epidemic). When faced with the unknown, multiple redundant prep-arations are appropriate; case investigation/isolation may control smallpox even if the vaccine does not work at all, but mass vaccination is useful in the event of an explosive epidemic for which case tracking becomes impossible.
17
Protection of pulmonary epithelial cells from oxidative stress by hMYH adenine glycosylase
BACKGROUND: Oxygen toxicity is a major cause of lung injury. The base excision repair pathway is one of the most important cellular protection mechanisms that responds to oxidative DNA damage. Lesion-specific DNA repair enzymes include hOgg1, hMYH, hNTH and hMTH. METHODS: The above lesion-specific DNA repair enzymes were expressed in human alveolar epithelial cells (A549) using the pSF91.1 retroviral vector. Cells were exposed to a 95% oxygen environment, ionizing radiation (IR), or H(2)O(2). Cell growth analysis was performed under non-toxic conditions. Western blot analysis was performed to verify over-expression and assess endogenous expression under toxic and non-toxic conditions. Statistical analysis was performed using the paired Student's t test with significance being accepted for p < 0.05. RESULTS: Cell killing assays demonstrated cells over-expressing hMYH had improved survival to both increased oxygen and IR. Cell growth analysis of A549 cells under non-toxic conditions revealed cells over-expressing hMYH also grow at a slower rate. Western blot analysis demonstrated over-expression of each individual gene and did not result in altered endogenous expression of the others. However, it was observed that O(2 )toxicity did lead to a reduced endogenous expression of hNTH in A549 cells. CONCLUSION: Increased expression of the DNA glycosylase repair enzyme hMYH in A549 cells exposed to O(2 )and IR leads to improvements in cell survival. DNA repair through the base excision repair pathway may provide an alternative way to offset the damaging effects of O(2 )and its metabolites.
Oxidative stress leading to the overproduction of free radicals in the lungs is present in many clinical situations. Such clinical settings include acute respiratory distress syndrome (ARDS), infants of prematurity going on to develop bronchopulmonary dysplasia (BPD), pathogenesis of chronic obstructive pulmonary disease (COPD), asthma, cystic fibrosis, ischemia-reperfusion injury, druginduced lung toxicity, cancer and aging [1] [2] [3] [4] . Although the use of oxygen may be clinically indicated in hypoxemic situations, one must consider the potential longterm toxic side effects. For example, we know that oxygen creates cellular damage by a variety of mechanisms. Normal cellular metabolism of oxygen involves the transfer of electrons from NADH to O 2 molecules to form water (H 2 O). At normal partial pressure, 95% of oxygen molecules (O 2 ) are reduced to H 2 O and 5% are partially reduced to toxic byproducts by normal metabolism in the mitochondria [5] . These metabolites include the superoxide anion (O 2 -), hydrogen peroxide (H 2 O 2 ), and hydroxyl radicals ( • OH) all of which make up what are known as Reactive Oxygen Species (ROS) [6] . Exposure to conditions of hyperoxia as well as ionizing radiation (IR) leads to increased amounts of these ROS and their damaging effects. ROS are known to attack the lipids, proteins, and nucleic acids of cells and tissues [5, 7] . Lipids, including pulmonary surfactant, react with ROS to produce lipid peroxides, which cause increased membrane permeability, inactivation of surfactant, and inhibition of normal cellular enzyme processes. Proteins reacting with ROS result in decreased protein synthesis due to inhibition of ribosomal translation or destruction of formed proteins. This ultimately leads to inactivation of intracellular enzymes and transport proteins resulting in impaired cellular metabolism and accumulation of cellular waste products. Lastly, ROS cause damage to nucleic acids by leading to modified purine and pyrimidine bases, apurinic (AP) / apyrimidinic sites, and DNA protein cross-links which can lead to single strand breaks [8] . Several defense mechanisms exist to combat the damaging effects of ROS. Intracellular enzymatic systems include superoxide dismutase which eliminates the superoxide anion, catalase which catalyzes the reduction of H 2 O 2 directly to H 2 O without the production of the hydroxyl radical, and glutathione peroxidase which directly reduces H 2 O 2 and lipid peroxides. Free radical scavengers, which stop free radical chain reactions by accepting electrons, include α-tocopheral (vitamin E), ascorbic acid (vitamin C), niacin (vitamin B), riboflavin (vitamin B 2 ), vitamin A, and ceruloplasmin [1, 2, 9] . These systems usually provide enough protection against oxygen metabolism under normal conditions, but may become depleted under conditions of increased oxidative stress [7, 10] . The defense mechanism of interest in this paper involves the repair of oxidative damage through the human DNA base excision repair pathway (BER). BER is the most important cellular protection mechanism that removes Base excision repair pathways for Oxidative DNA damage oxidative DNA damage [11] . Damaged bases are excised and replaced in a multi-step process. Lesion-specific DNA glycosylase repair genes initiate this process. After removal of the damaged base, the resulting AP site is cleaved by APendonuclease generating a 3'OH and 5'deoxyribose phosphate (dRP). β-polymerase, which possesses dRPase activity, cleaves the dRP residue generating a nucleotide gap and then fills in this single nucleotide gap. The final nick is sealed by DNA ligase [12] [13] [14] ( Figure 1A ). The oxidative repair genes that we have analyzed in this study include 8-oxoguanine DNA glycosylase (hOgg1), human Mut Y homologue (hMYH), human Mut T homologue (hMTH), and endonuclease III (hNTH) all of which are present in human cells and involved in the protection of DNA from oxidative damage. The repair enzyme hOgg1 is a purine oxidation glycosylase that recognizes and excise 8-oxoguanine lesions (GO) paired with cytosine. GO can pair with both cytosine and adenine during DNA replication [15] . If repair of C/GO does not occur, then G:C to T:A transversions may result [5, [15] [16] [17] . The repair enzyme hMYH is an 8-oxoguanine mismatch glycosylase that removes adenines misincorporated opposite 8-oxoG lesions that arise through DNA replication errors [5, [18] [19] [20] . The repair enzyme hMTH hydrolyzes oxidized purine nucleoside triphosphates such as 8-oxo-dGTP, 8-oxo-GTP, 8-oxo-dATP, and 2-hydroxy-dATP, effectively removing them from the nucleotide pool and preventing their incorporation into DNA ( Figure 1B ) [21] . Lastly, the repair gene endonuclease III (hNTH) is a pyrimidine oxidation and hydration glycosylase that recognizes a wide range of damaged pyrimidines [22] . hNTH has also been shown to have a similar DNA glycosylase/AP lyase activity that can remove 8-oxoG from 8-oxoG/G, 8-oxoG/A, and 8-oxoG/C mispairs [23, 24] . Subsequent steps following hNTH are identical to those following hOgg1 ( Figure 1A ). A previous study has shown that over-expression of the DNA repair gene hOgg1 leads to reduced hyperoxiainduced DNA damage in human alveolar epithelial cells [25] . The primary goal of our present study was to compare the protective effects of the four main lesion-specific DNA glycosylase repair genes by individually overexpressing each in lung cells and determining which of these provides the greatest degree of protection under conditions of increased oxidative stress. The human alveolar epithelial cell line A549 (58 year old Caucasian male), was purchased from ATCC Cat No CCL-185. The cells were grown in DMEM (Gibco, Grand Island, NY) supplemented with 10% fetal bovine serum (FBS) (HyClone, Logan, UT) and penicillin (100 U/ml)/ streptomycin (100 µg/ml) (Gibco, Grand Island, NY). Passaging of cells was performed every 3-4 days with cells grown to 80% confluency in a 10 cm cell culture dish (Corning Incorporated, Corning, NY). Cells were kept at 37°C in a humidified, 5% CO2 incubator. The retroviral vector pSF91.1, a gift from Dr. C. Baum from the University of Hamburg in Germany, was constructed with an internal ribosome entry site (IRES) upstream to the gene expressing enhanced green fluorescent protein (EGFP) as previously described [26] . Four DNA repair genes were individually ligated into the retroviral vector pSF91.1. hOgg1-6pcDNA3.1 was initially amplified by PCR by primers to introduce a kozak sequence at the 5' end [27] . Digestion of this product with EcoRI and SalI was performed and then hOgg1 was subcloned into digested plasmid vector pSF91.1, with T4 DNA ligase. DNA sequencing was performed to confirm integrity of the hOgg1 gene. hMYH/PGEX4T-1 and hMTH/PGEX4T-1 hMYH was a gift from Dr. A. McCullough (University of Texas Medical School, Galveston, TX) and hMTH was cloned in Dr. Kelley's lab. Plasmid DNA was prepared as above by digestion with EcoRI and SalI and ligated into pSF91.1 as above and sequencing was performed to confirm integrity of the genes. PGEX-6PI-hNTH1-wild type this gene was a gift from Dr. S. Mitra (University of Texas Medical School, Galveston, TX). Digestion with BamHI and SalI was performed and the hNTH1-wt fragment was ligated into the empty plasmid vector PUC18. The hNTH1-wt fragment was then excised with both sides flanked by EcoRI restriction sites and ligated into pSF91.1. Proper orientation of the gene was confirmed and sequencing was performed to determine the integrity of the gene. 2.5 × 10 5 A549 cells were suspended with the viral supernatant and plated in 1 well of a 6-well plate along with polybrene (Sigma, St. Louis, MO). This exposure was performed 6 hours per day for three days. At approximately five days from the beginning of the infection, the infected cells were analyzed using flow cytometry and sorted for EGFP expression. Cell pellets of sorted cells were resuspended in NuPage buffer (Invitrogen, Carlsbad, CA) and protein concentrations were determined using the DC protein assay (Bio-Rad, Hercules, CA). 20 ug of protein were loaded into individual lanes of a NuPage Bis-Tris Gel (Invitrogen, Carlsbad, CA). The gel was then transferred to nitrocellulose paper (Osmonics Inc, Gloucester, MA). The membranes were then blocked with 1% blocking solution (Roche Diagnostics, Indianapolis, IN) for 1 hour at room temperature and then incubated overnight at 4°C with rabbit polyclonal antibodies to hOgg1 (Novus Biologicals, Littleton, CO), hMTH (Novus Biologicals, Littleton, CO), hMYH (Oncogene Research Products, Darmstadt, Germany) and hNTH (Proteintech Group Inc, Chicago, IL) all at a dilution of 1:1000 except hNTH which was diluted 1:2500. They were then washed 2 times with TBST and 2 times with 0.5% blocking solution, 10 minutes per wash. The membranes were incubated with anti-rabbit secondary antibodies at 1:1000 for 1 hour at room temperature. Lastly, the membranes were washed 4 times with TBST, 15 minutes per wash. The membranes were briefly soaked in BM chemiluminescence blotting substrate (Roche Diagnostics, Indianapolis, IN) and then exposed to high performance autoradiography film (Amersham Biosciences, Buckinghamshire, England). Kodak Digital Science 1D Image Analysis software was utilized to quantify the region of interest (ROI) band mass of individual bands on films where visualized differences were detected. Sorted EGFP positive A549 cells infected with the above DNA repair genes were counted and seeded into 96-well plates at a density of 1000 cells/well, 6 wells per gene. Six hours after seeding, individual plates were placed into an oxygen chamber supplied by Dr. L. Haneline (Wells Center for Research, Indianapolis, IN) located in a 37°C incubator. The oxygen chamber was then infused with 95% O 2 and 5% CO 2 . Individual plates were removed after 12, 24, 48, and 72 hours of exposure. Control A549 cells were incubated in a normal 37°C humidified-5% CO 2 incubator. O 2 concentrations were monitored with a MAXO 2 analyzer (Maxtec, Salt Lake City, UT). Four days from the beginning of the exposure, cells were assessed for cell growth/survival using the sulforhodamine B assay (SRB assay). The SRB assay (Sigma, St. Louis, MO), developed by the National Cancer Institute, provides a sensitive measure of drug-induced cytotoxicity through a colorimetric endpoint that is non-destructive, indefinitely stable, and visible to the naked eye. This assay was used to assess the cell growth/survival of over-expressed cells [28] . Cold 10% TCA was used to fix the cells to the plate. After incubation for 1 hour at 4°C, the individual wells were rinsed with water. After air-drying, SRB solution was added to each well and cells were allowed to stain for 20-30 minutes. 1% acetic acid wash was used to rinse off unincorporated dye. Incorporated dye was then solubilized in 100 µl per well of 10 mM Tris. Absorbance was measured by a tunable microplate reader (Molecular Devices, Sunnyvale, CA) at a wavelength of 565 nm. Background absorbance measured at 690 nm was subtracted from the measurements at 565 nm. Sorted EGFP positive A549 cells were seeded into 96-well plates at a density of 1000 cells/well. Six hours after seeding, individual plates were then exposed to radiation at doses of 250, 500, 1000, and 1500 Rads or 0. well. All the plates were placed into a 37°C humidified-5% CO 2 incubator. Every 24 hours for 4 days, 1 plate was removed and the cells were fixed and analyzed by the SRB assay looking at cell growth under non-toxic conditions. Growth curves and exponential growth equations were determined to look at the doubling time (DT) of cells infected with each repair gene of interest compared to vector infected and uninfected wild type cells. All drug exposure experiments were performed at least three times and individual drug doses included 6-8 wells for each group of infected cells. Analysis of cell growth and exponential growth equations were determined using Microsoft Excel. All experiments involving drug exposures were normalized to the zero dose. Data are expressed as means ± SE. The significance of differences were calculated using the paired Student's t test with significance being accepted for p < 0.05. The DNA repair genes hOgg1, hMYH, hMTH, and hNTH were ligated into the retroviral vector pSF91.1 ( figure 2 ). This vector, derived from a murine stem cell virus backbone, along with each individual repair gene, was used for transfection of phoenix amphotropic cells. Viral supernatant was then collected and used to stably infect A549 Western analysis of A549 cells over-expressing individual repair genes and effect on endogenous glycosylase level contained the genes of interest integrated into their DNA (data not shown). Western blot analysis was performed on sorted cells in order to verify over-expression of the four genes of interest. hOgg1, hMYH, hMTH, and hNTH were all detected at their correct position on western blots (data not shown). Western analysis was also utilized to assess whether overexpression of each individual repair gene resulted in altered endogenous expression of the other repair genes under both non-toxic and toxic conditions (24 hrs of 95% O 2 and 1000 Rad). Cells over-expressing the repair genes hOgg1, hMYH, hMTH, and hNTH did not lead to altered expression of the other endogenous repair genes under the above conditions when compared to each other or pSF91.1 vector control cells ( Figure 3A ,3B,3C and 3D). hOgg1's endogenous expression was below the level of detection. The pattern of endogenous expression of hNTH was consistent for each condition when comparing cells over-expressing hOgg1, hMYH, hMTH, and pSF91.1. Reduced expression of hNTH after exposure to 95% O 2 was noted. Lastly, we assessed endogenous expression of each individual repair gene in cells infected with pSF91.1 following non-toxic and toxic conditions (24 hrs of 95% O 2 and 1000 Rad) at 24 and 48 hrs after the onset of exposure. Endogenous hMYH and hMTH were expressed to the same degree. hOgg1's endogenous expression was below the level of detection using western analysis (results not shown). When analyzing endogenous hNTH expression, it was noted that hyperoxia at 24 hrs and 48 hrs resulted in reduced protein expression by 93% and 64% respectively. There also was a small increase in expression of hNTH noted after 1000 Rad one day post exposure that was back to baseline by two days post exposure. ROI band mass quantification demonstrated this finding ( Figure 4A and 4B). Two or more replicates were performed for each western analysis to determine consistency of the results. A549 cells expressing hMYH demonstrated increased survival after exposure to conditions with elevated levels of oxygen compared to cells expressing only the pSF91.1 vector ( Figure 5A ). Results were highly significant at all time points except after 12 hours O 2 where it almost reached a highly significant value. The differences between pSF91.1 and hMYH varied from 12% after 12 hours O 2 exposure to 7% after 72 hours O 2 exposure. A549 cells expressing hMYH also demonstrated increased survival after exposure to all doses of radiation in comparison to pSF91.1 ( Figure 5B ). These results were also highly significant at all doses of radiation except at 250 Rads where it almost reached a highly significant value. The differences between pSF91.1 and hMYH varied from 12%-14% for all doses of radiation. Also noted in these experiments was that vector control cells demonstrated no Experiments looking at the effects of H 2 O 2 on cells expressing the repair genes did not demonstrate increased survival for any of these repair genes when compared to vector control cells ( Figure 5C ). This data demonstrates that over-expression of hMYH has the ability to improve cellular survival under conditions of hyperoxia and radiation but may not be able to overcome the toxicity of H 2 O 2 . Cell growth under normal conditions was ascertained to determine if over-expression of any of the repair genes caused an alteration in the growth of cells in the absence of oxidative stress. Wild type A549 cells and cells expressing pSF91.1, hNTH, hOgg1, and hMTH appeared to grow at similar rates with doubling times within the same range. A549 cells expressing hMYH did show a slower growth rate that resulted in significant differences in cell number by day 3. The calculated doubling time for the cells over expressing hMYH is > 3 hrs longer than the cells with the other repair genes and vector alone ( Figure 6 ). This slowing of growth may allow for more time to repair Cell survival analysis following O 2 , IR, and H 2 O 2 treatments Oxidative stress to the lung leads to cellular DNA damage as evidenced by the release of specific gene products known to regulate DNA base excision repair pathways such as p53 and p21 [29] [30] [31] . Alterations in pro-inflammatory mediators, transcription factors, and other related gene products are also observed [32] . This injury has been shown to be associated with features of both cellular necrosis and apoptosis [33] [34] [35] . The resultant cellular inflammation and death from oxidative stress has a dramatic impact on the outcome of patients in the clinical setting [7, 36] . Most of our current clinical therapy towards oxidative stress in the lung involves both supportive measures and prevention. Research dealing with oxidative lung injury has focused mainly on enhancing antioxidant enzymatic processes and free radical scavengers [37] [38] [39] [40] . The ability to alter cellular survival by increasing specific DNA repair mechanisms may add another approach to the treatment of oxidant-mediated lung injury. Many investigators have used hydrogen peroxide as a substitute for hyperoxia since it is known to be one of the metabolites produced by the metabolism of oxygen. ROS such as H 2 O 2 and those produced by hyperoxia clearly lead to DNA damage but questions exist as to whether H 2 O 2 leads to the same deleterious effects upon DNA as hyperoxia. Analysis of our growth curves after exposure to H 2 O 2 in comparison to hyperoxia and IR clearly indicates that cellular protection by oxidative DNA repair genes is specific to the agent used. Because no protection was observed with over-expression of any of the repair genes following exposure to H 2 O 2 , we speculate that the damage it causes is dissimilar. It may be that its damage not only involves oxidized bases, but may also include other forms of DNA, lipid, and protein damage that are not corrected by oxidative DNA repair genes. Alternatively, the amount and type of damage evoked by H 2 O 2 could be beyond that which can be corrected by over-expressing these repair genes. Another form of stress known to induce damage through the formation of ROS is IR. Radiation induced free radical damage to DNA has substantial overlap with that of oxidative damage [41] [42] [43] . The protection provided by specific oxidative DNA repair genes under conditions of IR, was notable throughout our experiments only with the repair enzyme hMYH. The primary agent utilized to induce the formation of ROS was an oxygen rich environment. The use of oxygen as a stressor leading to the formation of ROS, offers a distinct advantage over IR and H 2 O 2 by mimicking the clinical situation where constant exposure to hyperoxia leads to cumulative cellular damage which further compromises repair. We determined that survival of A549 cells was also enhanced to a small degree with increased expression of the repair enzyme hMYH. This was an unexpected finding as we anticipated the repair gene hOgg1 would demonstrate the greatest protection in response to oxidative stress based on previous studies, however these experiments utilized the colony forming assay (CFA) to detect improvements in survival [25] . Additionally, the CFA may provide different results compared to the SRB assay, which allows for growth analysis over a shorter window of time. Furthermore, their study did not look at the repair enzyme hMYH and its impact on survival. Another study has investigated the repair function of hMYH in MYH-deficient murine cells. It was demonstrated that transfection of the MYH-deficient cells with a wild-type MYH expression vector increased the efficiency of A:GO repair [44] . An interesting observation noted while doing our experiments lead us to look at individual growth characteristics of cells over-expressing each of the oxidative repair enzymes. Cells over-expressing the repair enzyme hMYH clearly grow at a slower rate when compared with the other enzymes. The mechanism behind this is not understood at this point in time. The repair action of Cell growth curve and associated doubling times (DT) Figure 6 Cell growth curve and associated doubling times (DT hMYH is known to remove adenines misincorporated opposite 8-oxoG lesions. This lesion occurs when a C/GO lesion is allowed to replicate before being corrected by hOgg1. Repair by hMYH is not a final corrective measure. The product of hMYH activity is the lesion C/GO, which allows hOgg1 to have another opportunity to remove 8-oxoG opposite cytosine. We know that A549 cells possess the hOgg1 gene based on a previous study demonstrating the presence of this gene after amplification by genomic PCR [45] . We also have demonstrated endogenous activity of hOgg1 in A549 cells by using an 8-oxoguanine bioactivity assay. Therefore, our explanation of these results is that the slowed growth created by hMYH may provide a wider window of opportunity for the repair process to take place, which ultimately grants endogenous hOgg1 another opportunity to remove the 8-oxoG lesion created by oxidative stress. As noted in the methods section, the SRB assay provides a sensitive measure of drug-induced cytotoxicity that is used to assess cell proliferation/survival. The reduced cell proliferation of A549 cells over-expressing hMYH under nontoxic conditions may likely underestimate the magnitude of the protective effect of this particular repair enzyme. This may in fact make the results even more significant. Recent studies have discovered hereditary variations of the glycosylase hMYH that may predispose to familial colorectal cancer [46, 47] . Others have looked for hMYH variants in lung cancer patients and have not identified any clear pathogenic biallelic hMYH mutations or an overrepresentation of hMYH polymorphisms [47] . The A549 cell line has not demonstrated somatic mutations in hMYH, but a single nucleotide polymorphism (SNPs) has been noted [45] . The impact on function by this SNP is unknown. It would appear that the function of hMYH is very important in preventing somatic mutations leading to cancer in the gastrointestinal tract. Although studies to date have not demonstrated this same relationship with lung cancer, we do know that the lungs are subjected to large quantities of ROS under certain conditions as discussed earlier. The formation of mutations from oxidative stress does have other deleterious effects on cells including cellular death by necrosis and apoptosis. Tissue viability is dependent upon mutation correction and replication of the surviving cells to replace those that have died. The ability to enhance cellular survival, after specific oxidative exposures, is evident after increased production of the hMYH repair gene in these experiments. We additionally wanted to determine the level of endogenous expression of the glycosylase repair genes in the pulmonary epithelial A549 cell line. Others have demonstrated how different stressors lead to alterations in the endogenous production of specific repair genes. For example, it has been shown that endogenous gene expression of hOgg1 was elevated following exposure to crocidolite asbestos which is known to cause an increase in 8-oxoG levels [48] . It has also previously been reported that treatment of A549 cells with sodium dichromate, a prooxidant, leads to a reduction of hOgg1 protein expression that was not observed with H 2 O 2 [49] . One additional study demonstrated a dose dependent down regulation of hOgg1 protein expression in rat lung after exposure to cadmium, a known carcinogen associated with the formation of intracellular ROS [50] . In our experiments we were able to demonstrate that both hyperoxia and IR do not appear to impact the endogenous expression of hOgg1, hMYH, and hMTH at 24 and 48 hours following exposure. It was noted that endogenous hNTH was reduced after hyperoxia at 24 and 48 hours after the onset of exposure. One would speculate that this reduction in endogenous hNTH secondary to hyperoxia is related to either decreased production or increased destruction in response to O 2 exposure. Over-expression of this repair enzyme did not result in improvements in survival after O 2 exposure based on our experiments. It may be that endogenous levels are adequate to correct this specific mutational burden for these experiments. Furthermore, no previous studies have determined how cells over-expressing specific repair genes may impact endogenous expression of the other oxidative BER genes under both normal and oxidative stress conditions. We were also able to demonstrate that endogenous expression of glycosylase repair genes were not altered under these conditions secondary to the over-expression of any of these genes. This is an important finding for interpretation of survival data; protection of cells is due to the overexpression of the specific gene and not due to enhancement of other endogenous repair enzyme levels, at least for the genes studied under these conditions. Some limitations may exist in using a lung carcinoma cellline, which likely differs both in proliferative properties as well as in response to oxidative stress in comparison to primary epithelial cells. The enhanced cell growth observed with cell lines may be more reflective of undifferentiated alveolar type II cells which are likely to replace terminally differentiated alveolar type I cells after injury/ death due to oxidative stress. This may not be a true reflection of growth under non-toxic conditions when very little cell division is occurring. This is an inherent problem observed when comparing cell lines with primary cells and results need to be interpreted in a way that considers this. It is difficult to know how this will translate to pulmonary epithelial cells in vivo at this stage. It certainly would appear that the protection observed is modest in degree in this pulmonary epithelial cell line. Further experiments assessing the function of the repair enzyme hMYH in this model will be important to perform in order to delineate the findings of slowed growth under normal conditions and improved survivability under conditions of O 2 and IR. More research looking at the potential for combination therapy, including DNA repair mechanisms in conjunction with other antioxidant defense mechanisms may be another approach to enhancing cell survival, which may lead to better clinical outcomes. Alternatively, cell survival may not be the most important end point for hyperoxia studies. Given that 8-oxoG, if left unrepaired, leads to G:C to T:A transversions, there may be an increase in mutational burden by these cells that isn't reflected in cell survival. Further experiments studying the impact on mutation production is underway. Ultimately, experiments need to be done in animal models to determine the translation to in vivo pulmonary cells. In summary, we have demonstrated that over-expression of the DNA glycosylase repair enzyme hMYH may enhance survival of a pulmonary epithelial cell line after exposure to conditions of IR and hyperoxia. We have also demonstrated that over-expression of hMYH leads to a slowing of growth of A549 cells under non-toxic conditions, which may in part play a role in this enhancement of survival by providing a wider window of opportunity for repair of oxidized lesions to occur. Lastly, we demonstrated that over-expression does not lead to altered endogenous expression of these repair genes. As the understanding of DNA repair mechanisms continues to grow and the evolution of gene therapy takes place, more treatment options may be available in the clinical setting to help with many disease processes including the damaging effects of oxygen and its metabolites.
18
Bioinformatic mapping of AlkB homology domains in viruses
BACKGROUND: AlkB-like proteins are members of the 2-oxoglutarate- and Fe(II)-dependent oxygenase superfamily. In Escherichia coli the protein protects RNA and DNA against damage from methylating agents. 1-methyladenine and 3-methylcytosine are repaired by oxidative demethylation and direct reversal of the methylated base back to its unmethylated form. Genes for AlkB homologues are widespread in nature, and Eukaryotes often have several genes coding for AlkB-like proteins. Similar domains have also been observed in certain plant viruses. The function of the viral domain is unknown, but it has been suggested that it may be involved in protecting the virus against the post-transcriptional gene silencing (PTGS) system found in plants. We wanted to do a phylogenomic mapping of viral AlkB-like domains as a basis for analysing functional aspects of these domains, because this could have some relevance for understanding possible alternative roles of AlkB homologues e.g. in Eukaryotes. RESULTS: Profile-based searches of protein sequence libraries showed that AlkB-like domains are found in at least 22 different single-stranded RNA positive-strand plant viruses, but mainly in a subgroup of the Flexiviridae family. Sequence analysis indicated that the AlkB domains probably are functionally conserved, and that they most likely have been integrated relatively recently into several viral genomes at geographically distinct locations. This pattern seems to be more consistent with increased environmental pressure, e.g. from methylating pesticides, than with interaction with the PTGS system. CONCLUSIONS: The AlkB domain found in viral genomes is most likely a conventional DNA/RNA repair domain that protects the viral RNA genome against methylating compounds from the environment.
The purpose of this study has been to identify domains with homology to AlkB in viral genomes, in order to get a better understanding of distribution and possible function of such domains. The AlkB protein of E. coli, and probably most of its homologues, is involved in repair of alkylation damage in DNA and RNA. It repairs 1-methyl-adenine and 3-methylcytosine by oxidative demethylation and direct reversal of the methylated base back to its unmethylated form. Recently the protein was identified as a member of the 2-oxoglutarate (2OG)-and Fe(II)dependent oxygenase superfamily [1] [2] [3] . The catalytic reaction requires molecular oxygen, Fe 2+ and 2-oxoglutar-ate, which is subsequently converted into succinate, CO 2 and formaldehyde [4] . The 2OG-FeII oxygenase superfamily is widespread in Eukaryotes and bacteria [1] , and is currently the largest known family of oxidising enzymes without a heme group [5] . The 3D structure of several of these oxygenases is known, and they share a common fold with a structurally conserved jelly roll β-sheet core with flanking α-helices. Very few residues are totally conserved across these structures, basically just the residues involved in coordination of the Fe(II) ion and the 2-oxoglutarate. AlkB-like genes are widespread in most types of organisms except Archaea. However, whereas bacteria normally have just one or at most two AlkB homologues [6] , multicellular Eukaryotes tend to have several homologues. In the human genome at least 8 different AlkB homologues (ABHs) have been identified [7] . These homologues seem to have slightly different properties with respect to substrate preference and subcellular localisation, and this may be a reason for the proliferation of ABHs e.g. in humans. However, a detailed functional mapping of all ABHs has not yet been carried out. A sequence alignment of known ABHs identifies very few residues as totally conserved, basically just a HxD motif, a H and a RxxxxxR motif. These residues are also conserved in the more general 2OG-FeII oxygenase superfamily as described above, except for the final R. The first three residues (HxD and H) are involved in Fe(II)-coordination, whereas the first R is involved in 2OG-coordination. The final R is most likely involved in AlkB-specific substrate binding. In addition to DNA repair, it has been shown that E. coli AlkB and the human AlkB homologue hABH3 may be involved in RNA repair. When expressed in E. coli both AlkB and hABH3 reactivate methylated RNA bacteriophage MS2 in vivo. This illustrates that direct repair may be an important mechanism for maintenance of RNA in living cells [4] . RNA repair proceeds by the same mechanism as DNA repair. Repair of damaged RNA was previously considered very unlikely, due to the natural redundancy of RNAs in a cell [8] . However, RNA is essential for cell function: unrepaired RNA can lead to miscoded or truncated proteins, and alkylated RNA could signal cell cycle checkpointing or apoptosis [9] . Consequently the occurrence of RNA repair does not come as a great surprise. The mechanism of direct reversal of methylation as used by AlkB homologues is particularly important for RNA repair, as it means that single-stranded regions may be repaired without introducing strand breaks. Repair of alkylation damage in DNA and RNA has recently been reviewed [10] . AlkB homologues have also been found in plant viruses. It has been suggested that methylation may be used in host-mediated inactivation of viral RNAs, and that AlkB homologues in some plant viruses may be used to counteract such defence mechanisms [1] . However, no detailed study of this has been published. The research project reported here has focused on a better understanding of the distribution and potential function of putative AlkB homology domains by using in silico mapping of viruses in which such domains have been found, as well as related viruses. The general mapping strategy of the project was to identify viral genomes with AlkB homology domains, identify common features of these genomes, and subsequently find additional genomes with similar features, but without AlkB homology domains. This data set could then be used to analyse the properties and distribution of AlkBlike domains in viruses, as a basis for generating hypotheses about the evolution and function of these domains. The PSI-Blast search for viruses in the NCBI nr protein sequence database was initiated with ALKB_ECOLI (NCBI gi113638), restricted to residues 110 to 210 and using the default inclusion threshold of 0.005 on E-values. The [11] . In all of these viruses the AlkB domain is a part of the replicase polyprotein, which normally consists of a viral Other Pfam domains -Peptidase_C21, C23, C33, C34, C35 and C41, A1pp and OTU -were also identified in subsets of sequences. A1pp is a member of the Appr-1-p processing enzyme family, and the domain is found in a number of otherwise unrelated proteins, including non-structural proteins of several types of ssRNA viruses. OTU is a mem-ber of a family of cysteine proteases that are homologous to the ovarian tumour (otu) gene in Drosophila. Members of this family are found in Eukaryotes, viruses and pathogenic bacteria. The MT, HEL and RdRp domains identified by Pfam as described above were extracted from the library sequences, aligned by ClustalX, and combined into a new alignment consisting of only these domain regions. This turned out to be necessary in order to get robust alignments. The intervening regions between the conserved domains are extremely variable in these sequences, and this tended to confuse alignment programs in the sense that conserved regions were not correctly aligned. The combined sequence alignment of domains from Closteroviridae, Flexiviridae and Tymoviridae was then used as input for building a phylogenetic tree with MEGA2. The final tree is shown in Figure 2 , with polyproteins containing AlkB-like domains indicated. A second alignment was generated from all sequences with AlkB-like domains, using only the regions corresponding to MT, AlkB, HEL and RdRp Pfam domains. The domains were aligned individually, and the combined alignment was used as input for MEGA2. However, this data set did not give a reliable phylogeny (data not shown), and the separate domains of this alignment were therefore analysed individually and compared. This analysis is summarised in Tymoviridae measures (including SJA) for comparison of random trees [12] . The SJA values shown in Table 2 for comparisons between MT, HEL and RdRp NJ trees were 14.2 -17.1 standard deviations from the expectation value of 0.665 for a tree with 22 nodes, whereas the corresponding values for the AlkB NJ tree were 4.4 -5.4 standard deviations from the expectation value. Similar ranges were observed for the ML trees as well as for alternative distance measures, e.g. the Symmetric Difference (SD) measure (data not shown). Although this means that the SJA value for comparing AlkB trees to MT, HEL and RdRp trees were significantly better than for random trees, it also shows that the MT, HEL and RdRp trees were clearly more similar to each other than to the AlkB tree. The alignment of the AlkB domain seemed to be of comparable quality to the other alignments. In fact the AlkB domain had the highest average pairwise sequence identity, as seen in Table 2 (see Figure 3 for the actual alignment). In other words, these AlkB domains were as similar to each other as the other three domains with respect to sequence identity, but they did not represent a consistent evolutionary history when compared to the other domains of this polyprotein. This may indicate that the AlkB domains have evolved separately from the other domains, and possibly as several independent instances. The degree of co-evolution was analysed by computing pairwise distances between sequence regions in the alignment of MT, AlkB, HEL and RdRp domains described above. In Figure 4 selected results are shown as scatter plots, where the Blosum 50 score value between e.g. the MT domains in a pair of sequences is plotted against the score value for AlkB domains in the same pair of sequences. Plots for the MT, HEL and RdRp domains show that they are strongly correlated for MT vs. RdRp (r 2 = 0.95), MT vs. HEL (r 2 = 0.87) and HEL vs. RdRp (r 2 = 0.81). The plot of the AlkB domain vs. these three domains for the same set of sequences shows a very low degree of correlation for AlkB vs. RdRp (r 2 = 0.10), AlkB vs. MT (r 2 = 0.12) and AlkB vs. HEL (r 2 = 0.16). As mentioned above the genome organisation of these replicase polyprotein sequences seems to be very flexible. In order to analyse domain organisation the location of identified Pfam domains were plotted for a number of sequences, as shown in Figure 5 . The results described above may indicate that the AlkB domains have been integrated into the replicase polyprotein relatively recently (see Discussion). In order to test for potential sources selected AlkB domains were compared to non-viral sequences. PSI-Blast was used to search the NCBI nr database, removing all viral hits in the final search report. Most of the remaining top-scoring hits were from bacteria. This included two different strains of Xanthomonas, X. axonopodis pv citri and X. campestris pv campestris. Xanthomonas attacks plants such as citrus, beans, grapevine, rice and cotton [13] . The search also returned high-scoring hits from another plant pathogen, Xylella fastidiosa. This bacterium infects a great variety of plants, including grapevine, citrus, periwinkle, almond, oleander and coffee [14] . Pfam searches obviously will only identify known domain types in protein sequences. In order to identify potential similarities in regions that were not recognised by Pfam, systematic PSI-Blast searches were performed, using the polyprotein regions between the MT and HEL domains and searching against the NCBI database of reference sequences [15] , excluding all viral entries. A maximum of 5 PSI-Blast iterations were allowed, with an inclusion threshold of 0.005. The expected homologues of the AlkBdomain were found with high confidence, as most of the E-values were < 1 × 10 -50 . Homologues of typical viral domains like the viral peptidases were obviously not found, as all viral database entries were excluded. Very few Multiple alignment of sequence regions corresponding to the AlkB domains Figure 3 Multiple alignment of sequence regions corresponding to the AlkB domains. The alignment was generated with ClustalX. The residues involved in coordination of the essential Fe 2+ ion are completely conserved, except in one of the Vitivirus sequences. These residues are the HxD motif, a single H, and the first R in the RxxxxxR motif. The function of the remaining conserved residues is unclear, but at least some of them may be involved in coordination of the substrate [10] . Pairwise distances between sequence regions corresponding to methyltransferase (MT), RdRp and AlkB domains. Each data point corresponds to e.g. RP-RP and MT-MT distances for the same pair of sequences, and sequences showing similar evolutionary distance in these two regions will fall on the diagonal. The pairwise distances were estimated from multiple alignments using the Blosum50 score matrix [47] . Trend lines were estimated with Excel. The trend line for AlkB vs. RdRp is heavily influenced by the point at (675, 670). It represents two Foveavirus sequences (NCBI gi3702789 and gi9630738), they are 98% identical over the full polyprotein sequence. Alignment score (AlkB) Alignment score (RdRp) r 2 = 0.10 new similarities were found by these searches. Pepper ringspot virus (Tobravirus, NCBI gi20178599) showed significant similarity to site-specific DNA-methyltransferase from Nostoc sp (E = 1 × 10 -74 ), as well as other cytosine 5Cspecific DNA methylases. Bamboo mosaic virus (Potexvirus, NCBI gi9627984) showed similarity to aggregation substance Asa1 from Enterococcus faecalis (E = 6 × 10 -34 ). A small number of additional similarities seemed to be caused by biased sequence properties (e.g. proline-rich regions), and were probably not significant. This included matches against mucin and cadherin-like proteins from Homo sapiens and multidomain presynaptic cytomatrix protein (piccolo) from Rattus norvegicus. In general the variable regions seemed to be truly variable, with very little similarity to other proteins, except for the Pfam domains already identified. As seen in Figures 2 and 5 , some closely related sequences are lacking specific domains in the sense that HMMER does not find a significant similarity to the Pfam entries for these domains. In order to understand the degree of sequence variation associated with this domain loss, as well as the general sequence variation in conserved vs. non-conserved regions of typical polyproteins, several dot plots were generated. The dot plot for two Carlavirus sequences, Potato virus M (NCBI gi9626090) and Aconitum latent virus (NCBI gi14251191), is shown in Figure 6 . The dot plot confirms that these two sequences are closely related in the MT, HEL and RdRp domains. However, there are significant differences in the region between MT and HEL. Potato virus M is lacking the AlkB domain whereas Aconitum latent virus is lacking the OTU domain. As seen from the dot plot, short regions of similarity close to the diagonal shows that both domains may have been present in an ancestral sequence. However, this region shows a high degree of sequence variation, and as indicated by the dot plot they are almost exclusively mutations. Non-essential or non-functional domains are probably rapidly lost. In this particular case, none of the typical AlkB motifs seem to be conserved in Potato virus M, indicating that this indeed is a non-functional AlkB domain. The N-terminal domains of Flexiviridae and Tymoviridae are methyltransferases As described above the Pfam methyltransferase motif (Vmethyltransf) did not match any of the putative methyltransferase domains of Flexiviridae and Tymoviridae, despite the fact that they had been identified via PSI-Blast searches starting with known methyltransferases. Therefore an additional Pfam-type profile was generated. It is obviously a possibility that these domains in Flexiviridae and Tymoviridae are not methyltransferases, and that they are false positives from PSI-Blast. However, the essential residues of a typical viral methyltransferase motif are conserved in the alignment of these domains (data not shown) [16] . In Bamboo mosaic virus, which belongs to Flexiviridae, the residues H68, D122, R125 and Y213 have been identified as putative active site residues with similarity to the Sindbis virus-like methyltransferase [17] , and it has been demonstrated that this region of the Bamboo mosaic virus has methyltransferase activity, as it catalyses the transfer of a methyl group from S-adenosylmethionine (AdoMet) to GTP or guanylylimidodiphosphate (GIDP). The corresponding sequence positions are almost completely conserved in the alignment of Flexiviridae and Tymoviridae N-terminal domains. This is most likely significant, as only 7 positions in total are completely conserved in this alignment, which means that the majority of the conserved positions are known to be essential for methyltransferase activity. Work e.g. by Hataya et al. seems to support the assumption that this sequence region is a methyltransferase domain [18] . It therefore seems likely that all the sequences with AlkB domains also contain functional MT, HEL and RdRp domains. The MT Location of Pfam domains in the variable region of Flexiviridae 2 sequences Figure 5 Location of Pfam domains in the variable region of Flexiviridae 2 sequences. The regions have been extracted directly from Pfam output, and sequences and regions are drawn to scale. The black bar at each end of a motif indicates that a fulllength motif has been found, for partial motifs the bar at the truncated end would be missing. domains are probably involved in capping of genomic and subgenomic RNA [19] . Based on the bioinformatic evidence generated here, it seems reasonable to assume that the viral AlkB domains identified by Pfam are functional. All the essential residues found in 2-oxoglutarate-and Fe(II)-dependent oxygenases are conserved, in particular the putative Fe 2+ coordinating H, D and H residues at alignment positions 19, 21 and 91 of Figure 3 , and the 2-oxoglutarate coordinating R at position 100. The conserved R at position 106 is also very characteristic of AlkB homologues [10] . The fact that all AlkB-like domains identified in these viral genomes are full-length, compared to the Pfam profile, also seems to support the hypothesis that these domains are functional. The Pfam searches show that AlkB domains are found only in a subset of the viral genomes. This subset is phylogenetically consistent (see Figure 2 ), as it is mainly restricted to the Flexiviridae, and in particular to a subset of the Flexiviridae consisting of Viti, Capillo, Tricho, Fovea and Carlavirus. This subset is well separated from the remaining Flexiviridae in the phylogenetic analysis. The split seems to be robust from bootstrap analysis, therefore this family will be discussed here as two subfamilies, Flexiviridae 1 and 2. The same split was observed by Adams et al. in their recent analysis of the Flexiviridae family [20] . Most of the AlkB domains (15) are found in Flexiviridae 2. The remaining AlkB domains are found in Flexiviridae 1 (5) and Closteroviridae (2) . In general, all the Flexiviridae 2 sequences have at least one extra domain in addition to MT, HEL and RdRp: either AlkB, OTU-like cysteine protease or a peptidase. Most other plant viruses that are included in this survey do not have additional domains, except for Tymoviridae where a peptidase domain seems to be common. For the remaining plant virus families included here (excluding Tymoviridae and Flexiviridae 2), only 14% seem to have additional domains. The observed distribution of AlkB domains could most easily be explained by assuming that an ancestral AlkB domain was integrated into the genome of the last common ancestor of the Flexiviridae 2 subfamily. Subsequent Figure 6 Dot plots for Potato virus M (NCBI gi9626090) and Aconitum latent virus (NCBI gi14251191). To the left the full sequences are shown, using the program default for similarity threshold, and to the right the region with AlkB, OTU and peptidase integration, using a slightly lower (more sensitive) threshold for sequence similarity. The Pfam regions corresponding to MT (magenta), AlkB (red), OTU (green), peptidase (blue), HEL (yellow) and RdRp (cyan) domains are indicated. virus generations derived from this common ancestor would then also contain an AlkB domain, except in those cases where the domain was lost again. This scenario could also include subsequent transfer to a small number of other virus families e.g. by recombination. If this scenario was correct, then one would expect the different domains of the polyprotein to have a similar evolutionary history. From the phylogenetic analysis (Table 2) this seems to be confirmed for the MT, HEL and RdRp domains, but not for the AlkB domain. This indicates that the AlkB domain may not have co-evolved with the other domains, at least until relatively recently. This seems to be confirmed by looking at the degree of co-evolution, which was analysed by computing pairwise distances between alignment regions representing the relevant domains ( Figure 4 ). In the case of perfect co-evolution all points should fall on a diagonal. This seems to be the case for the MT, HEL and RdRp domains. However, the plot of the AlkB domain vs. these three domains for the same set of sequences does not show a similar correlation. Only some of the closely related sequence pairs in the upper right quadrant of the plot in Figure 4 show some degree of correlation for AlkB vs. RdRp. The most likely explanation seems to be that most of the AlkB domains have not coevolved with the other domains for any significant period of time. This seems to rule out the possibility of ancient integration of the AlkB domain, except if we assume that an ancient viral AlkB domain has frequently recombined with other AlkB domains. However, it is difficult to distinguish a scenario with frequent recombination of AlkB domains from de novo integration, and the net effect on the properties observed here would be the same. As seen in Figure 4 , the range of score values is generally smaller for the AlkB domains than e.g. the RdRp domains, particularly if we exclude a couple of very high-scoring cases (see figure caption) . On the other hand, the degree of sequence variation within the collection of AlkB domains is significant, average sequence identity for pairwise alignments is 38%, and only 10% of the positions are totally conserved. This can be consistent with a recent integration if we assume that several different AlkB-type vectors have been used for integration (see below for details). An increased mutation rate after integration could also have contributed to sequence diversity in this region. Moving the AlkB domain into a novel structural and functional context would have removed many of the original evolutionarily constraints, as well as introduced some new ones. This could have created a "punctuated equilibrium" type of situation, potentially leading to a very rapid evolution that could have introduced significant differences between the AlkB domains, independent of the evolution in the other domains. A high mutation rate seems to be the case for this region in general, as indi-cated in Figure 6 . Although the MT, HEL and RdRp domains seem to be well conserved from the dot plot, there are very large sequence variations in the intervening region. One sequence in Figure 6 has a well conserved AlkB domain, the other an OTU domain. The fact that there are very weak sequence similarities in these two domains in the dot plot indicates that both sequences originally had both domains. However, the fact that this similarity now is very weak and without any of the typical AlkB active site motifs also indicates a high mutation rate where non-essential domains are rapidly lost. Therefore the conservation of AlkB domains is a strong indication that they are functional, as already mentioned. If we assume that AlkB domains have been integrated relatively recently, then either de novo integration or recombination (horizontal gene transfer) may have been the main driving force for spreading the AlkB domain to new genomes. In the first case a large number of individual integrations could have lead to the present situation. If horizontal gene transfer was the main driving force, the initial number of integrations might have been quite small. It is not easy to differentiate between these two situations. The map of Pfam motifs in the variable region between the MT and HEL domains in Flexiviridae 2 polyproteins ( Figure 5) shows that they have a very similar domain organisation, basically an AlkB domain followed by an OTU domain and a peptidase domain, located towards the C-terminal part of the sub-sequence. The relatively constant domain organisation seems to be consistent with a small number of initial integrations that were subsequently diffused to related genomes e.g. by homologous recombination. However, this is not fully consistent with the fact that the viruses with AlkB domains have been collected from hosts at very different locations, e.g. Canada, USA, Russia, Italy, Germany, France, India, Taiwan, China and Japan. Although import of virus-infected species or transmission by insects may transport viruses over significant distances, it is not obvious that this is enough to explain the observed distribution of AlkB-like domains. Therefore several independent integrations, mainly from closely related hosts, have to be considered as an alternative explanation. This explanation seems to be supported by the apparent lack of any consistent evolutionary relationships between the various AlkB domains, as seen in Table 2 . It is not easy to see how this model can be consistent with the observed similarities in domain organisation in Flexiviridae. Assuming that this region has a high degree of variability, one would expect the variability to affect localisation of integrated domains as well. However, it is possible that conserved regions e.g. in the polyprotein play a significant role in integration of novel domains. It may be relevant in this context that preliminary simulations indicate that e.g. the AlkB domains tend to form independent folding domains in the folded RNA structure of the polyprotein RNA (F. Drabløs, unpublished data). This property may possibly facilitate the insertion of such domains into the viral genome. There are many groups of organisms that can act as vectors and spread viruses, including bacteria, fungi, nematodes, arthropods and arachnids. The plant viruses may have acquired the AlkB domain either from the vector or from the host itself. As already mentioned, searching with viral AlkB domains in protein sequence databases resulted mainly in bacterial sequences, including the plant pathogens X. fastidiosa and campestris. It is therefore a reasonable possibility that AlkB domains in plant viruses have originated from bacterial mRNA. It is also possible that the mRNA originated from other vectors or from the host itself, but at the present time this is not easily verified or disproved because of the limited number of insect and plant genomes that have been sequenced. It has previously been suggested that the viral AlkB domain may be involved in protecting the virus against the post-transcriptional gene silencing (PTGS) system of the host [1] . PTGS is known as one of a plant's intrinsic defence mechanisms against viruses [21] . Gene silencing can occur either through repression of transcription (transcriptional gene silencing -TGS) or through mRNA degradation, PTGS. The PTGS-mechanism in plants shows similarities to RNA interference (RNAi) in animals [22] . This mechanism results in the specific degradation of RNA. Degradation can be activated by introduction of transgenes, RNA viruses or DNA sequences homologous to expressed genes [23] . Many viruses have developed mechanisms to counteract PTGS in order to successfully infect plants [24] . Two of these suppressors of PTGS have been identified as Hc-Protease and the 2b protein of Cucumber mosaic virus [25] . Although both proteins suppress PTGS, it is likely that they do so via different mechanisms. Could the AlkB-like domain found in some of the plant viruses also be a suppressor of PTGS? Previously reported research indicates that methylation of transcribed sequences is somehow connected with PTGS, and the methylation can be mediated by a direct RNA-DNA interaction [26] . This RNA-directed DNA methylation has been described in plants, and leads to de novo methylation of nearly all cytosine residues within the region of sequence identity between RNA and DNA [27] . Both RNA methylation and methylation of host proteins that are essential for viral replication would be detrimental to the virus. It has already been mentioned that AlkB repairs 1methyladenine and 3-methylcytosine by oxidative demethylation. It is therefore possible that AlkB demethylates the nucleotides methylated by the PTGS mechanism, helping the virus to overcome one of the major defence mechanisms of the plant. As shown here, only a subset of plant viruses have the AlkB domain. However, other viruses may be utilising naturally occurring AlkB proteins in the host. Viruses have to rely on a number of host proteins in order to replicate [28] . In some cases it is probably beneficial for the virus to integrate such genes into their own genome in order to ensure that they are accessible, although there will be a trade off between this advantage and the increased cost of maintaining a larger genome [29] . However, there is an alternative hypothesis with respect to the AlkB integration that also has to be considered. As discussed above, the AlkB domain seems to have been integrated relatively recently in viruses found at very different geographical locations, and the only obvious connection seems to be that most viruses belong to a subset of the Flexiviridae. However, the source of these viruses points at another common feature. As seen from the table given in Additional file 1, AlkB domains are often found in viruses associated with grapevine, apple, cherry, citrus and blueberry -crops where the usage of pesticides is common. It is known that several common pesticides (e.g. methyl bromide and some organophosphorus compounds) may cause methylation of DNA and RNA [30] [31] [32] [33] . An integrated repair domain for methylation damage as part of the viral replication complex would therefore give the virus a competitive advantage in a highly methylating environment. The application of such pesticides would probably also stimulate AlkB production e.g. in co-infecting bacteria, giving these viruses easy access to AlkB mRNA for integration into their RNA genome. It could be argued that a more active PTGS system in these plants would give a similar effect. However, in that case we would expect to see more ancient integrations of AlkB domains. It could also be argued that the presence of AlkB domains may be an artefact caused by promiscuous viral domains picking up available mRNA sequences during cultivation of viruses in the laboratory. However, given the large number of different laboratories involved, and the number of different hosts used (data not shown), this seems to be a very unlikely explanation. The hypothesis that environmental compounds, in particular pesticides, may have provoked the integration of AlkB domains into the viral genomes depends upon a high mutation rate and frequent integrations of non-viral domains. The integrations have to be recent, not only in relative terms, compared to other domains in the same genome, but also in absolute terms, compared to the progress of modern agriculture. The integrations also have to be frequent, in the sense that it is likely that integration could have happened several times, in different biotopes. It is difficult to estimate mutation rates in RNA viruses. They evolve very rapidly, and it is often difficult to assign reliable phylogenies. However, recent studies indicate that most ssRNA viruses have a mutation rate close to 10 -3 substitutions per site per year [34] , e.g. the SARS virus has 1.16-3.30 × 10 -3 non-synonymous substitutions per site per year, which is considered to be a "moderate" ssRNA mutation rate [34] . If we assume that most ssRNA viruses have effective mutation rates within the same order of magnitude, a realistic mutation rate for the viruses included here might be something like 2.0 × 10 -3 . In that case, the MT, HEL and RdRp trees shown in Additional file 2 represent approximately between 325 and 750 years of evolution. In general the NJ trees estimate a slightly shorter evolutionary history (between 325 and 450 years) compared to the ML trees (between 550 and 750 years). In this estimate the Ampelovirus sequences have not been included, as they seem to have diverged from the remaining AlkB-containing viruses at a much earlier stage. If we believe that the AlkB integrations happened after the divergence of most sequence included here, as indicated by the lack of co-evolution in Figure 4 , it does not seem unrealistic to assume that most of these integrations happened within the last 50 -100 years or so. This estimate is of course very approximate, in particular since we do not know the true mutation rate of these viruses. However, it shows that a likely time span for AlkB integration is compatible with the evolution of modern agriculture. Unfortunately, because of the lack of any robust phylogeny for the viral AlkB sequences it does not make sense to do a similar estimate for that domain. Although it is generally accepted that viruses frequently use recombination to acquire functionality [35] , it is less well known how often this includes nonviral sequences. However, there are some well-documented examples, and in particular the properties of the ssRNA positive-strand Pestivirus may be relevant in this context. There are two biotopes of the pestiviruses, cytopathogenic (cp) and noncytopatogenic (noncp). The host is infected by the noncp form which is converted into the cp form by integration of a fragment of a cellular gene into the viral genome [36] . This introduces a protease cleavage site in the polyprotein. However, the important point here is that this happens as part of the normal infection process. It has been suggested that the integration is facilitated by the viral polymerase undergoing two subsequent template switches during minus-strand synthesis [37] , although nonreplicative RNA recombination also may be a possibility [38] . Inte-gration of cellular sequences have also been observed in other viruses, e.g. in influenza virus [39] . This shows that at least some viruses do have efficient mechanisms for recruitment of host genes into the viral genome. Therefore a recent and rapid integration of AlkB domains into selected plant virus genomes does not seem to be an unlikely scenario. This study has focused on the AlkB domain, mainly as an attempt to get a better understanding of potential functions associated with this domain. However, it is likely that additional information about integration patterns and the relative importance of de novo integration vs. recombination can be achieved by a closer investigation of the other variable domains, e.g. by looking at how they correlate with the evolution of the AlkB domains. We believe that the viral AlkB-like domains are conventional repair domains targeted towards the viral RNA. The integration of AlkB domains into viral genomes may have been provoked by environmental methylating agents, e.g. the introduction of DNA/RNA-methylating pesticides in farming. The hypothesis [1] that the domain interferes with the PTGS system of plants can not be excluded, but seems to be less consistent with observed features of the AlkB integration. and Tymoviridae was generated from a ClustalX alignment, using hmmbuild and hmmcalibrate from the HMMER package. Visualisation of motif positions in viral sequences was generated directly from the HMMER output files using a local tool as an interface to the GNU [50] groff software. Systematic large scale searches with polyprotein subsequences were done locally with PSI-Blast and the NCBI reference sequence database [15] . Dot plots for comparison of viral protein sequences were generated with Dotter version 3.0 [51] .
19
Managing emerging infectious diseases: Is a federal system an impediment to effective laws?
In the 1980's and 1990's HIV/AIDS was the emerging infectious disease. In 2003–2004 we saw the emergence of SARS, Avian influenza and Anthrax in a man made form used for bioterrorism. Emergency powers legislation in Australia is a patchwork of Commonwealth quarantine laws and State and Territory based emergency powers in public health legislation. It is time for a review of such legislation and time for consideration of the efficacy of such legislation from a country wide perspective in an age when we have to consider the possibility of mass outbreaks of communicable diseases which ignore jurisdictional boundaries.
The management of infectious diseases in an increasingly complex world of mass international travel, globalization and terrorism heightens challenges for Federal, State and Territory Governments in ensuring that Australia's laws are sufficiently flexible to address the types of problems that may emerge. In the 1980's and 1990's HIV/AIDS was the latest "emerging infectious disease". Considerable thought was put into the legislative response by a number of Australian jurisdictions. Particular attention had to be given to the unique features of the disease such as the method of transmission, the kinds of people who were at risk, and the protections needed by the community and the infected population to best manage the care of those infected and to minimize new infections. Health workers and researchers began to find that "the most effective strategies that we have so far found to help promote reduction of the spread of HIV involve the adoption of laws and policies which protect the rights of people most at risk of infection" [1] . A good example of a legislative response which adopts this approach is found in section 119 and 120 of the Victorian Health Act 1958. These sections emphasize the need to protect the privacy of the infected individual and to undertake a staged response which is proportional to the risk presented by the infected individual. The legislation has been very effective with HIV and has been praised for its progressive approach [2] . In 2003 the community has been faced with the emergence of two new infectious diseases, SARS and Anthrax. Whilst there were no cases of either disease in Australia, the threat of a possible outbreak had to be acknowledged and a response planned. Anthrax is not a new infectious disease. Humans can become infected with anthrax by handling products from infected animals or by breathing in anthrax spores from infected animal products (like wool, for example). People also can become infected with gastrointestinal anthrax by eating undercooked meat from infected animals. However, its manufacture and use as a weapon for bioterrorism forces us to rethink its management in a new context. These two infectious diseases have very different features from HIV which spreads only via transmission of infected bodily fluids such as blood or semen. SARS, by contrast is transmitted via droplets from infected cases which, as a result of coughing, carry the virus to close contacts [3] Thus, the infection profile of SARS requires planning for the possible overrun of Intensive Care Units and the likely infection of a number of ICU staff affecting both morale and capacity to cope. Anthrax raised different problems. These include the possible investigation of terrorist suspects alongside investigation of the outbreak of the infectious disease. Difficulties are also raised by likelihood of public panic, and the flooding of public health officials with reports of suspicious white powder. In early 2004 the media reported the spread of avian influenza across South East Asia. This disease has different features from HIV/AIDS and SARS and an approach to an Australian outbreak would also be different. The main difference is in the source of transmission of the virus, that is, from infected birds to humans. There is very little difference [from ordinary influenza] in the symptoms (though these may vary in severity) or treatment of the virus [4] It is too early to predict whether this may be the next "emerging infectious disease", but its current spread has given rise to concern about such a possibility [5] Australia is a federal system. There are two parallel sets of laws in operation. The Commonwealth Constitution sets out the legislative powers of the Commonwealth. Specific powers are listed in the Commonwealth constitution but State constitutions have broad powers covering matters such as peace, order and good governance. As the Commonwealth has no specific power to legislate with respect to health, other than the quarantine power, national legislative schemes in public health which rely upon a cooperative approach from all States and Territories are cumbersome and difficult. Without a specific head of power, the Commonwealth has limited ability to legislate with respect to health. "That is, the legislative powers of the Commonwealth are specified in the Constitution and do not include expressly most of the activities that together comprise the field of public health" [6] For this reason, there are no Commonwealth emergency health powers except quarantine powers. Quarantine powers are currently restricted to isolation at the border of the country of people, plants, and animals to prevent the spread of disease. There is a real possibility that quarantine laws could have a broader scope. It depends on how widely the High Court would interpret section 51(x) of the Commonwealth Constitution. A quarantine law could override state laws as long as it remained a law "with respect to quarantine". However, "the power is potentially a colossus so far as the expansion of legislative authority in the fields of public health is concerned". [6] The quarantine power would be the most likely candidate for a head of power on which to base development of commonwealth laws for the management of public health emergencies. Another possibility may be the external affairs power, if there was a relevant treaty or international agreement which could be given effect to in domestic law. However the legislation would have to be limited to laws giving effect to the treaty. States and territories have a range of emergency powers available to them in their existing public health legislation. Some are relatively old. For example, the Health Act 1911 (WA), Public Health Act 1952 (NT) based on an 1898 Ordinance (Both these Acts are currently under review). Health emergency powers vary from one jurisdiction to another, but include powers to support disease surveillance, contact tracing and orders to restrict behavior or movement of individuals with an infectious disease in certain circumstances. There are also powers to recall food, search premises and seize property, close buildings and a range of other substantial and intrusive powers. It is suggested that it is time to consider whether state and territory public health legislation contains sufficient measures to manage the outbreak of an infectious disease in a modern environment which includes mass travel, swift spread of infection and additional complexity raised by fears of bioterrorism. Currently, in a public health emergency caused by the spread of an emerging infectious disease, Australia could need to rely on a patchwork of legislative measures to assist it to cope. Commonwealth quarantine laws and State and Territory powers in public health legislation may all be needed to address the problem. If an outbreak occurred on a border, or in some area where jurisdiction may be in doubt such as airspace or offshore and a state or territory response was required in addition to any quarantine measures, there could be confusion over jurisdiction for the application of State and Territory powers. State and Territory public health acts do not adequately provide for interjurisdictional communication and cooperation. There could also be difficulties if an infectious disease caused overseas deaths of people from more than one State or Territory in circumstances where an Australian coronial investigation was considered desirable. In such a situation, the jurisdiction of more than one Australian coroner would be triggered. Several State and Territory coronial laws could apply and there could be different inquests under different laws undertaken by different coroners into the same incident. It is suggested that it is time to look at the efficiency of the emergency powers laws of Australia as a whole: to map the laws in each jurisdiction and the Commonwealth quarantine laws and to consider their effectiveness in the face of the outbreak of a fast moving, easily spread infectious disease. The efficacy of Australia's laws should also be considered in relation to bioterrorism. While there were no infections from anthrax in 2003 despite a great deal of media coverage and infections and deaths in the US, a responsible legislature ought to acknowledge the possibility and ensure that the law is ready to support a swift and effective response. It is not enough to consider whether the individual pieces of legislation are up to the task of managing outbreaks of newly emerging infectious diseases. Indeed many of the jurisdictions are currently reviewing their public health legislation and will no doubt give proper consideration to this issue as part of the review. But who is thinking about how the legislation of all jurisdictions and the Commonwealth quarantine fits together? What powers enable communication and cooperation between jurisdictions about the outbreak of infectious disease? What kind of opportunity is there for a coordinated response? Can public health orders made in one jurisdiction travel to another jurisdiction when the infected individual travels? What arrangements can be made if an outbreak occurs on or close to a interstate border? What if there is an outbreak on a bus carrying passengers from Victoria, through South Australia to the Northern Territory? It is encouraging to note that, even without specific legislation, there has been a mechanism to achieve communication and cooperation between jurisdictions through the Communicable Disease Network of Australia (CDNA). This Network has in fact been quite successful in fostering regular communication between the Communicable Disease Units across the country and has been involved in coordinated actions during a number of multistate outbreaks. Despite the existence of this network and other good working relationships between government officials and various agencies in different jurisdictions, a serious outbreak of communicable disease would require the existence of legislative powers. Public health emergencies generate confusion, even panic. Clarity of powers and the way those powers interact with each other would be crucial in an emergency. It became apparent after the Bali tragedy in 2002 that coroner's jurisdiction was triggered differently in different jurisdictions and some acts did not support communication and cooperation when inquests might be needed for deaths of people ordinarily resident in several jurisdictions. The time to find the shortcomings in the legislation is well before the crisis. A review of the efficacy of how these laws work together to protect the public health of all Australians should be undertaken. It has been possible to overcome the hangovers of federation for the betterment of all Australians in relation to corporations law. When doubts were recently raised about the constitutional basis of the corporations law scheme, the States and Territories were able to cooperate and refer the necessary powers to the Commonwealth to provide certainty about the laws which govern our corporations. Is our public health any less important than governance of our corporations? Could we cooperate to give ourselves certainty, flexibility and a consistent approach which protects the rights of those subject to some very broad powers? The States and Territories are generally reluctant to refer powers to the Commonwealth. It may be time to seriously discuss referral of powers in the context of health emergency powers. At the very least, it is time that the Commonwealth, States and Territories recognised the need for the laws to work as a set of laws to protect the whole country, not simply individual laws to protect individual jurisdictions. There has been work done internationally in this area. A model State Emergency Health Powers Act has been developed in the US in 2001 [7] In the preamble to this Act a rationale for its development is set out: "In the wake of the tragic events of September 11, 2001, our nation realizes that the Government's foremost responsibility is to protect the health, safety and wellbeing of its citizens. New and emerging dangers including emergent and resurgent infectious diseases and incidents of civilian mass casualties -pose serious and immediate threats to the population. A renewed focus on the prevention, detection, management and containment of public health emergencies is thus called for." The US, like Australia, is a Federal system. The model was intended to be taken up by those US states which wished to do so. To date, it has been passed in over half the US states. This bill would be an excellent starting point for development of an Australian model. There are a number of legislative mechanisms which could be used to support a nationally uniform approach to health emergency powers legislation in Australia. The development and adoption of the model food legislation provides a useful model of a cooperative uniform approach. A model act was developed in consultation with all jurisdictions. It covered areas agreed to be core areas of the Act which ought to be the subject of a national approach and other provisions which were considered to be administrative and were to be adopted at the discretion of each jurisdiction. An intergovernmental agreement was signed as a mechanism to protect the uniformity of the legislation. The agreement sets up a Ministerial Council, supported by a Food Regulation Standing Committee. The Council has responsibility for deciding on proposals to amend the model [8] If a decision is made in favor of amendment, States and Territories will use their best endeavors to submit to their respective Parliaments, legislation which gives effect to the amendment. The law is an important tool in supporting the management of the outbreak of infectious diseases. The existence of our Federal system has meant that we have a different approach in each State and Territory together with Commonwealth control of quarantine. Newly emerging infectious diseases creating real threats to public health in an era of easy mass travel, and the present threat of bioterrorism mean that it is time Australia examined all laws to contain and manage infectious disease outbreak. The laws should be examined both for their effectiveness in the areas they cover, and as part of a whole which ought enable a response which protects the health of all Australians, and crosses borders as easily as SARS or avian influenza.
20
Protein secretion in Lactococcus lactis : an efficient way to increase the overall heterologous protein production
Lactococcus lactis, the model lactic acid bacterium (LAB), is a food grade and well-characterized Gram positive bacterium. It is a good candidate for heterologous protein delivery in foodstuff or in the digestive tract. L. lactis can also be used as a protein producer in fermentor. Many heterologous proteins have already been produced in L. lactis but only few reports allow comparing production yields for a given protein either produced intracellularly or secreted in the medium. Here, we review several works evaluating the influence of the localization on the production yields of several heterologous proteins produced in L. lactis. The questions of size limits, conformation, and proteolysis are addressed and discussed with regard to protein yields. These data show that i) secretion is preferable to cytoplasmic production; ii) secretion enhancement (by signal peptide and propeptide optimization) results in increased production yield; iii) protein conformation rather than protein size can impair secretion and thus alter production yields; and iv) fusion of a stable protein can stabilize labile proteins. The role of intracellular proteolysis on heterologous cytoplasmic proteins and precursors is discussed. The new challenges now are the development of food grade systems and the identification and optimization of host factors affecting heterologous protein production not only in L. lactis, but also in other LAB species.
Lactic Acid Bacteria (LAB) are anaerobic Gram positive bacteria with a GRAS (Generally Regarded As Safe) status. They are also food grade bacteria, and therefore, they can be used for the delivery of proteins of interest in foodstuff or in the digestive tract. A last advantage compared to other well-known protein producers is that L. lactis does not produce LPS or any proteases as Escherichia coli or Bacillus subtilis do, respectively. In the last two decades, genetic tools for the model LAB, Lactococcus lactis, were developed: transformation protocols, cloning-or screening-vectors [1, 2] , and mutagenesis systems [3] are now available. Moreover L. lactis genome is entirely sequenced [4] . Many protein expression-and targeting-systems have also been designed for L. lactis [5] [6] [7] . These systems have been used to engineer L. lactis for the intra-or extra-cellular production of numerous proteins of viral, bacterial or eukaryotic origins (Table 1) . To produce a protein of interest in fermentors, secretion is generally preferred to cytoplasmic production because it allows continuous culture and simplifies purification. To use L. lactis as a protein delivery vehicle in the digestive tract of humans or animals, secretion is also preferable because it facilitates interaction between the protein (e.g. enzyme or antigen) and its target (substrate or immune system). In LAB, like in other Gram positive bacteria, secreted proteins are synthesized as a precursor containing an N-terminal extension called the signal peptide (SP) and the mature moiety of the protein. Precursors are recognized by the host secretion machinery and translocated across the cytoplasmic membrane (early steps). The SP is then cleaved and degraded, and the mature protein is released in the culture supernatant (late steps). Sometimes, secreted proteins require subsequent folding and maturation steps to acquire their active conformation [8] . In most of the works describing heterologous protein production by recombinant lactococci, only one cellularlocation (i.e. cytoplasm, external media or surface anchored) is described. Only a few works report the production of a given protein in different locations using the same backbone vector, the same induction level and or promoter strength, allowing thus a rigorous comparison of the production yields of cytoplasmic and secreted forms. Here, six examples of different heterologous proteins produced in L. lactis in both secreted and cytoplasmic forms are reviewed and discussed. Our major conclusion is that the best production yields are observed in most of these cases with secretion (up to five-fold higher than with cytoplasmic production). Moreover, engineering the expres-sion cassette to enhance the secretion efficiency (SE, proportion of the total protein detected as mature form in the supernatant) resulted in increased overall amounts of the protein. L. lactis is able to secrete proteins ranging from low-(< 10 kDa) to high-(> 160 kDa) molecular mass through a Sec-dependant pathway. Altogether, these observations suggest that i) heterologous proteins produced in L. lactis are prone to intracellular degradation whereas secretion allows the precursor to escape proteolysis, and ii) conformation rather than protein size is the predominant feature that can impair SE. New perspectives are now opened in the studies of heterologous protein production in L. lactis. Indeed, there is a need for food grade systems and for a better understanding of the host factors influencing heterologous protein secretion in L. lactis . For example, HtrA-mediated proteolysis (HtrA is the unique housekeeping protease at the cell surface) is now well-characterized in L. lactis [9] and can be overcome by use of a htrA L. lactis strain designed for stable heterologous protein secretion [10] . However, intracellular proteolysis (involving Clp complex -the major cytoplasmic housekeeping protease-, and probably other cellular components) remains poorly understood and is also discussed here. Genetic tools to target a given protein in different cellular compartments were developed using several reporter proteins [6, [11] [12] [13] (Table 1 ). The staphylococcal nuclease (Nuc) is a well-characterized secreted protein whose activity is readily detectable by petri plate assay and it has been used as a reporter protein for secretion studies in several Gram positive hosts [14] [15] [16] . In L. lactis, Nuc was used to develop protein targeting- [6] and SP screening-systems [1, 2] . Nuc was chosen to develop the pCYT and pSEC vectors for controlled production in L. lactis of cytoplasmic or secreted forms of a protein of interest, respectively ( Fig. 1 ) [5] . The pCYT and pSEC plasmids, where expression is controlled by a nisin inducible promoter, should be used in L. lactis NZ9000 (hereafter referred to as NZ) strain bearing a nisR,K chromosomal cassette, required for the nisin signal transduction [17] . In each case described below, protein sample concentration was adjusted to the cell density of the producing culture (for details see [18] ). At similar induction levels in lactococcal strains containing pCYT:Nuc and pSEC:Nuc vectors, the highest production yields were observed with the secreted Nuc form ( Table 2) . Similar results were obtained with constitutive nuc expression cassettes for cytoplasmic and secreted forms. Nuc was the first heterologous protein where highest protein yields were obtained with the secreted form. Similar results were obtained for the production of a Brucella abortus ribosomal protein. B. abortus is a facultative intracellular Gram negative bacterial pathogen that infects Unpublished results Bacteriocins human and animals by entry through the digestive tract. The immunogenic B. abortus ribosomal protein L7/L12 is a promising candidate for the development of oral live vaccines against brucellosis using L. lactis as a delivery vector. L7/L12 was produced in L. lactis using pCYT and pSEC vectors [19] . Similarly to Nuc production, the production yield of secreted L7/L12 was reproducibly and significantly higher than that of the cytoplasmic form (Table 2) . Another example of higher protein yields in secreted vs cytoplasmic form is the production the human papillomavirus type 16 (HPV-16) E7 antigen, a good candidate for the development of therapeutic vaccines against HPV-16 induced cervical cancer. The E7 protein is constitutively produced in cervical carcinomas and interacts with several cell compounds. E7 was produced in a cytoplasmic and a secreted form in L. lactis [20] . Using similar induction level in exponential phase cultures, E7 production 1: protein samples were adjusted to the cell density and protein quantification was performed as described in the references either by western blot or by ELISA. *: E7 was not quantified but ratio was calculated by scanning the western blot signals and comparing their intensity as described in the corresponding reference. nd: not determined was higher for the secreted form than for the cytoplasmic form (Table 2) . This difference was even higher when induction occurred in late-exponential phase, where intracellular E7 was detected at only trace amount whereas secreted E7 was accumulated in NZ(pSEC:E7) culture supernatant (see below). Thus, production of E7 clearly illustrates the fact that secretion results in higher yields in L. lactis. Production of ovine interferon omega (IFN-ω) further illustrates this observation. In the case of poorly immunogenic antigens, co-delivery of an immuno-stimulator protein can enhance the immune response of the host. In order to optimize the use of lactococci as live vaccines, the production of cytokines was investigated in L. lactis [5, 21, 22] . IFN-ω is a cytokine able to confer resistance to enteric viruses in the digestive tract by reduction of viral penetration and by inhibition of intracellular multiplication of the viruses. Delivery of ovine IFN-ω in the digestive tract by recombinant L. lactis strains could therefore induce anti-viral resistance and could protect the enterocytes. Ovine IFN-ω cDNA was cloned into pCYT and pSEC plasmids for intracellular (pCYT:IFN) and secreted (pSEC:IFN) production respectively [5] . Induction of recombinant NZ(pCYT:IFN) and NZ(pSEC:IFN) strains were performed at equal level and IFN-ω production was measured. The levels of IFN-ω activity showed that i) an active form of IFN-ω was produced in both strains, and ii) the activity of IFN-ω found in the supernatant and cell fractions of NZ(pSEC:IFN) strain was about two-fold higher than that observed for the cytoplasmic form (Table 2) . Similarly to what was observed for Nuc and E7, secretion leads to higher heterologous protein yields. L. lactis has been engineered to secrete of a wide variety of heterologous proteins from bacterial, viral or eukaryotic origins (Table 1) . There are reports about secretion bottlenecks and biotechnological tools for heterologous secretion in model bacteria such as Escherichia coli and Bacillus subtilis [23, 24] , but only few data are available concerning this aspect in L. lactis. Protein size, nature of the SP and presence of a propeptide are parameters that may interfere with protein secretion. Some data available about these features are compiled here. To optimize secretion and thus production yields, the nature of the SP was the first parameter to modify on heterologous precursor as previously shown using Nuc as a reporter protein. The replacement of the native staphylococcal SP Nuc by the homologous lactococcal SP Usp45 to direct the secretion of Nuc in L. lactis led to an increased SE [25] (Table 3) . On the other hand, the replacement of SP Nuc by SP Usp45 did not enhance the SE of NucT (a truncated mature moiety of Nuc devoid of N-terminal propeptide) suggesting the importance of the propeptide in the SE for Nuc [25] (Table 3) . However, in several cases, the use of a homologous SP (and especially SP Usp45 ) allows a better SE compared to a heterologous one. Screening vectors were thus developed to search for new homologous secretion signals in L. lactis [1, 2] . These screening works offer now a panel of SPs that are suitable for heterologous secretion. However, when compared to SP Usp45, the newly described SPs were less efficient to direct secretion of Nuc [1] . Even after a direct mutagenesis on SP310, one of these new SPs identified using a screening strategy [1] , the enhanced SE was still lower than the one measured with SP Usp45 [26] . However, a recent study by Lindholm et al. showed that a Lactobacillus brevis SP (originated from a Slayer protein) drove the secretion of the E. coli FedF Schematic representation of Nuc cassettes for controlled and targeted production in L. lactis adhesin more efficiently than SP Usp45 [27] . High SE might thus result, at least in part, from good adequacy between the mature protein and the SP used to direct secretion. The fusion of a short synthetic propeptide between the SP and the mature moiety is another innovative biotechnological tool to enhance protein secretion. One such propeptide (composed of nine amino acid residues, LEISSTCDA) was developed and was shown to enhance the SE of several heterologous proteins in L. lactis: NucB, NucT, (Table 3 ) [18] , the B. abortus L7/L12 antigen (Table 3 ) [19] , and the α-amylase of Geobacillus stearothermophilus (Table 3 ) [18] . Directed mutagenesis experiments demonstrated that the positive effect of LEISSTCDA on protein secretion was due to the insertion of negatively charged residues in the N-terminus of the mature moiety [25] . Furthermore, the enhancement effect does not depend on the nature of the SP, since the secretion of NucB fused to either SP Nuc or SP Usp45 was enhanced by LEISSTCDA insertion [25] . Strikingly, the enhancement of SE was reproducibly accompanied by an overall increase of protein yields as determined in Western blot experiments. This observation suggests that heterologous precursors are degraded by intracellular proteases when they are not efficiently secreted and that a higher secretion could be a way to escape proteolysis. Proteins with molecular mass ranging from 165 kDa (size of DsrD, the Leuconostoc mesenteroides dextransucrase, [28] ) to 9.8 kDa (size of Afp1, a Streptomyces tendae antifungal protein; Freitas et al., submitted) have been successfully secreted in L. lactis. This suggests that protein size is not a serious bottleneck for heterologous protein secretion in L. lactis. In contrast to protein size, conformation may be a major problem for heterologous secretion in L. lactis as illustrated by some recent examples. The first example is the production of the non-structural protein 4 (NSP4) of the bovine rotavirus, the major etiologic agent of severe diarrhea in young cattle. In order to develop live vaccines against this virus, the NSP4 antigen was successfully produced in L. lactis [29] . Derivatives of pCYT and pSEC plasmids were constructed to target NSP4 into cytoplasmic or extracellular location. The highest level of production was obtained with the secreted form. However, no secreted NSP4 was detected in the supernatant and both SP Usp45 -NSP4 precursor and NSP4 mature protein were detected in the cell fraction. Two degradation products were detected in addition to the NSP4 precursor and mature protein. These results suggest that the cytoplasmic form of NSP4 was probably totally degraded inside the cell whereas fusion to the SP Usp45 protected NSP4 protein against intracellular proteolysis. Similar results were obtained when pCYT and pSEC vectors were used to produce the B. abortus GroEL chaperone protein: only pSEC:GroEL plasmid was obtained and subsequently the fusion SP Usp45 :GroEL was detected in Western blot experiments (V. Azevedo, unpublished data). In this case, B. abortus GroEL is likely to interact with lactococcal cytoplasmic proteins leading to severe cellular defects and thus to a lethal phenotype. On the other hand, fusion of SP Usp to GroEL might keep the chimeric protein in an unfolded and/or inactive state allowing thus its heterologous production. Another example is the production of the bovine β-lactoglobulin (BLG) in L. lactis [30, 31] . BLG, a 162 amino acid residues globular protein, is the dominant allergen in cow's milk and was produced in L. lactis to test the immunomodulation of the allergenic response in mice when BLG is delivered by a bacterial vector [30] . Western blot and ELISA showed that BLG production was significantly higher when BLG was fused to SP Usp45 although the SE was very low, with no detectable BLG in the supernatant of pSEC:BLG strains [30] . Further studies revealed that a fusion between the LEISS propeptide and BLG could not enhance the SE of BLG above ~5%, as determined by ELISA [31] . For rotavirus NSP4, B. abortus GroEL, and BLG (which are medium-sized compared to DsrD or Afp1), either very low secretion yields or absence of secretion was observed in L. lactis. In all cases, fusion to a SP stabilizes heterologous protein production even though they are not efficiently secreted. These results could be due either to the SP itself that reportedly acts as an intramolecular chaperone or to the protection of the chimeric precursor from intracellular proteolysis by the cytoplasmic chaperones of the Sec-machinery. GroEL (a cytoplasmic chaperone), NSP4 (a structural protein), and BLG (a globular protein) have dramatically different primary sequences. A higher affinity of intracellular housekeeping proteases for these particular sequences cannot be hypothesized since the fusion of a SP leads to the stabilization of the protein. Change of conformation is therefore the predominant criterion involved in the stabilization of the precursors and the higher yields observed. On the other hand, these proteins might undergo rapid folding right after their synthesis, which interferes with (or hampers) the secretion process. Such interferences between protein conformation and SE were previously shown in E. coli and B. subtilis [32, 33] . Altogether, these results suggest that protein conformation rather than protein size is a major problem for heterologous protein secretion in L. lactis as well. It was clearly demonstrated that the secreted form of E7, a reportedly labile protein, can be stabilized by fusion to Nuc [20, 34] . Nuc is reportedly a stable protein and its use, as a fusion partner, does not affect its enzymatic activity. The production of the resulting chimerical protein is thus easy to follow. The cytoplasmic form of E7 was stabilized by the fusion to Nuc even when the production was induced in stationary phase ( Fig. 2A) , whereas cytoplasmic E7 alone was degraded (see below; Fig. 3 ). Thus, fusion to the stable Nuc could rescue E7 production in L. lactis and allowed higher protein yields compared to E7 alone [20] . Stabilization by fusion to Nuc was observed for several secreted proteins as well. First, a Nuc-E7 fusion on a pSEC backbone resulted in higher production yield although the SE was altered (Fig. 2B) . Fusion to the synthetic propeptide LEISSTCDA in a pSEC:LEISS:Nuc:E7 construction restored an efficient secretion yield [34] . Second, in an attempt to increase the protein yield of the secreted L7/L12, a fusion to Nuc (pSEC:Nuc:L7/L12) resulted in a 2.5-fold increase in production yield (Fig. 2B ) [19] . Recent results concerning the production of BLG provide a third example of yield enhancement by fusion to Nuc. A pSEC:Nuc:BLG construction allowed a 2-fold increase in BLG yields compared to pSEC:BLG [31] . These results show that Nuc is a stable carrier protein and has a protective effect on labile heterologous chimerical proteins by reducing its sensitivity to intracellular proteolysis. To our knowledge, Nuc is the fusion partner most commonly tested so far for stabilization in L. lactis. Bernasconi et al (2002) fused the Lactobacillus bulgaricus proteinase PrtB to BLG, which was subsequently stabilized by the PrtB carrier [13] . It is thus difficult to postulate any rule concerning the stabilization effect. Different results (i.e. no stabilization) could perhaps be observed with a different partner and thus could help to determine the mechanism of the stabilization effect. In biotechnological use of recombinant L. lactis strains for protein production, fusions can also facilitate purification (e.g. His-tag strategy). Protein fusion has also been successfully used to optimize the production of the two subunits of heterodimeric complexes as demonstrated with murine interleukin-12 in L. lactis [22] or with heterodimeric enzymes in E. coli [35] . In both cases, the resulting fusion had the expected properties. In other cases however, such fusions might dramatically interfere with the conformation of one or both of the proteins, which might be deleterious for the expected activity. Nevertheless, when L. lactis is used as an antigen delivery vector, fusions can be envisioned since it was demonstrated that both moieties of the chimerical protein are still recognized by the corresponding antiserum [10, 20, 34] and are immunogenic [36] . Several of the results mentioned above suggest that secretion could be an efficient way to escape intracellular proteolysis. This hypothesis was particularly tested in E7 production [20] . E7 was indeed degraded when intracellular production was induced in late exponential or early stationary growth phase (Fig. 3) . E7 production was then tested in a clpP deficient strain (ClpP is reportedly the major house keeping protease in L. lactis; [37] ) and in a dnaK deficient strain (DnaK is an intracellular chaperone that may promote proteolysis by maintaining the protein in an unfolded state; [38] ). In exponential or stationary phase cultures, no significant difference in E7 patterns was observed between wild type and clpP - (Fig. 3 ) or dnaK -(not shown) strains: E7 was equally degraded in the cytoplasm and remained unchanged in supernatants samples. Altogether, these results indicate that E7 intracellular proteolysis is ClpP-and DnaK-independent. Until recently, only two cytoplasmic proteases, ClpP and FtsH [39] , have been identified in L. lactis. The existence of a third, as yet unidentified protease was postulated by studies of a clpP mutant suppressor [40] . E7 may thus be a useful screening target to identify a putative L. lactis protease that, as suggested by our data, is activated in stationary phase. Besides the features of the precursor itself, these results also rise that host factors are involved in protein stability and SE (Fig. 4) . Research efforts are now focusing on the analysis of host factors involved in protein production and secretion by either directed or random mutagenesis in L. lactis [41] . Although L. lactis possesses a wide range of enzymes (peptidases, housekeeping proteases) dedicated to intracellular proteolysis, it possesses only one extracellular housekeeping protease (HtrA) [9] and its major extracel-lular scavenger protease, PrtP, is plasmid encoded [42] . Thus, a plasmidless strain does not present any protease activity in the medium. Better production yields could then be expected when secretion is used versus cytoplasmic production. These results give clues and provide the research workers with target proteins to study intracellular proteolysis and protein stability inside and outside the host strain. Such studies already led to the development of htrA deficient L. lactis strains. Heterologous protein secretion and anchoring in a htrA deficient strain allowed Fusion to Nuc rescue E7 in intracellular production and increase protein yields for the secreted forms of E7 and L7/L12 Native E7 production in wt L. lactis depends on growth phase Figure 3 Native E7 production in wt L. lactis depends on growth phase. E7 production and secretion were analyzed by Western blot from cultures induced at different times so that, 1 hour after nisin induction, the samples are harvested at exponential (OD 600 = 0.5-0.6, upper panels) or stationary phase (OD 600 = 1.5, lower panels). wt/pCYT-E7, NZ(pCYT-E7) strain (encoding native E7, cytoplasmic form). wt/pSEC-E7 NZ(pSEC-E7) strain (encoding the precursor preE7). Positions of E7 mature and precursor forms are given by arrows. C, cell lysates; S, supernatant fraction. ClpP is not involved in the intracellular degradation of E7 in L. lactis. Analysis by western blot shows that a strain of L. lactis deficient in the intracellular protease ClpP cannot rescue cytoplasmic E7 production. Induced cultures samples of wt L. lactis or L. lactis clpP mutant strain containing pCYT-E7 (clpP/pCYT-E7) or pSEC-E7 (clpP/pSEC-E7) taken at exponential-(upper panel) or stationary-(lower panel) phase. Stationary-phase higher protein stability at the cell surface for several heterologous proteins [10] . Current research works are now focusing on other host factors that affect protein production and secretion in L. lactis. L. lactis complete genome sequence analysis revealed indeed that the Sec machinery comprises fewer components than the well-characterized B. subtilis Sec machinery. Notably, L. lactis does not possess any SecDF equivalent and complementation of the lactococcal Sec machinery with B. subtilis SecDF results in better secretion yields as determined for Nuc reporter protein (Nouaille et al., submitted) . Random mutagenesis approaches also revealed that features of some cell compartment, such as the cell wall, play an important role in the secretion process [41] . Similar approaches allowed the identification and characterization of genes of unknown functions specifically involved in production yields of the secreted proteins in L. lactis (Nouaille et al., in preparation) . Many molecular tools are now available to direct heterologous protein secretion in L. lactis and the list of heterologous proteins produced in this bacterium is regularly increased. The reports where cytoplasmic and secretion production can be compared mostly show that secretion allows better protein yields compared to intracellular Schematic presentation of the molecular tools and the cellular events that can affect the production yields of heterologous pro-tein in L. lactis Figure 4 Schematic presentation of the molecular tools and the cellular events that can affect the production yields of heterologous protein in L. lactis. Thicknesses of the arrows are proportional to the final production yields. All the host factors involved in the cellular events are not identified and or characterized yet. SP, signal peptide (encoded in pSEC constructions), +Nuc, fusion between the protein of interest and the stable Nuc protein. production; and allow a better understanding of the protein production and secretion process in L. lactis. Future works should investigate the L. lactis capacities for protein modifications. For example, we showed that proteins that require a disulfide bond (DSB) to acquire their native conformation can be efficiently produced and secreted in L. lactis [5, 22, 27] . However, no equivalent of E. coli dsb or B. subtilis bdb, the genes involved in DSB formation, was found by sequence comparison in L. lactis. Similarly, other folding elements (i.e. PPIases, so-called maturases...) are still to be identified and the L. lactis capacities for post-translational modifications are still to be investigated. Altogether, these works will contribute to the development and the improvement of new food-grade systems for L. lactis [43] and should lead, in a near future, to the construction of lactococcal strains dedicated to high-level production of proteins of interest. The GRAS status of L. lactis and LAB in general, is a clear advantage for their use in production and secretion of therapeutic or vaccinal proteins.
21
Detection and characterization of horizontal transfers in prokaryotes using genomic signature
Horizontal DNA transfer is an important factor of evolution and participates in biological diversity. Unfortunately, the location and length of horizontal transfers (HTs) are known for very few species. The usage of short oligonucleotides in a sequence (the so-called genomic signature) has been shown to be species-specific even in DNA fragments as short as 1 kb. The genomic signature is therefore proposed as a tool to detect HTs. Since DNA transfers originate from species with a signature different from those of the recipient species, the analysis of local variations of signature along recipient genome may allow for detecting exogenous DNA. The strategy consists in (i) scanning the genome with a sliding window, and calculating the corresponding local signature (ii) evaluating its deviation from the signature of the whole genome and (iii) looking for similar signatures in a database of genomic signatures. A total of 22 prokaryote genomes are analyzed in this way. It has been observed that atypical regions make up ∼6% of each genome on the average. Most of the claimed HTs as well as new ones are detected. The origin of putative DNA transfers is looked for among ∼12 000 species. Donor species are proposed and sometimes strongly suggested, considering similarity of signatures. Among the species studied, Bacillus subtilis, Haemophilus Influenzae and Escherichia coli are investigated by many authors and give the opportunity to perform a thorough comparison of most of the bioinformatics methods used to detect HTs.
It is now widely admitted that actual genomes have a common ancestor (LUCA, Last Universal Common Ancestor). Their current diversity results from events that have modified genomes during evolution. While some of these events happen at the nucleotide level (point mutation, indel of few nucleotides), others [strand inversion, duplications, repetitions, transpositions and horizontal transfers (HTs)] may concern significant parts of the genome. It has been postulated that HTs (exchange of genetic material between two different species) were very frequent during the first stages of evolution and are essentially subsisting nowadays in prokaryotes (1) (2) (3) (4) . As a consequence, the detection of HTs appears crucial to the understanding of the evolutionary processes and to the qualitative and quantitative evaluation of exchange rate between species (5) (6) (7) (8) (9) . The recent complete sequencing of several genomes allows to systematically search for the presence of DNA transfers in species, especially in prokaryotes where the probability of occurrence is higher (10) (11) (12) (13) (14) . It has been reported in particular that (i) HTs in bacteria account for up to 25% of the genome (8, (14) (15) (16) ; (ii) archaebacteria and non-pathogenic bacteria are more prone to transfers than pathogenic bacteria (15, 16) ; and (iii) operational genes are more likely transferred than genes dealing with information management (15) (16) (17) . The HT concept has been originally coined to explain the dramatic homologies between genes of unrelated species (18, 19) . An 'unusual' match is subsequently the criteria for the detection of HTs (20, 21) . While this approach allows detection of gene transfers with only a partial knowledge of genomes, it requires the sequencing of homologous genes in a number of species and consequently cannot be used for HT screening. Genes from a given species are very similar to one another with respect to base composition, codon biases and short oligonucleotide composition (15, 16, (22) (23) (24) . As a general rule, usage of oligonucleotides varies less along genomes than among genomes (24) (25) (26) (27) . In addition, it has been observed that transferred DNA retains (at least for some time) characteristics from its species of origin (8, 14) . These particularities are used alone or in conjunction to detect DNA transfers between species (8, 12, 13) . Transferred DNA is consequently detected on the basis of some of its singularities with respect to the sequence characteristics of the recipient species. However, these techniques suffer several drawbacks and weaknesses (28) (29) (30) that led us to consider generalizing the above approach for the screening of atypical regions in sequences. In fact, the genomic signature that accounts for all possible biases in DNA sequences has been shown to be speciesspecific (26, 27, 31, 32) . The signature is approximately invariant along the genome in such a way that the species of origin of DNA segments as small as 1 kb could be identified with a surprisingly high efficiency by means of their signatures (25, 27) . As a consequence, the sequence signature may be most often (at least in bacteria) considered a valuable estimation of the genomic signature. Assuming that (i) transferred DNA fragments exhibit signature of the species they come from and (ii) recipient and donor signatures are different, the screening of local variations of signature along genomes is expected to reveal regions of interest where HTs might be located. In addition, the status of HT is strongly suggested if the signatures of these regions of interest are found close to the signature of other species. The sequence signature is defined as the frequencies of the whole set of short oligonucleotides observed in a sequence (26, 31) . It can be easily obtained thanks to a very fast algorithm derived from the Chaos Game Representation (CGR) (33) , which allows coping with a 1 Mb sequence in a few seconds on a laptop computer. Signatures may be visualized as square images where the color (or gray level) of each pixel represents the frequency of a given oligonucleotide (called word thereafter) (31) (for examples of signatures, see Supplementary Materials 2, 4 and 6). DNA sequences are gathered from GenBank. The genomes of 22 prokaryotes are scanned for HTs, B.subtilis, E.coli and H.influenzae genomes being given a special attention to illustrate our approach. In particular, B.subtilis and E.coli provide valuable benchmark thanks to the set of previous works addressing that very issue (12, 14, 16, (34) (35) (36) (37) . Signatures of about 12 000 species are obtained from genomic sequences longer than 1.5 kb. Sequences derived from the same species are concatenated for accuracy purposes. Species from the three domains of life, archaea ($260 species), bacteria ($3950 species) and eukarya ($6750 species) as well as viruses ($1300 species), are represented for a total amount of 1.0 Gb. The detection of atypical regions is based on the observation of deviation of local signatures (i.e. signature of small fragments of DNA) from the genomic signature of the recipient species. Genomes are consequently sampled by means of a sliding window with an appropriate size. In fact, it would be interesting to have windows the smallest as possible for highest sampling accuracy. However, intra-genomic variability of signature increases for small windows. In addition, variability depends on species and word length. Base composition (1-letter word), 2-and 3-letter words are poorly speciesspecific: they do not allow a good discrimination between species (25, 27) . As a general rule, the longer the words (up to 9-letter long), the higher the specificity of the signature (25, 27, 31) . However, counts of long words in small windows are too low to allow a reliable estimation of the parameters. In our hands, the analysis of 4-letter words in a sliding window of 5 kb (with a 0.5 kb step) offers a good trade-off between reliability of count, file size and computational charge, whatever the species. In addition, a double-strand signature (called local signature thereafter) is computed for each window to get rid of variations induced by strand asymmetry (38) (39) (40) (41) (42) . For illustration purposes, local signatures are developed as vertical vectors and stacked together in genome order to give an overall picture of word usage variations along each genome. In such plots, horizontal lines show the variation in frequency of words along the genome, whereas local changes in word usage appear as vertical breaks ( Figure 1 ). Figure 1 . Signatures (4-letter words and 5 kb windows) along genome for Clostridium acetobutylicum, Deinococcus radiodurans and Mycobacterium tuberculosis. In this kind of displays, lines represent the frequency of words along genome, columns represent signature of windows. Considering that the greatest part of the genome is speciestypical, the signature of the recipient species might have been estimated from the analysis of the whole sequence. Although the vast majority of local signatures look mostly the same (believed to be instances of the recipient species signature), some of them may greatly differ. In order to avoid potential biases linked to these outliers, it has been subsequently decided to select typical local signatures on the basis of their similarities, observed after clustering. The underlying idea is that typical local signatures aggregate in few large groups, whereas outliers are found in small complementary groups at a great distance from the recipient genome signature. Groups were consequently determined with the K-means clustering tool, using every scheme of clusters between 3 and 8 for each species. Finally, the best scheme of clusters was obtained by a decision tree-based partition [CART algorithm (43) ]. The purpose of the CART algorithm is to predict values of a categorical dependent variable (clusters of local signatures in this work, each signature being characterized by its distance to the estimated genomic signature) from one or more continuous and/or categorical predictor variables [the different clustering schemes (3-8 clusters) in this work]. The CART algorithm thus provides an optimal split between groups collecting signatures close to the estimated recipient genome signature and the others groups. For each species, a clustering scheme is selected (e.g. the 5-group clustering) and a partition offered (continued example: group 2 and 3 on one side; 1, 4 and 5 on the other). The recipient species signature is subsequently calculated as the mean of the signatures of the groups belonging to the partition with the smallest distance to the estimated genomic signature. Comparison of signatures is made possible, thanks to an Euclidian metric, accounting for differences in word usage. It must be pointed out that distances between signatures are calculated for high dimensional data (256 dimensions corresponding to the 256 different 4-letter words) and are consequently subjected to the so-called 'concentration of measure phenomenon' (44) . All distances in a high dimension space seem to be comparable since they increase with the square root of the dimension of the space, whereas the variance of their distribution remains unchanged. In fact, the radius of the hyper sphere holding 99% of the signatures of our database is only seven times the nearest neighbor distance (smallest distance between two species). Small differences in distance may consequently be considered highly significant. For each species, a set of recipient-specific distances is obtained, every local signature belonging to the large clusters being given a distance to the host signature. In order to select outlying signatures, a cut-off distance is chosen on the basis of the distribution of distances observed for each species. It appears that the 99% percentile offered a good trade-off between sensibility and specificity for outlier detection (for impact of the threshold on detection of atypical regions, see Results). Most signatures from minority clusters are detected in this way. Isolated signatures are detected as well, while very few signatures from the recipient species clusters are selected (1%). Outliers together with the flanking regions on the genome are later on reanalyzed with smaller window and step (1/10 th of the original size typically) in order to more accurately determine their limits, when signal-to-noise ratio allows it. Finally, the gene content of all detected regions is analyzed with the help of species dedicated databases [Genome Information Broker, http://gib.genes.nig.ac.jp/]. A BlastN search (GenBank, default settings) is carried out for each atypical region in order to identify the origin of potential HTs if homology is high enough. Search for the origin of atypical regions About 12 000 species (including chromosomal, plasmidic, mitochondrial and chloroplastic DNA) from GenBank are found eligible for a genomic signature. Given the signature of an atypical DNA fragment, species with a close signature might be considered as potential donors. Such a screening is performed for every atypical region of the 22 species under consideration. The first five nearby species are retained when their distance to the outlier was donor-compatible. A total of 22 genomes are screened for atypical regions (Table 1 and Supplementary Material 1). On the average, the 6-cluster scheme offers the best partition. However, in a single case (Aeropyrum pernix), nine clusters are required. In general, a single cluster is devoted to rRNA. The mean distance of windows to host varies over species from 121 to 145 (mean = 132, coefficient of variation = 3%). It is tightly correlated (P-value for the Pearson correlation coefficient <10 À4 ) with the cut-off distance that varies from 178 to 289 (mean = 234, coefficient of variation = 14%). Such large variations can hardly be explained on the mere basis of statistical fluctuations. As already observed (31, 45, 46) , variation of oligonucleotides usage along genome depends on species and can consequently be considered as a species property. Segmentation quality of atypical regions can be tested using rRNA genes. About 94% of rRNA is detected as atypical ( Table 1) . Borders of rRNA genes are accurate to within 130 nt (0.5 kb window and 50 bp step, threshold 99%). Meanwhile, adjacent tRNAs are identified as well. As a general rule, it can be concluded that rRNA has a specific signature that is consistently at variance with the host signature. In this context, it is worth noticing that rRNA and the remaining outliers lie at comparable distances from the species they belong to, but they are clearly different from one another, rRNAs being consistently found in their own cluster. The percentage of RNA-free outliers (at the nucleotide level) varies from 1.3 to 13% as a function of species (threshold 99%, Table 1 ). B.subtilis shows the highest percentage of atypical regions, whereas Pyrococcus abyssi has the lowest. Percentages among species are found correlated with the cut-off distance: the higher the cut-off distance, the lower the percentage of outliers (P = 0.007). In fact, a high cutoff distance takes place in species that display a high intragenomic variability, also expressed by a high mean distance to the host (Table 1) . Whether the actual percentage of atypical DNA is an intrinsic property of the species or a mere consequence of the resolution power of nucleotide biases-based methods remains consequently an open question. In addition, as already observed (13, 14) , the percentage of outliers is significantly higher for longer genomes (P = 0.004), whereas the cut-off distance is not related to the length of the genome (P = 0.69). The mean cut-off distance for the 22 species is 234 (Table 1) . This value is chosen to select credible donors. About 50% of atypical regions are subsequently given credible donors (Supplementary Material 1). Each species has it own set of (Table 1) . Many plasmids and viruses are also found in agreement with the known molecular mechanisms of horizontal transfer (Table 1 and Supplementary Material 1). A clustering with three classes allows assessing the signature of B.subtilis. The most populated class (collecting 84% of the segments) is chosen to represent B.subtilis. For this subpopulation, the mean distance (arbitrary unit) to the recipient (centroid of the class) and the cut-off distance are 126 and 204, respectively ( Table 1 ). Runs of contiguous outlying windows sharing the same cluster are considered as single transfer events. As a consequence, 58 regions (Figure 2a and Supplementary Material 2) fall beyond the cut-off distance and are thus potential candidates for hosting foreign DNA (for a segmentation of the B.Subtilis genome in terms of genes, see Supplementary Material 3). Figure 2b illustrates the accuracy of segmentation of an atypical region obtained by using a sliding window of 0.5 kb with a 50 bp step. rRNA genes make up $1.1% of B.subtilis genome ( Table 1) . All rRNA genes are found in the outlier population. In addition, all windows containing rRNA are assigned to a specific cluster. In fact, it is known that rRNA has its own signature, which is at variance from the host signature (12) . rRNA genes account for 7% of the outliers (tRNAs are not considered in this study, because their size is too small to generate a significant deviation from the host signature if they are isolated). A total of 86% of the B.subtilis genome should be considered as B.subtilis typical (Table 1) . When looking for the origin of B.subtilis segments in the 12 000 signature database, B.subtilis appears in the 10 first potential donors for 84% of the whole set of 5 kb sequences that can be derived from its genome. This result confirms that segments having signatures belonging to the predominant clusters are good representatives of the recipient species signature. The 49 rRNA-free atypical regions vary in size from 1.5 to 135 kb and make up 13% of the total genome (Table 1) . About 50% of atypical regions are less than (or around) 6 kb long. Distances of outlier from first potential donor often fall within the intra-genomic range ( However, in some instances, the outlier-to-donor distance is too great to consider the 'closest' species as potential donor. In contrast, unusual small values deserve a specific attention. In particular, the very small distance between bacteriophage SPBc2 and '2150751-2285750' atypical region (d = 2) allows to spot the part of B.subtilis genome where bacteriophage SPBc2 is incorporated (12, 47) . Other regions in the genome are also found similar (in terms of signature) to bacteriophage SPBc2. Most of them correspond to bacteriophages, imbedded in B.subtilis genome, whose free forms are not sequenced (12, 47) . Observed similarities with SPBC2 are, however, expected since signatures of phages usually share some characteristics with the species they infect (48) . The SPBc2 sequence is the only foreign sequence identified in B.subtilis, using homology as criterion (BlastN, with parameters set to default). In fact, Blast analysis of B.subtilis outliers leads to contrasted results. Besides SPBc2 and 7 out of 9 prophages imbedded in the genome, the only atypical regions identified are those containing the 30 rRNA genes coded in B.subtilis genome. The only few genes that are homologous to parts of atypical regions are found in species belonging to the Bacillus genus. It is interesting to note that no house-keeping genes (except rRNA) are detected in atypical regions. In fact, a great number of genes in atypical regions (except bacteriophage genes and rRNA) have no known function. A clustering with five classes is required to determine the recipient species signature of H.influenzae. The three most populated classes (collecting 94% of the segments) are chosen to calculate the H.influenzae signature. Mean distance to host and cut-off distance is subsequently found equal to 130 and 239, respectively (Table 1) . Similarly to B.subtilis, one cluster (1.5% of H.influenzae genome) is devoted to the 18 rRNA gene copies (Table 1) . A total of 91% of rRNA is labeled atypical and account for 29% of the outliers. Analysis of Table 1 shows that 95% of the H.influenzae genome should be considered as H.influenzae typical. In fact, H.influenzae is one of the 10 first potential donors for 92% of all 5 kb sequences that can be derived from its genome. As already observed for B.subtilis, the concordance of these two percentages corroborates the partition procedure used for the selection of typical/atypical fragments. The 13 rRNA-free atypical regions vary in size from 1.5 to 19.5 kb and make up 3.3% of the genome (Table 1 , Annex 4 and Figure 3 , see Annex 5 for a segmentation of the H.influenzae genome in terms of genes). About 50% of atypical regions are less than (or around) 2.5 kb long. Numbers for H.influenzae are clearly at variance with those for B.subtilis: a smaller percentage of the genome qualifies as atypical and the average size of atypical regions is also smaller. This result is examined below in the context of intra-species signature variability (see Discussion). A clustering with six classes is required to determine the recipient species signature of E.coli. The main features are summarized in Table 1 . The potential donors of the 84 RNAfree atypical regions are given in Annex 6 (for a segmentation of the E.coli genome in terms of genes, see Annex 7). It is worth noticing that 56% of E.coli potential donors belong to the Enterobacteriales family. Segmentation in terms of genes is displayed in Annex 7. The analysis of this genome is particularly useful for the comparison with literature (see below). Numerous approaches for detecting horizontal gene transfers have been proposed in the last 2 decades. Phylogenetic trees of protein or DNA sequences, unusual distribution of genes, nucleotide composition (including codon biases) are some of the HT features that are considered within the framework of these models (16, 34) , Hidden Markov Models (HMMs) (12, 14, 35) and Factorial Correspondence Analysis (FCA) (37) are some criteria that are currently employed. Each of the resulting models has its own advantages and caveats (28) (29) (30) . As it has been recently pointed out by Ragan (49) and Lawrence and Ochman (50) , each approach deals with a particular subset of HTs, being for example more efficient for detecting recent transfers, or more effective for the detection of ancient HTs. Our approach, which is clearly based on oligonucleotide composition, assumes that different species have different signatures but does not rely on any other assumption. It is not surprising, therefore, that the genomic signature approach provides results (in terms of % of DNA transferred) in reasonable agreement with those proposed by Garcia-Vallve (16) and Nakamura et al. (14) for the 22 species that were analyzed in common. Correlations between percentages of HTs found by these three methods are highly significant Two species are extensively studied for HT content: B.subtilis (five methods including ours) and E.coli (six methods including ours). H.influenzae is also analyzed by Garcia-Vallve (16) and Nakamura (14) . Comparisons of methods are presented in Tables 2-4 and detailed in Supplementary Materials 3, 5 and 7. A voting procedure (majority rule) has been implemented to determine the status of genes with respect to atypicality. For that task, our initial analysis is converted in terms of genes (Supplementary Materials 3, 5 and 7). Degree of agreement between methods is subsequently observed using the statistical Kappa coefficient (51) . Kappa measures the degree of agreement on a scale from minus infinity to 1. A Kappa of one indicates full agreement, a Kappa of zero indicates that there is no more agreement than expected by chance and negative values are observed if agreement is weaker than expected by chance (a very rare situation). (14, 13, 11, 13 and 15%, respectively). The number of detected genes per method is close, ranging from 457 for Nakamura (14) to 599 for this work (median 537). Detailed votes are given in Table 2 . Among the 4100 genes of B.subtilis genome, 1011 genes are detected by at least one method (about 25% of B.subtilis genes). The number of 'single vote' genes ranges from 116 for Garcia-Vallve (16) to 47 for Nicolas (12) . A total of 470 genes make up the majority consensus set and we detected 453 of them, which is the best score of the five methods. The best agreement with the majority consensus (in terms of Kappas) is reached by Nicolas (12), followed by our method and Moszer (36) ( Table 2 ). Our method gets the best agreement with Nicolas (12) and the worst with the other HMM method used by Nakamura (14) (pairwise Kappa comparison, Table 2 and Supplementary Material 3). In fact, Nakamura approach is at variance with every other approach (14) . It gets the lowest Kappa with the Garcia-Vallve (16) Hayes (35) Lawrence (34) Nakamura (14) Medigue (49) This work majority consensus or with whatever other methods. From Table 2 , the probable number of HT genes in B.subtilis would range from 230 to 1011 with a 'reasonable' estimation around 470 corresponding to the majority consensus. It is to be noted that our method is unable to find two genes that are detected by every other methods (Supplementary Material 3) . These genes are 338 and 236 nt long, respectively, as compared with 2500 nt, the median size of atypical regions detected by our method (Table 1) . Clearly, our method is not appropriate for detecting short isolated atypical genes. H.influenzae. Garcia-Vallve (16), Nakamura et al. (14) and we are the voters concerned with the analysis of the H.influenzae genome (Supplementary Material 5 and Table 3 , H.influenzae). The originality of results obtained by Nakamura (14) is the salient feature of this comparison. The number of detected HT genes is more than twice higher for Nakamura et al., whereas the part belonging to the majority consensus is the smallest ( Table 3) . Eleven genes are detected both by Garcia-Vallve and Nakamura (14, 16) but not by our method; however, the small number of voters precludes any specific comment in this respect. The probable number of HT genes in H.influenzae would range between 11 and 273, with a 'reasonable' estimation around 60 (majority consensus of 57) ( Table 4 ). The results obtained by Hayes and Borodovsky (35) are clearly at variance with the others (Table 4 ). Although the proportion of claimed outliers is within the range of published numbers for E.coli (14, 16, 24, 34, 35, 37) , 37% of them are method-specific, and the agreement with other methods is weak (Table 4 ). Hayes and Borodovsky have obviously developed an approach based on HMM dealing with specific outliers. Lawrence and Ochman (34) also get a poor rating especially because they detect about twice as many genes as the other authors do (Table 4) . It is worth noting that if the cut-off distance for our method is lowered, i.e. 95% instead of 99% for instance, some of the 'single vote' genes are dug out (for details about the impact of the cut-off distance, see Supplementary Material 7). Meanwhile, the percentage of outliers as reported by our approach rises to 20% and the percentage of 'single vote' genes reaches 24%. As expected, a high cut-off distance provides few single vote genes at the risk of missing some potentially transferred genes. Lowering the cut-off increases the proportion of single vote genes with the advantage of detecting most of the potential transfers (Supplementary Material 7) . There is obviously a continuous grading in gene 'atypicality'. It is suggested to first consider most 'consensual' genes as potential HTs and then apply amelioration models to explain the grading. It is difficult to assess the relevancy of proposed donors, because genes detected as potential HT have generally undergone amelioration (8) . The comparison of recently diverged genomes (species or strains) provides the opportunity to find recent HTs, for which corresponding homologous genes in the donor species may be detected (52) . Such a study is performed for five E.coli strains (two K12 strains: E.coli MG1655, E.coli W3110, one uropathogenic strain: E.coli CFT073, two enterohaemorrhagic strains: E.coli O157-H7 RIMD 0509952, E.coli O157-H7_EDL933) and two Shigella flexneri strains (S.flexneri 2a 2457T, S.flexneri 2a 301). These seven strains/ species have recently diverged, genome sizes are different and the proportion of horizontally transferred genes varies from one strain/species to another (14, 52) . For instance, only $40% of the non-redundant set of proteins is common to E.coli strains CFT073, 0157-H7 EDL 9333 and MG1655 (53) . These strains/species can be clustered in four groups with respect to phylogeny (Table 5) . Two criteria are used to searching for 'recent horizontally transferred genes': atypical regions (window size 1 kb, step 0.5 kb) (i) must have a signature that differs greatly from that of the host [distance to host must be at least >325, 2.5 times the E.coli intrinsic mean distance (Table 1) ] and (ii) must be present in a limited number of strains/species to ascertain their recentness. In fact, outliers meeting the first criterion generally aggregate into several heterogeneous clusters (K-means clustering) that usually include samples from each strain/species. In some instances, however, some strains/species were absent from the cluster. It was subsequently considered that the corresponding regions might have been recently acquired by the relevant strains/ species. Table 5 shows a selection of potential recently transferred genes. Each cluster of atypical regions contains genes present in a specific set of strains. Some atypical genes are strainspecific, some are only absent in the non-pathogenic K12 strains and intermediate situations are also encountered. FASTA and Blast searches confirm that these genes are absent from some of the tested strains as already observed in the analysis complete genomes (53) (54) (55) . In a large number of cases, we are able to find a well-conserved homologous gene in another species (Table 5) . It is interesting to note that some of the suggested donors using our 12 000 signature database are in agreement with the species found by alignment methods. When no homologous gene is found, the proposed donors give credit to the known mechanisms of gene transfer (bacteriophages or plasmids) ( Table 5) . It is worth noticing that most of the selected genes that are absent in K12 strains are involved in the pathogenicity of the other strains (52) . E.coli 0157-H7 is the strain exhibiting the greatest number of genes absent in K12 strains [about 1400 (54) ]. It has the greatest number of genes for which no homolog can be found (Table 5) . Moreover, we are unable to propose a donor for a great part of these genes (Table 5) . Many selected genes for E.coli 0157-H7 lie in the Ter region of the genome (between positions 2 000 000 and 2 500 000) in agreement with the published results (56). We have observed that most genomic regions are typical of the genome they belong to, using the signature as endpoint. Considering that the genomic signature is species-specific, atypicality of a region in terms of oligonucleotide usage has been promoted as a criterion for the detection of HTs. However, atypicality-based methods suffer several caveats that reduce their effectiveness in such a way that only a part of HTs can be detected. In fact, transfers between species with close signatures cannot be detected: significant differences between characteristics of transferred DNA and recipient species DNA are required. For similar reasons, HTs that were drastically ameliorated following their introduction cannot be detected either (8, 14) . The most stringent constraint, however, results from the size of the screening window. On the one hand, ideally, the best signal-to-noise ratio would be obtained when windows and HTs have a comparable size. On the other hand, the window size must be large enough to provide significant word counts, a requirement that strengthens with the size of the words under consideration and the intrinsic variability of the genomic signature along the genome. All together, the trade-off that has been implemented in this paper allows detecting atypical regions as small as 1 kb. In fact, rRNA regions sharing this characteristic were consistently detected. It must be pointed out that smaller fragments can be eventually detected if their signatures are radically atypical. G+C% atypicality has often been considered as criterion for detecting HTs (8, 24) , but this approach suffered several drawbacks (28) (29) (30) . It is to be noted that our signature-based method detects regions for which the G+C% lies within one standard deviation from the mean G+C% of the species (for instance, regions 2675251-2676250 in B.subtilis or 534751-535250 in H.Influenzae, see also Supplementary Materials 2 and 4). As already observed by Nicolas et al. (12) for B.subtilis, rRNA has definitely an atypical signature. It is systematically classified as outlier, whatever the species (Table 1) . Although transfer of rRNA from one species to another is unlikely (11, 57) , it cannot be firmly ruled out. However, it is clear that the atypical signature of rRNA does not imply that they are horizontally transferred. The signature approach has an interesting property (that it shares with HMM) (7, 12, 28) : detection is not bound to any specific function in the genome. In contrast with most other methods, the signature approach not only detects genes, but whole transferred regions as well, in agreement with the described mechanisms of DNA exchange between species. It is to be noticed that the method allows detecting several atypical non-coding regions (Supplementary Materials 3, 5 and 7). One major difference between HMM and signature method lies beyond the time required for the learning process, in the few resources that HMM can mobilize to deal with a short 'one of its kind' HT. On the other hand, HTs shorter than 1 kb can hardly be detected by a signature-based approach. An innovative HT detector is likely to result from an adequate fusion of both methods. Several factors contribute to the efficiency of the search for donors. Of course, distance between putative HT and donor signatures is essential. Accuracy of signatures, linked to the length of available sequences, density of signatures in the 'vicinity' of HT, amount of amelioration sustained by HT during its presence in the host are also of importance [P. Deschavanne, S. Lespinats and B. Fertil, unpublished results; (25, 27, 31) ]. Distance between the signature of a putative HT and the closest species varies to a large extent, but usually the shortest ones fall within the intra-genomic range ( Table 1 , Supplementary Materials 1, 2, 4 and 6) . In some cases, the distance between the closest donor signature and the atypical segment signature is so great that no potential donor can be proposed (Supplementary Materials 1, 2, 4 and 6) . When strong similarities between a given DNA sequence and a foreign species are observed, the hypothesis for an underlying transfer is highly strengthen. However, the 'true' donor has to be previously sequenced and included in our bank of signatures to allow such a situation to occur. Moreover, we must take into account the intrinsic variability of short DNA segment signature (which is a function of their size, but also species-specific) when compared with the signature of a complete genome or any other large species sample (25, 27, 31) . In the present state, our signature database is in no way representative of the diversity and richness of life. However, it must be noticed that there is already an obvious structure (in terms of distances between signatures) expressing taxonomy relationships between species in our signature database (31, (58) (59) (60) (61) . Related species are often found close to one another. Clusters of potential donors may consequently provide pertinent information about the origin of HTs. The diversity of signatures of putative HTs that can be observed for most of the species analyzed in this paper reveals the multiplicity of transfer events and donors (Supplementary Materials 2, 4 and 6). However, several outliers, not necessarily neighbors in the genome, are given the same set of potential donors (Table 1 , Supplementary Materials 1, 2, 4 and 6). In general, the potential donors belong to few sets of taxonomically close species (Table 1 ) and share the biotope of the host (Supplementary Materials 1, 2, 4 and 6). For instance, B.subtilis, H.Influenzae and E.coli live in distinct biotopes; their potential donors do so as well. It is particularly encouraging to find that most of the potential donors that our approach has pointed out have had the opportunity to exchange DNA material with the recipient species. Numerous viruses and plasmids qualify as potential donors (Tables 1 and 5 , Supplementary Materials 1, 2, 4 and 6 ). It is not really surprising since they are known as HT vectors. They are often totally or partially inserted together with transferred genes in the host genome (14) . Some atypical DNA segments are particularly peculiar. They are isolated, have a specific signature (distances from neighbors are great), so that they cannot be given a credible set of donors (Supplementary Materials 1, 2, 4 and 6) . Lack of data in the search domain, shift of signature features after a substantial amelioration process, structural constraints serving special functions or roles (14,62) (as it is for rRNA coding regions) are some of the tracks that remain to explore in these circumstances. It would be interesting to localize the region the transfer may come from when the complete genome of the donor is available. However, homology (at the DNA level) is not a pertinent criterion for the comparison of sequences as soon as amelioration has taken place (8, 14) . In fact, homology is sometimes weak, e.g. between genes of Escherichia and Salmonella although these species have 'recently' diverged (34) . It is clear that a more powerful search for the origin of putative HTs would have to embody models of amelioration [such as the one designed by Lawrence and Ochman (8) ]. When searching for very recent horizontally transferred genes, in different strains of a species for instance, it was possible to find a great homology between detected genes and some genes from other species (Table 5 ). In numerous cases, the selection of donors is consistent with FASTA results ( Table 5 ). This confirms the pertinence, beyond the similarity of signature between putative HTs and donors, of the proposed method to retrieve the species of origin of a transferred region. It seems that the search for origin of HTs on the basis of genomic signature is a powerful approach to understand some of the mechanisms of evolution (13, 63) . Oligonucleotide usage is known to be species-specific and to suffer only minor variations along the genome (25, 27) . Considered together, these properties allow searching for atypical local signatures that may point out DNA transfers. Results obtained with the 22 genomes analyzed in this paper are found in good agreement with literature (Tables 2-4 , Supplementary Materials 3, 5 and 7) (12, (14) (15) (16) 24, 34, 35) . The species specificity of signature allows searching for donor species. Quite often, sets of donor species with common taxonomic features are obtained. With the help of environmental considerations, it is subsequently possible to identify (or collect clues about) potential donors. The search for donor makes use of non-homologous sequences. Partially sequenced species become consequently eligible, inasmuch 1.5 kb of the genome is available (25, 27) . Thanks to the exponentially growing rate of nucleotide databanks, the search for donor species by means of the sequence signature will turn more and more pertinent and fruitful in the future. In this context, it is worth noticing that computational power is clearly not an issue since the CGR algorithm described in this paper is fast and of 0 order (calculation time is proportional to the number of nucleotides). Several methods are proposed to look for HTs. The signature method, based on different hypotheses, is complementary to those already described. It seems that each method detects preferentially certain types of HTs (49, 50) . In agreement with many authors (1, 16, 49, 50, 64) , it appears that the conjunction of several methods is required to obtain an overview of HT extent in a genome. The signature method described in this paper generalized many approaches that ground the detection of outliers on the basis of the bias in oligonucleodides. The strong species specificity of the signature not only allows detecting various kinds of outliers but also provides clues about their possible origin. Obviously, the detection of HTs remains an open question; a consensus has still to emerge. Additional materials and experimentation with the genomic signature are available from the GENSTYLE site (http:// genstyle.imed.jussieu.fr).
22
Comparisons of substitution, insertion and deletion probes for resequencing and mutational analysis using oligonucleotide microarrays
Although oligonucleotide probes complementary to single nucleotide substitutions are commonly used in microarray-based screens for genetic variation, little is known about the hybridization properties of probes complementary to small insertions and deletions. It is necessary to define the hybridization properties of these latter probes in order to improve the specificity and sensitivity of oligonucleotide microarray-based mutational analysis of disease-related genes. Here, we compare and contrast the hybridization properties of oligonucleotide microarrays consisting of 25mer probes complementary to all possible single nucleotide substitutions and insertions, and one and two base deletions in the 9168 bp coding region of the ATM (ataxia telangiectasia mutated) gene. Over 68 different dye-labeled single-stranded nucleic acid targets representing all ATM coding exons were applied to these microarrays. We assess hybridization specificity by comparing the relative hybridization signals from probes perfectly matched to ATM sequences to those containing mismatches. Probes complementary to two base substitutions displayed the highest average specificity followed by those complementary to single base substitutions, single base deletions and single base insertions. In all the cases, hybridization specificity was strongly influenced by sequence context and possible intra- and intermolecular probe and/or target structure. Furthermore, single nucleotide substitution probes displayed the most consistent hybridization specificity data followed by single base deletions, two base deletions and single nucleotide insertions. Overall, these studies provide valuable empirical data that can be used to more accurately model the hybridization properties of insertion and deletion probes and improve the design and interpretation of oligonucleotide microarray-based resequencing and mutational analysis.
Oligonucleotide microarrays are a powerful technological platform for large-scale screens of common genetic variation and disease-causing mutations (1) (2) (3) (4) (5) . In most published studies (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) (21) , oligonucleotide microarrays are designed to screen specific sequence tracts, up to megabases in length (11, 15, 22, 23) , for all possible single nucleotide substitutions. With some exceptions (24) (25) (26) (27) (28) (29) (30) (31) , the same emphasis was not placed on identifying all possible small insertions and deletions in the heterozygous state. Nevertheless, it is crucial to detect such small insertions and deletions since they can play a major role in inactivating or altering gene function by disrupting functional elements (e.g. splice junctions, cis-acting elements and open reading frames) and also represent another class of common genetic variation. Two fundamental approaches are commonly used to analyze data sets from oligonucleotide microarrays tailored to identify genetic variation in specific DNA segments purely by hybridization (1, (3) (4) (5) 9) . One approach involves identifying statistically significant gains of target hybridization signal to oligonucleotide probes complementary to specific sequence variants (9) . In theory, the gain of signal approach has the advantage of both detecting the presence of genetic variation and identifying the nature of the sequence change in the target. However, it is not feasible to screen for virtually all possible insertions and deletions due to the overwhelming The online version of this article has been published under an open access model. Users are entitled to use, reproduce, disseminate, or display the open access version of this article for non-commercial purposes provided that: the original authorship is properly and fully attributed; the Journal and Oxford University Press are attributed as the original place of publication with the correct citation details given; if an article is subsequently reproduced or disseminated not in its entirety but only in part or as a derivative work this must be clearly indicated. For commercial re-use, please contact journals.permissions@oupjournals.org number of mutation-specific probes needed for this analysis. Furthermore, little effort has been made to systematically access the hybridization properties of probes complementary to these small insertions and deletions. The second approach involves identifying losses of hybridization signal to perfect match (PM) probes that are fully complementary to the DNA segment of interest (8, 25, 27, 30, 31) . In theory, the loss of signal approach allows one to screen for all possible sequence changes, including insertions and deletions, that cause a given target nucleic acid sequence to contain mismatches with specific PM probes. However, this necessitates the sequencing of specific DNA regions to identify the nature of the sequence changes (8, 25, 27, 30, 31) . Thus, a combination of the gain and loss of hybridization signal analysis could provide the most robust means of identifying and characterizing mutations using non-enzymatic oligonucleotide microarray assays. Here, we analyze the specificity and reproducibility of nucleic acid hybridization to oligonucleotide microarrays used in the large-scale mutational analysis of the ATM (ataxia telangiectasia mutated) gene that is responsible for autosomal recessive disorder involving cerebellar degeneration, immunodeficiency, radiation sensitivity and cancer predisposition and is also commonly mutated in certain lymphoid malignancies (32, 33) . These microarrays include 25mer oligonucleotide probes complementary to all possible single base substitutions and insertions as well as one and two base deletions on both strands of the ATM coding region. This provides the first comparative analysis of the hybridization properties of substitution, insertion and deletion probes in an oligonucleotide microarray-based mutational analysis of a large gene. A series of 120 DNA samples derived from biopsies of lymphoma patients were previously screened for all possible ATM mutations using oligonucleotide microarrays (30) . Here, we have selected a total of 68 samples that showed robust amplification signals in all 62 coding exons for further analysis (30) . A total of 17 unique mutations, each in a one-to-one mixture with wild-type sequence, occurred once in these samples. The impact of any given mutation in a single sample is minimal given that 67 other samples with wild-type sequences in the region encompassing a given mutation are included in this analysis. Several single nucleotide polymorphisms (SNPs) were present multiple times: 735 C/T, 2572 T/C and 4258 C/T in two samples; 3161 C/G in four samples; and 5557 G/A in five samples. Likewise, these SNPs have a minimal effect on our global analyses given the large number of samples and bases interrogated in this study. As previously described (30) , individual ATM coding exons were amplified from genomic DNA using primers containing T3 and T7 RNA polymerase tails, pooled, and then in vitro transcribed using T3 or T7 RNA polymerase to create biotinlabeled sense and antisense strand targets, respectively. Fluorescein-labeled reference target was made using genomic DNA from an unaffected individual. Reference and test sample targets were fragmented, diluted in hybridization buffer [3 M TMA-Cl (tetramethylammonium chloride), 1· TE, pH 7.4, 0.001% Triton X-100] and hybridized to the ATM microarrays as described previously (30) . Afterwards, the microarray was stained with a phycoerythrinstreptavidin conjugate and digitized hybridization images from both reference and test targets were acquired using the Gene Array Scanner (Hewlett Packard, Palo Alto, CA) equipped with the appropriate emission filters. Custom software was used to quantify hybridization signals for each probe and subtract background hybridization signals. We exclusively focused on raw data from the biotin-labeled test targets since they provide approximately seven times the hybridization signal of the fluorescein-labeled wild-type reference target in this system (28) . This enhanced signal provides greater sensitivity toward detecting weak hybridization. For each sample, for each base and for each potential type of mutation (i.e. substitution, one or two base deletion or one base insertion), the specificity was calculated as the ratio of the PM probe hybridization signal of the wild-type target to their cognate insertion, deletion or single base substitution probes on each strand. The logarithm of these ratios was plotted as a function of the position within the gene. To illustrate the special patterns and to smooth out random variation, running averages of data from 10 bases were used. To capture the variability, at each base, the sample-to-sample standard deviation was again calculated using data derived from a running average of 10 bases for each sample. To estimate the mean hybridization specificity for each type of mutation, the geometric mean (i.e. the antilog of the average of the logged ratios) over all bases and over all specimens was calculated (Table 1) . To further examine the variability of the specificity ratios, the coefficient of variation (cv) was calculated in two ways. The cv is the ratio of the standard deviation divided by the mean; it is useful for understanding the amount of variability relative to the magnitude of the mean or typical value. For the intra-sample cv, the cv was calculated for each of the 68 samples (using the running average of 10 at each ATM base) and the average of the 68 coefficient of variations was taken. For the inter-sample cv, at each of the bases, the cv a Hybridization specificity ratio is defined as the ratio of PM probe hybridization signal to that of the brightest mismatch probe within a given category. The global average of all hybridization specificity ratios for each base in all samples for a given probe type is provided. b Determined for hybridization specificity ratios averaged across windows of 10 bases either within (intra) or across (inter) samples. was calculated using the 68 samples, and the average of the coefficient of variations was taken. For both calculations, the moving average of 10 was used, instead of the original value, since the goal was to understand how the specificity varied over bases and across samples, rather than to estimate the experimental (or measurement) error. In order to determine the relative specificity of the hybridization of complex nucleic acid targets to oligonucleotide probes complementary to single base substitutions, insertions and deletions, we analyzed data generated from oligonucleotide microarray-based mutational analysis of the 9168 bp ATM coding region (30) . These studies used a pair of oligonucleotide microarrays (Affymetrix, Santa Clara, CA) containing over 250 000 probes (25 nt in length) specifically designed to screen the sense and antisense strands of the ATM coding region for genetic variation (27, 30) . Collectively, the ATM sense and antisense microarrays contain 55 008 probes complementary to all possible single base substitutions, 73 344 probes complementary to all possible one base insertions, and 18 336 probes complementary to all possible one base deletions and 18 336 probes complementary to all possible two base deletions in the ATM coding sequence (Figures 1 and 2 ). These microarrays have been used to screen for sequence variation in the ATM gene in over 100 DNA samples (30) . SNPs and gene inactivating mutations were uncovered by screening for localized losses of hybridization signal to PM probes complementary to every 25 nt segment of the ATM coding region (8, 25, 27, 30) . However, hybridization data from deletion and insertion probes were not relied upon in this analysis. Therefore, this data set provides a unique opportunity to examine the relative hybridization specificity of nucleic acid targets to each of these classes of mismatch probes. In order to gain a global overview of hybridization specificity, we determined the average ratio of PM probe hybridization signal of wild-type target (see Materials and Methods) to their cognate insertion, deletion and single base substitution probes on each strand (Table 1 ). In these calculations, we considered data for all 9168 interrogated bases in all 68 DNA samples (see Materials and Methods). For example, we report the ratio of the PM probe signal to the signal from its cognate 1 or 2 bp deletion probe. However, for single base substitutions, we report the ratio of the PM probe signal to that of the cognate substitution probe with the highest hybridization signal. This provides the most rigorous assessment of cross-hybridization to single base substitution probes. Likewise, for single base insertion probes, we report the ratio of the PM probe signal to that of the cognate insertion probe with the highest hybridization signal. For both sense and antisense strands, we found that the two base deletion probes had the highest average PM to cognate MM hybridization specificity ratio (3.26-fold sense and (Table 1) . To provide a finer-scale analysis of hybridization specificity, we determined the relative frequencies of hybridization specificity ratios in defined bins. There was a similar distribution of specificity ratios for single base substitution and two base deletion probes on both strands ( Figure 3 ). The overall lower hybridization specificities of single base deletion and insertion probes are reflected by the increased frequencies of probes within the lower specificity bins (i.e. <2-fold ratio) and decreased frequencies of probes within higher specificity bins (i.e. >3-fold ratio) on both strands. Next, we sought to uncover underlying trends in the hybridization specificity of different classes of mismatch probes across the entire ATM coding region within a given sample (intra-sample variation). This provides insights into sequence context effects that may influence the hybridization specificity of each class of mismatch probe. To approach this problem, we plotted the average hybridization specificity ratios of substitution, deletion and insertion probes for all 1168 bases across the 68 samples (Figure 4 and Supplementary Figure 1) . We analyzed data determined over running averages of 10 bases in order to maximize our ability to detect trends and minimize the effect of randomly dispersed confounding factors (e.g. intra-or intermolecular secondary structure) that may skew data for any given base. As expected from Table 1 and Figure 3 , the two base deletion probes consistently showed a higher average hybridization specificity ratio followed by single base substitution, single base deletion and single base insertion probes on both strands of exon 50 ( Figure 4) . Nevertheless, the hybridization specificity ratios for all classes of mismatch probes fluctuate across the exon 50 sequence (Figure 4 ). For example, two base deletion probes showed a peak value of 6.76 (unlogged) centered at base 7071 and a trough value of 1.90 (unlogged) centered at base 7002 on the sense strand. We also found similar fluctuations in specificity ratios for all mismatch probe types in the remaining 61 ATM coding exons (Supplementary Figure 1 ). To assess intra-sample variability in hybridization specificity by a different means, we determined the average cv for substitution, deletion and insertion probes within a given experiment (Table 1) . Again, we analyzed data from running average of 10 bases in order to maximize our ability to detect trends and maintain consistency in our data analysis. Substitution probes had the lowest average intra-sample cv, 0.31 and 0.23 for sense and antisense strands, respectively. One base deletion, two base deletion and insertion probes showed comparable intra-sample coefficients of variation on the sense strand, 0.37, 0.39, and 0.38, respectively. However, insertion probes showed relatively higher variability than the deletion probes on the antisense strand. Coupled with plots shown in Supplementary Figure 1 , it is evident that of all the mismatch probe types, the hybridization specificities of base substitution probes were least affected by target sequence context. Intrigued by the above observations, we next searched for specific target sequence tracts that produced the lowest hybridization specificity among and between the different classes of mismatch probes. To approach this problem, we determined how many mismatch probes within running windows of 10 bases gave poor hybridization specificity, previously defined as a hybridization specificity ratio <1.2 (26). In Table 2 , we report nucleotide tracts where at least 8 probes within a given 10 base window showed poor hybridization specificity ratios. A comprehensive listing of probes with poor hybridization specificity is provided in Supplementary Table 1 . Repetitive sequence tracts, including homopolymer, homopurine and homopyrimidine, are highly represented in Table 2 . Upon closer inspection, it became apparent why the cross-hybridization is strong for probes in homopolymeric regions. In these sequence contexts, substitution and deletion probes can form duplexes with wild-type target that are longer than 12 bp in length. For example, the probe designed to detect a single base deletion at position 633 is designed to form one 12 bp and one 13 bp duplex with wild-type target. However, this probe can form duplexes that range from 12 to 18 bp in length with wild-type sense strand target due to slippage ( Figure 5 ). This type of ambiguity leads to increased stability of these DNA-RNA heteroduplexes (34) . In principle, the homopurine and homopyrimidine tracts uncovered have the capacity to form higher order structures, such as triple helices (35) . These tracts are known to alter the conformation and stabilities of RNA-DNA heteroduplexes (36, 37) , such as those formed between RNA targets and DNA probes in our system. Finally, we expect the ATM target to be especially rich in such sequence tracts given that both strands of the 3 0 -splice acceptor sequences, typically containing homopyrimidine tracts, for all 62 coding exons are included in the ATM target. This increases the likelihood that highly related sequence tracts in the ATM target can cross-hybridize to probes interrogating a particular homopurine or homopyrimidine sequence tract and reduce the overall hybridization specificity in this region. Next, we screened for potential structures that can form in the PM probes listed in Table 2 or their targets that could explain their poor hybridization specificity. To do this, we used Mfold (38) to calculate Gibbs free energies for intramolecular structures that can form in these PM probes and targets. Based on these Gibbs free energy values, we classified the probes and targets as having strong (S) [DG < (À3 kcal/mmol)], medium (M) [(À1 kcal/mmol) > DG > (À3 kcal/mmol)] and weak (W) [G > (À1 kcal/mmol)] potential for secondary structure. We found that several target and probe sequences could form substantial secondary structures, as displayed in Figure 6 . This could artificially lower the affinity of target to PM probes and thus lower the hybridization specificity. It is more difficult to model intermolecular structure in the solution-phase complex target and in the solidphase oligonucleotide probes. However, it appears likely that such structures could also have a similar negative impact on hybridization specificity. The relative variability in hybridization specificity ratios across samples (inter-sample variability) represents another important issue that should be considered in resequencing analysis (9) . To uncover general trends in inter-sample variability for each type of mismatch probe, we calculated an average cv for mismatch probe hybridization specificity ratios determined over running windows of 10 bases (Table 1) . Interestingly, on both strands, the single base substitution probes showed the lowest inter-sample cv. The one and two base deletion probes showed at least 2-fold higher coefficients of variation on both strands, relative to the substitution probes. Surprisingly, the one base insertion probes showed significantly higher coefficient of variations than any of the other classes of mismatch probes across samples. In fact, they are 3.5-fold higher than the corresponding substitution probes on each strand. The relative levels of inter-sample variation for all mismatch probes across exon 50 are displayed graphically in Figure 4 . The error bars represent one standard deviation from the mean of the hybridization specificity ratio determined over a running window of 10 bases in each of the 68 samples. Note that the substitution probes show lower inter-sample variability than one base deletion, two base deletion and . Hybridization specificities of mismatch probes. A 10-base running window of the log 10 hybridization specificity ratios of substitution (red), one base deletion (green), two base deletion (blue) and one base insertion (black) was plotted for the sense (A) and antisense (B) strands of ATM exon 50. The light red, light green, light blue and gray shaded areas represent -1 SD of the log 10 hybridization specificity ratios for the substitution, one base deletion, two base deletion and one base insertion probes, respectively. one base insertion probes, in agreement with Table 1 . The variability in hybridization specificity measurements is consistent across all 62 ATM coding exons (Supplementary Figure 1) . Overall, our analyses indicate that, on average, single base insertion probes show substantially lower reproducibility across experiments than base substitution, one base deletion and two base deletion probes. The increased inter-and intrasample variability in hybridization specificity of single base insertion and deletion probes relative to single base substitution and two base deletion probes should be considered when designing and interpreting microarray-based screens for genetic variation. For a given microarray design, substantially more control hybridization experiments may be needed to determine baseline fluctuations in the hybridization specificities of insertion and deletion probes relative to those of substitution probes. In contrast to single nucleotide mismatches, detailed thermodynamic analyses of double helical nucleic acids with bulged nucleotides have only recently been conducted (34, (39) (40) (41) . In such cases, the bulged nucleotide is unpaired on only one of the nucleic acid strands. These studies are relevant to understanding the properties of the deletion and insertion probes since they can form duplexes containing bulges with target nucleic acid. For deletion probes, the bulged nucleotide is located on the target strand ( Figure 7) . Conversely, the insertion probes contain the bulged nucleotide in duplexes with wild-type target (Figure 7) . Although subject to sequence context effects, duplexes containing a single base bulge are predicted to be more stable than those containing single nucleotide mismatches (34, (39) (40) (41) . This is reflected in the lower average hybridization specificity of single base deletion and insertion probes relative to that of substitution probes (Table 1 and Figure 4) . Conversely, duplexes containing two base bulges are predicted to be generally less stable than those containing a single base mismatch (40, 41) . In part, this is due to the assumption that helical stacking is interrupted by bulges of two or greater bases in length while it is preserved for one base bulges (40, 41) . The higher average hybridization specificity ratios of two base (38) was used to predict the intramolecular structures with the lowest Gibbs free energy (DG) for either the 25-30 base stretches that encompass each listed sequence tract in the target or for the PM probes complementary to each sequence tract. We use these DG values to predict the stability of these structures. DG > (À1 kcal/mmol) = weak (W); (À1 kcal/mmol) > DG > (À3 kcal/mmol) = medium (M); and DG < (À3 kcal/mmol) = strong (S). c Type of mismatch probe that provided poor hybridization specificity ratios. d Low hybridization specificity found on both sense and antisense strands. e Immediately following the 3 0 end of this segment is a (T) 5 sequence tract. deletion probes relative to substitution probes are in agreement with the predicted properties of these probes ( Table 1) . The considerably lower average inter-sample variability of substitution probes relative to deletion and insertion probes was unexpected given that the same target was hybridized to all mismatch probes simultaneously in the same experiment. The sources of inter-sample variation include sample preparation, hybridization conditions and the microarrays themselves. It is reasonable to assume that the microarrays themselves are not the major source of variability since the combinatorial manufacturing processes should lead to roughly equivalent synthesis quality for all the arrayed probes (42, 43) . It seems more likely that the insertion and deletion probes are more sensitive to subtle changes in target preparation (e.g. amount of fragmentation and dye incorporation) and hybridization conditions (e.g. target concentration, temperature and wash conditions) than the substitution probes. However, a definitive explanation for our observations will require further investigations (44) (45) (46) (47) (48) (49) (50) (51) (52) . In addition to their potential value, it is important to note some of the caveats when relying upon mismatch probes for mutation detection. For example, it is important to screen for all possible sequence changes, including multiple base insertions and deletions, in mutational analyses of disease-related loci, such as the ATM, BRCA1 and BRCA2 genes. Given that 4 N probes per base per strand are needed to screen for insertions of length N in a mixed sequence, it is unlikely that oligonucleotides complementary to insertions of two or more base pairs will be represented on microarrays screening large sequence tracts for mutations in the near future. Deletions represent a more tenable situation since only one probe per base per strand is needed to screen for a deletion of a given length in a mixed sequence. Nevertheless, there will still be limitations as to the number of deletion probes that can be realistically represented in a given microarray. Finally, it is often critical to precisely determine the nature of a sequence change within a given sample in order to properly assess its functional significance. Thus, it is important to consider error rates when assigning the identity of a mutation based on mismatch probe data. When dealing with clinical samples, it will be especially important to confirm the identity
23
A Gene Encoding Sialic-Acid-Specific 9-O-Acetylesterase Found in Human Adult Testis
Using differential display RT-PCR, we identified a gene of 2750 bp from human adult testis, named H-Lse, which encoded a putative protein of 523 amino acids and molecular weight of 58 kd with structural characteristics similar to that of mouse lysosome sialic-acid-specific 9-O-acetylesterase. Northern blot analysis showed a widespread distribution of H-Lse in various human tissues with high expression in the testis, prostate, and colon. In situ hybridization results showed that while H-Lse was not detected in embryonic testis, positive signals were found in spermatocytes but not spermatogonia in adult testis of human. The subcellular localization of H-Lse was visualized by green fluorescent protein (GFP) fused to the amino terminus of H-Lse, showing compartmentalization of H-Lse in large dense-core vesicles, presumably lysosomes, in the cytoplasm. The developmentally regulated and spermatogenic stage-specific expression of H-Lse suggests its possible involvement in the development of the testis and/or differentiation of germ cells.
Sialic acids are a diverse family of acidic nine-carbon sugars that are frequently found as terminal units of oligosaccharide chains on different glycoconjugates in higher invertebrates and vertebrates [1, 2] . As a part of determinants in many glycoproteins [3, 4] , sialic acids play an important role in intercellular and/or intermolecular recognition [5] . The 9-O-acetylation and de-Oacetylation are the most common modifications of sialic acids found in mammalian cell surface sialoglycoconjugates, which can alter its size, hydrophobicity, net charge, and antigenicity [2, 6, 7] . These modifications can regulate a variety of biological phenomena, including endogenous lectin recognition, tumor antigenicity, virus binding, and complement activation [8, 9] . Enzymes specifically capable of removing O-acetyl esters from the 9-position of sialic acids are sialic-acidspecific 9-O-acetylesterase. The enzymes in mammals have two forms, one is cytosolic sialic-acid-specific 9-O-acetylesterase (Cse) in the cytosolic fraction and another is lysosome sialic-acid-specific 9-O-acetylesterase (Lse) in the lysosomal/endosomal compartment [10] . Lse is likely to participate in the terminal lysosomal degradation of 9-O-acetylated sialoglycoconjugates, while Cse is likely to salvage any 9-O-acetylated molecules that escape the initial action of the Lse enzyme. The process of de-O-acetylation of sialic acid, which is catalyzed by sialicacid-specific 9-O-acetylesterase, has been implicated in organogenesis and cellular differentiation [2, 5] . Spermatogenesis is a complicated process of germ cell differentiation in adult testis, which is established during testicular development. There are five types of germ cells, each at a specific developmental stage, found in the seminiferous tubules: spermatogonia, primary spermatocytes, secondary spermatocytes, spermatids and sperms. They can be divided into three groups according to their DNA content: 4N DNA content cells (4C cells), 2N DNA content cells (2C cells), and 1N DNA content cells (1C cells). The separation of these cells enables researchers to investigate the molecular mechanisms underlying testicular development and/or spermatogenesis. In the present study, we separated the 2C and 4C cells of seminiferous tubules in human adult testis by flow cytometry, and identified human H-Lse by differential display RT-PCR. The expression pattern of H-Lse was found to be developmentally regulated and stage-specific, indicating its possible role in testicular development and/or germ cell differentiation. Human testes were obtained from Donation Center of Nanjing Medical University with consent of relatives. The seminiferous tubules were collected in DMEM/F12, which contained collagenase, and washed to remove the Leydig cells as well as interstitial cells. Trypsin treatment and a brief treatment with DNase I were used to release the spermatogenic cells from seminiferous tubules. The suspension of cells was filtered with nylon mesh. Disaggregated spermatogenic cells were suspended at 1 × 10 6 cells/mL in 0.5 M sodium citrate solution (PH 2.35) with fresh 0.1% DEPC overnight at room temperature and at 4 • C for two days; they were centrifuged and resuspended in 0.5 M sodium citrate solution (PH 4.5) with fresh 0.1% DEPC for at least 1 day. The day before use, the cells were centrifuged and resuspended in PBS with 10 mM HEPES (PH 7.0), 0.1% BSA, and fresh 0.1% DEPC. Then the cells were spun down and resuspended in PBS with 100 µg/mL PI (propidium iodide) and fresh 0.1% DEPC. The cells were stained overnight at 4 • C [11] . The flow cytometry (FCM) used in this research was FACSVantage SE (Becton Dickinson, Calif) equipped with argon laser (power: 200 mW, wavelength: 488 nm); a 585 nm/42 nm filter set was used before the FL2 detector. Cellquest (Becton Dickinson) was used for sorting and the sorting mode was Normal-R. Drops per sort were 3 and drop delay was 13.6. The density of cells for sorting was about 1 × 10 6 cells/mL. Isolation of total RNA from 2C cells and 4C cells was performed with Trizol Reagent (Gibco BRL, Ontario). One hundred nanograms of total RNA was used for differential display RT-PCR [12] . The first chain cDNA was synthesized by using T12G, T12C, and T12A oligo (dT) primers, and then was used as template in PCR. PCR was performed as follows: 94 • C, 1 minute; 37 • C, 1 minute; 72 • C, 2 minutes for 40 cycles. Ten microlitres of the PCR products from the two cells were run on a 1.5% agarose gel. The fragments highly or specifically displayed in 4C cells were excised and purified. This DNA was reamplified with the same combination of primers and then subcloned into Pinpoint Xa1-T vector (Promega, USA). The colonies of full-length cDNA were screened by PCR. Human Testis Large-Insert cDNA Library (Clontech, Calif) was first converted into plasmid cDNA Library, and then an arrayed cDNA library in 96-well plates was made according to the method of Munroe [13, 14] . In this arrayed cDNA library 1.54 × 10 6 colonies were screened by PCR. Multiple tissue northern (MTN) blots (Clontech) were hybridized with the 32 P-labeled probes. The probe corresponding to 1378-1634 bp of H-Lse was used for hybridization. After stringent wash, the blot was placed on the storage phosphor screen (Packard, USA) and exposed for 3 hours in the dark. The signal was detected at the Cyclone storage phosphor system (Packard). The Stanford TNG Radiation Hybrid Panel (Research Genetics, Huntsville, Ala) was used to map the chromosomal localization of HSE with primers HSEmapF (5 -ATGAACACCGTCTCCACC-3 ) and HSEmapR (5 -AAATCTGAAGGACCCATC-3 ), according to the manufacturer's instructions. After 35 cycles of amplification, the reaction products were separated on a 1.5% agarose gel. The positive amplification was labeled as 1 and the negative one was labeled as 0. The results were analyzed through the Stanford genome center web server to determine the probable chromosomal location. RNA DIG-labeled probes were made by in vitro transcription. T7 and SP6 promoter sequences were incorporated into the two sides of the templates (195-553 bp of H-Lse) by PCR, sense and antisense probes were made using DIG-RNA labeling mix (Roche, USA) according to the manufacturer's instructions. After fixation, paraffin embedding, mounting, and sectioning, sections of human embryonic and adult testes were prehybridized in hybridization buffer (DIG Easy Hyb, Roche, Germany) at 42 • C for 2 hours. Hybridization was carried in hybridization buffer containing appropriate probes at 65 • C for 16 hours in humidity chamber. Subcellular localization of HSEI and HSEII was performed by the method of green fluorescent protein. pEGFP-C2-HSEI AND pEGFP-C2-HSEII were constructed using two sets of primers (HSEI: 5 -GGGGAATT CAATGATATGGTGCTGCAG-3 and 5 -GGGGTCGACAT TTAGCAACATTGCTCTG-3 ; HSEII: 5 -GGGGAATTCA TGGTCGCGCCGGGGCTTG-3 and 5 -GGGGTCGACA TTTAGCAACATTGCTCTG-3 ) and EcoRI/SalI restriction sites of pEGFP-C2. Recombinant vectors were transfected into BxPC-3 cells (BxPC-3 cell is a cell lineage of adenocarcinoma from pancreas) by Lipofectin reagent (Gibco BRL). Cells were imaged 40 hours after transfection on the fluorescence microscope. After being stained with PI and measured by the FCM, three groups of cells in seminiferous tubules of human adult testis were detected (Figure 1 ), 2C and 4C cells were subsequently sorted. A clone was identified by differential display RT-PCR, which was highly expressed in the 4C cells ( Figure 2 ) and with high homology (86%) to a mouse lysosome sialic acid 9-O-acetylesterase. The clone was named H-Lse. In the two rounds of screening in the arrayed cDNA library, the plasmid containing full-length H-Lse (GenBank accession number: AF303378) was found. H-Lse is 2750 bp in length, encoding a putative protein of 523 amino acids with a molecular weight of 58 kd. Its isoelectric point is 7.19. The N terminus (1-18 aa) of the protein is a region containing hydrophobic amino acid residues, which may be a signal peptide. By comparison of the protein sequences (Figure 3 ), we hypothesized that H-Lse is the human counterpart of mouse lysosome sialic acid 9-O-acetylesterase. After PCR amplification, the results can be shown as a pattern (00000000100010100000011000001000000011 000001000001000001001000000010000000000100100100 0001). Retrieving results from the Stanford genome center web server shows that HSE is localized in the human 11q24 ( Figure 4) . The distribution of H-Lse in various human tissues was analyzed by Northern blot ( Figure 5 ) and the results showed the presence of three distinct mRNA species at approximately 2.7 kb, 6.0 kb, and 7.5 kb. The expected transcript of H-Lse was approximately 2.7 kb and it was consistently expressed in all the tissues examined with high expression found in the testis, prostate, and colon. The transcript of approximately 7.5 kb was exclusively expressed in the colon. The transcript of approximately 6.0 kb was distributed in the testis, colon, small intestine, prostate, and thymus, with the highest level of expression found in the testis. To examine a possible role of H-Lse in testicular development and/or spermatogenesis, in situ hybridization experiments were conducted to compare H-Lse expression in human embryonic and adult testes since spermatogenesis is not initiated in the embryo and there is no meiosis in embryonic seminiferous tubules. The results showed that no signal was detected in the embryonic testis, while positive signals were detected in spermatocytes but not spermatogonia in the seminiferous tubules of adult testis. Signals were associated with germ cells but not other somatic cells in the testis, that is, Sertoli and Leydig cells. Negative control of sense probes confirmed the specificity of the results ( Figure 6 ). The subcellular localization of H-Lse fusion proteins was visualized by transiently transfecting H-Lse gene fused with GFP into BXPC-3 cells. As shown in Figure 7 , the control cells transfected with GFP protein alone exhibited fluorescence evenly distributed throughout the cytoplasm, while GFP-H-Lse fusion protein was compartmentalized in numerous large dense-core vesicles in the cytoplasm. Spermatogenesis is a developmental program that occurs in mitotic, meiotic, and postmeiotic phases. In the mitotic phase, spermatogonia proliferate to expand the quantity of germ cells; in the meiotic phase, spermatocytes accomplish chromosomal synapsis and genetic recombination before two meiotic divisions; and in the postmeiotic phase, haploid spermatids are remodeled into spermatozoa by the processes of acrosome formation, nuclear condensation, flagellar development, and loss of the majority of cytoplasm. Under the control of intrinsic and extrinsic factors, spermatogenesis is characterized by the expression of a spectrum of genes that are celltype-specific or stage-specific. They are thought to play an essential role in spermatogenesis at particular stages. For example, MutS homologue 5 is required for chromosome pairing, CPEB and SCP3 are required for synaptonemal complex assembly and chromosome synapsis in primary spermatocytes [15, 16, 17] . In the present study, we have identified a gene, H-Lse, from human adult testis with high homology to m-Cse 1 [19] . Similarly, it can inhibit binding of sialoadhesin, a macrophage-restricted and sialic-acid-dependent adhesion molecule [20] . On the other hand, 9-O-acetylation of sialic acids can form novel epitopes. Influenza virus C haemagglutinin specifically requires 9-O-acetylated sialic acids for binding to host cells [21] . Incubation of red blood cells with sialate 9-Oacetylesterase rendered the erythrocytes resistant against agglutination by influenza C virus [22] . O-acetylation of disialoganglioside GD3 by human melanoma cells has been reported to create a unique antigenic determinant [23] . Modifications of sialic acids may be an important mechanism underlying the interaction/cross-talk between different types of cells. The essential role of sialic acids modification in cellular communications may explain the presently observed wide distribution of H-Lse in all examined tissues. The present study suggests that the expression of H-Lse is developmentally regulated and spermatogenic stage-specific. The evidence for this includes: (1) lack of expression in embryonic testis; (2) association of high level of mRNA detected by DD-RT-PCR with the 4C but not 2C cells in adult testes; and (3) detection of in situ hybridization signal in spermatocytes but not spermatogonia or other somatic cells. In the absence of spermatogenesis, embryonic testis contains only two distinct cell types, spermatogonia and Sertoli cells, while the seminiferous epithelium of adult testis consists of germ cells at different stages of spermatogensis. The 4C cells found in adult testis include the primary spermatocytes and spermatogonia of G 2 /M stage, while 2C cells include spermatogonia of G0/G1 stage, secondary spermatocytes, and Sertoli cells. The absence of H-Lse mRNA in embryonic testis and the high level of its mRNA in the 4C cells of adult testis suggest that its expression is restricted to spermatocytes, particularly the primary spermatocytes. Together with the in situ hybridization results showing mRNA of H-Lse restricted to spermatocytes, but not spermatogonia, Sertoli cells or interstitial cells, these data suggest that H-Lse is likely to be involved in the process of spermatogenesis, although its role in testicular development cannot be entirely ruled out. Unfortunately, due to the deformation of the available human testes, we were not able to make further distinction between primary and secondary spermatocytes. What has been clearly shown by the present data is that H-Lse is only present at a stage beyond spermatogonia, suggesting its possible role in the differentiation of germ cells. G N F T Y M S A V C W L F G R Y L Y D T L Q Y P I G L V S S S W G G T Y I E V W S S R R T L K A C G V P N T 143 m-Lse 181 A G N L G H G N F T Y M S A V C W L F G R Y L Y D T L Q Y P I G L V S S S W G G T Y I E V W S S R R T L K A C G V P N T 240 h-Lse 181 S E N L G H G Y F K Y M S A V C W L F G R H L Y D T L Q Y P I G L I A S S W G G T P I E A W S S G R S L K A C G V P K Q 240 m-Cse 144 R D E R V G Q P E I K P M R N E C N S E E S S C P F R V V P S V R V T G P T R H S V L W N A M I H P L Q N M T L K G V V 203 m-Lse 241 R D E R V G Q P E I K P M R N E C N S E E S S C P F R V V P S V R V T G P T R H S V L W N A M I H P L Q N M T L K G V V 300 h-Lse 241 G S _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ I P Y D S V T G P S K H S V L W N A M I H P L O N M T L K G V V 274 m-Cse 204 W Y Q G E S N A D Y N R D L Y T C M F P E L I E D W R Q T F H Y G S Q G Q T D R F F P F G F V Q L S S Y M L K N S S D Y 263 m-Lse 301 W Y Q G E S N A D Y N R D L Y T C M F P E L I E D W R Q T F H Y G S Q G Q T D R F F P F G F V Q L S S Y M L K N S S D Y 360 h-Lse 275 W Y Q G E S N I N Y N T D Interestingly, the processes of 9-O-acetylation and de-O-acetylation of sialic acid have been implicated in organogenesis and cellular differentiation, since alteration of these processes could lead to interruption of cellular development such as embryogenesis. Transgenic mice constitutively overexpressed the 9-O-acetyl-sialicacid-specific esterase of influenza C that has been found to arrest embryo development at the two-cells stage. It has also been reported that in vitro development of embryonic stem cells shows that the expression level of Lse is low at the initiation of the development, and followed by an increase at later stages [24] . In transgenic mice with selective expression of 9-O-acetyl-sialic-acid-specific esterase in retina and the adrenal gland, these organs showed various abnormalities in organization, while all other tissues appeared normal [25] . Lse has also been considered to play a key role in the differentiation of B lymphocyte [2] , since it is expressed in late but not early B lymphocyte. The presently observed developmentdependent pattern of H-Lse expression is consistent with that found in other cell types: absence or low expression at early stage of differentiation but high at later stages. Taken together, 9-O-acetyl esters in sialic acids appear to be important for development or cellular differentiation. Spermatogenesis is a multiple-staged continuous progress of cellular differentiation. It has been reported that some cell surface glycoconjugates are modified during the early steps of spermatogenesis, and influence the differentiation of spermatogenic cells [26] . As ninecarbon sugars commonly found in many glycoproteins of spermatogenic cells, sialic acids represent a target for cell surface modification, that is, removal of 9-O-acetyl esters by enzymes such as Lse. Modification of sialic acids may result in alteration in cell-cell communication, that is, Sertoli cells and germ cells interaction, thereby influencing the differentiation of spermatogenic cells. Thus, future studies on the presently identified H-Lse may provide insight into molecular mechanisms underlying testicular development and/or germ cell differentiation during spermatogenesis in humans.
24
The role of mast cells in the pathogenesis of pain in chronic pancreatitis
BACKGROUND: The biological basis of pain in chronic pancreatitis is poorly understood. Mast cells have been implicated in the pathogenesis of pain in other conditions. We hypothesized that mast cells play a role in the pain of chronic pancreatitis. We examined the association of pain with mast cells in autopsy specimens of patients with painful chronic pancreatitis. We explored our hypothesis further using an experimental model of trinitrobenzene sulfonic acid (TNBS) -induced chronic pancreatitis in both wild type (WT) and mast cell deficient mice (MCDM). METHODS: Archival tissues with histological diagnoses of chronic pancreatitis were identified and clinical records reviewed for presence or absence of reported pain in humans. Mast cells were counted. The presence of pain was assessed using von Frey Filaments (VFF) to measure abdominal withdrawal responses in both WT and MCDM mice with and without chronic pancreatitis. RESULTS: Humans with painful chronic pancreatitis demonstrated a 3.5-fold increase in pancreatic mast cells as compared with those with painless chronic pancreatitis. WT mice with chronic pancreatitis were significantly more sensitive as assessed by VFF pain testing of the abdomen when compared with MCDM. CONCLUSION: Humans with painful chronic pancreatitis have an increased number of pancreatic mast cells as compared with those with painless chronic pancreatitis. MCDM are less sensitive to mechanical stimulation of the abdomen after induction of chronic pancreatitis as compared with WT. Mast cells may play an important role in the pathogenesis of pain in chronic pancreatitis.
Although pain is the presenting symptom of most patients with chronic pancreatitis, its neurobiological basis remains poorly understood [1] . In the past, investigators have focused on the role of anatomical abnormalities such as a strictured pancreatic duct or narrowed intraparenchymal ducts. However, mechanical decompression techniques such as endoscopic stent placement or even surgical pancreatojejunostomy frequently do not provide a permanent solution to the pain [1] . More recently, investigators have begun focusing on the role of neurotransmitters and neurotrophins such as substance P and nerve growth factor with known or suspected roles in nociceptive signaling and/or sensitization and have reported an increased expression of several of them in the pancreas of patients with painful chronic pancreatitis [2] . Mast cells are also increased in both acute and chronic pancreatitis [3, 4] but their role in the generation of pain in pancreatitis has not been investigated. We hypothesized that mast cells are involved in the pathogenesis of pain in chronic pancreatitis. This hypothesis is based on the following observations. First, mast cells have been associated with human conditions in which pain is a predominant symptom. Interstitial cystitis and irritable bowel syndrome are both conditions in which pain is out of proportion to the objective pathological findings [5, 6] . In both conditions, an increase in the number of mast cells has been described in the bladder and the colon, respectively [5, 6] . Further, mast cells are frequently found in close proximity to nerves in the intestinal mucosa and the bladder [7] [8] [9] . This has also been observed in the pancreas -the total number of mast cells was significantly higher in pancreatic tissue from patients with chronic pancreatitis than in the normal pancreatic controls [3] . One of the preferential locations of mast cells was around and within the perineurium of nerve fibers in tissue samples of patients with chronic pancreatitis, suggesting the potential for interactions between mast cells and the nervous system. Lastly, there is evidence for bi-directional functional communication between mast cells and nerves [10] [11] [12] . Mast cells can not only release mediators that increase excitability of neurons but in turn, neurotransmitters such as substance P can trigger mast cell degranulation [10] . Mast cells may therefore contribute to the pathogenesis of pain in pancreatitis through degranulation products that can sensitize pancreatic afferent neurons in an ongoing vicious circle of neuronally mediated mast cell degranulation. Our first aim was to analyze the presence and distribution of mast cells in autopsy specimens of chronic pancreatitis and study the correlation, if any, with historical documentation of pain. We then explored our hypothesis further using an experimental model of trinitrobenzene sulfonic acid (TNBS)-induced chronic pancreatitis in both wild type and Kit W /Kit W-v mice, a strain deficient in mast cells (MCDM). Autopsy records from the University of Texas Medical Branch from the years 1993 to 2000 were searched electronically for the term "pancreatitis." One-hundred sixtysix patients were identified of which 26 patients carried an autopsy diagnosis of chronic pancreatitis and 140 patients carried a diagnosis of acute pancreatitis. The medical charts from patients with an autopsy diagnosis of chronic pancreatitis were reviewed for documentation of a medical history of chronic pancreatitis. If no such documentation was present in the chart, patients were excluded from the study (12/26) . Thus, 14/26 patients with both a documented history and an autopsy based diagnosis of chronic pancreatitis, were included in the study. Patients were categorized as painful chronic pancreatitis (8/26) when they fulfilled one of the following criteria: a documented history of chronic abdominal pain clinically attributed to chronic pancreatitis that required the use of narcotics, and/or frequent admissions for recurrent abdominal pain clinically attributed to chronic pancreatitis, and/or a surgical or endoscopic procedure for refractory abdominal pain clinically attributed to chronic pancreatitis. Patients were categorized as non-painful chronic pancreatitis (6/ 26) if patients did not fit any of the criteria listed under painful chronic pancreatitis. In addition, the following data were collected: demographic factors (age and race), cause of death, comorbidities, clinical history of pancreatitis, etiology of pancreatitis, diagnostic studies supporting a diagnosis of pancreatitis (amylase, lipase, calcifications on abdominal plain film, CT-scan, ultrasound or ERCP). Human pancreatic control tissue was obtained from 8 arbitrarily chosen patients of whom the autopsy records recorded acute myocardial infarction as the cause of death. Their medical records were reviewed to ensure that they did not have a clinical history of pancreatitis. Therefore there were three categories of patients: one with painful chronic pancreatitis, one with non-painful chronic pancreatitis and non-pancreatitis controls. A pathologist, blinded to the group assignment, verified all histological diagnoses and counted mast cells on a Giemsa stained tissue section (average of 10 high-power randomly chosen (40X) fields per specimen). The protocol was approved by the Institutional Review Board of the University of Texas Medical Branch. All mice were purchased from the Jackson Laboratory (Bar Harbor, ME). Male mice were used from the following strain: WBB6F1/j-Kit W /Kit W-v (MCDM) and the respective littermate control mouse strain, Kit W-v -+/+ (WT). The mice were 3 months of age at the onset of the experiment with body weights of 25-30 gram. Experimental protocols involving mice were approved by our Institutional Animal Care and Use Committee (IACUC) in accordance with the guidelines provided by the National Institutes of Health. Mice were anesthetized with sodium Nembutal (50 mg/kg body weigh, i.p.) Following a midabdominal laparotomy, a canula was introduced into the common pancreato-biliary duct; the duct was ligated proximally and distally to ensure perfusion into the pancreas and prevent entry of the injected substance into the liver or duodenum. 0.1 ml of 1% TNBS in phosphate buffered saline (PBS)-10% ethanol, pH 8, was infused into the pancreas (modified after Puig-Divi [13] ). The canula was removed and the abdomen closed. Control mice were treated in the exact same fashion but were perfused with saline instead (Figure 1 ). Mice were sacrificed 8 weeks after surgery. VFF hairs consist of a series of filaments of increasing diameter that produce increasing sensations of touch when applied to the skin. When the tip of a fiber of given length and diameter is pressed against the skin, the force of application increases until the fiber bends. After the fiber bends, continued advance creates more bend, but not more force of application. This principle makes it possible to apply a reproducible force to the skin surface. VFF testing is an established behavioral pain assay used to determine mechanical pain thresholds in somatic pain. More recently, VFF testing has been used as a surrogate marker for visceral pain [14, 15] . Mice were placed in a cage with a mesh floor and habituated to the environment for 30-60 minutes. Measurements were taken from the abdomen and the plantar surface of both hindpaws over a period of three weeks prior to the surgery and for a total of three weeks after the surgery ( Figure 1 ). VFF filaments of various caliber were applied to the mid-abdomen in ascending order 10 times, each for 1-2 seconds with a 10 second interval. A response was defined as: a) sharp retraction of the abdomen; b) immediate licking or scratching of site of application of hair; or c) jumping. The response frequency was defined as the total number of responses out of 10 applications (expressed as a percentage) to the skin per filament. An investigator blinded to the different treatment groups performed the behavioral testing. Fresh specimens of the mouse pancreas were fixed in 10% formaldehyde in PBS pH 7.4 containing 1 mM MgCl 2 at 4°C overnight. Sections from paraffin-embedded specimens were stained with hematoxyline and eosin and observed under a light microscope. Pathological changes were scored based on a scale described by Tito et al. by a pathologist blinded to the different treatment groups [16] . Comparisons of the number of mast cells in autopsy specimen were analyzed using the Mann-Whitney U test. For each behavioral experiment (see figure 1), the average response frequency was calculated as the mean of the mean response frequencies for each mouse (across four measures). The "post-pre response frequency" was calculated by subtracting the pre-surgical average response frequency from the post-surgical average response frequency. To assess the independent effect of TNBS on VFF response (ie. to control for the effect of the surgery itself), the postpre response frequency for TNBS infusion was compared with the post-pre response frequency for saline infusion. This comparison was performed using analysis of variance for a two-factor experiment with repeated measures on time at each level of force for each type of mice (WT and MCDM). The two factors were induction of pancreatitis or not (TNBS or saline, respectively) and time (pre-surgical or post-surgical). TNBS infusion was considered to have had an independent effect on the VFF response if the postpre response frequency was greater for TNBS than for saline infusion. Fisher's least significant difference procedure was used for multiple comparisons of least squares means with Experimental design Figure 1 Experimental design All mice underwent pre and post surgical VFF testing. For the VFF testing, 4 measures were taken for each mouse. WT and MCDM were randomized to either saline or TNBS perfusion into the pancreatic duct. Patient demographics are summarized in Table 1 . Alcohol abuse was the most common cause for pancreatitis in both groups. Analysis of our results, using the Mann-Whitney U test, revealed significantly more mast cells in patients with a history of painful chronic pancreatitis (n = 8) when compared to patients with either non-painful chronic pancreatitis (n = 6) (33.8 vs 9.4 average mast cell count/10 high power fields; p < 0.01) or controls (n = 8), (33.8 vs 6.1 average mast cell count/10 high power fields; p < 0.01) ( Figure 2 ). The increased number of mast cells in patients with painful pancreatitis was noted predominantly in interstitial areas and, to a lesser degree, in the periacinar space. Figure 3 shows the post-pre surgical response frequency for both WT and MCDM. TNBS had a significant independent effect on abdominal VFF response in WT mice at the force levels 4 and 8 mN (p = 0.007 and 0.037, respectively) ( Figure 3A ). There was a trend towards a significant effect at the force level of 16 mN (p = 0.066). In contrast, for MCDM, TNBS had no significant effect on abdominal VFF response at any force level ( Figure 3B ). There was no significant TNBS effect on VFF response in the left A g e Pancreatic histology confirmed the presence of chronic pancreatitis in both WT and MCDM with marked fibrosis, inflammatory infiltrates and ductular proliferation mimicking changes seen in human chronic pancreatitis ( Figure 5A ). The pancreas of saline treated controls was normal. There was no significant difference in the overall inflammatory scores between the WT and MCDM ( Figure 5B ). An increased number of mast cells were counted in WT mice with chronic pancreatitis compared to saline Histology (Giemsa) of mice with chronic pancreatitis (Figure 6 ). As to be expected, no mast cells were present in pancreas of MCDM. Chronic pancreatitis has been defined as a continuing inflammatory disease of the pancreas characterized by irreversible morphologic changes that typically cause pain and/or permanent loss of function [17] . The pathogenesis of pain in this condition remains to be satisfactorily established. We examined the association, if any, of pain with mast cells as quantified in autopsy specimens of patients with a history of painful and non-painful chronic pancreatitis and normal controls. Significantly more mast cells were present in pancreatic tissue from patients with a history of painful chronic pancreatitis, indicating an association with this condition and a potential role for these cells in the pathogenesis of pain in painful chronic pancreatitis. There are clearly limitations to a retrospective, autopsybased study such as the one we report here. For instance, we do not know whether pain was present at the time of death and there was incomplete information on the different patterns of pain. Also, our findings pertain mainly to patients with a history of alcoholic pancreatitis. Nevertheless, our findings do suggest an association of painful chronic pancreatitis with an increased number of mast cells. This observation provided the rationale for further experimental testing, which we performed in mice. We first developed a model of chronic pancreatitis in mice following a modified protocol first described by Puig-Divi et al. [13] . Histological changes consisted of periductal and lobular fibrosis, duct stenosis, chronic inflammatory cell infiltrates, and gland atrophy, mimicking features of chronic pancreatitis in humans. Significantly more mast cells were present in WT mice with chronic pancreatitis, adding to the validity of this model for use in studies on the role of mast cells in pancreatitis. Both WT and MCDM developed histological changes consistent with chronic pancreatitis, indicating that the elimination of mast cells did not modulate the animals' ability to mount an inflammatory response. Therefore, any changes observed in pain behavior are unlikely to stem from differences in underlying inflammation. Next we determined whether this mouse model could be used to study behavioral differences associated with chronic pancreatitis. The assessment of spontaneous pain in a visceral organ presents significant difficulties. We have used a behavioral method to assess this, which relies on the association of visceral pain with sensitization of somatic regions of the body that share segmental innervation at the level of the spinal cord (referred pain). This somatic sensitization can be quantified using VFF to stim-ulate the somatotopically appropriate abdominal region and measuring the abdominal withdrawal response. Thus, VFF testing of the anterior abdominal wall can be used as a surrogate marker for visceral pain. Although this is the first time that this technique has been used for the measurement of referred visceral hyperalgesia in a mouse model of chronic pancreatitis, this method has previously been described and validated to assess the severity of referred visceral pain for models of colonic hypersensitivity [14] as well as rat models of acute necrotizing pancreatitis [15] and chronic pancreatitis [18] . The abdominal VFF response was compared to the hind paw response to assess the specificity of the interventions to the pancreas. TNBS treated mice, but not the saline control, developed increased abdominal wall withdrawal responses to VFF testing when compared to baseline, suggesting the development of force-dependent referred hyperalgesia of the abdominal wall in WT mice. There was no evidence of referred hyperalgesia in the hindpaws, suggesting that the measured effect on abdominal withdrawal is specific for an intra-abdominal origin of the pain. Vera-Portocarrero et al. previously described similar findings, increased withdrawal frequency after VFF stimulation to the abdominal area, in a rat model of chronic pancreatitis [18] . These behavioral changes were abrogated by morphine. Rats that demonstrated behavioral changes also expressed increased substance P expression in the nociceptive layers of the spinal cord, suggestive of central nociceptive changes. Mast cells produce a variety of degranulation products in the setting of inflammation that may activate and/or sensitize primary nociceptive neurons. The neurotrophin growth factor (NGF) is one such product [19] [20] [21] [22] . NGF is released in the setting of inflammation and can not only function as a chemoattractant for other mast cells, but it can also trigger mast cell degranulation [23] . We are speculating that NGF production in the inflamed pancreas is responsible for plastic changes in the sensory neurons by activating proalgesic receptors and channels such as the NGF receptor tyrosine kinase A (TrkA) and Transient Receptor Potential Family V receptor 1 (TRPV1; previously known as VR1) thereby contributing to the generation of pain [24] [25] [26] . Similarly, other mast cell degranulation products such as tryptase and histamine are capable of modulating neuronal function [27] [28] [29] [30] [31] [32] . Tryptase may directly activate the proteinase-activated receptor-2 (PAR-2), a G-protein coupled receptor expressed by pancreatic nerves, important in the pathogenesis of pain in pancreatitis [33, 34] . Although the role for mast cells in the mediation of visceral nociceptive signaling needs to be explored further, we speculate that mast cell products released in pancreatitis, contribute to the development of pain by direct effects on nociceptors located on pancreatic afferent neurons (Figure 7 ). However, before concluding a definite role for mast cells from our experimental data, it should be noted that MCDM carry a spontaneous mutation for tyrosine kinase receptor c-kit which not only produces a deficiency of mast cells but may have an independent effect on the function of sensory neurons, which are known to express it [35] . Therefore, it remains to be determined whether the detected differences in nociceptive responses is due to the absence of mast cells per se or a yet unknown change in the responsiveness of sensory neurons due to a congenital lack of the c-kit receptor. Reconstitution of mast cells into the MCDM mice should restore their nociceptive responses close to the wild type phenotype. Our data should increase awareness of the importance of mast cells in the pathogenesis of painful inflammatory Proposed involvement of mast cells in nociceptive signaling in pancreatitis Figure 7 Proposed involvement of mast cells in nociceptive signaling in pancreatitis In pancreatitis, mast cells may migrate to sites of inflammation, in response to release of mast cell chemoattractants. Mast cell degranulation products may modulate neurotransmission directly by activating proalgesic receptors and channels such as trka (NGF), TRPV1 (NGF) and PAR-2 (tryptase and trypsin). The pre-publication history for this paper can be accessed here: http://www.biomedcentral.com/1471-230X/5/8/prepub
25
Recombination Every Day: Abundant Recombination in a Virus during a Single Multi-Cellular Host Infection
Viral recombination can dramatically impact evolution and epidemiology. In viruses, the recombination rate depends on the frequency of genetic exchange between different viral genomes within an infected host cell and on the frequency at which such co-infections occur. While the recombination rate has been recently evaluated in experimentally co-infected cell cultures for several viruses, direct quantification at the most biologically significant level, that of a host infection, is still lacking. This study fills this gap using the cauliflower mosaic virus as a model. We distributed four neutral markers along the viral genome, and co-inoculated host plants with marker-containing and wild-type viruses. The frequency of recombinant genomes was evaluated 21 d post-inoculation. On average, over 50% of viral genomes recovered after a single host infection were recombinants, clearly indicating that recombination is very frequent in this virus. Estimates of the recombination rate show that all regions of the genome are equally affected by this process. Assuming that ten viral replication cycles occurred during our experiment—based on data on the timing of coat protein detection—the per base and replication cycle recombination rate was on the order of 2 × 10(−5) to 4 × 10(−5). This first determination of a virus recombination rate during a single multi-cellular host infection indicates that recombination is very frequent in the everyday life of this virus.
As increasing numbers of full-length viral sequences become available, recombinant or mosaic viruses are being recognized more frequently [1, 2, 3] . Recombination events have been demonstrated to be associated with viruses expanding their host range [4, 5, 6, 7] or increasing their virulence [8, 9] , thus accompanying, or perhaps even being at the origin of, major changes during virus adaptation. It remains unclear, however, whether recombination events represent a highly frequent and significant phenomenon in the everyday life of these viruses. Viruses can exchange genetic material when at least two different viral genomes co-infect the same host cell. Progeny can then become hybrid through different mechanisms, such as reassortment of segments when the parental genomes are fragmented [10] , intra-molecular recombination when polymerases switch templates (in RNA viruses) [11] , or homologous or non-homologous recombination (in both RNA and DNA viruses). Quantification of viral recombination in multicellular organisms has been attempted under two distinct experimental approaches: in vitro (in cell cultures) [12, 13, 14, 15] , and in vivo (in live hosts) [16, 17, 18] . The in vitro approach, which has so far been applied only to animal viruses, allows the establishment of the ''intrinsic'' recombination rate in experimentally co-infected cells in cell cultures [14, 15, 19] . However, it does not necessarily reflect the situation in entire, living hosts, where the frequency of coinfected cells is poorly known and depends on many factors such as the size of the pathogen population, the relative frequency and distribution of the different variants, and host defense mechanisms preventing secondary infection of cells. The in vivo experimental approach is closer to biological conditions and may thus be more informative of what actually happens in ''the real world.'' However, as discussed below, numerous experimental constraints have so far precluded an actual quantification of the baseline rate of recombination. First, many experimental designs have used extreme positive selection, where only recombinant genomes were viable (e.g., [13, 20, 21] ). Other studies did not use complementation techniques but detected recombinants by PCR within infected hosts or tissues [18, 22, 23, 24, 25] , which provides information on their presence but not on their frequency in the viral population. So far, no quantitative PCR or other quantitative method has been applied to evaluate the number of recombinants appearing in an experimentally infected live host. Finally, recent methods based on sequence analysis inferred the population recombination rate, rather than the individual recombination rate [1, 26, 27] . While results from these methods certainly take in vivo recombination into account, there are other caveats: isolates have often been collected in different hosts-sometimes in different geographical regions-and sometimes the selective neutrality of sequence variation on which these estimates are based is not clearly established. Estimates from such studies by essence address the estimation of the recombination rate at a different evolutionary scale. Taken together, the currently available information indicates that no viral recombination rate has ever been estimated directly at time and space scales corresponding to a single multi-cellular host infection, although this level is most significant for the biology and evolution of viruses. This study intends to fill this gap by evaluating the recombination frequency of the cauliflower mosaic virus (CaMV) during a single passage in one of its host plants (the turnip Brassica rapa). CaMV is a pararetrovirus, which is a major grouping containing hepadnaviruses (e.g., hepatitis B virus), badnaviruses (e.g., banana streak virus), and caulimoviruses (e.g., CaMV). Pararetroviruses are characterized by a non-segmented double-stranded DNA genome. After entering the host cell nucleus, the viral DNA accumulates as a minichromosome [28] whose transcription is ensured by the host RNA polymerase II [29] . The CaMV genome consists in approximately 8,000 bp and encodes six viral gene products that have been detected in planta ( Figure 1 ) [30] . Viral proteins P1 to P6 are expressed from two major transcripts, namely a 19S RNA, encoding P6, and a 35S RNA corresponding to the entire genome and serving as mRNA for proteins P1-P5 [31] . Using the pre-genomic 35S RNA as a matrix, the protein P5 (product of gene V) reverse-transcribes the genome into genomic DNA that is concomitantly encapsidated [30] . The detection of CaMV recombinants in turnip hosts has been reported numerous times. Some studies have demonstrated the appearance of infectious recombinant viral genomes after inoculation (i) of a host plant with two infectious or non-infectious parental clones [21, 32, 33, 34, 35] or (ii) of a transgenic plant containing one CaMV transgene with a CaMV genome missing the corresponding genomic region [36] . While the former revealed inter-genomic viral recombination, the latter demonstrated that CaMV can also recombine with transgenes within the host's genome. Another study based on phylogenetic analyses of various CaMV strains has clearly suggested different origins for different genomic regions and, hence, multiple recombination events during the evolution of this virus [37] . Indirect experimental evidence has indicated that, in some cases, CaMV recombination could occur within the host nucleus, between different viral minichromosomes, presumably through the action of the DNA repair cellular machinery [21, 35] . Nevertheless, the mechanism of ''template switching'' during reverse transcription, predominant in all retroviruses, most certainly also applies to pararetroviruses. For this reason, and on the basis of numerous experimental data, CaMV is generally believed to recombine mostly in the cytoplasm of the host cell, by ''legal'' template switching between two pre-genomic RNA molecules [21, 35, 36, 38, 39] , or ''illegal'' template switching between the 19S and the 35S RNA [36, 40] . Under this hypothesis, recombination in CaMV could therefore be considered as operating on a linear template during reverse transcription, with the 59 and 39 extremities later ligated to circularize the genomic DNA (position 0 in Figure 1 ). The above cited studies clearly demonstrate that CaMV is able to recombine. However, since these studies are based on complementation techniques, non-quantitative detection, or phylogenetically based inferences of recombination, they do not inform us on whether recombination is an exceptional event or an ''everyday'' process shaping the genetic composition of CaMV populations. In the present work, we aimed at answering this question. To this end, we have constructed a CaMV genome with four genetic markers, demonstrated to be neutral in competition experiments. By co-inoculating host plants with equal amounts of wild-type and marker-containing CaMV particles, we have generated mixed populations in which impressive proportions of recombinants-distributed in several different classes corresponding to exchange of different genomic regions-have been detected and quantified. Altogether, the recombinant genomes averaged over 50% of the population. Further analysis of these data, assuming a number of viral replications during the infection period ranging from five to 20, indicates that the per nucleotide per replication cycle [44] ) indicates the origin of replication via reverse transcription, which occurs in the direction indicated by the dotted outermost circle-like arrow. Reverse transcription is accomplished by the viral reverse transcriptase, using the 35S RNA as template [49] . DOI: 10.1371/journal.pbio.0030089.g001 recombination rate of CaMV is of the same order of magnitude, i.e., on the order of a few 10 À5 , across the entire genome. We thereby provide the first quantification, to our knowledge, of the recombination rate in a virus population during a single passage in a single host. From Figure 1 , and supposing that all marker-containing genomic regions can recombine, we could predict the detection and quantification of seven classes of recombinant genotypes: þbcd/aþþþ, aþcd/þbþþ, abþd/þþcþ, abcþ/þþþd, þþcd/abþþ, aþþd/þbcþ, and aþcþ/þbþd. Indeed, all classes were detected, and their frequencies in the ten CaMV populations analyzed are summarized in Table 1 . Altogether, the proportion of recombinant genomes found in the mixed viral populations was astonishingly high and very similar in the ten co-infected plants analyzed (Table 1 , last column), ranging between 44% (plant 5) to 60% (plants 7, 12, and 20), with a mean frequency (6 standard error) of 53.8% 6 2.0%. This result indicates that recombination events are very frequent during the invasion of the host plant by CaMV and represents, to our knowledge, the first direct quantification of viral recombination during the infection of a live multi-cellular host. The inferred per generation recombination and interference rates, assuming that CaMV undergoes ten replication cycles during the 21 d between infection and sampling, are given for each of the ten plants in Table 2 . Recombination rates between adjacent markers are large, on the order of 0.05 to 0.1. Taking the distance in nucleotides between markers into account yields an average recombination rate per nucleotide and generation on the order of 4 3 10 À5 . Interestingly, this recombination rate does not vary throughout the genome (Kruskal-Wallis test, p = 0.16). To relax the assumption of the number of replications during the 21 d, we calculated the recombination parameters assuming five or 20 generations. The effect of the number of generations on the estimates is linear: doubling the number of generations results in a halving of the recombination rate (detailed results not shown). For example, the average recombination rates r 1 , r 2 , and r 3 assuming 20 generations were equal to 0.05, 0.04, and 0.025, respectively (compare with values in Table 2 ), yielding per nucleotide per generation recombination rates of 1.9 3 10 À5 , 2.2 3 10 À5 and 1.6 3 10 À5 . Inspection of Table 2 also shows that first-order interference coefficients were in general negative, indicating that a crossing over in one genomic segment increases the probability that a crossing over will occur in another genomic segment, while the second-order coefficient parameter had an average value close to zero with a large variance. The mechanism leading to these results will be discussed in the following section. One major breakthrough in the work presented here lies in the space and time scales at which the experiments were performed. Indeed, the processes occurring within the course of a single infection of one multi-cellular host are of obvious biological relevance for any disease. Previous studies on viral recombination suffered from major drawbacks in this respect, basing their conclusions on experiments relying on complementation among non-infectious viruses or between viruses with undetermined relative fitness, on phylogenetically based analyses, or on experiments in cell cultures. For reasons detailed in the Introduction, the first two methods either do not provide information on the frequency of recombination, but only its occurrence, or address the question at a different temporal, and often spatial, scale. Results from cell cultures, on the other hand, impose cell coinfection by different viral variants, potentially overestimat- ing the frequency of recombination events. Our study circumvents these limitations by analyzing viral genotypes sampled from infected plants after the course of a single infection, and therefore the invasion and co-infection of cells in various organs and tissues is very close to natural. More than half of the genomes (53.8% 6 2.0%; see Table 1 ) present in a CaMV population after a single passage in its host plant were identified as recombinants, and these data allowed us to infer a per nucleotide per generation recombination rate on the order of 2 3 10 À5 to 4 3 10 À5 . The time length of one generation, i.e., the time required for a given genome to go from one replication to the next, is totally unknown in plant viruses. The only experimental data available on CaMV are based on the kinetics of gene expression in infected protoplasts, where the capsid protein is produced between 48 and 72 h [40] . The reverse transcription and the encapsidation of genomic DNA being two coupled phenomena [30] , we judged it reasonable to assume a generation time of 2 d and, thus, an average of ten generations during our experiments. In case this estimate is mistaken, we have verified a linear relationship between r and the number of generations, thereby allowing an immediate adjustment of r if the CaMV generation time is more precisely established. At this point, we must consider that all cloned genomes may not have been through all the successive replication events potentially allowed by the timing of our experiments. It was previously shown that about 95% of CaMV mature virus particles accumulate in compact inclusion bodies [41] , where they may be sequestered for a long time, as such inclusions are very frequent in all infected cells, including those in leaves that have been invaded by the virus population for several weeks. The viral population may thus present an age structure that could bias the estimation of the recombination rate. In order to minimize this bias, the clones we analyzed were collected in one young newly formed leaf, where the chances of finding genomes from ''unsequestered lines'' were assumed to be higher. In any case, our data analysis is conservative, since this age structure can only lead to an underestimation of the recombination rate. Our results show that interferences between pairs of loci are negative: a recombination event between two loci apparently increases the probability of recombination between another pair of loci. We believe that the most parsimonious explanation of these negative interferences is based on the way the infection builds up within plant hosts. Indeed, one can divide infected host cells into those infected by a single virus genotype and those infected by more than one viral genotype. In the former, analogous to clonal propagation, recombination is undetectable. In the latter, recombination is not only detectable but, as our results indicate, very frequent. Samples consisting of viruses resulting from a mixture of these two types of host cell infections will thus contain viruses with no recombination and viruses with several recombination events, thus yielding an impression of negative interference. These conceptual arguments are supported by mathematical models. It is indeed easy to show (detailed results not shown) that if a proportion F of the population reproduces clonally, analogous to single infections, while the remaining reproduces panmictically, negative interferences could be inferred even if they do not exist. For example, assuming a three-locus model with real recombination rates r 1 and r 2 and interference i 12 , the ''apparent'' recombination and interference parameters, would be r 1 = (1 À F)r 1 , r 2 = (1 À F)r 2 , and i 12 = À(F À i 12 )/(1 À F). Interestingly, this example also shows that our estimates of the recombination rate are conservative: that a fraction F of host cells are singly infected while others are multiply infected leads to an underestimation of the recombination rate. As judged by r 1 , r 2 , and r 3 , calculated between markers a-b, b-c, and c-d, respectively, we found evidence for recombination through the entire CaMV genome. The values for r 1, r 2 , and r 3 are remarkably similar, hence the recombination sites seem to be evenly distributed along the genome. We considered the template-switching model as the major way recombinants are created in CaMV. As already mentioned in the Introduction, hot spots of template switching have been predicted at the position of the 59 extremities of the 35S and 19S RNAs [21, 36, 42] . If other recombination mechanisms, such as that associated with second-strand DNA synthesis or with the host cell DNA repair machinery, act significantly, hot The various parameters are as follows: r1, recombination rate between markers a and b; r2, recombination rate between markers b and c; r3, recombination rate between markers c and d; i12, interference between crossovers in segments a-b and b-c; i23, interference between crossovers in segments b-c and c-d; i13, interference between crossovers in segments a-b and c-d; i123, second-order interference accounting for residual interference. The recombination rates are the maximum likelihood estimates (6 95% confidence intervals). The interference parameters were obtained numerically as explained in the Materials and Methods. DOI: 10.1371/journal.pbio.0030089.t002 spots would be expected at the positions of the sequence interruption D1, D2, and D3 [43] . Due to the design of our experiment and the position of the four markers, we have no information on putative hot spots at positions corresponding to the 59 end of the 35S RNA and to D1 (at nucleotide position 0). Nevertheless, the putative hot spots at the 59 end of the 19S RNA and at D2 and D3 (nucleotide positions 4,220 and 1,635, respectively) fall between marker pairs c-d, b-c, and a-b, respectively. Our results indicate that either these hot spots are quantitatively equivalent-though predicted by different recombination mechanisms-or, more likely, that they simply do not exist. Whatever the explanation, what we observe is that the CaMV can exchange any portion of its genome, and thus any gene thereof, with an astonishingly high frequency during the course of a single host infection. To our knowledge, the viral recombination rate has never previously been quantified experimentally for a plant virus [3] . In contrast, retroviruses and particularly HIV-1 have been extensively investigated in that sense. As we have already discussed for these latter cases, the quantification of the intrinsic recombination rate was carried out in artificially coinfected cell cultures. The estimated intrinsic per nucleotide per generation recombination rate in HIV-1 is on the order of 10 À4 [14, 15, 19] , less than one order of magnitude higher than our estimation for CaMV. Because for various reasons detailed above we probably underestimate the within-host CaMV recombination rate, we believe that the intrinsic recombination rate in CaMV is higher and perhaps on the order of that of HIV. Other pararetroviruses such as plant badnaviruses or vertebrate hepadnaviruses have a similar cycle within their host cells, including steps of nuclear minichromosome, genomic size RNA synthesis, and reverse transcription and encapsidation. Nevertheless, vertebrate hepadnaviruses (e.g., hepatitis B virus) infect hosts that are very different from plants in their biology and physiology, and this could lead to a totally different frequency of cell co-infection during the development of the virus populations. Thus, even though our results can be informative for other pararetroviruses because of the viruses' shared biological characteristics, they should not be extrapolated to vertebrate pararetroviruses without caution. Viral isolates. We used the plasmid pCa37, which is the complete genome of the CaMV isolate Cabb-S, cloned into the pBR322 plasmid at the unique SalI restriction site [44] . To analyze recombination in different regions of the genome, we introduced four genetic markers: a, b, c, and d, at the positions 881, 3,539, 5,365, and 6,943, respectively, thus approximately at four cardinal points of the CaMV circular double-stranded DNA of 8,024 bp ( Figure 1 ). All markers, each corresponding to a single nucleotide change, were introduced by PCR-directed mutagenesis in pCa37, and resulted in the duplication of previously unique restriction sites BsiWI, PstI, MluI, and SacI in a plasmid designated pMark-S. Because, in this study, we targeted the possible exchange of genes between viral genomes, all markers a, b, c, and d were introduced within coding regions corresponding to open reading frames I, IV, V, and VI, respectively. Another important concern was to quantify recombination in the absence of selection, i.e., to create neutral markers. Consequently all markers consist of synonymous mutations (see below). Production of viral particles and co-inoculation. To generate the parental virus particles, plasmids pCa37 and pMark-S were mechanically inoculated into individual plants as previously described [33] . All plants were turnips (B. rapa cv, ''Just Right'') grown under glasshouse conditions at 23 8C with a 16/8 (light/dark) photoperiod. Thirty days post-inoculation, all symptomatic leaves were harvested and viral particles were purified as described earlier [45] . The resulting preparations of parental viruses, designated Cabb-S and Mark-S, were quantified by spectrometry using the formula described by Hull et al. [46] . We fixed the initial frequency of markers to a value of 0.5, and a solution containing 0.1 mg/ml of virus particles of both Cabb-S and Mark-S at a 1:1 ratio was prepared. Plantlets were co-infected by mechanical inoculation of two to three leaves with 20 ll of this virus solution, using abrasive Celite AFA (Fluka, Ronkonkoma, New York, United States). The mixed CaMV population was allowed to grow during 21 d of systemic infection. Estimation of marker frequency within mixed virus populations. We designed an experimental protocol for quantifying marker frequency within a mixed Cabb-S/Mark-S virus population after a single passage in a host plant. Twenty-four individual plants, inoculated as above with equal amounts of Cabb-S and Mark-S, were harvested 21 d post-inoculation, when symptoms were fully developed. The viral DNA was purified from 200 mg of young newly formed infected leaves according to the protocol described previously [47] . After the precipitation step of this protocol, the viral DNA was resuspended and further purified with the Wizard DNA clean-up kit (Promega, Fitchburg, Wisconsin, United States) in TE 1X (100 mM Tris-HCl and 10 mM EDTA [pH 8]). Aliquots of viral DNA preparations were digested by restriction enzymes corresponding either to marker a, b, c, or d and submitted to a 1% agarose gel electrophoresis, colored by ethydium bromide and exposed to UV. Each individual restriction enzyme cut once in Cabb-S DNA and twice in Mark-S, thus generating DNA fragments of different sizes attributable to one or the other in the mixed population of CaMV genomes. After scanning the agarose gels, we estimated the relative frequency of the two genotypes in each viral DNA preparation and at each marker position, by densitometry using the NIH 1.62 Image program. The statistical analyses of the frequency of the four markers are described below. Isolation of individual CaMV genomes and identification of recombinants. To identify and quantify the recombinants within the CaMV mixed populations, aliquots from ten of the 24 viral DNA preparations described above were digested by the restriction enzyme SalI, and directly cloned into pUC19 at the corresponding site. In each of the ten viral populations analyzed, 50 full-genome-length clones were digested separately by BsiWI, PstI, MluI, and SacI, to test for the presence of marker a, b, c, and d, respectively. In this experiment, with the marker representing an additional restriction site, we could easily distinguish between the Cabb-S and the Mark-S genotype at all four marker positions, upon agarose gel (1%) electrophoresis of the digested clones. Clones with none or all four markers were parental genotypes, whereas clones harboring 1, 2, or 3 markers were clearly recombinants. Due to the very high number of recombinants detected, markers eventually appearing or disappearing due to spontaneous mutations were neglected. Statistical analysis. Here we present the different methods we used to quantify recombination in the CaMV genome. Because all these methods assume that the different markers are neutral, we first discuss assumption. We used two datasets to test the neutrality of markers, both resulting from plants co-infected with a 1:1 ratio of Mark-S and Cabb-S. The first consisted of viral DNA densitometry data derived from 24 plants (described above), where for each plant we have an estimate of the frequency of each marker in the genome population. The second consisted of the restriction of 50 individual full-genome-length viral clones obtained from one co-infected plant (described above), yielding an estimate of the frequency of each marker, and this was repeated on ten different plants. The frequencies of the different markers were 0.508, 0.501, 0.516, and 0.507 for markers a, b, c, and d in the first dataset and 0.521, 0.518, 0.514, and 0.524 in the second dataset. We tested whether these frequencies were significantly different from the expected value under neutrality, 0.5, using either t-tests, for datasets where normality could not be rejected (seven out of eight cases), or Wilcoxon signed-rank non-parametric tests otherwise (marker c in the first dataset). In all cases p-values were larger than 0.05. There are several cautionary remarks regarding these analyses. First, in all cases we found an excess of markers. Unfortunately, the two datasets cannot be regarded as independent because, even though the methods through which the frequency estimates were obtained were different, the plants used in the second dataset were a subset of the plants of the first. We thus have only four independent estimates in each case, and there is minimal power to detect significant deviations from neutrality with such a small sample size. It should be noted at this stage that deviations from the expected value could also be caused either by slight deviations from the 1:1 ratio in the infecting mixed solution, or by deviations from that ratio in the frequency of the viral particles that actually get into the plants. Second, because of the relatively small sample sizes and low statistical power, the tests presented above could have detected only large deviations. The results clearly show, however, precisely that the markers do not have large effects, if any, and that therefore recombination estimates would be affected only very slightly by any hypothetical selective effects of the markers. Because of this, along with the fact that the introduced markers provoke silent substitutions in the CaMV genome, we assumed that markers were effectively neutral in the rest of the analysis. The dataset used to estimate the recombination frequency consisted of the 500 full-genome-length viral clones (50 from each of ten co-infected plants) individually genotyped for each of the four markers. As discussed in detail in the Results, recombination was very frequent and concerned all four markers. Indeed, approximately half of the genotyped clones exhibited a recombinant genotype. It was therefore meaningful to try to obtain quantitative estimates of recombination from our data. Our aim was to analyze viral recombination in a live host. Consequently, we had to deal with the fact that more than one viral replication cycle occurred during the 21 d that infection lasted in our experiment (we had to wait that long for the disease to develop and to be able to recover sufficient amounts of viral DNA from each infection). Based on the kinetics of gene expression [40] , we postulate that each replication cycle lasts between 2 and 3 d, and that therefore seven to ten cycles occurred between infection and the sampling time. In case this assumption is incorrect, we did calculations assuming five, seven, ten, or 20 replication cycles during these 21 d. As shown, the results were not affected qualitatively, and only slightly quantitatively. It is important to note that we assumed that recombination occurred through a template-switching mechanism, and that therefore, from a recombination point of view, the CaMV genome is linear. The reverse transcription starts and finishes at the position 0 in Figure 1 , which is the point of circularization of the DNA genome. This implies that changes between contiguous markers a-b, b-c, and c-d can be considered as true recombination whereas those between a and d cannot, as they may simply stem from circularization of DNA, during the synthesis of which the polymerase has switched template once anywhere between a-b, b-c, or c-d. To estimate the recombination rate between markers, we wrote recurrence equations describing the change in frequency of each genotype over one generation, assuming random mating and no selection (i.e., the standard Wright-Fisher population genetics model). We then expressed the frequency of all possible genotypes n generations later as a function of their initial frequency and of the recombination parameters. Subsequently we calculated the maximum likelihood estimates of the recombination parameters and their asymptotic variances given initial frequencies (we assumed that the two ''parental'' genotypes, Mark-S and Cabb-S, had equal initial frequencies of 0.5 and that all other genotypes had initial frequencies of zero) and frequencies after n generations (the observed frequencies; as stated above we used different values of n). All algebraic and numerical calculations were carried out with the software Mathematica. The recombination parameters are the recombination rates between two adjacent loci, e.g., r 1 for the recombination rate between markers a and b, and the interference coefficients, e.g., i 12 for interference between recombination events in the segments between markers a and b and b and c. To define these parameters we followed Christiansen [48] , and in particular the recombination distributions for two, three, and four loci (respectively, Tables 2.7, 2.8, and 2.9 of [48] ). It is important to realize that given the definitions of these parameters, the estimator of the recombination rate between two loci is not affected by the number of loci considered. In other words, we obtain the same estimation of the recombination rate between markers a and b whether we consider genotypic frequencies at just these two loci, or the frequencies at these two loci plus a third locus, or the complete information to which we have access, the fourmarker genotypes. Information on additional loci only affects the estimates of the interference coefficients. It proved impossible to carry out the calculations for four loci algebraically. Instead, we used a computer program to calculate the expected genotypic frequencies at all four loci after n generations, given the above stated initial frequencies and specified recombination parameters. For each combination of recombination parameters we calculated a Euclidean distance between the vector of the expected genotypic frequencies and the observed genotypic frequencies, and considered that the estimated recombination parameters were those yielding the minimal Euclidean distance. In all cases, the estimated recombination rates between pairs of loci were equal to the second decimal to those estimated algebraically from data for three or two loci.
26
Torsional restraint: a new twist on frameshifting pseudoknots
mRNA pseudoknots have a stimulatory function in programmed −1 ribosomal frameshifting (−1 PRF). Though we previously presented a model for how mRNA pseudoknots might activate the mechanism for −1 PRF, it did not address the question of the role that they may play in positioning the mRNA relative to the ribosome in this process [E. P. Plant, K. L. M. Jacobs, J. W. Harger, A. Meskauskas, J. L. Jacobs, J. L. Baxter, A. N. Petrov and J. D. Dinman (2003) RNA, 9, 168–174]. A separate ‘torsional restraint’ model suggests that mRNA pseudoknots act to increase the fraction of ribosomes directed to pause with the upstream heptameric slippery site positioned at the ribosome's A- and P-decoding sites [J. D. Dinman (1995) Yeast, 11, 1115–1127]. Here, experiments using a series of ‘pseudo-pseudoknots’ having different degrees of rotational freedom were used to test this model. The results of this study support the mechanistic hypothesis that −1 ribosomal frameshifting is enhanced by torsional resistance of the mRNA pseudoknot.
The structure of an RNA molecule is widely recognized to play a role in many processes, including structurally organizing complex RNAs, the assembly of ribonucleoprotein complexes, and in translational recoding and regulation [reviewed in (1) ]. One common RNA folding motifs is a pseudoknot, the folding back of a single-stranded RNA onto itself to form two helical structures with single-stranded loops joining them (2) . Many such structures can be inferred from RNA sequences and frameshifting function has been demonstrated for some of these [reviewed in (3) (4) (5) ]. However, though much theoretical progress has been made in understanding how mRNA pseudoknots promote efficient À1 ribosomal frameshifting (6), a complete understanding of this mechanism remains untested. Programmed À1 ribosomal frameshift signals are typically divided into three components. From 5 0 to 3 0 these are (i) a 'slippery site' in the form N NNW WWH, where N must be a stretch of any three identical nucleotides, where W is either three A or U residues, and H is A, C or U (spacing indicates the unshifted zero frame), (ii) a spacer region and (iii) an mRNA structural element, most often a pseudoknot. The general model posits that upon encountering the mRNA pseudoknot, an elongating ribosome is forced to pause such that the anticodons of its A-and P-site tRNAs are base-paired with the zero-frame codons of the slippery site. The nature of the tRNA-mRNA interactions is such that a relative slip of À1 nucleotide still allows base-pairing in the non-wobble positions. The slippage occurs during the ribosomal pause, and it has been shown that changes affecting ribosome pause times affect frameshift efficiencies [reviewed in (7) ]. An important observation is that even though mRNA pseudoknots and energetically equivalent stem-loop structures appear to promote ribosome pausing with equal effectiveness, mRNA pseudoknots are more efficient at promoting À1 PRF (8). Our '9 Å ' model (6) provided a refinement of the original 'simultaneous slippage' (9, 10) model of frameshifting by suggesting that rather than the entire ribosome having to slip one base in the 5 0 direction, slippage could be accomplished by moving the small section of mRNA in the downstream tunnel by one base in the 3 0 direction. We have proposed that this is accomplished by the bulky and difficult to unwind mRNA pseudoknot structures becoming wedged in the downstream entrance tunnel of the ribosome, preventing the downstream region of the mRNA from being pulled into the ribosome by the equivalent of one base during the accommodation step of elongation. This blockage would introduce tension into the spacer region, which could be resolved by unpairing the mRNA from the tRNAs, allowing the mRNA to slip 1 nt backwards, resulting in a net shift of reading frame by À1 base. Though the 9 Å model provides a partial explanation for why mRNA pseudoknots promote programmed À1 ribosomal frameshifting (À1 PRF) more efficiently than simple *To whom correspondence should be addressed. Tel: +1 301 405 0981; Fax: +1 301 314 9489; Email: dinman@umd.edu ª The Author 2005. Published by Oxford University Press. All rights reserved. The online version of this article has been published under an open access model. Users are entitled to use, reproduce, disseminate, or display the open access version of this article for non-commercial purposes provided that: the original authorship is properly and fully attributed; the Journal and Oxford University Press are attributed as the original place of publication with the correct citation details given; if an article is subsequently reproduced or disseminated not in its entirety but only in part or as a derivative work this must be clearly indicated. For commercial re-use, please contact journals.permissions@oupjournals.org stem-loop structures, it does not answer the question of how the mRNA pseudoknot directs the ribosome to pause at the correct position along the mRNA. A complementary 'torsional restraint' model addresses this issue (11) . When a stem-loop structure is unwound by an elongating ribosome, unwinding of the stem forces the loop to rotate. Since a simple stem-loop is not restrained, the loop can rotate freely and only the base pairs within the stem resist ribosomal movement, and thus the potential energy of unwinding should be distributed along the length of Stem 1 ( Figure 1A ). However, if the loop is anchored or restrained, as it is in a pseudoknot by Stem 2, since the intrinsic ribosomal helicase is processive (12) , Stem 1 cannot be fully unwound until Stem 2 is first denatured. Mechanically, as the ribosome begins to unwind the base of Stem 1, Stem 2 forces the supercoiling in the remainder of Stem 1, providing extra resistance to ribosome movement. At some specific point, the resistance to ribosome movement provided by the supercoiling counteracts the forward movement of the ribosome, increasing the likelihood that the ribosome will stop at a precise point along the mRNA. Energetically, since full unwinding of Stem 1 is dependent on complete denaturation of Stem 2, the potential energy of unwinding of the pseudoknot structure should similarly be directed toward one point. Viewed either mechanically or energetically, this point is where ribosomes will be directed to specifically pause on the mRNA. If it occurs with the tRNAs in ribosomal A-and P-sites positioned at the slippery site, then frameshifting is stimulated. This is summarized in Figure 1B . The efficiency of À1 PRF can thus be viewed as a function of (i) the fraction of ribosomes paused over the slippery site and (ii) the rate at which the structure can be denatured. There is increasing evidence from single molecule experiments that unfolding occurs in quick 'rips' at a particular force (13) , suggesting that in the case of unfolding pseudoknots, frameshifting efficiency is related to both the energy barriers to unfolding the pseudoknot structure and the resistance of the structure against the force of the ribosome. In the context of the torsional restraint model, this resistance is dependent on the ability of Stem 2 to remain intact while Stem 1 is being unwound. There is experimental data that indirectly support this model: (i) disruption of the first 3 bp of Stem 1, which would displace the ribosome's pause site to a point 3 0 of the slippery site, has been shown to eliminate frameshifting (14) ; (ii) destabilizing Stem 2, which would allow it to be unwound more readily, has been shown to result in decreased frameshifting efficiency (15) (16) (17) ; (iii) replacing bulges in Stem 1 with base pairs would increase the energy required to unwind the first three bases, and a longer ribosomal pause over the slippery site would follow, yielding increased efficiencies in À1 frameshifting (15, 18) ; (iv) destabilizing the base of Stem 1 by replacing G:C base pairs with A:U pairs decreases À1 frameshifting efficiencies (19, 20) ; (v) the model eliminates the need for a 'pseudoknot recognizing factor', the evidence of which has not been forthcoming in either competition assays in in vitro translation systems (21) or by gel retardation assays (J. D. Dinman, unpublished data); and (vi) elimination of a potential torsion-restraining Stem 2, but not of a non-torsion-restraining Stem 2 in HIV-1, resulted in decreased À1 PRF efficiencies (22) . Though all of the cited studies support the torsional restraint model, none has directly addressed it. In the experiments presented in this study, a series of 'pseudo-pseudoknot' containing reporter constructs were used to test the torsional restraint hypothesis. In vitro frameshifting assays show that frameshifting can be significantly stimulated by limiting the rotational freedom of the loop region of a stem-loop structure, and that the degree of rotational freedom of Stem 1 is important in determining the extent of À1 PRF. Furthermore, mRNA toeprint analyses reveal a pseudo-pseudoknot-specific strong stop 16 nt 3 0 of the slippery site, consistent with this structure being able to direct ribosomes to pause with their A-and P-sites positioned at the slippery site. All synthetic DNA oligonucleotides were purchased by IDT (Coralville, IA). The modified L-A viral À1 PRF signal containing the G GGU UUA slippery site followed by a simple stem-loop was amplified from pJD18 (23) using the primers luc5 0 b (5 0 -CCCCAAGCTTATGACTTCTAGGCAGGGTTT-AGG-3 0 ) and luc3 0 b (5 0 -CCCCCCATGGGACGTTGTAAA-AACGACGGGATC-3 0 ). These were digested with HindIII and NcoI (restriction sites are underlined) and cloned into the firefly luciferase reporter plasmid pT7-LUC minus 3 0untranslated region-A50 (24) . In the resulting reporter construct (pJD214-18), expression of firefly luciferase requires a À1 frameshift, and the 5 0 sequence of the Stem 2 is not able to base pair with the 3 0 sequence, so that only a stem-loop rather than a pseudoknot is able to form. The same primers were used to amplify DNA from pJDRC (23) to make pJD214-Ry. In this construct, complementary mutations (5 0 -GCUGGC-3 0 to 5 0 -CGACCG-3 0 ) in the 3 0 acceptor sequence of the pseudoknot-forming region of Stem 2 allow the formation of an mRNA pseudoknot that has previously been shown to promote frameshifting at the same frequency as the wild type (23) . The primer luc5 0 CON (5 0 -CCCCAAGCTTATGACTTC-TAGGCAAGGGTTTAGG-3 0 ) contains an additional A nucleotide upstream of the slippery site and was used to make pJD214-0, the zero-frame control. To eliminate the possibility of internal initiation occurring at the luciferase initiation codon downstream of the frameshift signals, the AUG codon was changed to AUA. The Stratagene Quik-Change kit was used to mutate pJD214-18 and pJD214-Ry into pJD336-18 and pJD366-Ry, respectively, using the oligonucleotides 5 0 -GGCGTTCTTCTATGGGACGTTGTA-AAAACGGATC-3 0 and 5 0 -GATCCGTCGTTTTTACAACG-TCCCATAGAAGACGCC-3 0 (the mutated codon is underlined). pJD366-18 was further mutated to make a zero-frame control by placing an A upstream of the slippery site using the oligonucleotides 5 0 -TGACTTCTAGGCAAGGGTTTAGGAG-TG and 5 0 -CACTCCTAAACCCTTGCCTAGAAGTCA (the inserted base is underlined). A series of synthetic DNA oligonucleotides were designed to join the loop acceptor region of mRNA transcribed from pJD366-18 to the downstream region that forms the pseudoknot in the wild-type L-A À1 PRF signal. In the J-oligos, the 3 0 sequences base pair with the loop of the mRNA transcribed from pJD366-18, and the 5 0 regions of these oligos base pair with the downstream sequence. This orientation is reversed for the R-oligos. These general orientations are shown in Figure 3B . The naming of the oligonucleotide names refers to the number of additional residues placed between the regions of complementarity. The bases complementary to the pJD366-18 sequence are underlined. Plasmid DNAs were prepared using the Qiagen mini-prep kits and were linearized with DraI in a total volume of 20 ml. Proteins were eliminated by the addition of 2 ml of 1 mg/ml proteinase K and SDS to a final concentration of 0.5% followed by digestion at 50 C for 30 min. Volumes were then increased to 100 ml, extracted twice with phenol/chloroform, and DNA was precipitated with 10 ml NH 4 Ac and 250 ml ethanol. The purified DNA was resuspended in DEPCtreated H 2 O. To prepare synthetic mRNAs, 2 ml of purified linear DNAs were used for in vitro transcription using the Ambion T7 mMachine mMessage kit. RNAs were precipitated using 30 ml DEPC H 2 O and 25 ml LiAc. The RNA was resuspended in 11 ml DEPC H 2 O (1 ml in 500 would give an OD 260 of 0.05-00.1; 1-2 mg/ml). To anneal the oligonucleotides with the mRNA, J-oligos, R-oligos or the equivalent volumes of dilution buffer alone (20 mM Tris, pH 7.4, 2 mM MgCl 2 and 50 mM EDTA final concentration) were added to synthetic mRNA (0.5 mg), and the mixtures were first incubated in a 70 C heating block for 10 min; the block was then removed and allowed to cool to 37 C (30 min), after which they were briefly spun down and incubated on ice. In all experiments, the molar ratios of J-and R-oligos to synthetic mRNAs were 100:1. In experiments using the competing oligonucleotide (C-oligo), this was added to either 0.5:1 or 1:1 molar ratios with either J-or R-oligonucleotides. Reticulocyte lysates were thawed on ice, 15 ml of Àmet and 15 ml of Àleu master mixes plus 20 ml of H 2 O were added to 400 ml of lysate, and 19 ml of this was added to each annealed reaction to start the in vitro translation reactions. These were incubated at 30 C for 60 min (the reaction reached a plateau after 30-35 min where the greatest difference was seen between the zero-frame controls and the frameshifting plasmids) (data not shown), and the reactions were then placed on ice. An aliquot of 7.5 ml from each in vitro translation reaction was added to 50 ml of the prewarmed luciferase reagent, and luminescence readings were taken after a 3 s delay for 15 s in triplicate using a Turner 20/20 Luminometer. Synthetic transcripts generated from DraI-digested pJD366-18 ($1.7 kb) were 5 0 end labeled using [g-32 P]CTP. These RNAs (4 ml) were incubated with 1 ml of annealing buffer and either 1 ml of H 2 O or 1 ml of an oligo (0.25 ng) at 70 C. The heating block was allowed to cool at room temperature for 40 min before 8 ml of RNaseH buffer was added (20 mM HEPES, 50 mM KCl, 10 mM MgCl 2 and 1 mM DTT). An aliquot of 1 ml of enzyme was added (mung bean nuclease, RNaseH or RNaseT1) and the reactions incubated at 37 C for 1 h. The reactions were stopped by adding 4 ml of stop solution, the products separated through a 6% polyacrylamide-urea denaturing gel and visualized by autoradiography. mRNA toeprinting JD366-18 mRNA (1 mg in 8 ml) was annealed with 2 ml of 3 0 end-labeled toeprinting primer (5 0 -CGTACGTGATCTTCA-CC-3 0 , complementary to sequence 240 bp 3 0 of the slippery site) as described above. This was added to 15 ml of lysate (200 ml Ambion retic lysate, 7.5 ml of each master mix [Àleu and Àmet] and 70 ml of 250 mM KCl), except for 2 ml, which was added to 15 ml of RT buffer [50 mM Tris-HCl (25), 40 mM KCl, 6 mM MgCl 2 , 5 mM DTT and 575 mM dNTPs] to be used as a no-ribosome control. In vitro translation reactions were incubated at room temperature for 10 min, which was empirically determined to provide the optimum amount of time to allow ribosomes to initiate translation and pause at the frameshift signal. Subsequently, 15 ml of RT buffer containing RNasin inhibitor and cycloheximide (to a final concentration of 100 ng/ml) was added to stop translation. To this, 2 ml of Superscript II (Invitrogen) was added and the reaction incubated at room temperature for 10 min. Reactions were terminated by phenol:chloroform extraction and 15 ml of stop solution added. The toeprinting primer was also used in conjunction with pJD366-18 to produce sequencing ladders by standard dideoxynucleotide chain termination methods using Sequenase (USB). Products were separated though 6% polyacrylamide-urea denaturing gels and visualized using a Storm phosphorImager (Pharmacia). Pseudo-pseudoknots stimulate frameshifting, and frameshifting efficiency changes with the degree of pseudo-pseudoknot rotational freedom We previously showed in intact yeast cells that the pseudoknot containing mRNA produced from pJDRC was able to promote efficient À1 PRF, whereas one in which only a stem-loop can form, transcribed from pJD18, could not (23) . As a first step in this study, we tested the ability of synthetic mRNAs produced from pJD366-RC and from pJD366-18, two plasmids derived from these parental constructs, to promote À1 PRF. Total luciferase activities produced from these synthetic mRNAs were divided by the luciferase activity produced from the zero-frame control plasmid, pJD366-0, and multiplied by 100% to determine À1 PRF efficiencies. The results show that the trends observed in yeast were replicated in vitro, i.e. JD366-RC mRNA promoted $8% efficiency of À1 PRF as compared with $1.1% promoted by JD366-18 mRNA (Figure 2 ). The 'torsional restraint' model predicts that conditions that would inhibit the rotational freedom of the loop region of the pJD366-18-derived mRNA should result in enhanced À1 PRF efficiency. The strategy used in this study was to anneal this mRNA with synthetic oligonucleotides complementary to both the loop region and to the sequence downstream that is normally involved in pseudoknot formation. These 'pseudopseudoknots' would be predicted to restore a pseudoknot-like structure to the mRNA. This is diagrammed in Figure 3A . Two different classes of oligonucleotides having different orientations relative to the mRNA were used to this end: 'joining' (J-) and 'reverse' (R-) oligos. The orientation of the J-oligos promotes the formation of a structure containing the equivalent of a Loop 2 region, while that of the R-oligos promotes a Loop 1 equivalent. The model also predicts that pseudo-pseudoknots having different degrees of rotational freedom should promote different frequencies of ribosome pausing over the slippery site, resulting in different efficiencies of À1 PRF. In order to control this parameter, increasing numbers of nucleotides were inserted between the mRNA hybridizing regions of the J-and R-oligos. The additional non-complementary bases in the J-oligos are 3 0 to the stem-loop residues involved in Stem 2, thus effectively increasing Loop 2. Similarly, the additional non-complementary bases in the R-oligos are 5 0 to the loop acceptor residues and correspond to an increased Loop 1. The structure of the stem-loop of pJD366-18 and its maximum base-paired interactions with representative J-and R-oligos are shown in Figure 3B . To demonstrate that an oligonucleotide-mRNA hybrid was capable of forming under the assay conditions, the J1-oligo was incubated with 5 0 [ 32 P]labeled JD366-18 mRNA and subjected to RNaseH digestion. Digestion of the RNA-DNA hybrid resulted in a labeled 110 nt fragment, demonstrating that the oligonucleotide bound to the mRNA at the position of the pseudoknot (Figure 4) . Having demonstrated the utility of the in vitro frameshifting assay and that the J-and R-series of oligonucleotides were able to hybridize with synthetic mRNA produced from pJD366-18, the next step was to monitor frameshifting efficiencies promoted by these hybrid species. Significant increases in frameshifting were observed with the incubation of pJD366-18 mRNA with oligonucleotides J1 ($10%) and J2 ($35%), while only modest increases were seen with J3 and J4 ( Figure 5 ). These findings are consistent with the notion that changes in the degree of rotational freedom of the structure would affect the distribution of paused ribosomes in the vicinity of the slippery site. One potential complication with the J-oligos is the possibility that they could interact with the Loop 2-Stem 1 region. In the R-oligos, the additional bases are distal to any possible Loop 2-Stem 1 interactions and would be more analogous to increasing Loop 1. The R-oligos stimulated À1 PRF to an even higher extent than the J-oligos ( Figure 5 ). Importantly, increasing the length of the bridging regions in these oligonucleotides (R1 to R3), which is predicted to increase the rotational freedom of the stem-loop, resulted in decreased frameshifting activity as predicted by the torsional resistance model. However, addition of three residues between the two binding regions of the R-oligo (R4) resulted in an unexpected increase in frameshifting with a very large amount of variation. In a series of control experiments, 8 nt oligos complementary to the 5 0 (Loop 1) and 3 0 Stem 2 forming regions of the pseudo-pseudoknot were hybridized to the SL mRNA and À1 PRF assays were performed. Neither of these were able to stimulate À1 PRF, even at concentrations in 100-fold molar excess to the mRNA template (data not shown). Though supportive of our central hypothesis, it is also possible that these results were due to the thermodynamic instability of the RNA:DNA duplexes through the course of the experimental protocol. To determine whether the stimulation of frameshifting was specifically due to the bridging of the stem-loop with downstream sequence (the pseudo-pseudoknot), as opposed to the nonspecific presence of an RNA:DNA hybrid, the competing oligonucleotide (C-oligo) was designed to form a 15 bp duplex with JD366-18 mRNA, including the 3 0 Stem 2 forming region, which was expected to significantly out compete either the J-or R-oligos from binding to this site, thus disrupting formation of the pseudo-pseudoknot (see Figure 3A ). Additionally, in the presence of the C-oligo, the J-and R-oligos were still predicted to hybridize with the 5 0 Stem 2 forming region, enabling us to address the question of whether this interaction alone was able to stimulate frameshifting. The results demonstrate that the addition of the C-oligo severely inhibited the abilities of both the J-and R-oligos to promote efficient frameshifting ( Figure 6 ). These findings demonstrate that (i) frameshifting was specifically stimulated by bridging of the 5 0 and 3 0 Stem 2 forming regions by the J-and R-oligos, and (ii) that the presence of an RNA:DNA hybrid at the 5 0 Stem 2 forming region was not sufficient to stimulate frameshifting by itself. The torsional restraint model predicts that pseudoknots should direct elongating ribosomes to pause at one specific location 1 2 3 4 5 6 7 8 . Efficient frameshifting is stimulated by pseudo-pseudoknots. In vitro translation assays were performed in retic lysates with mRNAs derived from pJD366-18 (SL) to which J-or R-oligos were annealed. Luciferase activities were divided by those obtained using mRNAs generated from pJD366-0, and the resulting ratios were multiplied by 100 to calculate percent frameshifting. The averages of three independent experiments performed in triplicate are shown. Error bars denote standard deviation. on the mRNA, rather than being distributed along Stem 1. We used mRNA toeprint assays to test this hypothesis. In mRNA toeprint reactions, the movement of reverse transcriptase is blocked by paused ribosomes, resulting in a strong stop positioned $16-18 nt 3 0 of the P-site of eukaryotic ribosomes (25) . Synthetic JD366-18 mRNAs were annealed with the sequencing oligonucleotide and either J1, R1 or no second oligo, and these were then used for in vitro translation reactions. After a period of time (10 min were empirically determined to be optimal), elongation reactions were stopped by the addition of cycloheximide, and reverse transcription reactions were initiated on the sequencing oligonucleotides. In parallel, control reverse transcription reactions were carried out using synthetic JD366-18 mRNA and oligonucleotides, but without in vitro translation. The results are consistent with the model, showing that the J1-and R1-oligos specifically promoted one strong reverse transcriptase stop 16 nt 3 0 of the P-site of the slippery site only the in the in vitro translation reactions ( Figure 7 ). As further predicted by the model, a broad distribution of stops of equal intensities was observed in this region with JD366-18 mRNA alone (Figure 7, lane 1) . Importantly, the +16 stop was not observed when toeprint reactions were carried in the absence of ribosomes. Additional strong stops were also of interest. One corresponding to the 3 0 end of the base of Stem 1 was observed in all samples, consistent with the presence of this structure. Both J-and R-oligo-specific pauses were also observed. The reason for the strong pause in the J-oligo is unknown. The R-oligo-specific pause is perhaps more revealing. It occurs at the 3 0 end of the RNA:DNA hybrid formed by this oligo and the mRNA, a structure that should also promote pausing of reverse transcriptase. The results presented in this study provide strong support for the torsional restraint model of programmed À1 frameshifting. Specifically, we demonstrated that RNA:DNA hybrids that mimic mRNA pseudoknots can significantly stimulate frameshifting. As predicted by the model, changing the rotational freedom of the structure by altering the lengths in the J1-and R1-oligos between the 5 0 and 3 0 mRNA hybridizing regions resulted in changes in their abilities to stimulate À1 frameshifting. The demonstration that these 'pseudopseudoknot' structures cause elongating ribosomes to specifically pause with their A-and P-sites positioned at the slippery site provides independent evidence in support of the model. In the case of the J-oligo series, frameshifting was best stimulated by J2, suggesting the structure created and the rotational freedom allowed by it was optimal for À1 PRF. The experimental design is such that we assume a similar rate of unfolding for each oligo as the predicted maximum base pairing is the same for them all. However, we do note that the type of nucleotides separating the two, separately paired regions of the oligos, and their presentation, may play a role in À1 PRF efficiency. The recent NMR structural solution of the SRV-1 pseudoknot revealed a highly structured Loop 2-Stem 1 interface including base triples involving an A residue at the 3 0 end of Loop 2 (26) . The additional base in the J2oligonucleotide is also an A. Mutagenesis experiments in this region by other groups showed, for example, that replacing the 3 0 base in Loop 2 of IBV with an A residue promoted a significant increase in frameshifting efficiency (27) , and mutation or removal of the A residue at the 3 0 base in Loop 2 of the BWYV pseudoknot reduced frameshifting levels (17) . This part of the pseudoknot has been proposed to be important in a frameshifting model where differential transition state energy barriers (due to small differences in local structure, stability or dynamics) are the primary determinant of frameshifting efficiency (3). Indeed, a Loop 2-Stem 1 triplex interaction seen in smaller frameshifting pseudoknots from luteoviruses has been shown to be critical for À1 PRF, and that similar pseudoknots lacking the triplex are less efficient at frameshifting [(28) and references therein]. This extra structural feature would limit the rate of unfolding and provide extra anchoring of Stem 2 as the ribosome attempts to unwind Stem 1, i.e. it too would help to provide additional torsional restraint. It is also possible that although the J3-and J4oligonucleotides also help to form a pseudo-pseudoknot, the additional bases may interfere with the stabilization of Stem 1. With the R-oligos, a general correlation was observed between minimization of rotational freedom and frameshifting efficiency, though this was not the case of the R4-oligo. Since the stability of the pseudo-pseudoknot generated with R4 should be similar to that of the other oligonucleotides based on the base-pairing, this result suggests that there are additional considerations to be uncovered with regard to the Frameshifting (% stimulated by R1) Figure 6 . Competition for J-or R-oligo binding sites inhibits its ability to promote efficient frameshifting. mRNA transcribed from pJD366-18 (SL) was annealed with either J-or R-oligos alone, or in combination with different concentrations of competing (C-) oligos (in ratios of 2:1 or equimolar as indicated). Sample marked SL is mRNA alone. Luciferase activities generated from in vitro translation reactions in rabbit reticulocyte lysates were divided by those obtained using mRNAs generated from pJD366-0, and the resulting ratios were multiplied by 100 to calculate percent frameshifting. pseudoknot structure influencing frameshifting. Addition of residues in the R-oligos was analogous to lengthening Loop 1, which is typically short in À1 frameshifting pseudoknots. Limited and conflicting data are available on the importance of Loop 1 in À1 frameshifting pseudoknots. In one study, addition of three A bases to Loop 1 did not affect frameshifting efficiency (15) , while in another all the mutations made in this region were detrimental to frameshifting efficiency (17) . Given the complex interactions occurring between the helices and loops in this region, we cannot yet account for why the R4-oligo stimulated frameshifting so efficiently and with such variable results. Examination of the RNA toeprint data presented here reveals that both of the pseudo-pseudoknot structures formed by the J1-and R1-oligos promoted strong stops of the reverse transcriptase $16 nt 3 0 of the P-site codon of the slippery site, consistent with the hypothesis that the presence of Stem 2 forces ribosomes to pause with their A-and P-sites positioned over the slippery site. Previous studies mapping the lagging edge of paused ribosomes, i.e. mRNA heelprint studies, did not reveal any striking differences between the effects of pseudoknots versus stem-loops (8, 16) . Interestingly, using this method, the ribosomal pauses appeared distributed over a broader stretch of mRNA ($4 nt) than observed here. It is possible that some critical level of resolution is lost in the requirement for many additional manipulations of substrates using the mRNA heelprint as compared with the toeprint methods. A remaining question centers on whether the role of the RNA pseudoknot in À1 PRF is passive or active. In the '9 Å solution' (6), the frameshift mechanism is activated by movement of the A-site codon-anticodon complex by 1 base in the 5 0 direction upon accommodation. As currently described, the mRNA pseudoknot merely passively blocks entry of the downstream message into the ribosome, resulting in stretching of the segment of mRNA located between the codon-anticodon complex and the pseudoknot. By this model, all of the energetic input for the frameshift is derived from hydrolysis of GTP by eEF1A. However, it is possible that the pseudoknot may also actively contribute to the frameshift mechanism. Specifically, pulling the downstream message into the ribosome at accommodation could result in unwinding of Stem 1 of the pseudoknot by one additional base pair. The energetic cost of so doing would be to introduce an equivalent amount of torsional resistance into Stem 2. If Stem 2 were to release this resistance by 'pulling back', the base pair in Stem 1 would be re-formed, which in turn would contribute to the energy required to dissociate the A-and P-site codonanticodon complexes from the zero-frame. This would be followed by slippage of the mRNA by 1 base in the 3 0 direction relative to the ribosome, followed by the formation of À1 frame codon-anticodon complexes. As such, the proposed active role for the mRNA pseudoknot would further reduce the energetic barrier to À1 PRF. In sum, we suggest that the 'torsional restraint model' can be combined with the '9 Å solution' to mechanistically explain the original 'simultaneous 3' mRNA + Ribos. mRNA Figure 7 . Pseudo-pseudoknots direct ribosomes to pause over the slippery site. mRNAs generated from pJD366-18 (SL) were annealed with the sequencing oligonucleotide and either J1-, R1-or no oligo (lanes 1, 2 and 3, respectively), and these were then used for in vitro translation reactions. Reactions were stopped after 10 min by the addition of cycloheximide, and reverse transcription reactions were initiated on the sequencing oligonucleotides. In parallel, control reverse transcription reactions were carried out using synthetic JD366-18 mRNA and oligonucleotides, but without in vitro translation (lanes 4-6). The positions of the slippery site, loops and stems of the pseudo-pseudoknots are indicated next to a sequencing reaction. Arrowheads indicate positions of reverse transcriptase strong stops and these are mapped to a representation of the stem-loop structure of pJD366- 18. slippage' model of À1 PRF (9, 10) . In other words, the 9 Å solution + torsional restraint = simultaneous slippage. Two recent publications have also shown that oligonucleotide:mRNA duplexes can stimulate efficient À1 ribosomal frameshifting (29, 30) . These studies differed from the present one in a number of ways, particularly insofar as they examined the effects duplex structures immediately 3 0 of the slippery site rather than addressing mRNA pseudoknot related questions. The findings support the notion that the specific location of ribosome pausing on the mRNA plays a critical role in determining frameshifting, though they do come with caveats, e.g. neither study directly mapped ribosomal pausing, and the use of different slippery sites and downstream contexts likely contributed to disparate findings for the optimal distances between the 3 0 ends of slippery sites and 5 0 ends of frameshift-stimulating oligonucleotides. Although potentially useful therapeutically there are no known natural examples of frameshifting stimulated in this manner, and thus these results do not affect the hypothesis presented here. However, these studies are important in that they raise the possibility for a new role for micro-RNAs in regulating gene expression, and for therapeutic approaches to correcting inborn errors of metabolism due to the presence of frameshift mutations.
27
Correcting errors in synthetic DNA through consensus shuffling
Although efficient methods exist to assemble synthetic oligonucleotides into genes and genomes, these suffer from the presence of 1–3 random errors/kb of DNA. Here, we introduce a new method termed consensus shuffling and demonstrate its use to significantly reduce random errors in synthetic DNA. In this method, errors are revealed as mismatches by re-hybridization of the population. The DNA is fragmented, and mismatched fragments are removed upon binding to an immobilized mismatch binding protein (MutS). PCR assembly of the remaining fragments yields a new population of full-length sequences enriched for the consensus sequence of the input population. We show that two iterations of consensus shuffling improved a population of synthetic green fluorescent protein (GFPuv) clones from ∼60 to >90% fluorescent, and decreased errors 3.5- to 4.3-fold to final values of ∼1 error per 3500 bp. In addition, two iterations of consensus shuffling corrected a population of GFPuv clones where all members were non-functional, to a population where 82% of clones were fluorescent. Consensus shuffling should facilitate the rapid and accurate synthesis of long DNA sequences.
Methods for the automated chemical synthesis of oligonucleotides (1, 2) and their assembly into long double-stranded DNA (dsDNA) sequences by PCR (3, 4) and LCR (5) have enabled the chemical synthesis of genes and even entire viral genomes (6, 7) . These technological advances have helped spur the formation of the new field of synthetic biology (8) , which aims at defining the functional units of living organisms through the modular engineering of synthetic organisms. In addition, the demand for fully synthetic gene length DNA fragments of defined sequence has dramatically increased in recent years for use in applications such as codon optimization (9), construction of DNA vaccines (10) , de novo synthesis of novel biopolymers (11) , or simply to gain access to known DNA sequences when original templates are unavailable. The future demand for long synthetic DNA is likely to dramatically increase when it becomes cheaper/faster to synthesize a desired sequence than to obtain it by other means. The assembly of DNA is currently limited by the presence of random sequence errors in synthetic oligonucleotides that arise from side reactions during synthesis (incomplete couplings, misincorporations, etc.) and resulting in 1-3 errors/kb (7, 12, 13) . The deleterious impact of these errors becomes more significant as the desired lengths of synthetic DNA increase. Indeed, in the remarkable assembly of the PhiX 174 bacteriophage genome (5386 bp) using gel-purified, synthetic oligonucleotides, the products contained an average of $2 lethal errors/kb resulting in 1 plaque-forming genomes per 20 000 clones (7) . A functional selection (plaque formation) was required in this study to identify a clone with the correct sequence. Thus, error reduction/correction is a requirement for the efficient production of long synthetic DNA of defined sequence. However, the process of sequencing multiple clones and manual correction of errors is both costly and time consuming. Several methods have been reported for the removal of error-containing sequences in populations of DNA. These methods rely upon the selective destruction (14, 15) or physical separation (16, 17) of mismatch-containing heteroduplexes. Smith and Modrich (14) reported the selective destruction of error-containing sequences in PCR products by generating dsDNA breaks upon overdigestion with the Escherichia coli mismatch-specific endonuclease MutHLS (18) . Gel purification and cloning of the remaining full-length DNA resulted in an apparent 10-fold reduction in the error rate for PCR products. However, the existing approaches are not well suited for error removal in long synthetic DNA sequences where virtually all members in the population contain multiple errors. The online version of this article has been published under an open access model. Users are entitled to use, reproduce, disseminate, or display the open access version of this article for non-commercial purposes provided that: the original authorship is properly and fully attributed; the Journal and Oxford University Press are attributed as the original place of publication with the correct citation details given; if an article is subsequently reproduced or disseminated not in its entirety but only in part or as a derivative work this must be clearly indicated. For commercial re-use, please contact journals.permissions@oupjournals.org Error correction with MutS is outlined in Figure 1 . The population of DNA molecules containing random errors is first re-hybridized to expose synthesis errors as mismatches ( Figure 1A ). Duplexes containing mismatches can then be removed from the population by affinity capture with immobilized MutS ( Figure 1B) , a process we term coincidence filtering, since both strands of the duplex must match to pass this filtering step. For long synthetic DNA sequences or for sequences with high error rates, coincidence filtering is ineffective, since the likelihood of both strands being perfectly matched after re-hybridization is very low. To generalize MutS error filtering for application on synthetic DNA, the synthetic DNA is cleaved into small overlapping fragments before MutS filtering. Fragments containing mismatches are selectively removed through absorption to an immobilized maltose-binding protein (MBP)-Thermus aquaticus (Taq) MutS-His 6 fusion protein (MBP-MutS-H6) (18) (19) (20) . The remaining mixture of fragments (enriched with fragments of the correct sequence) serves as a template for assembly PCR to produce the full-length product ( Figure 1C ). This process can be iterated until the consensus sequence emerges as the dominant species in the population. This approach is equivalent to DNA shuffling (21) with additional mismatch exposure and removal steps. In this report, we assemble GFPuv from synthetic oligonucleotides and apply both coincidence filtering and consensus shuffling protocols to reduce errors in the resultant DNA populations. The error rates are characterized by gene function (fluorescence) and by DNA sequencing. We also provide a mathematical model describing the error reduction protocols to aid predictions about parameters influencing their effectiveness. Chemicals were from Sigma. Restriction enzymes were from Promega and New England Biolabs. KOD Hot Start DNA Polymerase was from Novagen. Amylose resin was from NEB (catalog no. E8021S). Ni-NTA resin was from Novagen (catalog no. 70666). Ultrafiltration device from Millipore (catalog no. UFC900524). Slide-A-Lyzer dialysis membrane was from Pierce (catalog no. 66415). Full-length Taq MutS was amplified from template pETMutS (22) with primers 5 0 -AAA AAA CAT ATG GAA GGC ATG CTG AAG G-3 0 and 5 0 -AAA AAT AAG CTT CCC CTT CAT GGT ATC CAA GG-3 0 and cloned into the Nde1/HindIII sites of vector pIADL14 (23) to give plasmid pMBP-MutS-H6. E.coli strain BL21(DE3) transformed with pMBP-MutS-H6 was grown to OD 600 $1.0 and induced using 1 mM isopropylb-D-thiogalactopyranoside for 4 h at 37 C. Cells from 4 l of culture were pelleted and resuspended in 60 ml of buffer A (20 mM Tris-HCl, pH 7.4, 300 mM NaCl, 1 mM EDTA, 1 mM DTT and 1 mM phenylmethlysulfonyl fluoride). Cell suspension was sonicated on ice and insoluble material was removed by centrifugation at 50 000 g for 10 min at 4 C. Supernatant was applied to 5 ml amylose resin pre-equilibrated in buffer A. Bound MBP-MutS-H6 was washed three times using 20 ml buffer B (20 mM Tris-HCl, pH 7.4, 300 mM NaCl) and stored The re-hybridized gene synthesis products are fragmented, and error containing fragments are precipitated by MBP-MutS-H6 immobilized on amylose support. Error reduced fragments (orange, blue and red) are reassembled into the full-length gene followed by PCR amplification to generate error reduced products. Primers: black lines. overnight at 4 C. MBP-MutS-H6 was eluted using 20 ml buffer B + 10 mM maltose. Eluate was applied to $4 ml of Ni-NTA resin pre-equilibrated in buffer B. Bound MBP-MutS-H6 was washed four times using 20 ml buffer B + 25 mM imidazole. Bound MBP-MutS-H6 was eluted using buffer B + 1 M imidazole. Eluate was concentrated via ultrafiltration using Amicon Ultra 5 kDa MWCO at 4 C. Concentrated sample was dialyzed extensively against 2· storage buffer (100 mM Tris-HCl, pH 7.5, 200 mM NaCl, 0.2 mM EDTA and 0.2 mM DTT) using a Slide-A-Lyzer 10 kDa MWCO cassette at 4 C. Protein concentration was determined using A 280 and a calculated extinction coefficient of 119 070 M À1 cm À1 . Dialyzed sample was diluted using an equal volume of glycerol and stored at À20 Oligonucleotides were purchased from Qiagen with 'salt-free' purification. Sequence 261-1020 of pGFPuv (GenBank accession no. U62636 with T357C, T811A and C812G base substitutions) was assembled using 40mer (37) and 20mer (2) oligonucleotides with 20 bp overlap (Supplementary Table 1 ). Assembly reactions contained the following components: 64 nM each oligonucleotide, 200 mM dNTPs, 1 mM MgSO 4 , 1· buffer and 0.02 U/ml KOD Hot Start DNA Polymerase. Assembly was carried out using 25 cycles of 94 C for 30 s, 52 C for 30 s and 72 C for 2 min. PCR amplification of assembly products contained the following components: 10-fold dilution of assembly reaction, 25 mM of 20 bp outside primers, 200 mM dNTPs, 1 mM MgSO 4 , 1· buffer and 0.02 U/ml KOD Hot Start DNA Polymerase. PCR was carried out using 35 cycles of 94 C for 30 s, 55 C for 30 s and 72 C for 1 min followed by a final extension at 72 C for 10 min. PCR products were purified using the Qiagen QIAquick PCR purification kit with elution in dH 2 O followed by speed-vac concentration. Assuming an error rate of 1 · 10 À6 /bp/duplication for KOD DNA polymerase (24) , 35 cycles of PCR would be expected to introduce $0.053 mutations per assembled GFPuv molecule. Assembled GFPuv was diluted to 250 ng/ml in 10 mM Tris-HCl, pH 7.8, 50 mM NaCl and heated to 95 C for 5 min followed by cooling 0.1 C/s to 25 C. Heteroduplex for consensus filtering was split into three pools and digested to completion with NlaIII (NEB), TaqI (NEB) or NcoI plus XhoI (Promega) for 2 h following the manufacturer's protocols. Digests were purified using the Qiagen QIAquick PCR purification kit with elution in dH 2 O. Samples were pooled and the concentration was determined by measuring A 260 . MBP-MutS-H6 binding reactions contained $11.5 ng/ml DNA and $950 nM MBP-MutS-H6 dimers in 1· binding buffer (20 mM Tris-HCl, pH 7.8, 10 mM NaCl, 5 mM MgCl 2 , 1 mM DTT and 5% glycerol). Reactions were allowed to incubate at room temperature for 10 min before incubation for 30 min with an equal volume of amylose resin pre-equilibrated in 1· binding buffer. Protein-DNA complexes were removed by low-speed centrifugation and aliquots of supernatant were removed for subsequent processing. Supernatant (50 ml) from consensus filtering experiments was desalted using Centri-Sep spin columns (Princeton Separations) and concentrated. Purified and concentrated DNA fragments were reassembled as above with aliquots removed at varying cycles. Aliquots of assembly reactions were resolved on 2% agarose gels to monitor the reassembly process. Aliquots showing predominantly reassembled fulllength GFPuv were PCR amplified as above. Aliquots of supernatant from coincidence filtering experiments were diluted 10-fold and PCR amplified as above. PCR products were digested with BamHI/EcoRI and ligated into the 2595 bp BamHI-EcoRI fragment of pGFPuv. Ligations were transformed into E.coli DH5 and fluorescent colonies were scored using a handheld 365 nm ultraviolet (UV) lamp. Preparation of substrate for consensus shuffling from 10 non-fluorescent GFPuv clones Ten non-fluorescent GFPuv clones were pooled in equal amounts. The nature and location of the mutations in these clones is shown in Figure 4 . The GFP coding region was PCR amplified from the mixture and submitted to the consensus shuffling protocol with and without the application of the MBP-MutS-H6 error filter. To create an error filter, we constructed a fusion protein between MBP (19) and the mismatch binding protein from T.aquaticus (22) with a C-terminal His 6 tag (MBP-MutS-H6). MBP-MutS-H6 was overexpressed and purified from E.coli to >95% purity (Supplementary Figure 1) . MBP-MutS-H6 immobilized on amylose resin was shown to selectively retain a 40mer heteroduplex containing a deletion mutation over wild-type homoduplex (Supplementary Figure 2) . To demonstrate error correction, unpurified 40mer oligonucleotides were assembled by PCR (3) to produce a 760 bp gene encoding green fluorescent protein (25) (GFPuv). Two independent preparations of GFPuv containing typical gene synthesis errors (Figure 3 and Table 1 ) were re-hybridized and subjected to two iterations of coincidence filtering or consensus shuffling. For consensus shuffling, the GFPuv assembly product was split into three pools and digested into sets of overlapping fragments using distinct Type II restriction enzymes ( Figure 2 ). The digests were pooled and subjected to error filtering with or without added MBP-MutS-H6. The unbound fragments were reassembled into full-length products and PCR amplified. For coincidence filtering, unbound fulllength GFPuv was PCR amplified following treatment with the error filter. After cloning in E.coli, error rates were estimated by scoring colonies for fluorescence under a handheld UV lamp (Figure 3) . The actual error rates of the input and consensus shuffled populations were determined by sequencing plasmid DNA from randomly selected colonies (Figure 3) . The results show that two rounds of consensus shuffling increased the percentage of fluorescent colonies from $60 to >90% and Table 1 . Sequence errors in input and consensus shuffled DNA Table 1 . Although DNA shuffling has traditionally been used to create diversity through the combinatorial shuffling of mutations in a population, DNA shuffling also creates a sub-population of sequences with a reduction in diversity, as correct fragments can recombine to produce error-free sequences. Indeed, with consensus shuffling, it is possible to start with a population of DNA molecules wherein every individual in the population contains errors and create a new population where the dominant sequence is error free. To demonstrate this, 10 nonfluorescent GFPuv clones, each containing a deletion mutation (Figure 4) , were pooled and subjected to either DNA shuffling alone or two iterations of consensus shuffling. Products were cloned in E.coli, and the percentage of fluorescent colonies was monitored as an indication of progress toward the consensus sequence. DNA shuffling alone (no MBP-MutS-H6) increased the percentage of fluorescent colonies to 30% (387 colonies total) similar to a previous report (26) . Two rounds of consensus shuffling gave a new population that was 82% fluorescent (551 colonies total), indicating that the dominant species was likely the consensus sequence of the input population. Both consensus shuffling and coincidence filtering protocols were effective in reducing errors in synthetic GFPuv populations ( Figure 3 ). In both cases, two iterations of either consensus shuffling or coincidence filtering increased fluorescent colonies from average values of $60 to >90%. Sequencing data from two independent experiments showed a 4.3-and 3.5-fold reduction in the error rate for the consensus shuffled populations compared with the input populations giving final error rates of 0.3 and 0.28 errors/kb, respectively. These results demonstrate the usefulness of the MBP-MutS-H6 error filter in both consensus shuffling and coincidence filtering protocols. Taq MutS has previously been shown to bind to deletion mutations with high affinity (27) , a mutation common in synthetic DNA. However, it is important to note that Taq MutS has lower affinity for specific point mutations and binds weakly to homoduplex DNA (27) . These factors may limit the stepwise efficiency of the error filter. Moreover, specific point mutations may be refractory to removal even after multiple rounds of consensus shuffling. Two rounds of consensus shuffling using the MBP-MutS-H6 error filter proved most effective in reducing deletions and G/C to A/T transitions, consistent with previous reports for the selectivity of Taq MutS (27) . However, it must be emphasized that each synthetic oligonucleotide point mutation would generate two heteroduplex DNA molecules containing unique mismatches after PCR amplification and re-hybridization ( Figure 1A and Table 1 ). For example, a G to A transition mutation in a synthetic oligonucleotide would generate heteroduplexes with G-T or A-C mismatches after PCR amplification and re-hybridization. For consensus shuffling, either of these mismatch containing heteroduplexes could evade precipitation by the MBP-MutS-H6 error filter and participate in the reassembly of full-length GFPuv. Therefore, Table 1 lists the pair of mismatches that could give rise to the observed transition or transversion mutation. These results show that the MBP-MutS-H6 error filter was most effective at removing insertion/deletion loops and G-T/A-C mismatches from the population. It should be possible to generalize the consensus shuffling protocol to a large number of synthetic DNA constructs. GFPuv was chosen as the synthetic construct in this study for its advantages as a fluorescent reporter gene. This allowed easy optimization of our protocol without the need to sequence thousands of base pairs of DNA. We expect the results reported here for consensus shuffling to readily translate to synthetic DNA constructs of varied sequence, greater overall length and/ or higher initial errors/kb. Synthetic DNA constructs of varied sequence can be digested into a defined set of fragments using Type II restriction enzymes or fragmented into any desired size range using controlled DNase I digestion (26) . Digestion and reassembly of a large number of different genes is expected to be as robust as the protocol of DNA shuffling (28) , which has been broadly applied to a variety of gene sequences. Synthetic DNA constructs larger than GFPuv are expected to be amenable to error correction by consensus shuffling, as the error filtering is conducted on gene fragments before reassembly of the full-length gene. Thus, the errors/kb data presented in this study are expected to translate to larger genes with similar initial errors/kb (excepting mutations introduced by PCR amplification following the final application of the error filter). Synthetic DNA constructs of higher initial errors/kb are expected to be amenable for error correction by consensus shuffling. However, these constructs will require digestion into smaller sized gene fragments that may affect the efficiency of error correction. In contrast to consensus shuffling, an increase in the size of the synthetic DNA product or an increase in errors/kb would preclude the use of the coincidence filtering protocol, as every molecule in the population would contain one or more errors. As proof of the utility of the consensus shuffling protocol, 10 non-fluorescent GFPuv clones containing one or more errors (Figure 4 ) were converted into a population where 82% of the clones were fluorescent. It is important to note that DNA shuffling alone shows an improvement in percent fluorescent colonies in this example (from 0 to 30%). For synthetic DNA populations, DNA shuffling alone shows no improvement in percent fluorescent colonies (see Figure 3 'no MutS' treatments). DNA shuffling alone improves the overall number of correct sequences only for small DNA populations with low error rates. For example, when shuffling 10 clones with a unique mutation in each clone, one would expect the fraction of correct products to be (9/10) 10 = 35% (26), very close to the value of 30% that we observed. A mathematical model describing the error rates for shuffling and error filtering of synthetic DNA populations is presented below. To estimate some parameters of consensus shuffling and coincidence filtering, a simple mathematical model (Equations 1-6) was constructed. An input population of dsDNA molecules of length N, containing E errors/base is re-hybridized, fragmented into shorter dsDNA fragments of average length S, error filtered and reassembled. P(F) is the probability a fragment of length S will have a correct sequence. We determine the probability that re-hybridized duplexes will have zero (C), one (H ) or both (I ) strands with errors. Equation 5 estimates the probability that a fragment will be correct after a cycle of MutS filtering, P(F 0 ), by applying a MutS selectivity factor (M ) to adjust the relative amounts of mismatch containing duplexes (I, H ) while accounting for the total fraction of correct strands in the re-hybridized duplexes. The probability of obtaining an error free assembly product, P(A), is then given by Equation 6 . From our consensus shuffling error rate data (Figure 3 ), we estimate the MutS selectivity factor M to be $2.2. Figure 5 shows some predictions that emerge from this model assuming typical length (2 kb), fragment sizes (200 bp) and error rates (1.8 errors/kb). Consensus shuffling is predicted to be most effective with smaller fragment sizes ( Figure 5A ). As mentioned above, smaller fragment sizes could be obtained by controlled digestion with DNase I (21) . In addition, multiple iterations of MutS filtering can have dramatic results on populations with few correct sequences ( Figure 5B ), although the model does not account for the differing specificity of MutS toward the various types of mismatches. The model also predicts that even modest improvements in the MutS selectivity factor through optimization of the MutS-DNA binding conditions and/or the use of a combination of MutS homologs with varying mismatch specificity (29) could dramatically improve the consensus shuffling protocol ( Figure 5C ). Coincidence filtering (N = S) is predicted to be effective for populations with low error rates per clone ( Figure 5D ) but becomes ineffective when the majority of re-hybridized duplexes contain mismatches. We have demonstrated consensus shuffling and coincidence filtering as experimental methods to significantly reduce errors in synthetic DNA. Consensus shuffling should be generally applicable for error correction on synthetic genes of typical lengths and error rates. Two iterations of consensus shuffling ($6 h/iteration) generated a population with $1 error/3500 bp. This reduction in error rate will allow the identification of a correct clone after sequencing DNA from a reduced number of colonies. Coincidence filtering is a simple and effective procedure to reduce errors in synthetic DNA populations with low error rates per clone. These methods should significantly increase the speed and decrease the cost of production of synthetic genes. Note: While this manuscript was under review, Carr et al. (30) independently reported the application of Taq MutS in protocols for error reduction on synthetic DNA.
28
Towards standardization of RNA quality assessment using user-independent classifiers of microcapillary electrophoresis traces
While it is universally accepted that intact RNA constitutes the best representation of the steady-state of transcription, there is no gold standard to define RNA quality prior to gene expression analysis. In this report, we evaluated the reliability of conventional methods for RNA quality assessment including UV spectroscopy and 28S:18S area ratios, and demonstrated their inconsistency. We then used two new freely available classifiers, the Degradometer and RIN systems, to produce user-independent RNA quality metrics, based on analysis of microcapillary electrophoresis traces. Both provided highly informative and valuable data and the results were found highly correlated, while the RIN system gave more reliable data. The relevance of the RNA quality metrics for assessment of gene expression differences was tested by Q-PCR, revealing a significant decline of the relative expression of genes in RNA samples of disparate quality, while samples of similar, even poor integrity were found highly comparable. We discuss the consequences of these observations to minimize artifactual detection of false positive and negative differential expression due to RNA integrity differences, and propose a scheme for the development of a standard operational procedure, with optional registration of RNA integrity metrics in public repositories of gene expression data.
Purity and integrity of RNA are critical elements for the overall success of RNA-based analyses, including gene expression profiling methods to assess the expression levels of thousands of genes in a single assay. Starting with low quality RNA may strongly compromise the results of downstream applications which are often labor-intensive, time-consuming and highly expensive. However, in spite of the need for standardization of RNA sample quality control, presently there is no real consensus on the best classification criteria. Conventional methods are often not sensitive enough, not specific for single-stranded RNA, and susceptible to interferences from contaminants present in the sample. For instance, when using a spectrophotometer, a ratio of absorbances at 260 and 280 nm (A 260 :A 280 ) greater than 1.8 is usually considered an acceptable indicator of RNA purity (1, 2) . However, the A 260 measurement can be compromised by the presence of genomic DNA leading to over-estimation of the actual RNA concentration. On the other hand, the A 280 measurement will estimate the presence of protein but provide no hint on possible residual organic contaminants, considered at 230 nm (3) (4) (5) . Pure RNA will have A 260 :A 230 equal to A 260 :A 280 and >1.8 (1) . A second check involves electrophoresis analysis, routinely performed using agarose gel electrophoresis, with RNA either stained with ethidium bromide (EtBr) (6) (7) (8) (9) , or the more sensitive SYBR Green dye (10) . The proportion of the ribosomal bands (28S:18S) has conventionally been viewed as the primary indicator of RNA integrity, with a ratio of 2.0 considered to be typical of 'high quality' intact RNA (1) . However, these methods are highly sample-consuming, using 0.5-2 mg total RNA and often not sensitive enough to detect slight RNA degradation. Today, microfluidic capillary electrophoresis with the Agilent 2100 bioanalyzer (Agilent Technologies, USA) has become widely used, particularly in the gene expression profiling platforms (11, 12) . It requires only a very small amount of RNA sample (as low as 200 pg), the use of a size standard during electrophoresis allows the estimation of sizes of RNA bands and the measurement appears relatively unaffected by contaminants. Integrity of *To whom correspondence should be addressed. Tel: The online version of this article has been published under an open access model. Users are entitled to use, reproduce, disseminate, or display the open access version of this article for non-commercial purposes provided that: the original authorship is properly and fully attributed; the Journal and Oxford University Press are attributed as the original place of publication with the correct citation details given; if an article is subsequently reproduced or disseminated not in its entirety but only in part or as a derivative work this must be clearly indicated. For commercial re-use, please contact journals.permissions@oupjournals.org the RNA may be assessed by visualization of the 28S and 18S ribosomal RNA bands ( Figure 1A and B); an elevated threshold baseline and a decreased 28S:18S ratio, both are indicative of degradation. A broad band shows DNA contamination ( Figure 1C ). As it is apparent from a review of the literature, the standard of a 2.0 rRNA ratio is difficult to meet, especially for RNA derived from clinical samples, and it now appears that the relationship between the rRNA profile and mRNA integrity is somewhat unclear (13) (14) (15) (16) (17) (18) (19) (20) (21) (22) (23) . On the one hand, this may reflect unspecific damage to the RNA, including sample mishandling, postmortem degradation, massive apoptosis or necrosis, but it can reflect specific regulatory processes or external factors within the living cells. Altogether, it appears that total RNA with lower rRNA ratios is not necessarily of poor quality especially if no degradation products can be observed in the electrophoretic trace ( Figure 1D ). For all these reasons, the development of a reliable, fully integrated and automated system appropriate for numeric evaluation of RNA integrity is highly desirable. Standardized RNA quality assessment would allow a more reliable comparison of experiments and facilitate exchange of biological information within the scientific community. With that prospect in mind, and with the aim of anticipating future standards by pre-normative research, we identified and tested two software packages recently developed to gauge the integrity of RNA samples with a user-independent strategy: one open source, the degradometer software for calculation of the degradation factor and 'true' 28S:18S ratio based on peak heights (24) and the freely available RIN algorithm of the Agilent 2100 expert software, based on computation of a 'RNA Integrity Number' (RIN) (25) . Both tools were developed separately to extract information about RNA integrity from microcapillary electrophoretic traces and produce a userindependent metrics. Using these tools, we assessed the purity and integrity of 414 RNA samples, derived from 14 different human adult tissues and cell lines, many of which representing tumors. Those results were compared with conventional RNA quality measurement approaches as well as with highly expert human interpretation. We evaluated the simplicity for users and examined the potential, accuracy and efficiency of each method to contribute to standardization of RNA integrity assessment upstream of biological assays. These procedures were further validated by real-time RT-PCR quantitation of the expression levels of three housekeeping genes, using the same RNA samples, at different levels of degradation. Total RNA was prepared from human cell lines (especially from the ATCC bio-resource center, N = 50) and tissue samples (clinical samples, N = 285) from 13 different human adult tissue types, i.e. blood, brain, breast, colon, epithelium, kidney, lymphoma, lung, liver, muscle, prostate, rectum and thyroid. RNA purification was performed by cesium chloride ultracentrifugation according to Chomczynski and Sacchi (26) , by phenol-based extraction methods (TRIzol reagent, Invitrogen, USA), or silica gel-based purification methods (RNeasy Mini Kit, Qiagen, Germany; Strataprep kit, Stratagene, USA or SV RNA isolation kit, Promega, USA) according to the manufacturer's instructions with some modifications. Material was maintained at À80 C with minimal handling. RNA extraction was carried out in an RNase-free environment (see Supplementary Table 1 online) . The commercially available RNA samples were the 'Universal Human Reference' (N = 75) distributed by Stratagene (USA), and human brain (N = 2) and muscle (N = 2) RNAs supplied by Clontech (USA). Once extracted, RNA concentration and purity was first verified by UV measurement, using the Ultrospec3100 pro (Amersham Biosciences, USA) and 5 mm cuvettes. The absorbance (A) spectra were measured from 200 to 340 nm. A 230 , A 260 and A 280 were determined. A 260 :A 280 and A 260 :A 230 ratios were calculated. For microcapillary electrophoresis measurements, the Agilent 2100 bioanalyzer (Agilent Technologies, USA) was used in conjunction with the RNA 6000 Nano and the RNA 6000 Pico LabChip kits. In total, 39 assays were run in accordance with the manufacturer's instructions (see Supplementary Notes online). To evaluate the reliability of the classifier systems described in this study, replicate runs were done on a set of 56 RNA samples loaded on different chips, resulting in 2 (N = 41), 3 (N = 12), 7 (N = 2) and 50 (N = 1) data points per sample. Human RNA integrity categorization RNA integrity checking was performed by expert operators who classified each total RNA sample within a predefined discrete category from 1 to 5, examining the integrity of the RNA from electropherograms (see Supplementary Table 2 online). A low number indicates high integrity. Reference criteria parameters include ribosomal peaks definition, baseline flatness, existence of additional or noise peaks between ribosomal peaks, low molecular weight species contamination and genomic DNA presence suspicion. A smearing of either 28S and 18S peaks, or a decrease in their intensity ratio indicate degradation of the RNA sample and results in the classification into the higher categories. To evaluate the robustness of this human interpretation, five highly experienced operators, trained in these cataloging steps, separately classified a subset of 33 samples from breast cancers. It included samples with varying levels of integrity: intact RNA (33%), low quality samples (20%) and a wide range of degradation (47%). Bioanalyzer electrophoretic data were exported in the degradometer software folder (.cld format). For comparison of samples, the original data were re-scaled by the classifier system, first along the time-axis to compensate for differences in migration time, then along the fluorescence intensity-axis to compensate for variation in total RNA amount. As a result, fluorescence curves that have the same shape will have the same peak heights after re-scaling. Then, Degradation Factors (DegFact) and corrected 28S:18S ratios were calculated (see Supplementary Table 3 online) using the mathematical model developed by Auer et al. (24) , examining additional 'degradation peak signals' appearing in the lower molecular weight range and comparing them to ribosomal peak heights. Calculation of the DegFact is based on a numbering of continuous metrics, ranging from 1 to ¥; increasing DegFact values correspond to more degradation, and a new group of integrity is defined after 8 graduation steps. Once the classification of the RNA samples is completed, 4 groups of integrity are displayed, 3 showing an alert warning indicative of some measurable degradation (Yellow: 8-16, Orange: 16-24 and Red: >24), while all non-reliable data come together and form the fourth group (Black). We introduced a fifth class labeled White (<8), when no alert was produced by the software. Software and manual are freely available at http://www. dnaarrays.org/downloads.php. Degradometer version 1.4.1 (released in May 2004) of the software was used. Bioanalyzer electrophoretic sizing files (.cld format) collected with biosizing software version A.02.12.SI292 (released in March 2003) were imported in the Agilent 2100 expert software (RIN beta release). The RIN algorithm allows calculation of RNA integrity using a trained artificial neural network based on the determination of the most informative features that can be extracted from the electrophoretic traces out of 100 features identified through signal analysis. The selected features which collectively catch the most information about the integrity levels include the total RNA ratio (ratio of area of ribosomal bands to total area of the electropherogram), the height of the 18S peak, the fast area ratio (ratio of the area in the fast region to the total area of the electropherogram) and the height of the lower marker. A total of 1300 electropherograms of RNA samples from various tissues of three mammalian species (human, mouse and rat), showing varying levels of degradation and an adaptive learning approach were used in order to assign a weight factor to the relevant features that describe the RNA integrity. A RIN number is computed for each RNA profile (see Supplementary Table 4 online) resulting in the classification of RNA samples in 10 numerically predefined categories of integrity. The output RIN is a decimal or integer number in the range of 1-10: a RIN of 1 is returned for a completely degraded RNA samples whereas a RIN of 10 is achieved for intact RNA sample. In some cases, the measured electropherogram signals are of an unusual shape, showing for example peaks at unexpected migration times, spikes or abnormal fluctuation of the baseline. In such cases, a reliable RIN computation is not possible. Several separate neural networks were trained to recognize such anomalies and display a warning to the user or even suppress the display of a RIN number. Combining the results of the neural network for the RIN computation and the neural networks to detect anomalies, the RIN algorithm achieves a mean square error of 0.1 and a mean absolute error of 0.25 on an independent test set. The beta release of the software and manual are freely available at http://www.agilent.com/chem/RIN. Agilent 2100 expert version B.01.03.SI144 (released in November 2003) of the software was used. Expression levels of three housekeeping genes (HKG)-GAPD, GUSB and TFRC-were measured by quantitative PCR using the TaqMan Gene Expression Assays according to the manufacturer's instructions (Applied Biosystems, USA). Sixteen aliquots of a unique batch of RNA sample (Universal Human Reference RNA, Stratagene, USA) of various levels of integrity (cf. Table 1 ) were used to test the influence of RNA quality on the relative expression of those three genes. In parallel, a 5 0 to 3 0 comparison was done using two separate GUSB and TFRC TaqMan probes. An homogeneous quantity (0.8-1 mg) of the RNA samples was subjected to a reverse transcription step using the highcapacity cDNA archive kit (Applied Biosystems, USA) as described by the manufacturer. Single-stranded cDNA products were then analyzed by real-time PCR using the TaqMan Gene Expression Assays according to the manufacturer's instructions (Applied Biosystems, USA). Single-stranded cDNA products were analyzed using the ABI PRISM 7700 Sequence Detector (Applied Biosystems, USA). The efficiency and reproducibility of the reverse transcription were tested using 18S rRNA TaqMan probes. Five assays were used, GAPDH-5 0 (Hs99999905_m1), GUSB-5 0 (Hs00388632_gH), GUSB-3 0 (Hs99999908_m1), TFRC-5 0 (Hs00951086_m1) and TFRC-3 0 (Hs00951085_m1). In each case, duplicate threshold cycle (Ct) values were obtained and averaged; then expression levels were evaluated by a relative quantification method (27) . The fold change in one tested HKG (target gene) was normalized to the 18S rRNA (reference gene) and compared to the highest quality sample (calibrator sample), using the following formula: Fold change = 2 ÀDDCt , where DDCt = (C t-target À C t-reference ) sample-n À (C t-target À C t-reference ) calibrator-sample . Sample-n corresponds to any sample for the target gene normalized to the reference gene and calibrator-sample represents the expression level (1·) of the target gene normalized to the reference gene considering the highest quality sample. Mean 2 ÀDDCt and SD were calculated, considering the samples either individually or grouped by quality metrics categories, based on RIN metrics or DegFact values, together with the lower and upper bound mean of 95% Intervals of Confidence (IC). Using this analysis, if the expression levels of the HKG are not affected by the RNA degradation, the values of the mean fold change at each condition should be very close to 1 (since 2 0 = 1) (27) . Descriptive statistics were executed using the XLSTAT software, version 7.1 (Addinsoft, USA), P = 0.05. Mean, SD and coefficient of variation (variation or CV) between and within groups of samples were calculated, together with a measure of the dispersion (range), inter-quartile range (1st and 3rd quartiles, Q1-Q3) and evaluation of the lower and upper bound mean of 95% Interval of Confidence (IC). Comparative statistical analyses between groups were completed, P = 0.05, using non-parametric statistical tests: two-independent Mann-Whitney U-test and k-independent Kruskal-Wallis test. We analyzed 414 total RNA sample profiles from various human tissues (69%) and cell lines (31%) of either tumoral (85%) or normal (15%) origin, with varying levels of RNA integrity. Supplementary Table 1 online for details). Significant differences in A 260 :A 280 ratios were observed between specific groups of samples (i.e. tumoral versus normal or tissues versus cell lines). For instance, RNA extracted from normal samples displayed an improved ratio of 1.97, with 97% falling within the desired range ( Figure 2A ). In contrast, the distribution of A 260 :A 280 ratios was not found to correlate with either purification methods or tissues of origin. RNA integrity was further assessed by resolving the 28S and 18S ribosomal RNA bands using the Agilent 2100 bioanalyzer and the RNA 6000 protocol. The analysis was done on 399 RNA profiles; data from 15 samples was not obtained due to device problems during the runs. The system automatically provided 28S:18S ratios for 348 (87%) of the 399 profiles. Figure 2B shows the distribution of the 28S:18S computed values, with a median ratio around 1.7 and a variation of 54% from the mean (IC 1.9-2.1 and Q1-Q3 1.4-2.5). In addition, a significant degree of variability of the 28S:18S ratio (19-24%) was found for identical samples from replicate runs (2-50 times). Among those RNA samples, 28S:18S ratios of 2.0 or greater were rare, less than 44% of the values measured being within the theoretically desired range, except for the samples prepared from cultured cells ( Figure 2B ). The integration failed in the remaining 51 cases, displaying an atypical migration, with no clear 28S and 18S rRNA bands, and no 28S:18S ratio was computed (data not shown). Expert operators categorized the set of RNA samples by inspecting the electrophoretic traces of successful assays. Over the 399 RNA profiles checked, 379 (95%) were scored within predefined categories ( Figure 2C ), namely good [Human Categorization (HC)-level 1], regular (HC-level 2), moderate (HC-level 3), low (HC-level 4) and degraded (HC-level 5). The remaining 20 (5%) were flagged as displaying a temperature-sensitive profile: RNA samples initially found intact became highly degraded when heated, although no RNase contamination was observed (data not shown). Estimation of the robustness of this cataloging was done through comparison of qualifying criteria using a set of 33 breast cancer samples (see Materials and Methods). Integrity of the samples was evaluated independently by five expert operators, and categorization was found highly reliable with a coefficient variation (CV) $16%. This is low considering that individual interpretation is involved, but can be explained by the fact that very experienced operators accomplished the scoring based on a clearly defined set of instructions, thus limiting frequently observed subjective visual interpretation and inconsistency of human categorization. Predictably, a 28S:18S ratio of 2.0 denoted high quality for a majority of RNA samples, 91% being classified in HC-levels 1 to 3. However, 83% of total RNAs with 28S:18S > 1.0 but a low baseline between the 18S and 5S rRNA or front marker were also classified in HC-levels 1-3 (see Figure 1D ) and could be considered suitable for most downstream applications. RNA degradation was first assessed using the degradometer software (see Materials and Methods). Over the 399 RNA profiles checked, all were scored in one of the five predefined classes ( Figure 3A) . Altogether, 334 (84%) Degradation Factors (DegFact) values were computed, the remaining 65 RNA samples (16%) displaying profiles that could not be interpreted reliably; no DegFact values could be scored, and samples were flagged in the Black category ( Figure 3A ). Most of them (80%) correspond to samples previously classified by our operators as degraded (HC-level 5). The remaining cases had an average degradation factor of 7.5 (IC 6.7-8.3) with large variations over the entire set of samples (over 103% from the mean, range 1-52). A lower variability was persistently found when identical samples from replicate runs were considered, resulting in observed DegFact values with a 26-32% CV. In addition, statistically significant differences were found between DegFact values of samples sorted by types. The highest DegFact values were found characteristic of tissue samples, 41% of them displaying a DegFact > 8, as compared with 6% for the cell lines (data not shown). Remarkably, we found a significant linear relationship between the DegFact values distribution and the explicit human categorization. Most HC classes corresponded to an unambiguous DegFact distribution ( Figure 3B ), while HClevels 2 and 3 form a single class: HC-level 1, mean DegFact of 3.3, SD of 2.8 (IC 2.8-3.7); HC-level 2 and 3, mean Deg-Fact of 8.8, SD of 6.8 (IC 7.5-10.2); HC-level 4, mean DegFact of 15.9, SD of 7.8 (IC 12.7-19.1); HC-level 5, mean DegFact of 26.0, SD of 7.5 (IC 21.9-30.1). It is worth mentioning that the normalized heights of 18S and 28S peaks, and the interval between them after rescaling gradually decrease and then reverse with increasing degradation ( Figure 3B ). Integrity of RNA samples was measured in parallel based on the RNA Integrity Number metrics using an artificial neural network trained to distinguish between different RNA integrity levels by examining the shape of the microcapillary electrophoretic traces (see Materials and Methods). Over the 399 RNA profiles checked, 363 (91%) were scored successfully ( Figure 4A) , with an average RIN of 7.7 (IC 7.4-8.0). The remaining 36 (9%) samples were associated with various unexpected signals, disturbing computation of the RIN using default anomaly detection parameters. In each case, a flag alert was added corresponding to critical anomalies including unexpected data in sample type, (or) ribosomal ratio, (or) baseline and signal in the 5S region (data not shown). RIN categorization was found regular, variability between replicate runs, compared to the other methods, being consistently very small (CV 8-12%). As expected, the highest RIN were characteristic of cell line samples, 72% of them displaying a RIN > 9, as compared with 47% for the tissue samples (data not shown). A first group, corresponding to 295 (82%) of the 363 RNA profiles, was analyzed using the default settings of the RIN system, but with a lower threshold of RNA quantity loaded (20 ng) for reliable detection of anomalies than that recommended by the manufacturer (50 ng). A significant linear relationship was found between the RIN number and both the explicit human classification provided by our operators, Figure 3 . RNA degradation characterization. Integrity of 399 RNA sample profiles was scored using the degradometer software. (A) A total of 334 RNA profiles were successfully categorized into 5 predefined alert classes using a mathematical model that quantifies RNA degradation and computes a degradation factor (DegFact). Four classes (White, Yellow, Orange and Red) are associated with different levels of degradation. A fifth class, Black alert corresponds to samples that the system was not able to qualify with accuracy (n.d.). The distribution is represented by the number of records in each class. (B) Comparative analysis was done using human evaluation (x-axis) based on electrophoresis analysis as a reference for RNA integrity classification; observations of rRNA peak heights and DegFact values were taken at each of the 5 HC levels. Histograms refer to the mean 28S and 18S rRNA peak heights and 95% confidence intervals (fluorescence intensities; left scale). Mean DegFact values and 95% confidence intervals (arbitrary unit, right scale) are plotted with the means joined. and the DegFact values calculated by the degradometer software ( Figure 4B ). Each distinct HC class corresponds to an explicit RIN number, with HC-levels 2 and 3 forming once again a single class: HC-level 1, mean RIN of 9.6, SD of 0.7 (IC 9.5-9.7); HC-level 2 and 3, mean RIN of 8.6, SD of 0.9 (IC 8.4-8.9); HC-level 4, mean RIN of 6.1, SD of 1.5 (IC 5.2-7.1); HC-level 5, mean RIN of 3.7, SD of 2.0 (IC 2.9-4.5). For the remaining 68 samples (assay done with <20 ng of RNA), two separate groups were considered: 41 samples with a computed RIN below 5.0, and 27 above 7.0. All samples in the first group were derived from RNA 6000 Nano assays, with mean RNA quantities loaded below 10 ng (Q1-Q3, 5-12 ng), i.e. below the lower limit of quantitation indicated by the manufacturer. All but 8 of these samples were estimated by our operators to be of poor quality (HC-level 4; N = 3) or degraded (HC-level 5; N = 30), and all but 4 were flagged Black by the degradometer software and no DegFact values were scored. These RNA profiles could not be interpreted reliably, possibly due to either the low RNA concentration or the unusual migration behavior and shifted baseline values of degraded samples. Thus, the two automated systems were in disagreement for these samples; while human interpretation was in most cases in agreement with the RIN system, with less than 20% of inconsistency. In the second group of 27 samples, 20 of the profiles were derived from RNA 6000 Pico assays with RNA quantities loaded being on average below 4 ng (Q1-Q3, 0.5-0.8 ng), which is within the manufacturer specifications. All but 3 of them were estimated by our operators to range from high (HC-level 1; N = 12) to correct (HC-level 2 and 3; N = 12) quality levels. In addition, all RNA profiles except 1 were scored by the degradometer software, most of them displaying an alert flag (N = 20); some slight degradation was detected, associated to a low mean DegFact value of 9.7 (IC 8.1-11.3; Q1-Q3, 6.2-12.6). Thus, both automated systems and human interpretations agreed in most of these cases, with <11% of inconsistency. The influence of RNA quality categorization obtained with both user-independent classifiers on gene expression profiling was explored using real-time RT-PCR. The expression levels of three housekeeping genes (HKG)-GAPDH, GUSB and TFRC-were measured in 16 aliquots of a unique RNA displaying various integrity metrics ( Table 1 ). The mean correlation coefficient (r) between the threshold cycle (Ct) among the 16 samples and both quality metrics was found high: r = À0.87 considering the RIN metrics and r = 0.85 considering the DegFact values. The values of the mean fold changes, calculated according to the 2 ÀDDCt quantification method (see Materials and Methods), were found lower than 1.0, corresponding to the expression level (1·) in the sample exhibiting the highest RNA quality (Table 2 and Figure 5 ). Considering that HKG expression was measured relative to the reference sample, an obvious decline of the relative expression levels was observed, up to 24, 70 and 82%, in samples categorized according to the RIN metrics ( Figure 5A) and DegFact values ( Figure 5B ). These results indicate that 2-to 7-fold differences may be expected in the relative expression levels of genes in samples that differ only by their quality (Table 2 ). These fold differences are much larger than those measured for RNA samples of comparable integrity, consistently lower than 1.6 (Table 2 and Figure 5 ). In addition, an unambiguous gap in the distribution may be defined ( Figure 5A and B) , distinguishing the RNA samples of the higher quality categories (RIN > 8 and DegFact values < 7) from those of the lower categories (RIN < 8 and DegFact values > 12). It would be expected that measuring expression of an intact mRNA would yield approximately equal results regardless of the region being probed, and if mRNA fragmentation had occurred, then some sequences may be more abundant than others. We thus tested the effect of PCR probe location on the RNAs. The 5 0 and 3 0 GUSB probes, separated by 1209 nt, were associated with highly similar threshold cycle (Ct) measures (r = 0.98, b parameter = 0.88) ( Figure 5C ). Similar results were obtained for TFRC, with probes separated by 2066 nt (r = 0.84, b parameter = 0.92, data not shown). It seems therefore that the region being probed is not a source of variation in our results. It is universally accepted that RNA purity and integrity are of foremost importance to ensure reliability and reproducibility of downstream applications. In the biomedical literature (PubMed, November 2004), from the 485 090 articles that relate to RNA, and the 287 515 or 40 395 including respectively the 'quality' or 'integrity' term, less than 100 were found to contain 'RNA quality' or 'RNA integrity' terms. Interestingly, half of them were published between 2001 and 2004; but none is proposing a standard operational procedure for RNA quality assessment to the scientific community. Except for two studies (24, 25) , those reports are based on 10 to 15 years old methods (1), indicating that they represent the established and currently mostly used methods. Our results strongly challenge the reliability and usefulness of those conventional methods, demonstrating their inconsistency to evaluate RNA quality. First, the A 260 :A 280 and A 260 :A 230 ratios are reflecting RNA purity, but are not informative regarding the integrity of the RNA. Available RNA extraction and purification methods yield highly pure RNA with very little DNA or other contaminations, resulting most often in both ratios )1.8, although 18% of the samples were found degraded and 7% more of poor quality. The high A 260 :A 280 ratios are indicative of limited protein contaminations, whereas high A 260 :A 230 ratios are indicative of an absence of residual contamination by organic compounds such as phenol, sugar or alcohol, which could be highly detrimental to downstream applications. Nonetheless, samples displaying low A 260 :A 230 ratios ((1.8) did not exhibit any inhibition during downstream applications, such as cDNA synthesis and labeling or in vitro transcription (data not shown). Second, due to a lack of reliability, the 28S:18S rRNA ratios may not be used as a gold standard for assessing RNA integrity. When ribosomal ratios were calculated from identical samples but through independent runs, a large degree of variability (CV 19-24%) was observed. Moreover, using the biosizing software, we found 28S:18S rRNA ratios evaluation compromised by the fact that their calculation is based on area measurements and therefore heavily dependent on definition of start and end points of peaks. In 13% of the cases, the system was unable to localize the ribosomal peaks, and therefore no 28S:18S ratios were computed. For the remaining samples, no clear correlation between 28S:18S ratios and RNA integrity was found although RNAs with 28S:18S >2.0 were usually of high quality. Most of the RNAs we studied (83%), displaying a 28S:18S > 1.0, could be considered of good quality. Interestingly, Auer et al. (24) in a study on 19 tissues from seven organisms, reported that an objective measurement of the RNA integrity may possibly be done through comparison of re-scaled 28S and 18S peak heights, but not of the corresponding areas. Actually, we observed a linear relationship between RNA integrity and differences in normalized 28S and 18S peak heights. Increased degradation resulted in a significant decrease in the scaled corrected heights of the ribosomal peaks, with inversion of the ratio at the highly degraded stages (cf. Figure 3B ). In comparison to the area computation, 28S:18S rRNA re-scaled peak height measurement produced more consistent values, with a CV reduced to 12-14%, and displayed clear concentration-independent values (see Supplementary Tables 1 and 3 online) . Human evaluation of the integrity of RNA through visual inspection of the electrophoresis profiles provided very consistent data. Variability between classifications produced by five independent expert operators (CV 16%) was lower than with automated management of more conventional control 28S:18S area values (CV 19-24%). It is, however, very time-consuming and strongly dependent on individual competence. Even with highly trained specialists, 5% of the set of RNA samples could not be allocated to any of the five predefined categories; their corresponding profiles were considered by our experts as atypical, displaying a temperature-sensitive shape (data not shown). These strategies appear unsuitable for standardization and quality control of RNA integrity assessment, which require simple but consistent expert-independent classification, facilitating information exchanges between laboratories. The N-value corresponds to the number of samples by category. The mean quality metrics, i.e. RIN and DegFact and the mean fold change (2 ÀDDCt ) relative to the reference sample are indicated, together with the 95% confidence intervals. Observed technical variation (IC-rep, P = 0.05) is also specified, considering duplicate (two per gene per target sample) and replicate (six per gene per calibrator sample) measures. The reference sample exhibits a RIN of 9, a DegFact value of 4.9 and by default mean fold change set to 1. The observed decrease in the expression (relative expression, %) relative to the reference sample is calculated. The fold differences refer to the fold-ratios that are expected in the expression levels for a gene, across categories (between categories), given that the samples only differ by their quality, and within each category (within categories), considering RNA of comparable integrity. The fold-ratios (technical variation) that may be expected by chance in the gene expression levels, P = 0.05, from some technical reasons, are also considered. We therefore investigated the performance of two recently developed user-independent software algorithms (24, 25) . The degradometer software provided a reliable evaluation of RNA integrity based on the identification of additional 'degradation peak signals' and their integration in a mathematical calculation together with the ribosomal peak heights. It allowed characterization of the integrity of 84% of the samples tested, one-third with an alert flag, which was first found to be fairly informative, as it strongly reduces the complexity of the metrics by introducing three distinct classes labeled Yellow, Orange and Red, and can be used as a first straightforward simple filtering step. However, degradation factors (DegFact) metrics yield precise measures with less than 32% CV and are much more valuable than flag alerts for the purpose of standardization. The same is true for the RNA Integrity Number 'RIN' software which allowed the characterization of the integrity of 91% of the RNA samples tested, with a RIN value for 363 RNA sample profiles with less than 12% CV. In general, there was a good agreement between the human classification, the degradation factor and the RIN (see Figure 4B ). This provided a cross-validation of the user-independent qualification systems tested. Both resulted in the refinement of human interpretations, validating four statistically relevant classes of samples, namely good (HC-level 1), regular/ moderate (HC-level 2 and 3), poor (HC-level 4) and degraded (HC-level 5). Moreover, the 5% RNA samples previously flagged by the operators as displaying an atypical temperature-sensitive shape were unambiguously assigned to one or the other category of samples [RIN = 7.3 (IC 6.8-7.8); DegFact = 11.9 (IC 9.5-14.2); data not shown]. Altogether, we found the degradometer and RIN algorithms to be highly reliable user-independent methods for automated assessment of RNA degradation and integrity. The RIN system is a slightly more informative tool, able to compute assessment metrics for 91% of the RNA profiles, compared to 84% with the degradometer software; the remaining being flagged respectively as N/A or Black alert. For samples available below a low limit of 20 ng (N = 80) the RIN system provided Figure 6 . Workflow of operational procedure for RNA quality assessment. Integrity of the RNA, once extracted and purified from cell lines, clinical or biological tissues samples, is controlled from the widely used bioanalyzer electrophoretic traces. As standard part of the Agilent analysis software (25), a RIN metrics is first calculated, scoring each RNA sample into 10 numerically predefined categories of integrity (RIN, from 1 to 10; N is a threshold value). As an independent control, a degradation factor metrics (DegFact, from 1 to ¥; N 0 is a threshold value) may optionally be allocated to each RNA sample using the bioanalyzer-independent degradometer software (24) . In a standard operating procedure, RIN and/or DegFact metrics will first be used as a standard exchange language to document RNA integrity and degradation, second to classify the RNA in homogeneous groups, and finally to select samples of comparable RNA integrity to improve the scheme of meaningful downstream experiments. The standard operating procedure will benefit from feedback information that will help users to define threshold integrity metrics values based on the results of RNA-based analyses. metric values for 85% of them, compared to only 46% with the degradometer software. Similarly, the RIN system was able to provide metric values for 81% of poor quality samples (including low quality and degraded samples; N = 96), whereas the degradometer software could classify only 44% of them. Another advantage with the RIN classifier is that, if there are critical anomalies detected (including genomic DNA contamination, wavy baseline, etc.), threshold settings may be changed and a reliable RIN value computed. This was the case for 25 of the 363 RNA sample profiles successfully classified by the system. While intact RNA obviously constitutes the best representation of the natural state of the transcriptome, there are situations in which gene expression analysis may be desirable even on partially degraded RNA. Some studies report collection of reasonable microarray data from RNA samples of impaired quality (28) , leading to meaningful results if used carefully. Moreover, Auer et al. (24) recently concluded that degradation does not preclude microarray analysis if comparison is done using samples of comparable RNA integrity. We confirmed the direct influence of the RNA quality on the distribution of gene expression levels, by detecting using Q-PCR a significant (up to 7-fold) difference in the relative expression of genes in samples of slightly decreased RNA integrity, which is much larger than the variation within comparable RNA quality categories (cf. Figure 5 and Table 2 ). This may correlate with ratio discrepancies in gene expression experiments, and therefore with false positive and false negative rates of differential gene expression when comparing two samples. Therefore, computing reliable metrics of RNA integrity, even if the RNA is found to be partially degraded, may be highly valuable. The straight and unambiguous relationships established between human interpretations and both RIN and DegFact distributions indicates that, using these metrics, it should be possible to distinguish specific samples that are too disparate to be included in comparative gene expression analyses without compromising the results. Although the information provided by these user-independent classifiers is not a guarantee for successful downstream experiments, it gives a more comprehensive picture of the samples and can be used as a safeguard against performing useless and costly experiments. Thus, the RIN system may be used as simple metrics that can be easily integrated in any sample tracking information system for definition of standard operating procedures under quality assurance following a scheme such as the one described in Figure 6 . In this context, we suggest that the growing number of laboratories performing RNA Quality Control by microcapillary electrophoresis should be offered the option to report objective RNA quality metrics as part of the 'Minimum Information About a Microarray Experiment' MIAME standards (29) . Through registration of RNA profiles in a public electronic repository, such standardized information should enable and facilitate comparisons of RNA-based bioassays performed across laboratories with RNA samples of similar quality, in much the same way as sequencing traces are compared.
29
Factors affecting translation at the programmed −1 ribosomal frameshifting site of Cocksfoot mottle virus RNA in vivo
The ratio between proteins P27 and replicase of Cocksfoot mottle virus (CfMV) is regulated via a −1 programmed ribosomal frameshift (−1 PRF). A minimal frameshift signal with a slippery U UUA AAC heptamer and a downstream stem–loop structure was inserted into a dual reporter vector and directed −1 PRF with an efficiency of 14.4 ± 1.9% in yeast and 2.4 ± 0.7% in bacteria. P27-encoding CfMV sequence flanking the minimal frameshift signal caused ∼2-fold increase in the −1 PRF efficiencies both in yeast and in bacteria. In addition to the expected fusion proteins, termination products ending putatively at the frameshift site were found in yeast cells. We propose that the amount of premature translation termination from control mRNAs played a role in determining the calculated −1PRF efficiency. Co-expression of CfMV P27 with the dual reporter vector containing the minimal frameshift signal reduced the production of the downstream reporter, whereas replicase co-expression had no pronounced effect. This finding allows us to propose that CfMV protein P27 may influence translation at the frameshift site but the mechanism needs to be elucidated.
The principal mechanism of translation is the accurate decoding of the triplet codon sequences in one reading frame of mRNA. Specific signals built into the mRNA sequences can cause deviations from this rule. Viruses exploit several translational 'recoding' mechanisms, including translational hopping, stop codon readthrough and programmed ribosomal frameshifting (PRF) [reviewed in (1, 2) ], for regulating the amount of proteins produced from their polyproteins. For positive-stranded RNA viruses, À1 PRF is the prevailing recoding mechanism and an essential determinant of the stoichiometry of synthesized viral proteins. Most viral À1 PRF signals are regulating the production of replication-associated proteins. Depending on the virus, the efficiency of À1 PRF can vary between 1 and 40% (3) , and changes in the efficiency can inhibit virus assembly and replication (4) (5) (6) . Therefore, À1 PRF can be regarded as a potential target for antiviral agents (4, 7) . However, the development of efficient antiviral drugs is still hindered, since little is known about the trans-acting factors and the biophysical parameters affecting the À1 PRF efficiencies. Database searches have identified putative frameshift signals from a substantial number of chromosomally encoded eukaryotic mRNAs (8) . Thus, À1 PRF may also have an impact on the complexity of the proteome of several eukaryotic organisms. Two cis-acting signals, a slippery heptamer X XXY YYZ (the incoming reading frame indicated) and a downstream secondary structure, direct the slippage and are therefore essential for this event (9) . À1 PRF takes place after the accommodation step in the slippery sequence by simultaneous slippage of both tRNAs into the overlapping À1 frame XXX YYY (9, 10) . The sequence of the heptamer allows postslippage base-pairing between the non-wobble bases of the tRNAs and the new À1 frame codons of the mRNA. Downstream RNA secondary structures [reviewed in (11) ] force the ribosomes to pause, and place the ribosomal A-and P-sites correctly over the slippery sequence (12) . However, the pausing of the ribosomes is not sufficient for À1 PRF to occur (13) ; in fact, the duration of the halt does not necessarily correlate with the level of the À1 PRF observed (12) . Crystallographic, molecular, biochemical and genetic studies suggest that a pseudoknot restricts the movement of the mRNA during the tRNA accommodation step of elongation by filling the entrance of the ribosomal mRNA tunnel (14) . This restriction can be eased either by unwinding the pseudoknot, which allows the mRNA to move forward, or by a slippage of the mRNA one nucleotide backwards. Chemical agents such as *To whom correspondence should be addressed. Tel: +358 9 19158342 ; Fax: +358 9 19158633; Email: kristiina.makinen@helsinki.fi ª The Author 2005. Published by Oxford University Press. All rights reserved. The online version of this article has been published under an open access model. Users are entitled to use, reproduce, disseminate, or display the open access version of this article for non-commercial purposes provided that: the original authorship is properly and fully attributed; the Journal and Oxford University Press are attributed as the original place of publication with the correct citation details given; if an article is subsequently reproduced or disseminated not in its entirety but only in part or as a derivative work this must be clearly indicated. For commercial re-use, please contact journals.permissions@oupjournals.org antibiotics, certain mutations in the translation apparatus, and in translation elongation factors that change the translation fidelity and kinetics, have been shown to influence À1 PRF efficiency [(10,15) ; reviewed in (16) ]. The parameters known to contribute to the efficiency of À1 PRF are the sequence of the slippery heptamer, the downstream secondary structure, and the length and sequence of the spacer between the two cis-acting signals. Up-and downstream sequences such as termination codons in the vicinity of the À1 PRF signals, or even several kilobases away from them, can affect the À1 PRF efficiencies (3, (17) (18) (19) (20) (21) (22) . A specific sequence in the Barley yellow dwarf virus (BYDV) 3 0 untranslated region (UTR), 4 kb downstream from the slippage site, is vital for À1 PRF (6, 19) . A stimulating effect is achieved through the formation of a tertiary structure, where complementary nucleotides from the 3 0 UTR base pair with a single-stranded bulge in the cis-acting stem-loop (6). Human immunodeficiency virus (HIV) was also shown to require a more complex secondary structure instead of a simple stemloop for optimal À1 PRF in vivo (21, 22) . These investigations suggest that À1 PRF studies carried out with minimal frameshift signals may lead to inaccurate estimates of the stoichiometry of synthesized viral protein products during infection. Cocksfoot mottle virus (CfMV; genus Sobemovirus) infects a few monocotyledonous plant species such as barley, oats and wheat. It has a monopartite, single-stranded, 4082 nt long, positive-sense RNA genome (23, 24) . The polyprotein of CfMV is translated from two overlapping open reading frames (ORFs) 2A and 2B by a À1 PRF mechanism (25) . In this study, we wanted to determine the in vivo À1 PRF efficiency guided by the CfMV U UUA AAC heptamer and the stem-loop structure. In addition to the minimal signal (18) , we decided to test the effect of flanking CfMV sequences for their ability to contribute to À1 PRF. We found that the surrounding viral sequences promoted more efficient À1 PRF than the minimal signal sequence in vivo when measured with the dual reporter vector system developed by Stahl et al. (26) . Therefore, we carried out an expression pattern and deletion analysis to understand the molecular basis of the observed upregulation. In addition, we critically analysed the suitability of the implemented experimental system for this type of a recoding study. An interesting possibility is that the viral proteins produced via À1 PRF could regulate À1 PRF. This hypothesis was tested by co-expressing the CfMV proteins P27 and replicase together with the dual reporter vectors. Three regions from the CfMV polyprotein ORFs ( Figure 1A) were cloned into the NheI and BclI sites between the lacZ and the luc ORFs in pAC74 (26) . This dual reporter vector was a generous gift from Dr J. Rousset of the Universite Paris-Sud, France. The inserted sequences 1602-1720 (A region), 1386-2137 (B region) and 1551-1900 (C region), were amplified by PCR using pAB-21 as a template (18) . Primers were used to introduce NheI and BglII sites to the flanking ends of the inserts. Since NheI digestion removed lacZ ORF, it was reintroduced into the plasmids as a final cloning step. The resulting plasmids were named pAC-A, pAC-B and pAC-C. Corresponding inframe controls, where one nucleotide was added in front of the slippery heptamer, were generated by PCR-based mutagenesis (Exsite, Stratagene) and named as pAC-Am, pAC-Bm and pAC-Cm, respectively. Deletion plasmids pAC-AB/ABm (1602-2137), pAC-AC/ACm (1602-1900), pAC-BA/BAm (1386-1720) and pAC-CA/ CAm (1551-1720) were also generated. The target sequences are shown in Figure 1B . The base numbering refers to the CfMV genome as in (23) . Transcription was driven from SV40 promoter. Plasmids encoded leucine (LEU2) and b-lactamase (ampicillin resistance) as selective markers. Plasmids were transformed into Saccharomyces cerevisiae H23 [MATa hsp150::URA3 ura3-1 his3-11 15leu2-3 112trp1-1 ade2-1 can100]. Dual reporter plasmid pAC1789 and the inframe control pAC1790 containing a 53 bp sequence from the HIV-1 frameshift region (26) were used as a positive control for monitoring the À1 PRF efficiency. To analyse the proteins produced during À1 PRF, lacZ-A/ Am/B/Bm/C/Cm-Fluc fragments were cloned inframe with the N-terminal 6xhistidine-tag in pYES2/NT KpnI and XhoI sites (Invitrogen). Reporter fusions were amplified by PCR using pAC-A/Am, pAC-B/Bm or pAC-C/Cm as templates. The resulting plasmids were named pYES2/NT-A/ Am, B/Bm and C/Cm. Protein expression was regulated from GAL1 promoter. Two CfMV encoded proteins, P27 (C-terminal end of ORF2A) and replicase (ORF2B), were cloned into pYES2 (Invitrogen). Translation initiation codons were introduced within the oligonucleotides during PCR. The resulting plasmids were named pYES-P27 and pYES-Rep. Control plasmids, which lacked the translation initiation codons were prepared by PCR-based mutagenesis (pYES-P27DAUG and pYES-RepDAUG) and the resulting plasmids were verified by sequencing. Plasmids encoded auxotrophic marker for uracil (URA3). All cloning steps were performed using standard protocols. Plasmids were amplified either in Escherichia coli DH5a or JM110, and purified with Qiagen columns. Inserts were verified by sequencing. Yeast transformations were done using the LiAc method (27) , and transformants were selected on a synthetic minimal defined medium (SC) lacking the corresponding auxotrophic marker(s) encoded by the used plasmid(s). Bacteria (E.coli DH5a) were grown in LB-medium containing ampicillin, whereas yeast cells were grown either in YPD, or in an SC medium. Protein expression from GAL1 promoter was repressed during growth at SC medium containing 2% glucose. Expression was induced by replacing glucose with 2% galactose and 1% raffinose. Reporter fusions were expressed in S.cerevisiae INVSc1 (his3D1/his3D1, leu2/leu2 trp1-289/trp1-289 ura3-52/ura3-52) (Invitrogen) overnight. Protein fusions were purified in denaturing conditions using Ni-NTA agarose (Qiagen), and analysed in 6% SDS-PAGE gels. Proteins were visualized either by Coomassie staining, or by using antisera raised against the CfMV polyprotein region 1386-1724 encoding CfMV VPg (28) . Protein antibody complexes were visualized with horseradish peroxidase-conjugated anti-rabbit antibodies (Sigma) and ECL chemiluminescent reagents (Amersham). Plasmids pYES-P27, pYES-Rep, pYES-P27DAUG, pYES-RepDAUG, or empty pYES2 were co-expressed with pAC-A or with the corresponding pAC-Am inframe control in S.cerevisiae EGY48 strain (MATa, ura3, trp1, his3, 6lexAop-LEU2) (Invitrogen). Transformants were grown overnight in SC-Leu-Ura media in non-inducing conditions, and used to inoculate induction medium. Cells were harvested at late logarithmic phase. Expression of the CfMV proteins was confirmed by western blotting using polyclonal antisera against the CfMV ORF 2a and 2b proteins (28) . Determining the enzymatic activities as described below monitored the effect of CfMV P27 and replicase on À1 PRF. For the in vitro analysis, the lacZ-gene of pAC-A/Am, -B/Bm and -C/Cm vectors was replaced with PCR-amplified Renilla luciferase (Rluc) gene from pRLnull vector (Promega). The resulting pACRF plasmids were used as templates for PCR in order to add T7 promoter upstream of the Rluc gene. These PCR products were used for RNA synthesis with RiboMax kit (Promega). Transcripts were treated with RQ1-DNase (Promega), purified with Qiagen RNeasy columns, and quantified spectrophotometrically. The integrity of the transcripts was checked in agarose gels. In vitro translations were carried CfMV À1 PRF test and control sequences were cloned between the b-galactosidase (LacZ) and firefly luciferase (Luc) genes into a dual reporter vector pAC74 (26) . Inframe control constructs had one extra nucleotide inserted in front of the slippery heptamer, which fused the reporters into the same reading frame. Thus, translation of the inframe control results in the production of a b-galactosidase-CfMV-firefly luciferase fusion. Translation of the test constructs in the incoming 0-frame yields a b-galactosidase-CfMV fusion, whereas À1 PRF produces a b-galactosidase-CfMV-firefly luciferase fusions identical to those produced from the inframe controls. À1 PRF efficiencies were calculated from the firefly luciferase activities after b-galactosidase normalization with the given formula. (B) CfMV polyprotein is encoded by two overlapping ORFs, 2A and 2B via À1 PRF. Sequence regions tested in the dual reporter vectors for their activity to promote À1 PRF are indicated. The numbering refers to the CfMV RNA sequence as published in (23) . out in wheat germ extract (WGE) according to the manufacturer's protocols (Promega). Reactions were incubated in room temperature for 60 min, and stopped on ice prior to enzymatic measurements. Cell cultures were started from at least three independent clones and grown until the late exponential phase. Cells were collected by centrifugation, frozen in liquid nitrogen and stored at À70 C. Bacterial cells were lysed by sonication (3 · 15 s), and yeasts by vortexing with glass beads (0.5 vol) in +4 C for 30 min. Lysates were cleared by centrifugation, and enzymatic activities were determined immediately. Total protein concentrations were measured by using a Bradford protein assay reagent (Bio-Rad). b-Galactosidase (LacZ) and firefly or Renilla luciferase (LUC or RUC) activities were measured with commercial kits from Promega according to the manufacturer's instructions. LacZ activity was determined as the colour intensity at A414 nm. Luciferase activities were measured as relative light units (RLUs) with luminometer (Biohit or ThermoLabsystems). À1 PRF efficiencies were calculated from normalized firefly luciferase activities with the following formula: [(LUC activity from the test construct)/(LacZ or RUC activity from the test construct)]/[(LUC activity from the inframe control)/(LacZ or RUC activity from the inframe control)] · 100%. In CfMV, the motif for À1 PRF is the slippery heptamer U UUA AAC and a stem-loop structure 7 nt downstream (25) . The efficiency of À1 PRF directed by CfMV cis-acting signals was assayed in vivo using a dual reporter vector system ( Figure 1A ). Since reporters are produced from one single mRNA, factors that affect the stability of the mRNA as well as the rate of translation initiation have a similar influence on the expression of both reporters, and these variations can be monitored as changes in the activity of the upstream reporter. We quantified À1 PRF by comparing the b-galactosidase normalized firefly luciferase activities derived from the test constructs via À1 PRF to those obtained from the inframe controls, in which identical b-galactosidase-CfMV-firefly luciferase fusions are produced without À1 PRF due to the added nucleotide in front of the slippery heptamer (see Figure 1A ). Similar vectors have been shown to detect even small changes in the recoding efficiencies resulting from alterations in the cis-or trans-acting factors (26, (29) (30) (31) (32) . Three inserts of varied lengths from the CfMV polyproteinencoding region (ORF2A/2B) were introduced between the two reporters ( Figure 1B) . The A-region, which at 119 bp was the shortest, represented approximately the minimal frameshift signal proven to be functional in vitro (18) . The longest region was the B-insert. At 752 bp, it started from the 5 0 -terminus of the 12 kDa viral genome-linked protein (VPg) gene and continued to the end of ORF2A. This region encodes CfMV protein P27 with an unknown function (28) . Since the minimal requirements for the functional frameshift signal in vivo were not known, an intermediate 349 bp C-sequence was also selected for the analysis. A well-characterized 53 bp frameshift cassette derived from HIV-1 RNA was used as a positive control. Our results regarding the HIV À1 PRF efficiency, 0.7 -0.1% in bacteria, and 4.5 -1.1% in yeast (Figure 2) , are corroborated by those published earlier (26, 33, 34) indicating that our dual reporter system was fully functional. b-Galactosidase has been shown to retain its specific activity well, irrespective of the C-terminal fusions (35) . This is important, since the first reporter serves to control the variations among the abundance and translation rates of the studied mRNAs (26, 30) . In addition to changes in specific activities, heterologous fusions can cause alterations in the solubility and conformation, which can expose cryptic protease target sites and reduce the stability of the proteins (36) . Therefore, for a reliable quantification of À1 PRF, it was important to test that equimolar amounts of fusions produced from the corresponding test and control constructs had similar enzymatic activities. Most inframe controls and the analogous test constructs had equal absolute b-galactosidase activities ( Table 1) . Comparable results were obtained, if activities were normalized with total protein concentration (data not shown). These results indicated that the length of the fusion as such did not affect the specific activities. The b-galactosidase activity from pAC-Am inframe control was also comparable to activity obtained from an empty pAC74, where this enzyme has no fusion (data not shown). This further supported the view that the few observed variations in the b-galactosidase activities more likely resulted from the changes in translatability or stability of the transcripts. In addition to pAC-Cm, two inframe controls pAC-Am and pAC-ACm showed $25% lower b-galactosidase activities when compared to the equivalent test constructs ( Table 1 ), indicating that the productivity from these constructs was reduced. Taken together, b-galactosidase seemed to fit well to be used as the first reporter and thus normalization factor in the in vivo experiments of this study. CfMV frameshift signals generated significant À1 PRF in yeast. À1 PRF level measured from pAC-A was 3-fold higher than from HIV RNA (Figure 2A and B) . The extent of À1 PRF directed by the minimal region A in yeast, 14.4 -1.9%, was at the same level as that reported for the CfMV minimal frameshift signal in vitro (12.7%) (18) . In contrast to our earlier in vitro observations (18) , the longer CfMV sequences upregulated À1 PRF in vivo. In yeast, the level of upregulation was 2-fold for pAC-B, the À1 PRF frequency being 26.3%, and almost 5-fold for pAC-C resulting in efficiency close to 70% (Figure 2A) , which is an extremely high value, if compared to the other values published earlier (3). CfMV frameshift signals directed À1 PRF at a lower level in bacteria than in yeast ( Figure 2B ). The extent of À1 PRF directed by region A in bacteria was 2.4 -0.7%. As in yeast, the longest B region stimulated À1 PRF 2-fold in bacteria when compared with pAC-A. However, region C did not further improve À1 PRF, but programmed À1 PRF to similar levels as pAC-B, the percentages being 4.7 -1.6% for pAC-C and 5.5 -1.5% for pAC-B. To identify the sequence(s) responsible for the enhancement of À1 PRF in vivo, a deletion analysis was carried out. The 5 0 -or the 3 0 -sequences flanking the A-region were deleted from pAC-B/Bm or pAC-C/Cm as indicated in Figure 1B , which generated vectors pAC-AB/ABm, pAC-BA/BAm, pAC-AC/ ACm and pAC-CA/CAm. À1 PRF frequencies were determined in yeast ( Figure 2C ). Increased À1 PRF was observed in all deletion constructs in comparison to the À1 PRF directed by the A region. The BA and AB regions promoted À1 PRF as efficiently as the B region, whereas regions CA and AC were better than region A, but not as good as region B. In other words, the presence of nucleotides 1386-1720, or downstream nucleotides 1602-2137, was sufficient to increase À1 PRF to the level directed by the region B. Thus, the deletion analysis did not identify single specific sequence region as being responsible for the increased À1 PRF frequencies. The expression pattern of the test and control constructs was analysed to understand the basis for the observed upregulation in yeast. Cassettes containing the reporters and the studied intercistronic sequences were expressed and purified as N-terminal histidine fusions. This allowed us to capture all the N-terminally intact products. The affinity-purified proteins were separated in SDS-PAGE gels, and visualized either by Coomassie staining (data not shown), or by western blotting with the CfMV-specific anti-VPg antibodies. The expected b-galactosidase-CfMV fusions terminating at the end of the 0-frame in the test constructs were detected. Also, the longer transframe b-galactosidase-CfMV-firefly luciferase fusion proteins were present in both the test and the inframe constructs ( Figure 3) . Comparison of the Coomassie-stained gels with the western blots revealed that the antisera recognized the products terminating at the CfMV-encoding regions better than the transframe products. Furthermore, the small size of the CfMV-specific region in the pYES2/NT-Am decreased the binding of the antibodies to these inframe control fusions. Thus, this data were not suitable for quantitative analysis of À1 PRF. Interestingly, an additional protein, which reacted with CfMV-specific antisera, was co-purified from the cells expressing pYES2/NT-Bm and pYES2/NT-Cm inframe controls ( Figure 3 ). The size of these fusions suggested that translation had terminated approximately at the site for À1 PRF signals. If such putative termination products were also present in cells expressing the test constructs, the correctly terminated 0-frame products in the western blots masked these products. A closer look at the absolute b-galactosidase and firefly luciferase activities revealed that firefly luciferase expression from pAC-Cm was clearly reduced (data not shown). In fact, expression from the inframe control was comparable to the corresponding pAC-C test construct. This was also obvious when the firefly luciferase activities were normalized with the total protein amount. After setting the activity from pAC-Am to a relative value of one, the corresponding values from pAC-Bm and pAC-Cm were 0.80 and 0.28. Although the b-galactosidase measurements (Table 1) suggested that the overall translatability of the pAC-Cm mRNA was also reduced to some extent, it explained the decrease in firefly luciferase expression only partially. In the light of these findings, the extremely high À1 PRF frequency estimate calculated for the C-region could be explained with more frequent translation termination at the frameshift signals of the pAC-Cm mRNA, which reduced firefly luciferase activity in relation to b-galactosidase. À1 PRF was also assayed in vitro in WGE. Although LacZ-encoding gene is suitable for the in vivo studies, it is an unsuitable first reporter for the in vitro determination of À1 PRF efficiencies due to its big size (30) . In good agreement with this, we observed several unexpected products in the in vitro translations programmed with LacZ-CfMV-luc mRNAs (data not shown). Renilla luciferase has been shown to retain its specific activity irrespective of the C-terminal fusions (30) . Therefore, we decided to use Rluc-CfMV-luc transcripts to determine the À1 PRF efficiencies in the cellfree system. First, we verified the suitability of Renilla luciferase for the intended in vitro experiments as described in (30) . Transcripts encoding monocistronic Renilla luciferase and Renilla luciferase fused to firefly luciferase (Rluc-Am/ Cm-luc) were mixed in different ratios and used to program the in vitro translations. Increasing concentrations of transcripts encoding the Rluc-Am-luc fusion resulted in linearly growing firefly luciferase activities. At the same time Renilla luciferase activities remained constant, which showed that its enzymatic activity was not sensitive to the C-terminal fusions ( Figure 4A ). Similar results were obtained with Rluc-Cm-luc mRNA (data not shown). À1 PRF efficiencies were then determined with transcripts that contained CfMV regions A, B and C, and their corresponding inframe controls. In all cases, slightly higher À1 PRF frequencies were obtained than in vivo. In nice correlation with the in vivo results, enhanced À1 PRF was observed with the region B, although the effect was weaker than in vivo. In this context, region C did not differ from the minimal region A in its capacity to program À1 PRF ( Figure 4B) . The ratio between the CfMV P27 and replicase is regulated by À1 PRF during CfMV infection (28) . We studied whether these proteins could regulate the À1 PRF process. P27, replicase, or an empty expression vector was co-expressed in yeast together with the dual reporter vectors containing the minimal À1 PRF test and inframe control regions as intergenic sequences (pAC-A and -Am). P27 and replicase expression was verified by a western blot analysis ( Figure 5) . A faint band having nearly the same mobility as the replicase was detected in cells grown under repressing conditions. However, due to the small size difference, this protein was not regarded as replicase. Enzymatic activities were measured from yeast lysates prepared from induced cultures. Measurements showed comparable levels of b-galactosidase in all the samples, indicating that P27 or replicase expression did not affect the stability of the dual reporter mRNA or the translatability of the first reporter ( Table 2 ). The effect of P27 or replicase expression was monitored by comparing the reporter activity ratios to those measured from cells harbouring the empty expression plasmids (Table 2) . Co-expression of CfMV replicase did not affect the normalized firefly luciferase expression (LUC/LacZ) from the inframe control, whereas slightly increased luciferase expression from the test construct was observed. In contrast, P27 expression reduced firefly luciferase expression both from the test and the inframe constructs. The effect was stronger in the presence of inframe control as normalized firefly luciferase levels reached only 54% of expression measured from the empty vector control. To verify that the observed differences in firefly luciferase production depended on the studied CfMV proteins, we co-expressed the dual reporter vectors with plasmids having the first translation initiation codons of P27 and replicase deleted (pYES-P27DAUG and pYES-RepDAUG). Western blot analysis with antisera against ORF2A or 2B did not detect any proteins produced from these vectors (data not shown). The obtained LUC/LacZ ratios were compared to those measured from cells expressing the CfMV proteins (pYES-P27 or pYES-Rep). LUC/LacZ ratios measured from cells expressing replicase were slightly lower than the ratios calculated from cells harbouring pYES-RepDAUG plasmids, being $90% when co-expressed with pAC-A and $84% when co-expressed with pAC-Am. In the presence of P27, LUC/ LacZ ratio of pAC-A reached $81% of expression measured from cells transformed with pYES-P27DAUG. Again the effect of P27 expression was more evident with pAC-Am inframe control as P27 expression reduced LUC/LacZ ratio to half ($48%) when compared to the corresponding value measured from the cells harbouring pYES-P27DAUG. This verified that CfMV P27 was able to reduce the downstream reporter expression from dual reporter mRNAs. Since CfMV P27 had a proportionally stronger effect to firefly luciferase production from the inframe control mRNAs in comparison to the test mRNAs (Table 2) , the calculated À1 PRF efficiency increased from 14.7 to 22.4%. Since À1 PRF studies are affected by a huge number of different parameters, it is not an easy task to determine the real ratio between the proteins produced via this mechanism in vivo. However, in viral systems, the efficiency of À1 PRF is an essential determinant of the stoichiometry of synthesized viral protein products, which must be rigidly maintained for efficient propagation of the virus. For example, frameshifting in retroviruses determines the ratio of structural (Gag) to enzymatic (Gag-Pol) proteins, and plays a critical role in viral particle assembly (5) . In this study, the capacity of CfMV frameshift signals to direct efficient À1 PRF was analysed in vivo by using dual reporter vectors. The length of the CfMV sequence clearly affected the actual efficiency percent in vivo. The PRF efficiency was elevated when longer viral sequences were directing the À1 PRF, but the deletion analysis did not identify any specific region as being solely responsible for the enhancement. Up-and downstream sequences nearby or far away from the cis-acting signals have been reported to enhance À1 PRF in other viruses, such as HIV, human T-cell leukaemia virus and BYDV (6, 19, 20) . Also out-of-frame stop codons have been shown to influence À1 PRF frequency in vitro in retroviruses (17) and in CfMV (18) . A study on the spacer sequences located between the cis-acting signals showed that high slippage frequencies were obtained when the first three nucleotides were G/U, G/A and G/A, the first two being the most important (37) . In CfMV, the spacer starts with UAC, which partially explains the capacity of the CfMV sequence to promote high slippage levels. In this study, the observed enhancement of À1 PRF was, however, caused by sequences that were not in the immediate vicinity of the Figure 5 . Co-expression of CfMV P27 or replicase simultaneously with the minimal frameshift signal construct pAC-A or the corresponding inframe control pAC-Am in yeast. Yeast total protein samples were separated in 12% SDS-PAGE gels, transferred onto PVDF membranes, and immunocomplexes detected by ECL chemiluminescent system. CfMV P27 expression was verified by western blotting with antisera raised against ORF2A (A), and CfMV replicase expression was detected with antisera raised against ORF2B (B). Abbreviations: À, repressed; +, induced; C1, pMAL-VPg $53 kDa; and C2, baculovirus expressed CfMV replicase. slippery sequence thus indicating that CfMV sequences further away also have an influence on the level of frameshifting in vivo. We conclude that the most reliable estimates for À1 PRF and consequently for the amount of replicase versus the 0-frame translation product P27 can be obtained only by using the full-length viral sequences. In reality, such a study would however be hampered by the non-quantitative nature of the western blot analysis, the presence of different polyprotein processing intermediates, and the differences in the stabilities of the end products in the infected cells. The overall competence of CfMV signals to direct À1 PRF was high, when compared to related plant viruses, such as Potato leaf roll virus and BYDV. À1 PRF values of $1% have been reported for these viruses when measured with reporter-based assays (6, 38) . We can hypothesize that one reason for the high efficiency is the slippery tRNA Asn encoding the AAC triplet of the CfMV heptamer. Equal U UUA AAC slippery heptamer has been measured to induce 20-40% of À1 PRF in a diversity of animal viruses [(39); reviewed in (3)]. The low fitness of CfMV À1 PRF signals in bacteria is in agreement with the poor functioning of the eukaryotic slippery heptamers of the order X XXA AAC in prokaryotes (40) (41) (42) . IBV RNA, having an identical shifty heptamer, has been shown to direct À1 PRF at similar 2-3% level in bacteria (41) . A recent study reported that XXXAAAC heptamers dictate À1 PRF to occur via the slippage of two adjacent tRNAs placed over the heptamer, irrespective of whether the host is an eukaryote or a prokaryote (42) . Therefore, the inability of prokaryotic translation systems to direct efficient À1 PRF from this heptamer is not an inherited property of prokaryotic tRNA Asn , but results from differences in the ribosomes (42) . Paused ribosomes can pass the À1 PRF site by À1 frameshifting, resumption of 0-frame translation, or termination (43) . Transient polypeptide intermediates that result from the pausing of ribosomes in the slippery sequences have been observed during IBV and S.cerevisiae L-A virus polyprotein synthesis (12, 13, 43, 44) . A pseudoknot structure formed by IBV mRNA causes a translational pause at fixed position upstream the secondary structure regardless of whether the slippery heptamer is present or absent (12) . Based on the findings of this study, we propose that also here a certain percent of ribosomes stalled at the secondary structure of the frameshift site in our inframe control and test mRNAs in yeast, and this led to the prematurely terminated products observed with the inframe control constructs pAC-Bm and -Cm. Although not unambiguously proven by this study, high frequency of termination of translation especially at the frameshift site of the pAC-Cm mRNA would nicely explain the extremely high calculated À1 PRF efficiency. Factors that change the translation fidelity and kinetics have been shown to influence À1 PRF efficiency [ (10, 15) ; reviewed in (16) ]. Autoregulation of +1 frameshifting by mammalian ornithine decarboxylase antizyme has been reported (45) . This mechanism allows modulation of frameshifting frequency according to the cellular concentration of polyamines. One could speculate that such a regulation mechanism could also be useful to adjust the amounts of the replicationassociated proteins to match the requirements of different phases in viral replication cycle. This hypothesis was studied by expressing CfMV proteins P27 and replicase together with pAC-A and pAC-Am in yeast cells. Since b-galactosidase production remained constant regardless of the presence or absence of CfMV proteins, they did not interfere with translation initiation from pAC-A/Am mRNAs per se. However, P27 expression caused a reduction in the firefly luciferase production especially from the inframe control, whereas replicase production only slightly increased the firefly luciferase production from pAC-A, but not from pAC-Am. Since replicase expression had only a faint effect on the normalized firefly luciferase production via À1 PRF, our conclusion is that CfMV replicase had no pronounced effect on translation at the frameshift site. Co-expression of the non-translatable form of P27 with the dual reporter vectors verified that P27 truly affected firefly luciferase expression on the protein level. Therefore, we propose that CfMV protein P27 may influence translation at the frameshift site. If CfMV P27 indeed interferes with viral protein synthesis during CfMV infection, the mechanism, its specificity and the possible biological role needs to be elucidated in the future.
30
Australian public health policy in 2003 – 2004
In Australia, compared with other developed countries the many and varied programs which comprise public health have continued to be funded poorly and unsystematically, particularly given the amount of publicly voiced political support. In 2003, the major public health policy developments in communicable disease control were in the fields of SARS, and vaccine funding, whilst the TGA was focused on the Pan Pharmaceutical crisis. Programs directed to health maintenance and healthy ageing were approved. The tertiary education sector was involved in the development of programs for training the public health workforce and new professional qualifications and competencies. The Abelson Report received support from overseas experts, providing a potential platform for calls to improve national funding for future Australian preventive programs; however, inconsistencies continued across all jurisdictions in their approaches to tackling national health priorities. Despite 2004 being an election year, public health policy was not visible, with the bulk of the public health funding available in the 2004/05 federal budget allocated to managing such emerging risks as avian flu. We conclude by suggesting several implications for the future.
Public health is a small component of the health system, both in terms of budgetary allocation at either state or national level and in terms of the number of practitioners. It incorporates a myriad of activities; legislation and regulation for health protection, preventive services directed at specific diseases and populations, and health promotion programs geared towards particular risk factors and vulnerable groups in the community. As such, it looks like a disparate collection of programs and investments. In Australia, there is also confusion about the very terminology of 'public health'. Despite its extensive history and global understanding, in Australia the term is used variously; to refer to publicly funded health services, and interventions (regardless of the funding source) which are aimed at primary prevention and the promotion and protection of the public health ('rats and drains'). This has led to an increasing number of jurisdictions adopting the label 'population health'. Renovation of the public health system has been on the international agenda for some years. In the US, the Institute of Medicine released reports during 2003 about the public health workforce required for 21 st century challenges [1], as well as re-visited and updated its landmark report, The Future of Public Health in the 21st Century [2] . In the UK, following the path-breaking review of the NHS by Derek Wanless [3] the Treasury commissioned him, in 2003, to undertake a review of whole-of-government effort in public health. Arising in part from the challenges that confronted Canada during the outbreak of sudden acute respiratory syndrome (SARS) in 2003, a new public health agency, at arms length from government, is being created. Public health in Australia, meanwhile, remained fragmented -by programs, across jurisdictions (particularly the states and territories) -and without a systematic approach to funding, organisation, or conceptualisation. In 2003/04, the gap between rhetoric and funding continued to be noticeable, along with the tension between framing priorities for popular appeal versus the technical language of the evidence base. This article will examine some of the indicative developments of public health in Australia in 2003/04. The key developments are identified, and a number of them are selected for in-depth analysis. In this article, we use the traditional meaning of the term 'public health' and focus on activities which are usually designed to promote and protect the health of the population. The drivers for these developments, their short term implications and some signposts for the future are suggested. While early global anxiety over SARS occupied headlines between February and May, the more persistent popular headline in 2003 focused on obesity. Summits were held in NSW and Victoria, while the National Obesity Taskforce was convened under the auspice of the Australian Health Ministers Council (AHMC). When Kay Patterson was the Federal Health Minister, she declared that prevention was the fourth pillar of Medicare and she wanted to be 'Minister for Prevention'. Indeed, the 2003/04 federal budget, although limited, contained a bundle of initiatives entitled "Prevention on the Health Agenda". In particular, a number of immunisation and health promotion programs were included. Significant amongst the funding initiatives for public health announced in 2003/04 was government support for the meningococcal vaccine. Although this was the culmination of many months of careful planning, a perception existed that this only occurred after considerable public interest in and anxiety about deaths from outbreaks of this disease. Further changes to the recommended schedule in 2003 were made by the Australian Technical Advisory Group on Immunisation (ATAGI), in particular the inclusion of pneumococcal and varicella vaccines; however, these did not result in similar prescribed vaccine programs or in similar funding. These three developments are reviewed in greater detail in the next section. The National Public Health Partnership (NPHP) and the AHMC adopted the influenza pandemic plan in October 2003, and with the advent of the newly-identified disease SARS, as well as outbreaks of meningococcal disease, management and prevention of communicable diseases was prominent. Following on from the significant funding boost for bioterrorism preparedness in 2002/03, public health preparedness became a more generic theme. The arrival of SARS occupied the national popular and political imagination as well as tested the infrastructure capacity of public health. Australia fared well during the outbreak. Apart from escaping with only six Australian cases, it provided an opportunity to establish a coordinated approach between the Commonwealth and the states/territories and also contributed to the global epidemiological investigation and prevention effort. SARS also prompted amendments to the Quarantine Act [4] . While the recall following the Pan Pharmaceutical crisis put the Therapeutic Goods Administration (TGA) under the spotlight, it also managed to conclude negotiations that had been in train for several years on a Trans-Tasman regulatory regime and authority. Also on the regulatory front, the Australian New Zealand Food Regulation Ministerial Council endorsed a nutrition, health and related claims policy guidelines and established a review of genetically modified (GM) labeling of foods [5] . All these developments pointed to the global nature of public health, and the intersection between public health activities and the economy. Policy development in public health has never been confined to a set of health programs, and in 2003/04, the lead was often taken from outside the health sector. Most significant was the adoption of the National Agenda for Early Childhood [6] , pushed by public health advocates for child health since the mid 1990s. The National Public Health Partnership responded by coordinating a scoping of child health strategies across Australia. Elsewhere in Government, "Promoting and Maintaining Good Health" was adopted as one of the National Research Priorities [7] . Healthy ageing also emerged as a policy theme in Ageing Research. Public health workforce development was pursued outside the mainstream education and training arrangements for public health in universities. The Community Services and Health Training Board commissioned a consultative process to develop population health competencies for the Vocational Education and Training (VET) sector [8] . New population health qualifications and competencies were proposed for incorporation into the Health Training Package -including certificates in population health and in environmental health, and diplomas in population health and in indigenous environmental health. The release in 2003 of the report "Returns on Investments in Public Health: an epidemiological and economic analysis" [9] (often referred to as the Abelson report), may have a significant impact in subsequent years. Commissioned several years earlier by the Population Health Division of the Department of Health and Ageing (DoHA), the report experienced a relatively low profile until Derek Wanless visited from the UK. Having chaired a review that contributed to a significant budgetary increase for the NHS, Wanless had been commissioned by the British Treasury to examine prevention across government. In September 2003, at a meeting in Canberra with senior officials across key agencies, Wanless marveled at the value of the Abelson report, described in more detail below. Although 2004 was an election year, public health policy was neither visible during the campaign or in policy development more generally. The Federal Government's initiative to wind up the National Occupational Health and Safety Commission received little publicity and comment, even though it indicated the Commonwealth's increasing tendency to pursue its own pathway, separate from states and territories, and to bring the functions of statutory bodies into departments. Jurisdictional and annual reports show that across the states and territories, there were multiple plans, draft guidelines, meetings, episodic training and programs across a broad range of areas. Some health issues are being taken up across jurisdictions -particularly tobacco control, sexually transmitted infections, Aboriginal health, and vaccination. Innovative activities were reported in some jurisdictions, such as a new Health Impact Assessment Branch and a new public health training program in Western Australia. There was, however, no apparent consistency in health priorities across the nation, and an apparent divergence in the interests of the states/territories and the federal government. While the "prevention and management of overweight and obesity" agenda may have appeared to many observers as a new issue in 2003, its arrival was preceded by several years of intensive work. The NHMRC had released Acting on Australia's Weight: Strategic plan for the prevention of overweight and obesity in 1997 [10] , the same year the ABS published the findings from the 1995 National Nutrition Survey, revealing that 45% of men and 29% of women in Australia were overweight, with an additional 18% of men and women classified as obese [11] . Furthermore, overweight and obesity were more common in lower socio-economic groups, in rural populations, in some immigrant groups, and in Aboriginal and Torres Strait Islander (ATSI) peoples. Despite longstanding national cooperation on nutrition (since the days of the National Better Health Program in the late 1980s), and even more recent national cooperation on physical activity, public and political imagination was not captured until the same issues were recast as 'obesity', with a focus in particular on childhood obesity. Following from the NSW Childhood Obesity Summit in late 2002, the Australian Health Ministers agreed that a national approach was required and established a National Obesity Taskforce [12] . In 2003, NSW Health released it's response to the Summit recommendations and supported the vast majority of the 145 resolutions [13] . The Victorian Department of Human Services also held a summit [14] , while Healthy Weight 2008 -Australia's Future was released by the Commonwealth [15] . The NHMRC joined in with release in late 2003 of clinical practice guidelines for general practitioners and other health professionals [16] . While the specifics vary, the major themes and strategies are captured in Healthy Weight 2008. These are summarised in the Table 1. The Commonwealth strategy is, however, relatively weak on intersectoral policy and regulatory measures. As an illustrative example of the contrast at the state level, implementation in NSW now ranges from school physical activity and nutrition survey, to a school canteen strategy, to negotiating with Commercial Television Australia about their code of practice on advertising in peak children's viewing hours. The Commonwealth apparently chose not to consider how it might exercise its relevant taxation or legislative powers, despite the history of health promotion pointing to the importance of public policy measures beyond the health system. An examination of the manner in which the obesity issue was framed, and the details contained in the national strategy, raises a number of issues and questions: -Why was framing the issues as 'obesity' more successful than the focus on 'nutrition' and 'physical activity'? Why did 'obesity' gain traction while the other terms did not? -Why did the Commonwealth opt for the softer programmatic approach, rather than tackle obesity with stronger public policy measures (such as taxation and regulation), and demonstrate its national leadership capacity? -Was the absence of stronger public policy measures because 'obesity' is regarded as largely a health issue, rather than a whole-of-government issue? Or was the Government waiting to see if the US opposed the WHO Global Strategy on account of the strength of the industry lobby? -After a number of years of public concern about eating disorders and whether they arise in part because of promotion of certain types of body image, was the 'obesity' label a backward step for mental health and a return to traditional images of beauty? -Is there a risk that people, including children, who are labeled as 'overweight and obese' will be stigmatised? To what extent have the voices of affected communities been incorporated into the development of national strategies, if at all? -Given the correlation between obesity and socioeconomic disadvantage, how would the proposed strategy not exacerbate those inequalities? -Were children targeted because they are a "captive audience" and therefore easy targets or did the evidence suggest the best return on investment (in terms of health gain and managing demand on the health care system) would come from a focus on children? -Was the move to appeal to a populist agenda, while simultaneously progressing the longer-term agenda of tackling health inequalities through multi-sectoral partnerships, a triumph for public health advocates? These complex threads are interwoven. For the moment, the publicly enunciated agenda represents a confluence of a number of rationales. During 2003-4 three new vaccines were added to the schedule of recommended vaccines for Australians (an additional change to the schedule, recommending that polio immunisation be changed from oral to injected (IPD) vaccine, will not be discussed here). These vaccines protect against serogroup C meningococcal disease, some strains of Pneumococcal disease, and chicken pox [17] . For the first time, not all of these recommended vaccines will be funded by Government. Prior to the introduction of these vaccines, the quality of information about the epidemiology and burden of disease caused by these three infections was extremely variable. Meningococcal disease has been notifiable for many years, and in Australia almost all is caused by serogroups B and C. Whilst serogroup B predominantly occurs in young children, a new strain of serogroup C [18] was causing increasing anxiety amongst public health professionals, microbiologists, staff of accident and emergency departments, intensive care units and of course the public and media. The cause of anxiety amongst health professionals was based on the fact that this new strain carried a high fatality rate with severe after-effects in a high proportion of survivors. The attack rate, although still small, was increasing exponentially each year and reaching an important trigger point, and the majority of cases were now healthy teenagers and young adults. Although an initial accelerated catch-up programme was introduced for teenagers (the major risk group), the new conjugated vaccine was also introduced to the childhood schedule at age one, as from that age, only one dose (at a cost of $30-$60) was considered necessary for full protection from serogroup C disease. Pneumococcal disease became notifiable in 2001, however, with such a short surveillance history, not much is certain locally, epidemiologically speaking, about risk groups and effects (although there is no reason to suppose that it has a different epidemiological pattern from other developed countries). Pneumococcal disease is thought to occur at least four times as often as meningococcal disease, is known to carry major sequelae and has a high case fatality rate. For some time it has been known to be even more common amongst the indigenous Australian population with attack rates of up to 1 in 500 each year, knowledge which underpinned the 1999 decision to target Aboriginal people for free vaccination as soon as the new vaccines became available. Unfortunately at about $120 per dose, conjugate pneumococcal vaccine is very expensive and, for the protection of the very young children who bear the brunt of this disease, it is licensed only to be given as a three dose course, making provision of this vaccine to all Australian children prohibitively expensive. Varicella, predominantly a childhood disease, is caused by a Herpes virus known as herpes virus 3 or varicella-zoster virus or VZV. It is not notifiable in Australia; therefore no epidemiological population data are available. A reliable varicella vaccine has been available since the mid 1990s in the USA and is part of American routine immunisation schedule. This vaccine became available in Australia in 2000, at a cost of about $75-$90 per dose, with two doses being required for full protection. In 2003 the Commonwealth provided its periodic update on the Australian Standard Vaccination Schedule, the list of vaccines it provides as appropriate at no cost to all Australians [19] . For the first time it differed from the National Immunisation Program recommendations in that besides meningococcal serogroup C conjugate vaccines, pneumococcal vaccine, varicella vaccine and also inactivated polio (injected) vaccine were also recommended: however, funding was only secured for meningococcal conjugate vaccines, with a continuation of the provision of pneumococcal vaccines for indigenous children. As a result, although recommended, pneumococcal and varicella vaccines were not funded and parents would have to decide whether or not to pay for them. These funding decisions had important implications. Vaccines protect most of their recipients from unpleasant and sometimes life-threatening disease. One view, subscribed to in the UK, is that ethically, children should not be denied access because of their parents' inability to pay. These vaccines have been the subject of several cost-benefit studies, with generally favourable to extremely favourable pro-vaccination results. Table 2 summarises the various models for framing policy. The policy of funding meningococcal serogroup C vaccine was built on a sustained program of epidemiological evidence, ethical decision-making and public support (and was arguably honed by public pressure). Pneumococcal disease and varicella vaccination programs however, were neither supported by good local epidemiological evidence nor respectable levels of public awareness about these diseases. There had not been a similar program of sustained policy building to support or drive a decision to fund these vaccines. As a funding policy, this was noteworthy in that it marked a departure from previous policies where all recommended vaccines were fully funded by governments. National vaccination policy is designed to advise vaccination policy makers and practitioners of the most up-to-date thinking about optimal vaccination schedules for Australian children, and is not therefore proscriptive, unlike the United Kingdom (UK). Changing or adding vaccines to the recommended schedule is therefore an advisory matter, and the question of funding the vaccination program is decided separately. Cost benefit studies indicate pneumococcal polysaccharide and conjugate vaccines can be cost-effective although vaccine costs clearly affect ratios of cost to benefit greatly [20, 21] . Varicella vaccine is more contentious, because this disease is more severe in older cases, and it is possible that one result of a vaccination program could be an increase in older cases (and therefore severe disease). Whilst the vaccine undoubtedly works, there is no consensus about precisely who should be vaccinated for maximum population health as well as cost benefit, and again potential financial savings are highly dependant upon vaccine costs [22, 23] . The costs of preventive vaccine programs and curative medicine are funded from different sources. Vaccines are currently funded by the Commonwealth and subsidised through the states according to local vaccination policies, whilst the costs of curing cases of these diseases is broadly funded through the Medicare and private health insurance systems. Savings to Medicare and health insurance funds, as a result of successful vaccination programs, are not automatically transferred to the Commonwealth to fund the vaccine programs. Savings -or costs -in one area are of little interest or importance to other program areas. In 2004 the Government revised this funding policy, providing funding for conjugate pneumococcal vaccines population immunisation program for all children under seven years of age (as well as specific people in other risk groups) to commence in January 2005. The Australian Technical Advisory Group on Immunisation (ATAGI) completed Ministerial reports on both varicella and polio (injected as well as oral) vaccination late in 2004, and it is possible that programs for these vaccines will also be funded in the future. The 2002/2003 Federal Budget papers stated that "the Government is committed to making disease prevention and health promotion a fundamental pillar of the health system": however, this was not evident in the subsequent 2003/2004 budget. The Government's Focus on Prevention Package in 2002/03 aimed to incorporate disease prevention into the core business of the primary health care system and was reflective of how the public health agenda was evolving at the national level [24] . The package was comprised largely of a range of measures directed at specific diseases, plus a bundle of initiatives for general practitioners, also referred to as the "primary health care system". Amongst health conditions affecting Australians, breast cancer received the most attention, with the National Breast Cancer Centre being funded to develop a partnership approach to the review and dissemination of new information, along with information, support and management initiatives for rural women diagnosed with breast cancer. Hepatitis C also received some attention, with funding for national education and prevention projects. Financial support was offered for the SARS efforts that had been undertaken by states and territories, in particular for providing medical personnel at international airports. A clear process for assessing priorities under the broad banded National Public Health Program was also flagged. For purposes of the budget, primary health care was defined as general practitioners, and the measures funded included: • "Lifestyle prescriptions" to help GPs "raise community awareness and understanding of benefits of preventive health"; • Collaborative approach to learning, training education and support systems; • Coordinated care plans for people with chronic or terminal conditions; and • Involvement in multidisciplinary case conferencing. The budget did not adopt a comprehensive approach to the primary health care system, perhaps because many community health services, which represent the other important arm for delivery of public health services, are the responsibility of states. The timetable for renewing Public Health Outcome Funding Agreements (PHOFAs) between the Commonwealth and states and territories in 2004 raised in the minds of some stakeholders, the possibility that the Commonwealth might adopt a more comprehensive and strategic approach, linking public health and primary health care funding streams. Judging by the actual quantum of funds made available in the 2003/2004 budget, it would seem that most elements from the package did not actually receive additional funding, as shown in Table 3 . Indeed, many of the GP initiatives, previously cast as improving primary health care, were subsequently packaged as 'prevention'. The combination of these measures reflected a tight fiscal climate, with little growth in the overall health budget, as well as that of other portfolios. It was also a package that demonstrated relatively limited imagination, with support for established issues (such as breast cancer) and repackaging general practice measures that were already in train. With Medicare spending "uncapped" (and targeted public health programs "capped"), attaining more prevention dollars through the GP sector may appear to be one of the few ways to 'grow' dollars for prevention. Although this could be considered to be consistent with the Ottawa Charter of "reorienting health services", many GPs are not trained in a population-based approach to practice, and simply providing new for payments to all represents an undifferentiated, uncoordinated and untargeted approach to prevention. If there is limited support to GPs, and little monitoring, then these measures are unlikely to translate into improved health outcomes. Funding for the Tough on Drugs strategy was announced outside the Focus on Prevention package; perhaps due to Source: [28] the fact that the Tough on Drugs was the responsibility of the Parliamentary Secretary therefore requiring a separate communications strategy, or because the Prime Minister has a strong personal interest in the illicit drug strategy. The range of measures funded (which included introduction of retractable needle and syringe technology, addressing problems related to increased availability and use of psycho stimulants, establishing a research fund, supporting alcohol and drug workforce development needs, promoting access to drug treatment in rural areas, and tackling problems faced by drug users with concurrent mental health problems) certainly suggested more serious government interest and commitment to illicit drugs. During the course of the Howard Government, there has been a gradual process of re-casting the "landscape" of interest groups and policy constituencies. Strong support for breast cancer and zero-tolerance on illicit drugs contrasts sharply with the delays experienced in renewal of the National HIV/Hepatitis C Strategy. The new prominence given to meningococcal vaccine, child health and obesity creates space for other interest groups: even if the re-framing was shaped by nutrition and physical activity lobbies, other clinical interests have been brought into the picture. These developments illustrate how 'political' considerations are important in determining 'public health policy'. It was interesting however, to observe the interest in prevention from outside the health portfolio, particularly from Treasury. This was motivated in part by the Intergenerational Report and concerns about both the sustainability of Medicare as well as the social and economic cost burden arising from an ageing society. This helped to ensure interest in the Abelson Report [9] . Few countries have conducted research on return of investment from prevention efforts. Australia was praised by Derek Wanless at a high-level consultation for completing such an analysis, during his visit to Canberra while conducting a review for the UK Treasury, "Securing Good Health for the Whole Population" [26] . His final report pointed to Australia and Netherlands as two countries that were increasingly using economic evaluation in public health programs. It will be interesting to see if public health policy analysts and Treasury officials draw on this report in future years. In the future it will be interesting to see if the focus on high-visibility programs can demonstrate short-term economic returns. Given 2004 was an election year, the "political economy" of prevention programs could arguably have become a focus of future public health policy, with the 2003/4 agenda providing the Government with the opportunity to gauge public reaction to this new positioning and design their election campaign appropriately. This was, however, not the case. The American emphasis on 'preparedness' appears not to resonate with the Australian public in the same way. From the perspective of public health policy advocates, some lessons that can be drawn from 2003/04 are: • Government's response to public health proposals are shaped by its understanding of the popular interest and desire to communicate directly with the general public; • Longer term public health issues which have struggled to gain support can be progressed if they are cleverly shaped to fit the Government's "formula"; • Develop and nurture new advocates, particularly in seeking to engage with the broader health system; and • Work with the media as partners rather than adversaries These lessons need to be learned well and quickly, to assist with moving the forum for public health policy debate more into the public domain; beyond an essentially "in house" discourse between politicians, researchers and public health advocates. If a more engaged and informed community takes up a public health issue, government will be more likely to respond.
31
GIDEON: a comprehensive Web-based resource for geographic medicine
GIDEON (Global Infectious Diseases and Epidemiology Network) is a web-based computer program designed for decision support and informatics in the field of Geographic Medicine. The first of four interactive modules generates a ranked differential diagnosis based on patient signs, symptoms, exposure history and country of disease acquisition. Additional options include syndromic disease surveillance capability and simulation of bioterrorism scenarios. The second module accesses detailed and current information regarding the status of 338 individual diseases in each of 220 countries. Over 50,000 disease images, maps and user-designed graphs may be downloaded for use in teaching and preparation of written materials. The third module is a comprehensive source on the use of 328 anti-infective drugs and vaccines, including a listing of over 9,500 international trade names. The fourth module can be used to characterize or identify any bacterium or yeast, based on laboratory phenotype. GIDEON is an up-to-date and comprehensive resource for Geographic Medicine.
As of 2005, the world is confronted by 338 generic infectious diseases, scattered in a complex fashion across over 220 countries and regions. Each new day confronts health care workers with unexpected outbreaks, epidemics and heretofore unknown pathogens. Over 2,000 named bacteria, viruses, fungi and parasites are known to cause human disease; and are confronted by 328 anti-infective agents and vaccines. Experts working in Health Geographics share an obvious and immediate need for comprehensive and timely data on the status of infection around the globe. A recent outline of GIDEON addressed uses for the Infectious Diseases clinician [1] . This review will focus on the Global Health aspect of the program. In 1990, we initiated a project to design computer systems to follow all diseases, outbreaks, pathogens and drugs. The initial DOS-based program was written in Paradox for floppy disks, later evolving through a compact disk-based program in Windows. A commercial web-based program was eventually released under the name, GIDEON (Global Infectious Diseases and Epidemiology ON-line, Gideon Informatics, Inc, Los Angeles, California) at http:/ /www.GideonOnline.com. The current version is available on CD (updated every three months) or web subscription (updated every week). The program consists of four modules: Diagnosis, Epidemiology, Therapy and Microbiology. Program modules of peripheral interest in Health Geographics (Therapy and Microbiology) will be discussed only briefly. The Diagnosis module is designed to generate a ranked differential diagnosis based on signs, symptoms, laboratory tests, incubation period, nature of exposure and country of disease origin. Figure 1 depicts the data entry screen for a patient suffering from fever and joint pain following a trip to Indonesia. The lower 'Personal notes' box is used to record additional case data, and can be written in the user's own language. The differential diagnosis list for this case (figure 2) indicates that this patient may be suffering from Chikungunya. The appearance of many diseases on the list indicates that the user failed to enter all positive, and negative findings. For example, the fact that cough was absent would have reduced the likelihood of the second disease listed (Mycoplasma infection) and increased the statistical probability of Chikungunya. At this point, the user can generate a hard copy or e-mail report, access a table comparing the clinical features of the diseases listed, or examine the ranking or omission of specific diseases. If the user clicks on a specific disease name, clinical and epidemiological data on the disease in question are depicted (figure 3). The differential diagnosis list is generated by a Bayesian formula which compares the product of disease-incidence and symptom incidence, for all compatible infectious diseases. In the above example, a number of diseases known to occur in Indonesia were capable of producing fever, and joint pain. The statistical likelihood of Chikungunya in this case can be computed by a simple Bayesian formula, as follows: Two spreadsheets in the GIDEON database respectively follow the incidence of all symptoms for every disease, and the incidence of all diseases for every country. When a clinical case is "entered" into GIDEON, the program identifies all compatible diseases and ranks their relative likelihoods as determined by the above formula, ie: P-(C/ S) vs. P-(D2/S) vs. P-(D3/S) ... vs. P-(Dn/S). A blinded study of 500 cases conducted by this author found that the correct diagnosis was listed in the differential list in 94.7% of cases, and was ranked first in 75% [2] . A second study of hospitalized patients in Boston found that the correct diagnosis was listed in only 69%, and was ranked first in 60% [3] . It is likely that inclusion in the differential diagnosis list may be more important than disease ranking in such systems [4] . A "Bioterrorism" option generates the differential diagnosis for diseases associated with suspected bioterror scenarios. In Figure 4 , "<bioterrorism simulator>" has been substituted for Indonesia, given the above constellation of fever, joint pain, etc. The resulting differential diagnosis lists Ebola (42.9% probability), followed by Crimean-Data entry screen for a bioterrorism scenario Figure 4 Data entry screen for a bioterrorism scenario. Congo hemorrhagic fever (12.6% probability). A similar "Worldwide" option can be used to explore all of the worlds diseases consistent with given clinical features, and access text on the global status for individual diseases. In theory, data entry by users can be monitored at the server level for purposes of surveillance. For example, if one or more users in China were to enter cases of fatal pneumonia, a "red-flag" at any monitoring agency (i.e., the World Health Organization) could indicate the possible appearance of SARS -long before submission of specimens or reporting of the case to local authorities. Similarly, the appearance of multiple cases of "dysentery" by users in a given community could indicate a possible outbreak of shigellosis. The Epidemiology module presents detailed country-specific information on the status of each disease, both globally and within each relevant country. The current version contains over two million words in 12,000 notes. All data are derived from Health Ministry publications, peer-review journals, standard textbooks, WHO and CDC websites and data presented at conferences. The user may also access over 30,000 graphs which follow disease incidence, rates and other numerical data. The main Epidemiology screen is shown in Figure 5 . Note that the user can append custom "personal notes" -in any national language or font-regarding the status of every disease in their own institution. Such notes would be accessible by all colleagues using GIDEON on the local network. Maps which depict the global distribution of each disease can be accessed through the 'Distribution' tab ( Figure 6 ). Epidemiology module, main screen Figure 5 Epidemiology module, main screen. The 'images' tab has been pressed, to access thumbnail images of Plague. These can be maximized and copied to PowerPoint, etc. Note addition of 'Personal notes' by the user, at lower right. Text outlining country-specific data for the disease ( Figure 7) is available through either a list of countries displayed in this module, or by clicking the relevant 'red dot' on the map. These text boxes also include data sets which automatically generate incidence / rate graphs (Figure 8) , a chronological account of all disease outbreaks, and numbered reference links to relevant journal publications and reports of ongoing outbreaks from ProMed http:// www.promedmail.org. A separate 'Graphs' option allows the user to generate custom-made graphs comparing multiple disease rates, or rates in multiple countries. (Figure 9 ). Additional tabs access the descriptive epidemiology and clinical background of each disease. Synonym tabs generate lists of alternative terms for diseases and countries in Spanish, German, Norwegian, etc. Historical data record the incidence of individual diseases and significant outbreaks spanning decades. An additional "Fingerprint" option generates a list of diseases compatible with any set of epidemiological parameters. For example, in Figure 10 we see that ten parasitic diseases are transmitted by fish in Japan. The Therapy module follows the pharmacology and application of all drugs and vaccines used in Infectious Diseases. The current version contains 264 generic drugs and 64 vaccines. Various sub-modules present the mechanism of action; pharmacology, dosages, drug-drug interactions, contraindications, spectrum, and susceptibility testing standards. An international synonym lists contains over 9,500 trade names. As in other modules, users may add Epidemiology module Figure 6 Epidemiology module. Map depicting the global distribution of plague. Specific map areas can be expanded, and all elements can be copied for reproduction as necessary. Country-specific notes regarding plague appear when corresponding red dots are clicked. custom notes in their own language for each drug or vaccine: prices, resistance patterns, local trade names, etc. The Microbiology option is similar to the Diagnosis module. Users may enter any combination of phenotypic tests, and obtain a ranked probability list of compatible bacteria. The current version incorporates more than 1,300 taxa. The Microbiology module is also designed to list the phenotype, prior names, ecology and disease association for any organism, or compare the phenotypes of any combination of organisms selected by the user. Since the graphic and mapping functions of GIDEON treat individual countries as whole units, data presentations lack a certain degree of "granularity." Thus, the dif-ferential diagnosis of fever in Venezuela will include malaria, even if the patient is living outside of the endemic, southern region. This problem is corrected to a large extent by text in the associated country-specific notes and the general knowledge base of the treating physician. In theory, the manufacturer could follow the incidence of each disease for every state, district, province and oblast; but variability would still exist according to occupation, rural vs. urban setting, season, etc. An additional problem relates to the availability and quality of valid epidemiological data. Disease reporting varies widely from country to country. For example, AIDS reporting statistics from sub-Saharan Africa are generally inadequate. Where necessary, the spreadsheets used by GIDEON record published true estimates rather than questionable reports. In other instances, Health Ministry Figure 7 Plague in Tanzania. Clicking on relevant data sets will generate incidence and rates graphs. Note several numbered links to journal publications. Plague -Worldwide incidence and rates per 100,000 Figure 8 Plague -Worldwide incidence and rates per 100,000. data conflict with reports of the World Health Organisation, a fact which is recorded in relevant GIDEON country notes. Occasionally, major diseases are not reported at all. For example, several recent cases of cholera in Japan originated from Thailand; but Thailand has not officially reported a single case in many years. Where possible, the GIDEON data base relies on published best estimates, and at times 'educated guesses' when data are entirely lacking. Thus, there are few published data for disease incidence in Togo, and the program is forced to rely on publications for neighboring Ghana. The reader is referred to the GIDEON website http:// www.GideonOnline.com for an extensive listing of data sources, published reviews, technical background and pricing information. Graph contrasting AIDS rates among user-selected countries Figure 9 Graph contrasting AIDS rates among user-selected countries. Publish with Bio Med Central and every scientist can read your work free of charge
32
Globalization and Health
This debut editorial of Globalization and Health introduces the journal, briefly delineating its goals and objectives and outlines its scope of subject matter. 'Open Access' publishing is expected to become an increasingly important format for peer reviewed academic journals and that Globalization and Health is 'Open Access' is appropriate. The rationale behind starting a journal dedicated to globalization and health is three fold: Firstly: Globalization is reshaping the social geography within which we might strive to create health or prevent disease. The determinants of health – be they a SARS virus or a predilection for fatty foods – have joined us in our global mobility. Driven by economic liberalization and changing technologies, the phenomenon of 'access' is likely to dominate to an increasing extent the unfolding experience of human disease and wellbeing. Secondly: Understanding globalization as a subject matter itself needs certain benchmarks and barometers of its successes and failings. Health is one such barometer. It is a marker of social infrastructure and social welfare and as such can be used to either sound an alarm or give a victory cheer as our interconnectedness hurts and heals the populations we serve. And lastly: In as much as globalization can have an effect on health, it is also true that health and disease has an effect on globalization as exemplified by the existence of quarantine laws and the devastating economic effects of the AIDS pandemic. A balanced view would propose that the effects of globalization on health (and health systems) are neither universally good nor bad, but rather context specific. If the dialogue pertaining to globalization is to be directed or biased in any direction, then it must be this: that we consider the poor first.
Secondly: Understanding globalization as a subject matter itself needs certain benchmarks and barometers of its successes and failings. Health is one such barometer. It is a marker of social infrastructure and social welfare and as such can be used to either sound an alarm or give a victory cheer as our interconnectedness hurts and heals the populations we serve. And lastly: In as much as globalization can have an effect on health, it is also true that health and disease has an effect on globalization as exemplified by the existence of quarantine laws and the devastating economic effects of the AIDS pandemic. A balanced view would propose that the effects of globalization on health (and health systems) are neither universally good nor bad, but rather context specific. If the dialogue pertaining to globalization is to be directed or biased in any direction, then it must be this: that we consider the poor first. I am pleased to introduce 'Globalization and Health', a peer reviewed, open access (free to the end user) journal. In this, the début editorial, I will briefly outline the purpose and scope of this journal highlighting our intention to publish a balanced mixture of opinion on the subject. That the journal be 'Open Access' is entirely appropriate. Knowledge, at its best utility, is a 'public good' i.e. nonrival, non-excludable. While this journal will deal with the subject matter of creating 'global public goods for health', it will also by virtue of its very existence, contribute toward that process. Globalization and Health's 'Open Access' policy changes the way in which articles are pub-lished. First, all articles become freely and universally accessible online, and so an author's work can be read by anyone at no cost. Second, the authors hold copyright for their work and grant anyone the right to reproduce and disseminate the article, provided that it is correctly cited and no errors are introduced [1] . Third, a copy of the full text of each Open Access article is permanently archived in an online repository separate from the journal. Globalization and Health's articles are archived in PubMed Central [2], the US National Library of Medicine's full-text repository of life science literature, and also in repositories at the University of Potsdam [3] in Germany, at INIST [4] in France and in e-Depot [5], the National Library of the Netherlands' digital archive of all electronic publications. Importantly, the results of publicly funded research will be accessible to all taxpayers and not just those with access to a library with a subscription. As such, Open Access could help to increase public interest in, and support of, research. Note that this public accessibility may become a legal requirement in the USA if the proposed Public Access to Science Act is made law [6]. Added to this, a country's economy will not influence its scientists' ability to access articles because resource-poor countries (and institutions) will be able to read the same material as wealthier ones (although creating access to the internet is another matter [7] ). The rationale behind starting a journal dedicated to globalization and health is three fold: Firstly: Globalization is reshaping the social geography within which we might strive to create health or prevent disease. The determinants of health -be they a SARS virus or a predilection for fatty foods -have joined us in our global mobility. Driven by economic liberalization and changing technologies, the phenomenon of 'access' is likely to dominate to an increasing extent the unfolding experience of human disease and wellbeing. Secondly: Understanding globalization as a subject matter itself needs certain benchmarks and barometers of its successes and failings. Health is one such barometer. It is a marker of social infrastructure and social welfare and as such can be used to either sound an alarm or give a victory cheer as our interconnectedness hurts and heals the populations we serve. And lastly: In as much as globalization can have an effect on health, it is also true that health and disease has an effect on globalization as exemplified by the existence of quarantine laws and the devastating economic effects of the AIDS pandemic. A balanced view would propose that the effects of globalization on health (and health systems) are neither univer-sally good nor bad, but rather context specific. The extent to which individual states are able to engage the process of globalization on their own terms differs widely from one country to the next. Child mortality, for example, changes quickly in response to subtle changes in purchasing power in impoverished communities. In affluent communities however, a small change in income has little effect on utility in either direction. As we consider the effects of globalization on wellbeing it becomes apparent that we need to consider both the long term scenarios for populations as a whole, and the immediate effects for the more vulnerable within those populations who are dependent on fragile local economies. If the dialogue pertaining to globalization is to be directed or biased in any direction, then it must be this: that we consider the poor first.
33
The 'polysemous' codon--a codon with multiple amino acid assignment caused by dual specificity of tRNA identity.
In some Candida species, the universal CUG leucine codon is translated as serine. However, in most cases, the serine tRNAs responsible for this non-universal decoding (tRNA(Ser)CAG) accept in vitro not only serine, but also, to some extent, leucine. Nucleotide replacement experiments indicated that m1G37 is critical for leucylation activity. This finding was supported by the fact that the tRNA(Ser)CAGs possessing the leucylation activity always have m1G37, whereas that of Candida cylindracea, which possesses no leucylation activity, has A37. Quantification of defined aminoacetylated tRNAs in cells demonstrated that 3% of the tRNA(Ser)CAGs possessing m1G37 were, in fact, charged with leucine in vivo. A genetic approach using an auxotroph mutant of C.maltosa possessing this type of tRNA(Ser)CAG also suggested that the URA3 gene inactivated due to the translation of CUG as serine was rescued by a slight incorporation of leucine into the polypeptide, which demonstrated that the tRNA charged with multiple amino acids could participate in the translation. These findings provide the first evidence that two distinct amino acids are assigned by a single codon, which occurs naturally in the translation process of certain Candida species. We term this novel type of codon a 'polysemous codon'.
(termed tRNA Ser CAG), and revealed its decoding mechan-Bioscience and Biotechnology, Tokyo Institute of Technology, ism by means of an in vitro translational assay system Nagatsuta, Midori-ku, Yokohama 227, Japan (Yokogawa et al., 1992; Suzuki et al., 1994) . Furthermore, 2 Corresponding authors when we investigated the distribution of this non-universal genetic code in fungi, as well as C.cylindracea, eight other In some Candida species, the universal CUG leucine Candida species-C.albicans, C.zeylanoides, C.lusitaniae, codon is translated as serine. However, in most cases, C.tropicalis, C.melbiosica, C.parapsilosis, C.guilliermonthe serine tRNAs responsible for this non-universal dii and C.rugosa-were found to utilize the codon CUG decoding (tRNA Ser CAG) accept in vitro not only serine, for serine instead of leucine, all having tRNA Ser CAG as but also, to some extent, leucine. Nucleotide replacethe mediator in the unusual decoding (Ohama et al., 1993 ; ment experiments indicated that m 1 G37 is critical for Ueda et al., 1994) . Several other investigators have also leucylation activity. This finding was supported by the shown that the codon CUG is actually translated as serine fact that the tRNA Ser CAGs possessing the leucylation in vivo in C.albicans and C.maltosa (Santos and Tuite, activity always have m 1 G37, whereas that of Candida 1995a; Sugiyama et al., 1995; Zimmer and Schunck, 1995) . One of the most remarkable structural features observed A37. Quantification of defined aminoacetylated tRNAs in most of these tRNA Ser CAGs is that the nucleotide 5Јin cells demonstrated that 3% of the tRNA Ser CAGs adjacent to the anticodon (position 33) is occupied not by possessing m 1 G37 were, in fact, charged with leucine the conserved U residue (U33) but by a G residue (G33). It has been speculated that U33 is necessary for forming of C.maltosa possessing this type of tRNA Ser CAG also the U-turn structure of the anticodon loop in all tRNAs suggested that the URA3 gene inactivated due to the reported so far (Quigley and Rich, 1976 ; Sprinzl et al., translation of CUG as serine was rescued by a slight 1996) . Moreover, the nucleotide at position 37, 3Ј-adjacent incorporation of leucine into the polypeptide, which to the anticodon CAG, is 1-methyl guanosine (m 1 G) in demonstrated that the tRNA charged with multiple almost all tRNA Ser CAGs except for that of C.cylindracea amino acids could participate in the translation. These (A37), while all the serine tRNAs in fungi corresponding findings provide the first evidence that two distinct Introduction Normanly and Abelson, 1989; Shimizu et al., 1992; McClain, 1993; Schimmel et al., 1993) . This line of study The universality of the genetic code was once considered began with the artificial conversion of leucine tRNA of to be one of the essential characteristics of life, which led Escherichia coli to serine tRNA by Abelson's group 10 to the conception of the 'frozen accident theory'. This years ago (Normanly et al., 1986) . Recently, tRNA identity theory proposes that all extant living organisms use the elements of Saccharomyces cerevisiae leucine tRNA were universal genetic code, which was born by accident and elucidated using unmodified variants synthesized by T7 'frozen', and that they originate from a single, closely RNA polymerase (Soma et al., 1996) , indicating that in interbreeding population (Crick, 1968) . However, in recent addition to the discriminator base, A73, the second letter years a number of non-universal genetic codes have been of the anticodon, A35, and the nucleotide 3Ј-adjacent to reported in various non-plant mitochondrial systems, as the anticodon, m 1 G37, are important for recognition by well as in several nuclear systems (reviewed in Osawa leucyl-tRNA synthetase (LeuRS). The majority of Candida et al., 1992; Osawa, 1995) , which contradict the frozen tRNA Ser CAGs have A35 and m 1 G37, while the discriminaccident theory. Among these deviations from the universal codes, ator is occupied by a nucleoside other than adenosine (mostly G73). In this respect, tRNA Ser CAG seems to be a potentially chimeric tRNA molecule capable of being recognized not only by seryl-but also by leucyl-tRNA synthetases. Previously, we showed that these tRNA Ser CAGs would have originated from the serine tRNA corresponding to codon UCG . This suggests an evolutionary pathway in which conversion from A to m 1 G would have taken place at position 37 just after the emergence of tRNA Ser CAG had brought about a change in the universal code. Since such a mutation at position 37 might potentially result in the leucylation of tRNA Ser CAG, we attempted to elucidate the charging properties of these tRNA Ser CAGs both in vitro and in vivo. Based on the results of in vitro aminoacylation reactions using tRNA variants constructed by the microsurgery method, the direct analysis of aminoacylated tRNAs in cells and a genetic approach, we demonstrate here that these serine tRNAs are actually leucylated both in vitro and in vivo. Furthermore, m 1 G at position 37 was found to be indispensable for the leucylation of tRNA Ser CAGs. In fact, the tRNA Ser CAG of C.cylindracea, which has A at position 37, exhibits no leucylation activity. C.cylindracea has a high GϩC content (63%) and utilizes CUG as a major serine codon. However, the other Candida species have no such high GϩC content and utilize the CUG as a minor serine codon (Kawaguchi et al., 1989; Lloyd and Sharp, 1992; our unpublished observation) . Considering the relationship between the usage of the codon CUG as serine and the leucylation properties of tRNA Ser CAG, it seems that only Candida species with a genome in which the incidence of the CUG serine codon is very low possess serine tRNA Ser CAG that can be leucylated. Furthermore, such tRNA Ser CAGs charged with heterogeneous amino acids should be utilized equally in the translation process. This is the first demonstration that a single tRNA species is assigned to two different amino acids in the cell. We propose designating this type of codon having multiple amino acid assignment as a 'polysemous codon'. The correlation between the dual-assignment state and the pathway of genetic code diversification is also discussed. and C.cylindracea (Yokogawa et al., 1992; Ohama et al., 1993 ). The numbering system and abbreviations for modified nucleotides conform Candida zeylanoides tRNA Ser CAG is leucylated to Sprinzl et al. (1996) and Crain and McCloskey (1996), respectively. in vitro (B) Time-dependent aminoacylation with SerRS or LeuRS from First the leucylation of tRNA Ser CAGs from C.zeylanoides C.zeylanoides cells. Aminoacylation reactions were carried out with and C.cylindracea was examined using LeuRS partially 0.7 µM tRNAs and with same amounts of enzyme activities calculated using cognate tRNAs. Serylation and leucylation are shown by dotted purified from C.zeylanoides, since it is known that leucine and solid lines, respectively. The right-hand frame shows the solid tRNAs of yeast have one of their identity determinants at curves from left-hand frame plotted with an enlarged ordinate. The position 37 (Soma et al., 1996) and tRNA Ser CAGs of aminoacylation of C.zeylanoides tRNA Ser CAG (s) and of C.zeylanoides and C.cylindracea have different nucleo-C.cylindracea tRNA Ser CAG (u) are compared; C.cylindracea tRNA Ser GCU (j), having no leucylation activity, is shown as a tides at this position (m 1 G and A, respectively) (Figure control. (C) TLC analysis of acetylleucyl-tRNA fragments derived 1A). Both tRNAs showed almost full serylation activity from leucylated tRNA Ser CAGs. After leucylation with [ 14 C]leucine, (~1200-1500 pmol/A 260 unit), as shown in Figure 1B . The leucyl-tRNAs were acetylated with acetic anhydride. Acetyl-tRNA Ser CAG of C.zeylanoides was evidently leucylated leucylated at all, as was the case when another species tRNA Ser CAG by gel-electrophoresis under acidic con- observed with LeuRSs from both C.cylindracea and S.cerevisiae (data not shown). of serine tRNA specific for codon AGY (Y: U or C) m 1 G37 is responsible for recognition by (tRNA Ser GCU) was employed as a control substrate leucyl-tRNA synthetase ( Figure 1B , right-hand graph). The K m value of C.zeylan-Among the tRNA Ser CAGs of several Candida species, oides LeuRS towards tRNA Ser CAG (5.0 µM) is only one that of C.cylindracea is unique because it alone possesses order of magnitude larger than that of the serylation of no leucylation capacity. A sequence comparison of these this tRNA (0.22 µM) as well as that of leucylation toward tRNAs ( Figure 1A ) prompts us to speculate that the the cognate leucine tRNAs of S.cerevisae (0.34 µM; Soma nucleotide at position 37 is strongly associated with et al., 1996) . leucylation, because all tRNA Ser CAGs possessing leucyl-In order to verify that the leucylation activity observed ation activity have m 1 G in common, while only the for the tRNA Ser CAG of C.zeylanoides actually came from tRNA Ser CAG of C.cylindracea, which possesses no leucylthe tRNA Ser CAG itself, and not from a trace amount ation activity, has A at this position. of leucine tRNA contaminating the tRNA sample, the To examine the validity of this speculation, a series of leucylated 3Ј-terminal RNA fragment derived from leucyl-tRNA Ser CAG variants was constructed by the in vitro tRNA Ser CAG was analyzed in the following manner. 14 Ctranscription method using T7 RNA polymerase, as well leucylated tRNA Ser CAG from C.zeylanoides was first as by the microsurgery method, and the leucylation activity acetylated with acetic anhydride to prevent deacylation, of each variant was measured. When the tRNA Ser CAG of and then digested with RNase T1. The resulting 3Ј-C.zeylanoides synthesized by in vitro transcription was terminal fragment with 14 C-labeled acetylleucine was employed as a substrate, no leucylation activity was analyzed by cellulose TLC. The results are shown in detected, not even for the tRNA transcript having G37 Figure 1C . If leucylated tRNA Ser CAG were digested ( Figure 3A ). On the other hand, as shown in Figure 3A , with RNase T1, 14 C-labeled acetylleucyl-CCA should be serylation activity exceeded 1000 pmol/A 260 unit. These released as a labeled fragment ( Figure 1C , lane 3), because results strongly suggested that some nucleoside modifica-G is located at position 73 of the tRNA Ser CAG (Figure tion is necessary in tRNA Ser CAG for recognition by 1A, left-hand structure). Any contaminated leucine tRNAs, LeuRS. We thus attempted to replace the m 1 G37 of if they exist, will give some 14 C-labeled fragments larger C.zeylanoides tRNA Ser CAG with G (the variant is symbolthan the tetramer ( Figure 1C , lane 4), because all the ized as m 1 G37G) or A (m 1 G37A), by the microsurgery leucine tRNAs of yeasts so far analyzed (Sprinzl et al., method (Figure 2A and B; for details, see Materials 1996) including those of C.zeylanoides (T. Suzuki, unpuband methods) to examine the contribution of m 1 G37 to lished result) are known to have A73 at their 3Ј-ends, leucylation and the contribution of A37 of C.cylindracea which are resistant to RNase T1. The mobility of the tRNA Ser CAG to the prevention of leucylation. acetylleucyl-oligonucleotide derived from tRNA Ser CAG When aminoacylation of m 1 G37A and m 1 G37G was from C.zeylanoides ( Figure 1C , lane 1) was identical to examined ( Figure 3A ), the results indicated that both that of acelylleucyl-CCA prepared from the RNase U2 substitutions lead to complete loss of leucylation (Figure digests of leucyl-tRNA Leu s from C.zeylanoides (lane 3). 3A, right-hand graph), although no apparent influence was This observation clearly demonstrates that leucine is observed on serylation ( Figure 3A , left-hand graph). These definitely attached to the tRNA possessing G73; the tRNA findings strongly indicate that the methyl group of m 1 G37 therefore must be tRNA Ser CAG and not tRNA Leu . Thus, plays a crucial role in enhancing the leucylation activity it is concluded that the tRNA which incorporated leucine of tRNA Ser CAG. in vitro is in fact tRNA Ser CAG. This deduction is supported The slight reduction in leucylation activity observed in by the results of an additional experiment: incorporation the control variant z-G33G ( Figure 2A ) compared with of [ 14 C]leucine into the tRNA Ser CAG sample with LeuRS native tRNA ( Figure 3A , right-hand graph) was found to was reduced by the addition of SerRS and non-labeled have resulted from the partial deacetylation of 4-acetyl serine to the reaction mixture (data not shown), which cytidine (ac 4 C) due to acid treatment of the 5Ј-half clearly indicates that the same tRNA molecule is comfragment of tRNA Ser CAG (see Materials and methods). petitively aminoacylated by these two enzymes. This is considered further in the Discussion. To conclude that tRNA Ser CAG is aminoacylated with leucine, we carried out a further experiment. The G33 acts as a modulator of leucylation tRNA Ser CAG was charged with serine and serylated In addition to m 1 G37, another unique feature of the serine tRNA Ser CAGs in these Candida species is the presence tRNA Ser CAG was separated from non-aminoacylated The effect of mutation at position 33 in these two each variant was confirmed to have been replaced as expected (shown tRNAs was found to be quite different. In the case of the by arrows). C.cylindracea tRNA, none of the mutations at position 33 caused leucylation of the tRNA, as was observed with the native tRNA Ser CAG, and there was no reduction in of G at position 33, where a pyrimidine (mostly U) is completely conserved in usual tRNAs (Sprinzl et al., serylation activity ( Figure 3C ). In contrast, the replacement of G33 by pyrimidines in C.zeylanoides tRNA Ser CAG 1996). Since we considered it is possible that this notable feature may be in some way related to the unusual considerably enhanced the leucylation activity ( Figure 3B , right-hand graph), while no significant difference was aminoacylation characteristics described above and/or to the translation of non-universal genetic code, we examined observed in the serylation activity ( Figure 3B , left-hand graph). The kinetic parameters of leucylation for the the effect of residue 33 on the aminoacylation and transla- show the spots corresponding to acetylleucine and acetylserine as markers, respectively. (D) Analysis of acetylamino acids attached to tRNA fragments on a TLC plate. Lane 2 shows the spot corresponding to the acetylamino acids derived from the RNase T1 fragment of C.zeylanoides tRNA Ser CAG. Lanes 1 and 3 indicate the spots corresponding to acetylleucine and acetylserine, respectively. Ten micrograms of [ 14 C]acetylaminoacyl-tRNA Ser CAG from C.zeylanoides was digested with RNase T1 and developed on cellulose TLC plates under the same conditions as (C). CCA fragments with [ 14 C]acetylamino acids were scraped from the plate from which the fragments were eluted with H20 and desalted by Sep-pak C18 under the conditions described in the literature (Wang et al., 1990) . [ 14 C]acetylamino acids discharged from the fragments were developed on TLC and visualized by an imaging analyzer (BAS-1000, Fuji Photo Systems). variants of C.zeylanoides tRNA are shown in Table I . It Evidence for leucylation of C.zeylanoides tRNA Ser CAG in vivo is notable that the K m values of the two pyrimidine At this point, we had established that the tRNA Ser CAG mutants, z-G33U (1.4 µM) and z-G33C (1.3 µM), are of C.zeylanoides is actually able to accept leucine in vitro. clearly lower than those of the two purine mutants, z-G33A However, considering the facts that SerRS and LeuRS (6.7 µM) and z-G33G (5.6 µM). The V max value of z-G33U coexist in cells and, judging from their K m values, that (1.2 pmol/min) is 39% of that of z-G33C (3.1 pmol/min), the affinity of tRNA Ser CAG toward SerRS is one order of which could explain why z-G33U shows lower leucylation magnitude higher than that toward LeuRS, we needed to activity than z-G33C despite having nearly the same K m ascertain whether the tRNA Ser CAG of C.zeylanoides is in value ( Figure 3B , right-hand graph). Judging from the fact leucylated in vivo. For this purpose, we adopted a sequence analysis (data not shown), the slight reduction newly developed method for quantifying an individual in the leucylation of z-G33G (5.6 µM) compared with aminoacyl-tRNA in cells (Suzuki et al., 1996) . that of the native tRNA Ser CAG (5.0 µM) is probably due Aminoacyl-tRNAs separately prepared from cells of to the partial deacetylation of ac 4 C at position 12, as C.zeylanoides and C.cylindracea were immediately submentioned above. This was confirmed by the observation jected to acetylation using [1-14 C]acetic anhydride to label of a slight reduction in leucylation activity also in acidthe amino acids as well as to stabilize the aminoacylated treated native tRNA Ser CAG (data not shown). It is thus tRNAs. From each of the acetylated aminoacyl-tRNA concluded that replacement of a pyrimidine by a purine mixtures, tRNA Ser CAGs from C.zeylanoides and C.cylindat position 33 has a repressive effect on leucylation of the racea were fished out by a solid-phase-attached DNA tRNA Ser CAG of C.zeylanoides. probe as described previously (Tsurui et al., 1994 ; Wakita The translation efficiencies of the variants with a muta et al., 1994) . A single band for each of the aminoacyltion at position 33 were also examined in a cell-free tRNAs was detected by staining ( Figure 4A ) with which translation system of C.cylindracea (Yokogawa et al., the radioactivity coincided in each case ( Figure 4B ). 1992; Suzuki et al., 1994) , to evaluate the effect of G33. Acetylated amino acids attached to these tRNAs were A change from G to U at position 33 apparently enhanced deacylated by alkaline treatment and analyzed by TLC. the translation activity 2.5-fold, although their decoding As shown in Figure 4C , acetylserine was observed as a properties did not change at all (data not shown). We thus major amino acid derivative in both tRNA Ser CAGs, but consider that G33 serves as a modulator of leucylation of acetylleucine was detected only in the C.zeylanoides tRNA Ser CAG, despite a slight disadvantage in transla-tRNA Ser CAG; the acetylserine and acetylleucine spots were identified as described previously (Suzuki et al., tion activity. 1996) . The radioactivities remaining on the origins probably came from the direct acetylation of some nucleotides in the tRNAs, as discussed previously (Suzuki et al., 1996) . From comparison with the radioactivity of acetylserine, it was calculated that~3% of the tRNA Ser CAG was attached with acetylleucine. These results were reproducible. Digestion of purified acetyl-aminoacyl tRNA Ser CAG with RNase T1 also gave only a 14 C-labeled CCA fragment, as shown in Figure 1C . When the acetylated amino acid released from the fragment purified from the corresponding spot on TLC was analyzed by TLC, the ratio of acetylleucine to acetylserine was also found to be 3% ( Figure 4D ), indicating that acetylleucine is covalently attached to the tRNA Ser CAG fragment with G73. It thus became clear that the tRNA Ser CAG of C.zeylanoides was in fact charged with leucine by 3% of the amount of serylation of the same tRNA Ser CAG in C.zeylanoides cells. Aminoacylation has generally been considered to be the final stage determining translational accuracy (reviewed in Parker, 1989; Kurland, 1992; Farabaugh, 1993) . However, in the case of tRNA Gln charged with glutamate in the chloroplast, Glu-tRNA Gln is rejected by an elongation factor so that the chloroplast translation machinery does not employ the mischarged aminoacyl-tRNA (Stanzel et al., 1994) . It is likely that this is an exceptional case due to the lack of glutamyl-tRNA synthetase in the chloroplast. In order to prove that leucylated tRNA Ser CAGs actually participate in the translation process in Candida cells without such a rejection mechanism, we utilized a URA3 gene expression system derived from S.cerevisiae in C.maltosa, which was developed by Sugiyama et al. (1995) . Candida maltosa utilizes the codon CUG as serine and possesses the relevant tRNA Ser CAG gene (Sugiyama et al., 1995; Zimmer and Schunck, 1995) . Since the et al., 1995) . In the present study, this URA3 gene, with the CTG codon replaced by various leucine or serine codons, was utilized as a marker gene ( Figure 5A ). First, ADE1/ura3::C-ADE1) (Ohkuma et al., 1993) , the growth of which was monitored on minimal medium SD plates a plasmid in which the S.cerevisiae URA3 gene was inserted downstream of a C.maltosa-specific promoter in the presence and absence of uracil. When uracil was supplied to the SD plate for the (C-p) was constructed and designated as pCSU-CTG (Sugiyama, 1995) . As controls, mutant plasmids of pCSU-positive control experiments, all the transformants grew normally ( Figure 5B , middle row). However, in the absence CTG, in which the codon CTG was replaced by either the serine codon TCT or the leucine codon CTC, were of uracil, cells harboring pCCU and pCSU-CTC showed normal growth, whereas no growth was observed in those constructed and named pCSU-TCT and pCSU-CTC, respectively. In addition, a plasmid (pCCU) consisting of harboring pCSU-TCT and pUTH18 that contained no URA3 gene insertion. Cells harboring pCSU-CTG showed the URA3 gene of C.maltosa having a CTT leucine codon at the corresponding site, combined with the C.maltosa-weak but significant growth ( Figure 5B, uppermost row) . These results demonstrate that if the codon at position 45 specific promoter, was also used as a positive control. These variant plasmids were introduced into a URA3-is translated as leucine, active ODCase will be produced and the cells will be able to grow, but translation of the defective C.maltosa strain CHU1 (his5, ade1, ura3::C-codon with serine will produce inactive ODCase and the We believe that tRNA Ser CAG is the only molecule responsible for the leucine insertion corresponding to cells will be unable to grow. The result with cells harboring pCSU-CTG clearly demonstrates that the URA3 mutation codon CUG in C.maltosa cells, based on the following obervations. We have purified and sequenced a number on the C.maltosa chromosome was in some way complemented by the introduced pCSU-CTG plasmid, suggesting of leucine and serine tRNAs from Candida species, in which codon CUG is translated as serine, and failed in that the CTG codon was read at least partially as leucine in C.maltosa cells possessing tRNA Ser CAG. finding tRNA with the anticodon sequence potentially complementary to codon CUG other than tRNA Ser CAG In order to quantify the growth rate of the cells harboring pCSU-CTG, the viability of the cells was examined in (Yokogawa et al., 1992; Ohama et al., 1993; Suzuki et al., 1994; Ueda et al., 1994; our unpublished observation) . liquid medium without uracil. As shown in Figure 5C , whereas translation of the CTG codon as serine completely Futhermore, tRNA genes for serine and leucine from these Candida species were sequenced following the blocked cell growth in the case of pCSU-TCT, and full complementation was observed in the case of pCSU-CTC amplification by cloning and/or PCR methods, and we found that only tRNA Ser CAG is able to translate codon in which the CTC codon was read as leucine, intermediate cell growth was observed in the case of pCSU-CTG, CUG (Yokogawa et al., 1992; Ohama et al., 1993; Suzuki et al., 1994; Ueda et al., 1994 ; our unpublished observ-indicating that ODCase was expressed in an active form, albeit at a low level, when there was a slight incorporation ation). Thus, it could be concluded that only the tRNA Ser CAG species inserts leucine into polypeptide of leucine at the CTG codon. The slow growth of the cells harboring pCSU-CTG was not due to the spontaneous corresponding to codon CUG. reversion of the CTG codon to another leucine codon or due to any other mutation, because the cells harvested Discussion from the colony on the SD-plate show the same growth phenotype. These results are unlikely to reflect the different The observations presented here clearly demonstrate that, in certain living organisms, a single codon can be simul-expression levels of the URA3 gene variants because the URA3 mRNA level is not altered by mutations at position taneously assigned to two distinct amino acids. Most codons in the genetic code degenerate, but our findings 45 (Ohkuma, 1993) . Furthermore, the possibility that the URA3 gene with CTG at position 45 is translated more show that some amino acids are also able to degenerate with respect to a particular codon. Such codon ambiguity efficiently than the gene with TCT at the same site due to codon preference (Ikemura, 1982) is excluded by the is governed by a tRNA acceptable to two amino acids simultaneously, as described above. We propose to desig-fact that the TCT codon is the most preferred of all the serine codons, including the CUG codon, in C.maltosa nate a codon corresponding to multiple amino acids a 'polysemous codon'. (Sugiyama et al., 1995) . ODCase activity resulting from the translation of the A high degree of accuracy in tRNA aminoacylation has been considered crucial for preserving fidelity in protein URA3 gene was examined in the presence of a pyrimidine analog, 5-fluoroorotic acid (5FOA), an inhibitor in synthesis. It has been established that aminoacyl-tRNA synthetase is able to discriminate precisely its cognate pyrimidine biosynthesis. Incorporation of 5FOA with ODCase results in the formation of 5-fluorouridylate, amino acid from other structurally related amino acids at the adenylation reaction step, and its cognate tRNAs from which is harmful to cell propagation (Boeke et al., 1984) . Thus, URA3-defective strains grow normally on a medium non-cognate ones (reviewed in Parker, 1989; Kurland, 1992) . The misacylation error in this process has been containing 5FOA, whereas cells possessing the active URA3 gene are unable to grow on this medium. Cells estimated to range between 10 -4 and 10 -5 (Lin et al., 1984; Okamoto et al., 1984) . Discrimination of cognate harboring the respective plasmids were cultivated in the presence of 5FOA in addition to uracil. tRNA from non-cognate tRNAs is mediated by positive and negative identity determinants localized on the tRNA As shown in the bottom row of Figure 5B , cells harboring pCSU-CTG exhibited similar growth on the molecule (Yarus, 1988; Normanly and Abelson, 1989) . The only exception reported so far is that tRNA Gln is agar plate to those with pCSU-TCT and pUTH18, although the transformants with pCSU-CTC and pCCU were unable aminoacylated with glutamate in Gram-positive bacteria and in some organelles (Lapointe et al., 1986 ; Schön to grow. These results indicate that the CTG codon at position 45 was mainly translated as serine in C. maltosa, et al., 1988) . However, this differs from misaminoacylation in that this process is indispensable to compensate for the so as to produce the inactive ODCase. However, when the liquid medium was supplied with 5FOA, a slight lack of glutamyl-tRNA synthetase in these organisms. In general, high fidelity in the aminoacylation process is reduction in the growth rate was observed in the case of pCSU-CTG, compared with pCSU-TCT ( Figure 5C ), considered to be indispensable for translating genes into functionally active proteins with a high degree of accuracy. while very slow growth was observed in the case of pCSU-CTC used as a control. In order to detect a low The discovery of a polysemous codon in a Candida species contradicts the established notion of aminoacyl-level of ODCase activity arising from a slight incorporation of leucine at the CUG codon in the 45th position, we ation with high fidelity. We have shown that a single tRNA is acceptable to two different amino acids, and adjusted the ratio of 5FOA and uracil as shown in Materials and methods. This growth rate reduction clearly suggests that it can therefore transfer two different amino acids corresponding to a particular codon. The expression that the slow growth observed in the SD medium was due to low expression of active ODCase. Thus it is concluded experiment using the ODCase-encoding URA3 gene containing codon CUG at the site essential for its activity that the CUG codon is partially translated as leucine in C.maltosa cells. (see also Sugiyama et al., 1995) suggested that leucine could be incorporated into the gene product corresponding on experiments using an artificial mutation, and it does not reflect experimental observation in an extant living to codon CUG in C.maltosa, as judged from the complementation tests with the URA3 mutation. Although the organism. On the basis of peptide sequences, several research amount of leucine incorporated per CUG codon was not quantitatively determined, it is clear that the incorporation groups have reported that codon CUG corresponds only to serine in C.maltosa (Sugiyama et al., 1995) and was mediated by the leucyl-tRNA Ser CAG. We thus concluded that codon CUG was simultaneously assigned to C.albicans (Santos and Tuite, 1995a; White et al., 1995) . No leucine-inserted peptide was detected in these studies. serine and leucine in the normal translation process in C.maltosa. A quantitative analysis of the amino acids However, we consider that any peptide with a leucine which was inserted for the codon CUG might have been attached to the tRNA indicated that 3% of tRNA Ser CAG is leucylated in C.zeylanoides cells. Such a high level of missed during purification or was undetectable in the peptide sequencing, because the amount of leucine-inserted leucylation is far beyond conventional misacylation, whose rate is estimated to be less than 10 -4 . Unless a proofreading peptide (~3%) would have been too low to be positively identified in sequencing experiments. mechanism exists on the ribosome, incorporation of leucine at CUG codon sites may reflect the relative ratio of We have shown that tRNA Ser CAG in Candida species is a chimera of tRNA Ser CAG and tRNALeuCAG in so far tRNA Ser CAG leucylation, which is two orders of magnitude higher than that of conventional mistranslation. as it is the substrate for both SerRS and LeuRS. The K m value for LeuRS is 5.0 µM, which is only one order of To date, artificial manipulations of molecules participating in the translation process, such as the overproduction magnitude larger than that for SerRS (0.22 µM). In an in vitro aminoacylation experiment Ͼ30% of tRNA Ser CAG of aminoacyl-tRNA synthetase (Swanson et al., 1988) , mutations of tRNAs etc. and/or control of growth condi-subjected to the reaction could be converted to leucyl-tRNA Ser CAG using an increased amount of LeuRS and a tions, such as deprivation of amino acids in the medium (Edelmann and Gallant, 1977; O'Farrell, 1978; Parker and longer incubation time (data not shown). We observed that while the presence of SerRS and non-radioactive Precup, 1986), have been found to increase the error rate in translation (reviewed in Parker, 1989) . However, our serine reduced leucylation, complete loss of leucylation could not be achieved (data not shown), indicating that the observation is based on experiments using wild-type cells grown in a rich medium suitable for high viability. In affinity of LeuRS toward tRNA Ser CAG is relatively high. In proliferating cells of C.zeylanoides, the leucyl-these respects, the polysemous codon is a phenomenon completely different from these artificial translational tRNA Ser CAG in the cells was estimated to be 3% of the seryl-tRNA Ser CAG, which is much lower than that errors. It is known that many examples exist for alternative decoding of universal codons-initiation codons other obtained in the in vitro experiments. We consider that such a reduction in leucylation is due to the competition than AUG (Gold, 1988; Kozak, 1983) , leaky stop codons caused by nonsense suppresser or native tRNAs (Murgola, for the tRNA Ser CAG between SerRS and LeuRS in the cells. Despite this competition, the distinct detection of 1985), the UGA codon used for incorporation of selenocysteine (Leinfelder et al., 1988) and so on. However, leucylated tRNA Ser CAG in vivo supports the existence of an ambiguous aminoacylation reaction toward the single because of strong dependence on the context effects or possible secondary structures of mRNAs, these recoding tRNA Ser CAG species. The polysemous codon results from the coexistence of events are those which are programed in the mRNAs (Gesteland et al., 1992) . We have sequenced several genes tRNA identity determinants for serine and leucine in a single tRNA molecule. Construction of tRNA Ser CAG in Candida genomes, but we could not find any secondary structure around the codon CUG in these genes. Con-variants by the microsurgery method led to the finding that a single methyl moiety of m 1 G at position 37 sidering that the polysemous codon is mediated by a single tRNA, it is unlikely that a polysemous codon occurs under is involved in the leucylation process. In contrast, the tRNA Ser CAG of C.cylindracea, which has A at the same the influences of the neighboring regions in mRNAs. Alternative decoding of a polysemous codon CUG is position, is deprived of such leucine-accepting activity. Himeno and his co-workers noted that three nucleotides possible, assuming that LeuRS is overexpressed under a certain physiological condition. Depending on the of leucine tRNAs were strongly recognized by S.cerevisiae LeuRS using unmodified variants transcribed by T7 RNA increased amount of the LeuRS in cells, incorporation of leucine corresponding to codon CUG may occur fre-polymerase (Soma et al., 1996) . Although the discriminator base, A73, is the strongest recognition site among them, quently, which causes the production of polypeptides with new functions. This possibility should be examined in A35 and G37 in the anticodon loop also play roles as determinants in tRNA. They were able to compare the further experiments. The idea of a polysemous codon also differs from the activities of variants mutated at position 37 with A or G using the variants with A at the discriminator position 'near-cognate' concept proposed by Schultz and Yarus (1994) . They claimed that ambiguous decoding may occur which effectively elevates leucylation activity. In our work, we utilized serine tRNA with a modified nucleoside as a consequence of an irregular codon-anticodon interaction induced by the 27-43 base pair at the anticodon and with G at the discriminator position as a substrate for LeuRS, because the T7 transcript of tRNA Ser CAG showed stem of the tRNA, resulting in a genetic code change transition state. The polysemous codon found in our study no activity for leucylation. Our experiments using microsurgery methods indicated that m 1 G is of great importance is caused by the tRNA aminoacylation process of tRNA with codon-anticodon interaction proceeding precisely in in leucylation, despite the fact that the presence of G at the discriminator position is unsuitable for the recognition the conventional manner . Furthermore, since the hypothesis of Schultz and Yarus is based of LeuRS. Some modified nucleotides in tRNA are known to be involved in recoginition of some synthetases (Muramatsu et al., 1988) . Pütz et al. (1994) showed that m 1 G at position 37 of yeast tRNA Asp is one of the negative determinants for arginyl-tRNA synthetase. We have also demonstrated that the nucleotide at position 33, where only tRNA Ser CAG uniquely possesses G, modulates the leucine-accepting activity. G33 may prevent tRNA Ser CAG from excessive leucylation, which S.cerevisiae cells, but that the viability of the cells decreased substantially. This finding suggests the polysemous state may be tolerated only when the ambiguous recognizes its cognate leucine tRNA from the 3Ј-side of the anticodon loop, which is afforded by the uridine-turn translation is under a strict constraint. We consider that G33 functions as a negative modulator in the leucylation structure due to U33 ( Figure 6A ). The methyl moiety of m 1 G37 is directly recognized by LeuRS. In the case of tRNA Ser CAG, thereby controlling the relative seryl-to leucyl-tRNA Ser CAG ratio. of C.zeylanoides, the anticodon loop distorted by G33 decreases the affinity toward LeuRS, judging from the Several lines of experiment have suggested that U33 is involved in the tRNA function on ribosomes, such as in observation that G33 increased the K m value for leucylation approximately 4-to 5-fold in comparison with that with rigid codon-anticodon interaction, proper GTP hydrolysis of the ternary complex and the efficient translation of prymidine bases at the position ( Figure 6B ). In C.cylindracea, m 1 G is replaced by A, which means that the tRNA termination codons (Bare et al., 1983; Dix et al., 1986) . Indeed, the replacement of G33 by U in C.cylindracea has lost the two major determinants for LeuRS, m 1 G and the discriminator base ( Figure 6C ). Consequently, LeuRS tRNA Ser CAG increased the efficiency of in vitro translation by 2-to 3-fold (data not shown). The negative effect of is unable to recognize tRNA Ser CAG at all, and G33 concomitantly loses its function as a modulator. LeuRS G33 on translation may indicate involvement in some mechanism for decoding the polysemous codon. This is, of course, unable to recognize other serine isoacceptor tRNAs corresponding to universal codons, because they possibility needs to be clarified by further study. Nevertheless, we have shown here that one of the roles of G33 is have modified A at position 37. How did this interaction between LeuRS and the suppression of leucylation, and we consider that the nucleotide at position 33 is not directly involved in tRNA Ser CAG evolve? Candida species utilizing CUG as serine can be classified into two distinct groups: group 1 recognition by LeuRS. On the basis of our observation that no leucylation was detectable in the C.cylindracea contains the species that have tRNA Ser CAG with leucylation activity, and includes C.zeylanoides, C.maltosa and tRNA Ser CAG variants in which G33 was replaced by a pyrimidine base (c-G33U and c-G33C), we speculate that others (see Figure 6B ); group 2, which is represented solely by C.cylindracea, contains species that have tRNA Ser CAG G33 influences the location and/or conformation of m 1 G37, accompanied by the alteration of the anticodon loop without leucylation activity ( Figure 6C ). A plausible evolutionary process is that group 1 would have arisen structure, decreasing the affinity of LeuRS toward tRNA Ser CAG. prior to group 2 after the genetic code change, which is speculated on the basis of the following observations. It has been generally considered that reconstructed tRNA does not lose its activity during the several reaction First, the homology between tRNA Ser CAGs in group 1 and its isoacceptor tRNAs for codon UCG is higher than steps needed in the microsurgery method, such as cleavage of the tRNA strand and ligation of tRNA fragments that between the tRNA Ser CAG from C.cylindracea and its isoacceptor . Second, C.cylindracea (Ohyama et al., 1985) . However, a slight reduction of leucylation activity was observed in the control variant, (group 2) possesses high copy numbers of the tRNA Ser CAG genes (~20 copies) on the diploid genome (Suzuki z-G33G, compared with that of the native tRNA ( Figure 3A , right-hand graph), which turned out to result from et al., 1994) , while low copy numbers (two or four copies) are observed for group 1 tRNA Ser CAG genes (Santos the partial deacetylation of 4-acetyl cytidine (ac4C) due to acid treatment of the 5Ј-half fragment (see Materials et al., 1993; Sugiyama et al., 1995; T.Suzuki, personal observations) . Third, the codon CUG is utilized as a major and methods). Nevertheless, it is reasonable to deduce the effect of base replacement on the aminoacylation activity serine codon on several genes in C.cylindracea, such as lipase (Kawaguchi et al., 1989) and chitin synthase by comparing the activities of these reconstructed tRNAs, because the same 5Ј-half fragments were used for all the (unpublished results) , while CUG appears infrequently on the genomes of other species belonging to group 1 (Lloyd manipulated tRNA molecules of C.zeylanoides. A plausible mechanism by which LeuRS could recog-and Sharp, 1992; Sugiyama et al., 1995; T.Suzuki, personal observations) . During the course of the change in the nize cognate leucine and serine tRNAs specific for codon CUG is illustrated in Figure 6 . LeuRS contacts and genetic code, the genome should pass through a state group 2 (Ohama et al., 1993) . Fourth, the phylogenetic dextrose) and minimal medium SD [0.67% yeast nitrogen base without tree of these species and relatives constructed by using amino acids (Difco) and 2% dextrose] supplied with 24 mg/ml uracil were used for the cultivation of yeast cells. several genes also supports this evolutionary pathway SD-plates with or without uracil were prepared by adding agar at a (manuscript in preparation (Boeke et al., 1984) . for codon UCG Pesole et al., 1995) . Thus, the nucleotide at position 37 seems likely to have In order to introduce mutation at the 45th codon in the reading frame mutated in the direction modified A→m 1 G- Alternative splicing generates a multiple protein pUTH18 containing an autonomously replicating sequence of C.maltosa (Takagi et al., 1986) and C-HIS5 (Hikiji et al., 1989) were used as sequence from a single gene at the mRNA level. In when the codon appears infrequently, as observed in group was mutated from CTG to CTC, and pCCU (Sugiyama et al., 1995) synthesis caused by a polysemous codon. We speculate instruction manual. The electrified cells were spread on a SD-plate that such ambiguity could have given rise to proteins containing uracil and incubated at 30°C. with multiple amino acid sequences in non-house-keeping genes, which may have conferred multifunctionality on In vitro aminoacylation assay Seryl-or leucyl-tRNA synthetases were partially purified from C.zeylan-the proteins. Since the C.cylindracea strain was developed oides cells as described previously , both of the industrially for the production of lipase, such multifunc- Large-scale purification of tRNA Ser CAGs from C.zeylanoides and C.cylindracea mmol) and leucine (11.5 MBq/mmol) were from Amersham. 5-fluoroorotic acid monohydrate (5FOA) was from PCR inc. 3Ј-Biotinylated Candida cylindracea cells (3.1 kg) were treated with phenol, from which 150 000 A 260 units of unfractionated tRNA were extracted. Eighty DNA probes were synthesized by Sci. Media, Japan. Synthetic RNA oligomers and a chimeric oligonucleotide composed of DNA and 2Ј-O-thousand A 260 units of tRNA mixture were obtained by DEAE-cellulose chromatography with stepwise elution, which was then applied onto a methyl RNA were synthesized by Genset Co. Ltd. Most of the enzymes used for the microsurgery were from Takara Shuzo (Tokyo, Japan). DEAE-Sephadex A-50 column (6ϫ100 cm). Elution was performed with a linear gradient of NaCl from 0.375 to 0.525 M in a buffer Other chemicals were obtained from Wako Chemical Industries. consisting of 20 mM Tris-HCl (pH 7.5) and 8 mM MgCl 2 . The fraction as CZE-37 (5ЈGmCmCmCmAmAmUmGmGmAmAmdCdCdTdG-CmAmUmCmCmAmUm3Ј), possessing a cleavage site between posi-rich in tRNA Ser was applied onto a RPC-5 column (1ϫ80 cm) and eluted with a linear gradient of NaCl from 0.4 to 1 M NaCl in a buffer tions 37 and 38 of C.zeylanoides tRNA Ser CAG. Two hundred micrograms of purified tRNA Ser CAG from C.zeylanoides was incubated at 65°C for consisting of 10 mM Tris-HCl (pH 7.5) and 10 mM Mg(OAc) 2 . As a result of these chromatographies, 300 A 260 units of purified tRNA Ser CAG 10 min with 14.4 nmol CZE-37 in a buffer consisting of 40 mM Tris-HCl (pH 7.7), 0.5 mM NaCl, 0.1 mM DTT, 0.0003% BSA and 0.4% were finally obtained. One hundred and fifty thousand A 260 units of tRNA from C.zeylanoides glycerol (500 µl), and then annealed at room temperature. Magnesium chloride was added to the mixture up to a final concentration of 4 mM cells (3.7 kg) were fractionated on DEAE-Sepharose fast-flow column (3.5ϫ130 cm) with a linear gradient of NaCl from 0.25 to 0.4 M in a and the reaction was carried out at 30°C for 2 h by the addition of 600 units of RNase H (Takara Shuzo). About 60 µg of the cleaved 3Ј-half buffer consisting of 20 mM Tris-HCl (pH 7.5) and 8 mM MgCl 2 . About 300 A 260 units of C.zeylanoides tRNA Ser CAG were finally obtained by fragment was obtained by purification using 10% PAGE containing 7 M urea. Either of two synthetic oligo-RNAs, pCAGAp or pCAGGp, was further column chromatography with Sepharose 4B in a reverse gradient of ammonium sulfate from 1.7 to 0 M with a buffer consisting of ligated with the same 5Ј-half fragment digested by RNase T1 as the variants mutated at position 33 under the conditions described above. 10 mM NaOAc (pH 4.5), 10 mM MgCl 2 , 6 mM β-mercaptoethanol and 1 mM EDTA. The ligated and dephosphorylated 5Ј-half fragments were annealed and ligated with the 3Ј-half fragment digested by RNase H. About 50 µg of each of the two variants from C.zeylanoides mutated at position 37-Construction of tRNA variants with mutation at position 33 The microsurgery procedures were basically carried out according to the m 1 G37A and m 1 G37G-was obtained by the phosphorylation of the 5Јend and purification by 12% PAGE containing 7 M urea. literature (Ohyama et al., 1985 (Ohyama et al., , 1986 . Limited digestion of 4 mg purified tRNA Ser CAG from C.zeylanoides with RNase T1 was performed at 0°C for 30 min in a reaction mixture containing 50 mM Tris-HCl (pH 7.5), Identification of amino acids attached to tRNA Ser CAGs in 100 mM MgCl 2 , 0.5 mg/ml of the tRNA and 25 000 units/ml RNase the cells T1 (Sigma). After phenol extraction, the resulting fragments were treated Identification of aminoacyl-tRNA Ser CAG from Candida cells was carried with 0.1 N HCl at 0°C for 12 h in order to cleave the 2Ј, 3Ј cyclic out by a new method developed recently by us (Suzuki et al., 1996) . phosphate of the 3Ј-end of the fragments formed in the limited digestion, The experimental conditions were the same as those reported. To and then the 5Ј-and 3Ј-half fragments were separated by 10% PAGE fish out the aminoacyl-tRNAs, we designed two 3Ј-biotinylated DNA containing 7 M urea (10ϫ10 cm). Four hundred and thirty micrograms probes: 5ЈAGCAAGCTCAATGGATTCTGCGTCC3Ј for C.cylindracea of the 5Ј-half and 520 µg of the 3Ј-half fragments were recovered from tRNA Ser CAG and 5ЈGAAGCCCAATGGAACCTGCATCC3Ј for the gel. The purified 5Ј-half fragment was dephosphorylated with C.zeylanoides tRNA Ser CAG. These probes were immobilized with strepbacterial alkaline phosphatase (Takara Shuzo), and G33 at the 3Ј-end of tavidin agarose (Gibco BRL) as reported previously (Wakita et al., 1994) . the 5Ј-half fragment was removed by oxidation with sodium periodate as described in the literature (Keith and Gilham, 1974) . After dephos-
34
A universal BMV-based RNA recombination system—how to search for general rules in RNA recombination
At present, there is no doubt that RNA recombination is one of the major factors responsible for the generation of new RNA viruses and retroviruses. Numerous experimental systems have been created to investigate this complex phenomenon. Consequently, specific RNA structural motifs mediating recombination have been identified in several viruses. Unfortunately, up till now a unified model of genetic RNA recombination has not been formulated, mainly due to difficulties with the direct comparison of data obtained for different RNA-based viruses. To solve this problem, we have attempted to construct a universal system in which the recombination activity of various RNA sequences could be tested. To this end, we have used brome mosaic virus, a model (+)RNA virus of plants, for which the structural requirements of RNA recombination are well defined. The effectiveness of the new homomolecular system has been proven in an experiment involving two RNA sequences derived from the hepatitis C virus genome. In addition, comparison of the data obtained with the homomolecular system with those generated earlier using the heteromolecular one has provided new evidence that the mechanisms of homologous and non-homologous recombination are different and depend on the virus' mode of replication.
RNA recombination is a very common phenomenon. It has been observed in all types of viruses using RNA as a carrier of genetic information: in positive-sense, single-stranded RNA viruses (1) (2) (3) (4) , in negative-sense, single-stranded RNA viruses (5, 6) , in double-stranded RNA viruses (7, 8) and in retroviruses (9) (10) (11) . Moreover, it has been shown that RNA recombination enables the exchange of genetic material not only between the same or similar viruses but also between distinctly different viruses (12) . Sometimes it also permits crossovers between viral and host RNA (13) (14) (15) (16) (17) . Taking into account the structure of viral genomic molecules and the location of crossover sites, three basic types of RNA recombination were distinguished: homologous, aberrant homologous and non-homologous (3, 4, 18) . The former two occur between two identical or similar RNAs (or between molecules displaying local homology), while the latter involves two different molecules. Most of the collected data suggest that RNA recombinants are formed according to a copy choice model (4, 18) . A viral replication complex starts nascent RNA strand synthesis on one template, called RNA donor and then switches to another template, called RNA acceptor. Accordingly, two main factors are thought to affect RNA recombination: the structure of recombining molecules and the ability of the viral replicase to switch templates. To gain more knowledge of the mechanism of RNA recombination, several model experimental systems have been created. They provided us with some specific data describing homologous and/or non-homologous recombination in particular viruses, e.g. in poliovirus, (19) mouse hepatitis virus (20, 21) , brome mosaic virus (BMV) (4, 22) , turnip crinkle virus (23, 24) or tomato bushy stunt virus (25) . As a result, the involvement of viral replicase proteins in recombination has been demonstrated (26, 27) and a wide spectrum of RNA motifs supporting recombination have been identified (4, 23, (28) (29) (30) . In general, the collected data suggest that there exist two major types of RNA structural elements that induce recombination events: (i) universal ones mediating template switching by different viral replicases, e.g. regions *To whom correspondence should be addressed. Tel: +48 61 8528503; Fax: +48 61 8520532; Email: marekf@ibch.poznan.pl The online version of this article has been published under an open access model. Users are entitled to use, reproduce, disseminate, or display the open access version of this article for non-commercial purposes provided that: the original authorship is properly and fully attributed; the Journal and Oxford University Press are attributed as the original place of publication with the correct citation details given; if an article is subsequently reproduced or disseminated not in its entirety but only in part or as a derivative work this must be clearly indicated. For commercial re-use, please contact journals.permissions@oupjournals.org of local homology (28, 31) or complementarity (32) (33) (34) (35) and (ii) virus-specific ones, e.g. promoter-like structures (36, 37) . Unfortunately, up till now there has been no in vivo recombination system that could be used to test the recombination activity of any given RNA sequence and consequently to verify the above hypothesis and find some general laws governing the studied process. In our studies on genetic RNA recombination we have used the well-characterized in vivo system developed in BMV (30, 33) . BMV is a model (+)RNA virus of plants (38) . Its genome is composed of three segments called RNA1, RNA2 and RNA3. RNA1 and RNA2 encode BMV replicase proteins 1a and 2a, respectively. RNA3 encodes movement (3a) and coat proteins (CP) (38) . All three BMV RNAs possess an almost identical 3 0 -untranslated region (3 0 -UTR). The first BMV-based recombination system was created by Nagy and Bujarski (33) . They constructed a recombinationally active BMV mutant whose genome is composed of wtRNA1, wtRNA2 and modified RNA3 (PN0-RNA3 called the recombination vector, for details see Figure 1 ). Only 3 0 -UTR was modified in PN0-RNA3, while its 5 0 -UTR, intergenic and coding regions were unchanged. Despite the introduced changes, the recombination vector is stable and replicates when used together with wtRNA1 and wtRNA2 to infect plants. It starts to recombine if a recombinationally active sequence (RAS) is introduced just between the CP coding sequence and the modified 3 0 region (into the RAS-cloning site). Non-homologous recombination was observed when a 140-60 nt sequence complementary to RNA1 between positions 2856 and 2992 was inserted into PN0-RNA3 (a sequence from the 3 0 -portion of RNA1 was introduced in antisense orientation) (30, 33) . Interestingly, the same RNA1 fragment inserted in sense Figure 1 . The BMV-based recombination system. White, black and gray boxes represent coding, noncoding and recombinationally active sequences, respectively. The location of the primers (A and B) used for specific RT-PCR amplification of the 3 0 -portion of BMV RNA3 (parental or recombinant) is indicated by arrows. (A) BMV genome. The BMV genome consists of three RNA segments: RNA1, RNA2 and RNA3. All three BMV RNAs share an almost identical 3 0 -noncoding region with a tRNA-like structure at the very end. (B) Recombination vector. The PN0-RNA3 vector is a wtRNA3 derivative with a modified 3 0 -noncoding end [(for details see ref. (33) ]. The latter includes (i) the RAS cloning site, (ii) a 197 nt sequence derived from the 3 0 -noncoding region of cowpea chlorotic mottle virus RNA3 (marked as CCMV), (iii) the sequence of wtRNA3 between nt 7 and 200 (counting from the 3 0 end-marked as the region B) and (iv) the last 236 nt from the 3 0 end of BMV wtRNA1 (marked as A). (C) Non-homologous recombination. Non-homologous recombination was observed if an 140 nt sequence, called RAS1as (shown in A), complementary to wtRNA1 between positions 2856 and 2992 was inserted into the PN0-RNA3 vector. The presence of the RAS1as sequence in RNA3 derivative (called Mag1-RNA3) allows local RNA1-RNA3 hybridization that mediates frequent non-homologous crossovers. It is thought that polymerase starts nascent strand synthesis on the 3 0 end of RNA1 and then switches to RNA3 within a local double-stranded region. (D) Non-homologous recombinant. Non-homologous recombination repairs Mag1-RNA3 by replacing its modified 3 0 end with the 3 0 -noncoding fragment coming from RNA1. orientation did not support homologous crossovers (33) . Non-homologous recombination repaired the RNA3 vector by replacing its highly modified 3 0 end with 3 0 -UTR derived from RNA1. The resultant recombinants replicated and accumulated better than the parental RNA3 molecule, and so the latter was out competed from the infected cells. The above system is extremely efficient, since it employs selection pressure to support the accumulation of RNA3 recombinants. Because RAS is placed in two different segments of the BMV genome, we have proposed to name this system heteromolecular. Unfortunately, Nagy and Bujarski's BMV-based recombination system has one serious limitation. It was designed in such a manner that viable RNA3 recombinants can easily form only if a sequence derived from the 3 0 -portion of RNA1 or RNA2 is used as a RAS. Consequently, the heteromolecular system could not be applied for testing the recombination capacity of various RNA motifs. Olsthoorn et al. (39) attempted to solve that problem by inserting examined sequences into the 3 0 -noncoding region of BMV RNA2 and RNA3. This system was not further developed, since any changes in RNA2, which encodes BMV polymerase, could strongly affect the studied process. Here, we describe a new BMV-based recombination system. It has been constructed in such a way that both tested RASes are placed in the same segment of the BMV genome (in the modified RNA3 molecule); therefore, we have called this system homomolecular. To prove the usefulness of the homomolecular system, we have employed it to examine the recombination activity of sequences derived from the hepatitis C virus (HCV) genome. The examined sequences have been inserted into RNA3 as direct or inverted repeats. This demonstrated that the 101 nt hypervariable region of HCV efficiently supports both homologous and non-homologous crossovers, while the most conservative 98 nt portion of HCV's 3 0 -UTR induces only non-homologous recombination events. Moreover, a direct comparison of the hetero-and homomolecular systems revealed crucial differences between the mechanisms of homologous and non-homologous recombination. The former involves preferentially two different segments of the BMV genome and the latter occurs more easily between the same genomic RNAs. Plasmids pB1TP3, pB2TP5 and pPN0-RNA3 containing fulllength cDNA of BMV RNA1, RNA2 and modified RNA3 (recombination vector), respectively, were the generous gift from J. J. Bujarski (Northern Illinois University, DeKalb, IL). Restriction enzymes (EcoRI, SpeI and XbaI) T7 RNA polymerase, RNasine, RQ DNase RNase free, MMLV-reverse transcriptase, Taq polymerase and pUC19 cloning vector were from Promega. The following primers were used for the construction of pMatNH-pMatH-, pMatNH-HVR-, pMatH-HVR-, pMatNH-X-, pMatH-X-RNA3: Plasmids pMag1-and pMagH-RNA3 contain full-length cDNA of the RNA3 vector carrying the recombinationally active sequence RAS1 inserted in antisense or sense orientation, respectively. Both plasmids were constructed in the same way: pPN0-RNA3 was linearized with SpeI endonuclease and ligated with SpeI cut RAS1 cDNA. Then plasmids carrying RAS1 in antisense (pMag1-RNA3) and sense (pMagH-RNA3) orientation were identified (30) . To prepare pMatNH-RNA3 and pMatH-RNA3 plasmids (containing full-length cDNA of MatNH-RNA3 and MatH-RNA3), pMag1-RNA3 and pMagH-RNA3 were digested with KpnI and EcoRI endonucleases. Then, the deleted fragment was replaced with a KpnI-EcoRI cut 379 nt cDNA fragment corresponding to the BMV RNA1 3 0 end (containing the entire 3 0 -UTR and RAS1). The latter were obtained by PCR involving primers 1, 2 and pB1TP3 as a template. To construct pMat0-RNA3, i.e. a plasmid containing cDNA of the universal recombination vector Mat0-RNA3, the following modifications were introduced into pMatNH-RNA3. First, it was digested with SpeI endonuclease and religated. This way 5 0 RAS1as was removed and the 5 0 RAS cloning site (including only one restriction site SpeI) was created. Next, the plasmid was cut with KpnI and EcoRI to remove RNA1 3 0 -UTR and 3 0 RAS1s. Instead, a 295 nt fragment of RNA1 3 0 end (between positions 2940 and 3234) followed by the 3 0 RAS cloning site (including KpnI, MluI, BamHI and EcoRV restriction sites) was ligated into pMatNH-RNA3. The inserted sequence was obtained by PCR using primers 2, 3 and pB1TP3 as a template and digested with KpnI and EcoRI, prior to ligation. To test the recombination activity of HCV-derived sequences (hypervariable region 1, abbreviated HVR and sequence X, abbreviated X) cDNA of the corresponding fragments of the virus' genome was obtained by RT-PCR method (40, 41) and cloned into the pUC19 vector. Then, both tested sequences were amplified by PCR with primers introducing an SpeI restriction site. Primers 4, 5 and primers 6, 7 were used to obtain HVR and X cDNA, respectively. PCR products and pMat0-RNA3 were digested with SpeI and ligated. Then pMat0-RNA3 derivatives bearing HVR and X in sense and antisense orientation were identified. HVR and X were amplified again by PCR involving primers introducing MluI and EcoRV restriction sites (primers 8, 9 and 10, 11 to amplify HVR and X, respectively). PCR products were cut with MluI and EcoRV and ligated into the 5 0 RAS cloning site of previously identified pMat0-RNA3 derivatives (carrying HVR and X in sense and antisense orientation). As a result four plasmids were obtained: (i) pMatH-HVR-RNA3-containing cDNA of MatH-HVR-RNA3 in which two HVRs were inserted in sense orientation; (ii) pMatNH-HVR-RNA3-containing cDNA of MatNH-HVR-RNA3 possessing two HVRs, 5 0 HVR in antisense and 3 0 HVR in sense orientation; (iii) pMatH-X-RNA3containing cDNA of MatH-X-RNA3 in which two Xes are in sense orientation; (iv) pMatNH-X-RNA3-containing cDNA of MatNH-X-RNA3 carrying two Xes, 5 0 X in sense and 3 0 X in antisense orientation. Their structure was confirmed by sequencing. To test the recombination activity of the BMV mutants, the previously described procedure was applied (30, 33) . Infectious BMV genomic RNAs were obtained by in vitro transcription for which EcoRI linearized plasmids pB1TP3, pB2TP5, pMag1-RNA3, pMagH-RNA3, pMatNH-RNA3, pMatH-RNA3, pMat0-RNA3, pMatH-HVR-RNA3, pMatH-HVR-RNA3, pMatHN-X-RNA3 and pMatH-X-RNA3 were used. Five-leaf C.quinoa plants (local lesion host for BMV) were mechanically inoculated with mixtures containing BMV RNA1, RNA2 and one of the RNA3 derivatives. Two weeks post-inoculation, the number of lesions developed on each inoculated leaf was counted to establish the infectivity of the tested BMV mutant. Then, individual local lesions were excised and total RNA was extracted separately from every lesion. The isolated RNA was subjected to RT-PCR involving primer A (the first strand primer) and primer B (the second strand primer) specific for RNA3 3 0 fragment amplification (the region where recombination crossovers occur). As a control identical reactions involving either parental RNA3 transcript (positive control) or water (negative control) were carried out. RT-PCR products were analyzed by electrophoresis in a 1.5% agarose gel. The formation of 800 nt or shorter 500 nt products indicated that parental or recombinant RNA3 accumulated in the analyzed lesion, respectively. Next, RT-PCR products were cloned into the pUC19 vector and sequenced to determine the location of recombinant junction sites. Finally, the presence of recombinants in the selected local lesions was additionally confirmed by northern blot analysis. The main question that we had to answer during our studies was how to design a vector that could be used for examining the recombination activity of any RNA sequences in vivo. As a result, the idea arose to construct a BMV-based homomolecular recombination system. In such a system, both tested sequences are supposed to be present within the same segment of the BMV genome (either in RNA1, RNA2 or RNA3). Thus, a new vector should possess two separately located RAS cloning sites, be replicable and stable during infection. It has to be capable of generating viable recombinants, which have selective advantage over parental RNA molecules. Consequently, recombinants ought to be able to out compete the vector with inserted RASes. Assuming that RNA recombination occurs according to a copy choice mechanism, we decided that RNA3, being dispensable for BMV replication, is the best candidate for a new vector. Any changes in RNA1 and RNA2, which encode BMV replicase proteins, would strongly affect the studied process. The next important question was whether the location of RASes within the same (homomolecular system) or within two different segments of the BMV genome (heteromolecular system) influences the recombination activity of the examined RNA sequence. To address both issues, we decided to construct a so-called mixed system, homo-and heteromolecular at the same time. To this end two RNA3 molecules, prototypes of a new vector carrying two RASes, were prepared. To obtain them we used PN0-RNA3, described earlier, and a well-characterized recombinationally active sequence from BMV RNA1 (RAS1, see Figure 1 ). The 137 nt RAS1 corresponding to RNA1 between positions 2856 and 2992 was inserted into the PN0-RNA3 RAS cloning site, in antisense (RAS1as) and sense (RAS1s) orientations ( Figure 2 ). As a result, we obtained Mag1-and MagH-RNA3 derivatives (30) . Then the 356 nt portion of Mag1-and MagH-RNA3 3 0 end was replaced with a 379 nt sequence representing the wtRNA1 3 0 end (fragment encompassing the entire 3 0 -UTR and RAS1s sequence) ( Figure 2 ). In addition, a marker mutation (called DXho) was introduced within the RNA1-derived fragment to make it distinguishable from an analogous region present in wtRNA1. To this end, the XhoI restriction site (2988-2994) was disrupted by a 4 nt insertion (GATC) between C-2991 and G-2992. Resultant RNA3 derivatives, called MatNH-and MatH-RNA3, have unchanged 5 0 -UTR, intergenic and coding regions and a highly modified 3 0 -UTR. The latter includes 3 0 -UTR coming from wtRNA1 and two RAS1 sequences (3 0 RAS1 and 5 0 RAS1) separated by a 338 nt spacer (sequence CCMV and B1). In MatNH-RNA3, 3 0 RAS1 is located in sense and 5 0 RAS1 in antisense orientation, while in MatH-RNA3 both RAS1 sequences are in sense orientation ( Figure 2 ). Having these two RNA3 derivatives, we were able to construct two variants of the mixed system: one for homologous (MatH-BMV mutant) and the other for non-homologous (MatNH-BMV mutant) recombination studies. The MatH-BMV genome is composed of wtRNA1, wtRNA2 and MatH-RNA3 and the MatNH-BMV genome of wtRNA1, wtRNA2 and MatNH-RNA3 (Table 1 ). In genomes of both BMV mutants three copies of RAS1 are present: two in the RNA3 derivative (RAS1s-RAS1s or RAS1as-RAS1s in MatH-and MatNH-RNA3, respectively) and one in wtRNA1 (RAS1s). Thus, in the mixed systems two identical RASes or RAS and its complementary counterpart were capable of supporting, respectively, homologous or non-homologous (heteroduplex-mediated) recombination between the same or between different BMV genomic RNAs. As a result, we could directly compare homo-and heteromolecular recombination systems in one in vivo experiment and examine whether our presumptions concerning the new recombination vector are correct. Homologous recombination in the mixed homo-heteromolecular system Earlier, Nagy and Bujarski (33) demonstrated that the 66 nt portion of RAS1 did not support homologous recombination in heteromolecular system. We repeated this experiment using MH-BMV mutants. Recombinants also did not form although the entire RAS1 sequence was present in wtRNA1 and MagH-RNA3 molecules (Table 1 and Figure 3 ). To test RAS1 activity Figure 1 , white, black and gray boxes represent coding, noncoding and recombinationally active sequences, respectively, in sense (RAS1s) or antisense (RAS1as) orientation, dashed line squares encompass replaced parts of wtRNA1 and modified RNA3 molecules. Mag1-and MagH-RNA3 were created by inserting the RAS1 sequence from wtRNA1 into the RAS cloning site of PN0-RNA3 in antisense (Mag1-RNA3) or sense (MagH-RNA3) orientation. To construct MatNH-RNA3 and MatH-RNA3, the 356 nt very 3 0 end of Mag1-RNA3 or MagH-RNA3 (between KpnI and EcoRI sites) was replaced with a 379 nt portion of the wtRNA1 3 0 end (fragment containing the entire 3 0 -UTR and RAS1). Thus, both constructs contain two copies of RAS1 sequence-MatNH-RNA3 includes 5 0 RAS1as and 3 0 RAS1s, while MatH-RNA3 comprises 5 0 RAS1s and 3 0 RAS1s. Furthermore, a marker mutation DXho (marked as a white dot), removing the XhoI restriction site, was introduced into the 3 0 end of MatNH-and MatH-RNA3, to make it distinguishable from an analogous region present in wtRNA1. in the mixed homologous recombination system, a previously used, well-established procedure was applied (30, 33) . C.quinoa plants (local lesion host for BMV) were inoculated with a mixture containing in vitro transcribed wtRNA1, wtRNA2 and MatH-RNA3. After 2 weeks, when infection symptoms were well developed, the number of lesions formed on every inoculated leaf was counted to determine the infectivity of the MatH-BMV mutant. Individual local lesions were excised and total RNA was extracted separately from each of them. Then, the 3 0 -portion of RNA3 progeny accumulating in examined lesions was selectively amplified by RT-PCR involving RNA3 specific primers A and B (for their location see Figure 1 ). Reaction products were separated in a 1.5% agarose gel and their length was determined. The formation of an 800 or 400-500 nt DNA fragment indicated that the lesion contained parental or recombinant RNA3, respectively. In this way, we were able to determine the number of lesions in which a viable RNA3 recombinant was generated. The presence of recombinants in analyzed lesions was confirmed by standard northern blot hybridization. DNA fragments obtained during selective RT-PCR amplification of RNA3 were cloned and sequenced. Finally, the results obtained with our new mixed homologous recombination system were compared with analogous data previously got using the heteromolecular system (33) (see Table 1 and Figure 3 ). The data presented in Table 1 indicate that the exchange of MagH-RNA3 (carrying a single RAS1) into MatH-RNA3 (bearing two RAS1 sequences) did not affect the infectivity of the BMV mutants. The average numbers of lesions appearing on the leaves inoculated with MH-and MatH-BMV were similar: 18 and 17, respectively. Interestingly, although RAS1 did not support homologous crossovers in the heteromolecular system represented by MH-BMV, it was very active in the mixed system. About 85% of the local lesions developed during MatH-BMV infection accumulated the RNA3 recombinant instead of parental MatH-RNA3. In all of them, one RAS1 and a spacer were deleted. This indicates that crossovers occurred either within 3 0 end 5 0 RAS1 present in MatH-RNA3 (inter-or intramolecular crossovers) or within RAS1 and 5 0 RAS1 located in wtRNA1 and MatH-RNA3, respectively. Recombinant junction sites were placed in identical regions; therefore, their location could not be precisely established. The data presented till now also could not answer which molecules, exclusively MatH-RNA3 or wtRNA1 and MatH-RNA3, participated in recombination. In the heteromolecular system, only RAS1-mediated crossovers between wtRNA1 and MagH-RNA3 were permitted. The situation seems to be more complicated in the mixed homologous system where three copies of RAS1 are present, all in sense orientation: two of them in MatH-RNA3 (3 0 RAS1 and 5 0 RAS1) and one in wtRNA1. Consequently, RAS1-mediated homologous recombination may happen according to four different scenarios ( Figure 3 ). It can engage MatH-RNA3 only and occur as intra-or intermolecular process or it can involve wtRNA1 and MatH-RNA3. In the latter case, recombination can be mediated by RAS1 present in wtRNA1 and either 5 0 -or 3 0 RAS1 located in MatH-RNA3. To learn according to which scenario homologous recombination occurred, we checked whether mutation DXho introduced into MatH-RNA3 (just behind 3 0 RAS1) is still present in RNA3 recombinants. In this way, we were able to determine if their 3 0 -UTR was derived from MatH-RNA3 or wtRNA1 molecules. The undertaken analysis revealed that the mutation was present in 20% of recombinants. This result suggested that homologous crossovers preferentially occur between wtRNA1 and MatH-RNA3. However, there are other explanations why DXho was absent in a large fraction of recombinants. It is possible that the mutation was removed either from MatH-RNA3, due to homologous recombination between its 3 0 RAS1 and wtRNA1, or from the RNA3 recombinant (carrying a single copy of RAS1) because it could also have recombined with wtRNA1. To examine the first possibility, progeny RNA3 extracted from the local lesions accumulating MatH-RNA3 (lesions in which recombinant was not generated) was analyzed. About 800 nt RT-PCR products obtained during selective amplification of RNA3's 3 0 -portion were cloned and sequenced. In all of 20 analyzed clones DXho was present. To test the second possibility, a full-length cDNA clone of RNA3 recombinant containing DXho (RNA3-DXhoR) was obtained. It was inserted into the pUC19 vector under the T7 polymerase promoter. The resultant plasmid named pRNA3-DXhoR was used after linearization to produce an infectious RNA3-DXhoR molecule by in vitro transcription. Then, RNA3-DXhoR was used together with wtRNA1 and wtRNA2 to inoculate C.quinoa plants. After 2 weeks, total RNA was extracted from individual lesions and a 3 0 -portion of the progeny RNA3 was amplified by RT-PCR. Obtained products were cloned and sequenced. As described previously, DXho was present in all of the analyzed 20 clones. These two experiments proved that homologous recombination between either 3 0 RAS1 of MatH-RNA3 or RAS1 present in RNA3 recombinant and wtRNA1 does not occur frequently enough to explain why most recombinants lack DXho. Altogether, these results supported our initial thesis that DXho was removed from 80% of homologous recombinants, since most of the crossovers occurred within 5 0 RAS1 from MatH-RNA3 and RAS1 from wtRNA1. Earlier we showed that RAS1 can effectively support nonhomologous recombination if inserted into PN0-RNA3 in antisense orientation (30, 33) . The heteromolecular system used in our experiment was composed of wtRNA1, wtRNA2 and Mag1-RNA3 (M1-BMV mutant). Crossovers occurred within the local double-stranded region (local heteroduplex), which wtRNA1 and Mag1-RNA3 were capable of forming. In order to test RAS1 activity in the mixed non-homologous recombination system, C.quinoa plants were inoculated with the MatNH-BMV mutant (its genome is composed of wtRNA1, wtRNA2 and MatNH-RNA3). Two weeks later progeny RNA3 were analyzed as described above. The number of lesions developed on each leaf was counted and total RNA was extracted from individual local lesions. After RT-PCR amplification, the 3 0 -portion of BMV RNA3 accumulating in each lesion was analyzed in an agarose gel, cloned and sequenced. The presence of recombinants was confirmed by a standard Northern blot. The results obtained were compared with analogous data we had got using the heteromolecular system (30) (Table 1 and Figure 4) . As described previously, we observed that the exchange of Mag1-RNA3 (with a single RAS1as sequence) for MatNH-RNA3 (with two sequences: RAS1as and RAS1s) did not influence the infectivity of the BMV mutants. The average numbers of lesions developed on each leaf during infection with M1-BMV and MatNH-BMV were 19 and 18, respectively. There was also no difference between the recombination activity of M1-BMV and MatNH-BMV. RAS-1 (in fact RAS1s and RAS1as) supported non-homologous recombination . Non-homologous recombination in the heteromolecular and mixed homo-heteromolecular systems. As in Figure 3 , light and thicker lines represent viral genomic RNAs and nascent recombinant RNA, respectively. WtRNA1 is red (the recombinationally active sequence RAS1s which it contains is additionally boxed), wtRNA2 is black and modified RNA3 is blue. When inserted into RNA3 the wtRNA1-derived RAS1s sequence is also shown as a red box, whereas the complementary RAS1as sequence is shown as a green box. The portion of the RNA3 recombinant synthesized on wtRNA1 is red, the portion synthesized on the RAS1as sequence is green and the fragment synthesized on RNA3 is blue. The black dot symbolizes the DXho mutation present in MatNH-RNA3. RF, recombination frequency. Dashed line squares encompass the region identical in both systems [the region where crossovers occur, shown in detail in (E)]. (A) Heteromolecular system (M1-BMV). A detailed description of the M1-BMV genome is presented in Figure 1 . All nascent RNA3 molecules accumulating in M1-BMV infected plants were recombinants (RF = 100%). (B) Mixed system (MatNH-BMV). Three copies of RAS1 are located in the MatNH-BMV genome, two in MatNH-RNA3 (3 0 RAS1s and 5 0 RAS1as) and one in wtRNA1 (RAS1s). The recombination frequency observed during infection with MatNH-BMV was 95%. Of the identified recombinants, 10% were without the DXho marker, and 90% with the DXho marker. (C) Putative scenario of RAS1s/RAS1as-mediated non-homologous recombination in the heteromolecular system. Owing to the presence of RAS1s and RAS1as sequences in wtRNA1 and Mag1-RNA3, respectively, they are capable of forming a local double-stranded structure supporting non-homologous crossovers (for details see Figure 1 ). (D) Putative scenarios of RAS1s/RAS1as-mediated non-homologous recombination in the mixed system. The presence of 3 0 RAS1s and 5 0 RAS1as sequences in MatH-RNA3 and RAS1s in wtRNA1 creates several opportunities of heteroduplex formation: between wtRNA1 RAS1s and MatNH-RNA3 5 0 RAS1as (intermolecular), between two pairs of RAS1s/RAS1as sequences of two different MatNH-RNA3 molecules (intermolecular) and between 5 0 RAS1as and 3 0 RAS1s of the same MatNH-RNA3 molecule (intramolecular). Recombinants are generated if BMV replicase initiates nascent strand synthesis at the 3 0 end of wtRNA1 or MatNH-RNA3 and then switches to MatNH-RNA3 within the local double-stranded region. (E) Recombinants identified during M1-and MatNH-BMV infection. Boxed fragments of recombining wtRNA1/Mag1-RNA3, wtRNA1/MatNH-RNA3 and MatNH-RNA3/MatNH-RNA3 molecules are practically identical in both systems (except for the DXho mutation present in MatNH-RNA3). The locations of the junction sites are marked with arrows and letters. The numbers indicate how many recombinants of the same type were isolated. Upper case letters refer to M1-BMV, lower case letters refer to MatNH-BMV. equally in both systems. Recombination events occurred with a similar frequency (100 and 95% for M1-and MatNH-BMV, respectively) and recombinant junction sites were located within the same region of the heteroduplexes, which recombining molecules were capable of forming. In the heteromolecular system, only one type of heteroduplex supporting non-homologous crossovers could possibly form: between wtRNA1 and Mag1-RNA3. In the mixed system, recombining molecules were capable of forming three types of heteroduplexes: (i) intermolecular, between 5 0 RAS1as from MatNH-RNA3 and RAS1s from wtRNA1, (ii) intermolecular, between 5 0 RAS1as and 3 0 RAS1s located in two MatHN-RNA3 molecules and (iii) intramolecular, between 5 0 RAS1as and 3 0 RAS1s located in the same MatNH-RNA3 molecule (Figure 4) . To determine the molecules that participated in non-homologous recombination, the RT-PCR amplified 3 0 -portions of RNA3 were checked for DXho. It was present in 90% of recombinants. This result clearly showed that non-homologous crossovers almost always involve one (intramolecular recombination) or two (intermolecular recombination) MatNH-RNA3 molecules. The results presented above indicated that the homomolecular system can provide new interesting data concerning the mechanism of RNA recombination, especially if it could be used for testing the recombination activity of RNA sequences derived from other RNA-based viruses. Consequently, we attempted to construct a universal BMV RNA3-based recombination vector called Mat0-RNA3 (for details see Materials and Methods and Figure 5A ). In Mat0-RNA3, as in the former PN0-RNA3 vector, only 3 0 -UTR was modified. It is composed of the 295 nt very 3 0 end of RNA1 followed by the 3 0 RAS cloning site, a 338 nt spacer and the 5 0 RAS cloning site. To determine the infectivity and stability of the new vector, C.quinoa plants were inoculated with a mixture containing wtRNA1, wtRNA2 and Mat0-RNA3 (Mat0-BMV mutant). After 2 weeks, the number of lesions developed on inoculated leaves was counted and then standard analysis of progeny RNA was carried out. Twenty separate lesions were excised, the total RNA was isolated and used for the selective RT-PCR amplification of the 3 0 -portion of progeny RNA3. The length of RT-PCR products was established by electrophoresis in a 1.5% agarose gel. In addition, reaction products were cloned and sequenced. This demonstrated that the Mat0-BMV mutant is infectious (usually 20 lesions were developed on each leaf, see Table 2 ) and Mat0-RNA3 is stable during the whole period of infection and thus it can be used as a recombination vector. In order to demonstrate that Mat0-RNA3 can be used as an effective tool in recombination studies, we applied it to examine the recombination activity of two specific sequences derived from the HCV genome. The first, 101 nt sequence is placed within the 5 0 -portion of the HCV genome (within the fragment encoding E2 protein) and is named HVR (40, 42) . The second, called the sequence X (X) constitutes a 98 nt 3 0 end of HCV genomic RNA. It has been shown that X represents the most conservative fragment of HCV genome (43) . Both sequences were obtained by a standard RT-PCR method involving viral RNA isolated from the blood of infected patients as a template (40, 41) . Amplified fragments were inserted into the 5 0 -cloning site of Mat0-RNA3 in two different orientations (sense and antisense), then only in sense orientation into the 3 0 -cloning site (for details see Materials and Methods). As a result, four different Mat0-RNA3 derivatives were generated: (i) MatH-HVR-RNA3, possessing two copies of HVR in sense orientation (3 0 and 5 0 HVRs); (ii) MatH-X-RNA3, with two copies of X in sense orientation (3 0 and 5 0 Xs); (iii) MatNH-HVR-RNA3, with two copies of HVR, the 3 0 -copy in sense and the 5 0 in antisense orientation (3 0 HVRs and 5 0 HVRas); (iv) MatNH-X-RNA3, with two copies of X located in different orientation (3 0 Xs and 5 0 Xas) ( Figure 5B ). The former two were applied to test HVR's and X's ability to support homologous crossovers while the latter two to examine Xs/Xas' and HVRs/HVRas' capacity to induce nonhomologous, heteroduplex-mediated recombination. Unlike previously tested mutants (MatNH-and MatH-BMV), in MatH-HVR-, MatNH-HVR-, MatH-X-and MatNH-X-BMV, the examined sequences were present only in the recombination vector. They were absent in the two other genomic RNAs, so that recombination crossovers could involve only RNA3 molecules. To determine the recombination activity of HCV-derived sequences, C.quinoa plants were inoculated with four BMV mutants: MatH-HVR-BMV, MatNH-HVR-BMV, MatH-X-BMV and MatNH-X-BMV. Their genomes were composed of wtRNA1, wtRNA2 and one of the newly generated Mat0-RNA3 derivatives (either MatH-HVR-, MatNH-HVR-, MatH-X-or MatNH-X-RNA3) ( Table 2) . After 2 weeks, the standard procedure of BMV RNA3 progeny analysis was applied. The number of lesions developed during each infection was counted. The 3 0 -portion of progeny RNA3 was amplified by the RT-PCR method. The length of RT-PCR products was established (by electrophoresis in a 1.5% agarose gel), then they were cloned and sequenced. We found that BMV mutants carrying HVRs/HVRs and HVRs/HVRas sequences are as infectious as Mat0-BMV; usually they developed 14-18 lesions on each inoculated leaf. The two others, MatH-X-and MatNH-X-BMV mutants, are visibly less infectious and developed 3-5 and 4-8 lesions/leaf, respectively. Despite differences in their infectivity, BMV mutants carrying 3 0 HVRs and 5 0 HVRas as well as 3 0 Xs and 5 0 Xas supported non-homologous, heteroduplex-mediated crossovers very efficiently. An RNA3 recombinant was generated in 100 and 90% of lesions developed during infection with MatNH-HVR-and MatNH-X-BMV, respectively. Recombinant junction sites were located within the left portion of the local double-stranded region that could potentially be formed either by HVRs and HVRas or by Xs and Xas. As a result, both sequences supporting non-homologous crossovers were almost entirely deleted, together with the whole spacer ( Figure 5D and E) . Interestingly, homologous recombinants were generated only during infection involving MatH-HVR-BMV. Fiftyfive percent of analyzed lesions contained the RNA3 recombinant. In all sequenced recombinants, one HVR and the spacer were deleted ( Figure 5C ). Their 3 0 -UTR was composed of a 295 nt RNA1-derived sequence and HVRs followed by To test the recombination activity of HCV-derived sequences X and HVR the following Mat0-RNA3 derivatives were prepared: MatH-X-RNA3-with two copies of X in sense orientation (3 0 and 5 0 Xs), MatNH-X-RNA3-with two copies of X located in different orientation (3 0 Xs and 5 0 Xas), MatH-HVR-RNA3-containing two copies of HVR in sense orientation (3 0 and 5 0 HVRs) and MatNH-HVR-RNA3-with two copies of HVR, 3 0 -copy in sense and 5 0 in antisense orientation (3 0 HVRs and 5 0 HVRas). MatH-X-and MatH-HVR-RNA3 were applied to test X's and HVR's ability to support homologous crossovers while MatNH-X-and MatNH-HVR-RNA3 were applied to examine X's and HVR's competence to induce non-homologous, heteroduplex-mediated recombination. RF, recombination frequency observed during infection involving each RNA3 derivative. (C). Homologous recombinants generated during infection involving MatH-HVR-BMV. RNA3 recombinants were only formed during MatH-HVR-BMV infection. In all of them, one recombinationally active sequence HVRs and a spacer were deleted and their 3 0 -UTR was composed of RNA1 derived sequence and HVRs followed by the CP coding region. Because crossovers occurred within identical regions, the location of recombinant junction sites could not be precisely The presence of homologous and non-homologous recombinants in the examined lesions was always confirmed not only by RT-PCR but also by northern blot analysis ( Figure 6 ). This revealed the same tendency as that observed earlier using the heteromolecular system (30, 33) . BMV accumulated to a very low level in lesions containing parental RNA3 (original molecules with duplicated sequences- Figure 6A , lane 5 and Figure 6B , lane 2). However, this changed in lesions where a recombinant was generated ( Figure 6A , lanes 1-4 and Figure 6B, lanes 1 and 3) . Earlier it was shown that the BMV-based heteromolecular system can be used as an effective tool for investigating the mechanism of homologous and non-homologous recombination, although it is only suitable for testing the recombination activity of the sequences derived from the 3 0 -portion of BMV RNA1 or RNA2 (30, 33) . To overcome this problem, we attempted to create a new universal recombination in vivo system. The collected data suggested that a BMV RNA3-based homomolecular system would best fulfill our expectations. To confirm the correctness of the above presumption, to determine the efficacy of the homomolecular system and to compare it with the heteromolecular one, two mixed homoheteromolecular systems were constructed-one to study homologous (MatH-BMV) and the other non-homologous (MatNH-BMV) recombination. The mixed systems were prepared in such a way that two identical or two complementary sequences were capable of supporting homologous or non-homologous crossovers, respectively, either between molecules representing the same segment of the BMV genome (modified RNA3) or between molecules representing two different segments of the BMV genome (wtRNA1 and modified RNA3). Experiments involving MatH-and MatNH-BMV showed that recombination can occur both in homo-and heteromolecular systems and proved that the former should be at least as effective as the previously utilized heteromolecular one. Interestingly, the RAS1s sequence did not support homologous recombination during infection with MH-BMV (heteromolecular system) and it was Infectivity was defined as the average number of lesions per leaf. b Recombination frequency was defined as the ratio between the number of lesions that developed recombinants and the total number of analyzed lesions. quite active in the mixed system. This clearly demonstrates that not only primary and secondary structure but also the location of RAS within the viral genome affects its ability to mediate homologous crossovers. The undertaken experiments also revealed that homologous recombination occurs more often between two different RNAs (RNA1 and RNA3), while non-homologous recombination usually involves molecules representing the same segment of the BMV genome (RNA3). The obtained results constitute yet another piece of evidence that the mechanisms of homologous and nonhomologous recombination are different. The same conclusion was reached by us earlier while studying the influence of specific mutations in BMV-encoded protein 2a (26, 27) . We identified among other the mutation in the 2a protein, which inhibits non-homologous crossovers without affecting the frequency of homologous ones. At present, it is difficult to judge at which stage of the recombination process the observed differences occur. One can only suppose that the structural requirements of transfer of the replicase-nascent strand complex from the donor to the acceptor molecule must be different in homologous and nonhomologous recombination. In the case of the former, a basic factor facilitating this process is complementarity between the acceptor and the nascent strand. Consequently, there is no necessity for the replication complex to be stable during homologous crossovers (22) . Replicase can leave the donor template alone or together with a nascent strand. Then, the 3 0 end of the newly synthesized RNA molecule can function as a guide; it can find a complementary sequence in the acceptor RNA, hybridize and serve as a primer allowing viral replicase to reinitiate RNA synthesis. In non-homologous heteroduplexmediated recombination, a factor enhancing crossover seems to be the interaction between the donor and the acceptor (the formation of a local double-stranded region) (30, 33) . Considering that BMV genomic RNAs are copied within spherules (44) , intramolecular hybridization between RAS1s and RAS1as, located in MatNH-RNA3, is much more likely. Thus, the results presented here indicate that the way the virus replicates can strongly affect the recombination process. However, further detailed studies are necessary in order to explain this phenomenon. Based on results obtained using the heteromolecular (30, 33) and mixed systems, we constructed a new homomolecular one. Its most crucial element is the Mat0-RNA3 vector, into which both tested sequences can be introduced. In order to show the usefulness of this system, we employed it to test the recombination activity of two distinctly different sequences deriving from the genome of an RNA virus not related to BMV. Our choice was the 98 nt sequence X and an HVR both from HCV genome. There are many reasons as to why the two sequences can be deemed drastically different. The most important of them are (i) sequence X is placed in a noncoding region, while HVR in a coding one, (ii) sequence X is the least variable and HVR the most variable fragment of the HCV genome (42, 43, 45) , (iii) unlike to HVR, sequence X possesses a very stable and well-defined secondary and tertiary structure (42, 45) . We ascertained that the introduction of HVR into the Mat0-RNA3 vector (in sense/sense and antisense/sense orientation) does not influence BMV infectivity. The latter was, however, reduced if HVR was replaced with sequence X. We found that HVR supports homologous recombination and HVRs and its complementary counterpart HVRas mediate non-homologous crossovers. The frequency of homologous recombination amounted to 55% and of non-homologous to 100%. Sequence Xs did not support homologous crossovers but Xs and complementary sequence Xas were capable of inducing non-homologous ones (their frequency reaching 90%). The obtained results testify that the local double-stranded structures induce non-homologous recombination crossovers very efficiently. This may reflect the capacity of RNA viruses to remove inverted repeats from their genomes. Viruses lacking such ability would be an easy target for double-stranded RNA-induced RNA silencing, which is known as the plant antiviral mechanism (46) . Moreover, the data presented suggest that sequence X, which adopts a very compact and stable structure (45), is not able to mediate homologous recombination. It occurs efficiently within AU-rich HVR sequences whose structure is more labile and dynamic (42) . Earlier research on homologous recombination in BMV led to similar conclusions. It was shown that homologous recombination occurs effectively in AU-rich regions (47) and is not observed within highly structured 3 0 -and 5 0 -UTR (48) . These observations concur with the proposed mechanism of homologous RNA recombination (22) . It assumes that AU-rich regions facilitate the detachment of the polymerase-nascent strand complex from donor RNA. On the other hand, it is thought that the stability of RNA structure makes the hybridization of the nascent strand and/or replicase to the acceptor difficult. Currently, it is becoming increasingly clear that RNA recombination plays a very complex role in a virus' life cycle. Not only does it permit the exchange of genetic material between viruses (3, 4, 22) , frequent homologous crossovers between molecules representing the same segment of the virus genome also stabilize genetic information (48) . Moreover, here we showed that homologous and non-homologous recombination might control the organization of the virus genome by removing direct or inverted repeats, which affect the virus' ability to replicate or accumulate in the infected cells. Interestingly, we observed that complementary sequences are more effectively deleted than homologous ones. It seems that some of the latter can prevail in the viral genome probably due to their compact stable structure that prevents recombination events. Altogether, the data presented here prove that the newly created BMV-based homomolecular recombination system can be used to examine in vivo recombination activity of various RNA sequences derived from the genomes of related or unrelated viruses. However, there are other factors which, in addition to RNA structure, can affect the course of the studied process. Specific properties of the viral replicase and the host proteins that are necessary for recombination events can be of equally great importance. Therefore, there is a need to create similar universal recombination systems in other viruses. We believe that these systems will be very helpful in finding some general rules in RNA recombination and will provide us with knowledge which is indispensable to understand how new RNA viruses or retroviruses are generated.
35
Neutrophil elastase, an acid-independent serine protease, facilitates reovirus uncoating and infection in U937 promonocyte cells
BACKGROUND: Mammalian reoviruses naturally infect their hosts through the enteric and respiratory tracts. During enteric infections, proteolysis of the reovirus outer capsid protein σ3 is mediated by pancreatic serine proteases. In contrast, the proteases critical for reovirus replication in the lung are unknown. Neutrophil elastase (NE) is an acid-independent, inflammatory serine protease predominantly expressed by neutrophils. In addition to its normal role in microbial defense, aberrant expression of NE has been implicated in the pathology of acute respiratory distress syndrome (ARDS). Because reovirus replication in rodent lungs causes ARDS-like symptoms and induces an infiltration of neutrophils, we investigated the capacity of NE to promote reovirus virion uncoating. RESULTS: The human promonocyte cell line U937 expresses NE. Treatment of U937 cells with the broad-spectrum cysteine-protease inhibitor E64 [trans-epoxysuccinyl-L-leucylamido-(4-guanidino)butane] and with agents that increase vesicular pH did not inhibit reovirus replication. Even when these inhibitors were used in combination, reovirus replicated to significant yields, indicating that an acid-independent non-cysteine protease was capable of mediating reovirus uncoating in U937 cell cultures. To identify the protease(s) responsible, U937 cells were treated with phorbol 12-myristate 13-acetate (PMA), an agent that induces cellular differentiation and results in decreased expression of acid-independent serine proteases, including NE and cathepsin (Cat) G. In the presence of E64, reovirus did not replicate efficiently in PMA-treated cells. To directly assess the role of NE in reovirus infection of U937 cells, we examined viral growth in the presence of N-Ala-Ala-Pro-Val chloromethylketone, a NE-specific inhibitor. Reovirus replication in the presence of E64 was significantly reduced by treatment of cells with the NE inhibitor. Incubation of virions with purified NE resulted in the generation of infectious subviron particles that did not require additional intracellular proteolysis. CONCLUSION: Our findings reveal that NE can facilitate reovirus infection. The fact that it does so in the presence of agents that raise vesicular pH supports a model in which the requirement for acidic pH during infection reflects the conditions required for optimal protease activity. The capacity of reovirus to exploit NE may impact viral replication in the lung and other tissues during natural infections.
Mammalian reoviruses are the prototypic members of the Reoviridae family, which also includes the pathogenic rotaviruses, coltiviruses, seadornaviruses and orbiviruses. These viruses share elements of their replication cycle as well as structural features, including a non-enveloped multi-layered capsid that surrounds a segmented dsRNA genome. In humans, mammalian reoviruses are typically associated with mild and self-limiting enteric and respiratory infections. However, studies in neonatal mice reveal that reoviruses can spread to distant tissue sites in immunocompromised hosts (reviewed in [1] ). The factors that determine reovirus cellular host range are poorly understood. Because reovirus attaches to cells through interactions with broadly expressed receptors, one or more subsequent steps in the viral life cycle must help to regulate host range and pathogenesis. Our recent studies suggest that one such step is proteolysis of the capsid protein σ3 [2, 3] . In cell culture, the first step in infection is attachment to cellular receptors through interactions with the viral protein σ1 [4, 5] . σ1 interacts with two known receptors: sialic acid and junctional adhesion molecule 1 [6] [7] [8] . Following binding, virions are internalized by receptor-mediated endocytosis [9] . Endocytosis is an essential step in the viral life cycle under standard infection conditions [10] . Within the endosomal and/or lysosomal compartment, proteases convert virions into particles that resemble in vitro-generated intermediate subvirion particles (ISVPs) [10] [11] [12] [13] [14] . These uncoating intermediates, typically prepared using chymotrypsin or trypsin, lack σ3 and have a cleaved form of µ1. Studies using ISVPs and ISVPs recoated with recombinant outer capsid proteins reveal that σ3 plays a key role in regulating reovirus cell entry by interacting with, protecting, and controlling the conformational status of the underlying penetration protein µ1 [15] [16] [17] [18] . In cells that cannot efficiently mediate σ3 degradation during uncoating, reovirus infection is slow or blocked; these cells can be productively infected by particles that lack σ3 [2] . In vitro, ISVP-like particles can be generated by a variety of proteases in addition to chymotrypsin and trypsin, including proteinase K, thermolysin, endoproteinase lys-C, Cat L, Cat B and Cat S [3, [19] [20] [21] . Recent work has provided insight into the cellular determinants of reovirus uncoating. In murine fibroblasts, where reovirus entry has been best studied, the cysteine proteases Cat L, and to a lesser extent Cat B, are required for σ3 removal, whereas the aspartyl protease Cat D is not [14, [21] [22] [23] [24] [25] . Virion disassembly in murine fibroblasts also requires acidic pH [10, 26, 27] . Recently, we demonstrated that reovirus uncoating in the macrophage-like cell line P388D is mediated by the acid-independent lysosomal cysteine protease Cat S [3] . This finding revealed that in different cell types, distinct proteases can facilitate reovirus uncoating. Our results suggested a model in which infection in some cells is acid-dependent because the proteases that mediate σ3 removal in those cells require acidic pH for maximal activity. Thus, in fibroblasts or other cells in which the acid-dependent proteases Cat L and Cat B mediate σ3 removal, infection is acid-dependent [21, 23, 28] , whereas in Cat S-expressing cells it is not [3] , because Cat S maintains its activity at neutral pH [29] . Insight from the analysis of reovirus cell entry facilitated the recent discovery that activation of the Ebola virus glycoprotein also depends on the activity of the acid-dependent endosomal proteases Cat B and Cat L [30] . The role that specific intracellular and extracellular proteases play in regulating reovirus tropism, spread, and disease in animals is largely unknown, except in the murine intestinal tract where pancreatic serine proteases have been shown to mediate σ3 removal [31, 32] . Reovirus also naturally infects hosts via the respiratory tract [33] [34] [35] . One protease with well-described effects in the respiratory tract is elastase 2 (GenBank NM_001972), an inflammatory serine protease of the chymotrypsin family, which is predominantly expressed by neutrophils [36] . NE plays a prominent role in wound repair [37] [38] [39] and in controlling microbial infections [38] [39] [40] . NE expression can also promote pathogenesis; it has been implicated in smokeinduced emphysema [41] , respiratory syncytial viral bronchiolitis [42] and in the respiratory syndrome ARDS [4] . The fact that reovirus replication in the rodent lung causes an influx of neutrophils [35, 43] and that reovirus infection can recapitulate ARDS [44] , led us to ask whether NE could mediate productive reovirus uncoating. We investigated reovirus infection in the monocyte-like cell line U937, because it is known to express NE [45] . Experiments described in this report demonstrate that reovirus infection in U937 cells does not require cysteine protease activity and is not blocked in the presence of agents that raise vesicular pH. Studies using protease inhibitors suggest that, in the absence of cysteine protease activity, NE is largely responsible for productive infection of U937 cells. NE can directly mediate σ3 removal from reovirus virions; the resultant particles are infectious and do not require additional intracellular proteolysis. Our data raise the possibility that NE is involved in reovirus replication in the respiratory tract. Analysis of viral replication in L929 and U937 cells treated with E64 Figure 1 Analysis of viral replication in L929 and U937 cells treated with E64. A. 3 × 10 6 L929 and U937 cells were untreated (-; black) or treated (+; grey) with 300 µM E64 for 3 h or 3 d. Cysteine protease activity was assessed using the fluorogenic substrate Z-Phe-Arg-MCA (Sigma) and plotted in arbitrary units. Activity levels in treated cells were so low (in L939 cells, 254 units at 3 h and 231 units at 3 d; in U937 cells, 200 units at 3 h and 115 units at 3 days) that they cannot be visualized on this graph. B. L929 (L; black bars) and U937 (U; grey bars) cells were treated with 300 µM E64 for 3 h prior to infection. Cells were then infected with reovirus strain Lang virions or ISVPs at an MOI of 3. Infectious virus present at 3 d p.i. was determined by plaque assay on L929 cell monolayers. Each time point represents the mean (+/-SD) derived from three independent samples. Virion ISVP we first established conditions under which lysosomal cysteine protease activity was inhibited. Cells were treated with 300 µM E64, a broad-spectrum cysteine protease inhibitor [46] , and protease activity was assessed using the Cat L and Cat B-specific fluorogenic substrate Z-Phe-Arg-MCA. We analyzed enzyme activity at two time points: first after 3 h of treatment, because we typically pre-treat cells with inhibitors for 3 h prior to infection, and second at 3 d, the time point at which viral yield would be quantified. As shown in Fig. 1A , treatment with 300 µM E64 completely abolished cysteine protease activity in U937 cells. Consistent with our previous findings [3] , E64 also completely blocked cysteine protease activity in L929 cells. Raw values are provided, to illustrate the relative difference in Cat L/B enzyme activity levels between U937 cells and L929 fibroblasts. In the absence of inhibitor, Cat L and B activity was significantly lower in U937 cells than in L929 cells. This may be a consequence of high expression in U937 cells of cystatin F, an intracellular cysteine protease inhibitor with specificity for Cat L and papain [47] . Next, we compared reovirus replication in E64-treated U937 and L929 cells. Cells were pre-treated for 3 h and infected with Lang virions or ISVPs at a multiplicity of infection (MOI) of 3. The results of a representative experiment are shown in Fig. 1B . In the absence of E64, both L929 and U937 cells supported reovirus replication, consistent with the fact that these cells express Cat L. As expected, E64 blocked virion infection of L929 cells; however, viral yields in E64-treated U937 cells were only slightly reduced relative to untreated cells. ISVPs, which lack capsid protein σ3, replicated efficiently in treated cells, indicating that 300 µM E64 was not toxic to either cell type. These results demonstrate that productive infection of U937 cells by Lang virions does not require the activity of E64-sensitive, papain-like cysteine proteases. Acidic pH is required for productive reovirus infection of murine L929 fibroblasts [10, 27] , in which the aciddependent proteases Cat L and Cat B mediate uncoating [21, 23] . Serine proteases, including NE, and metalloproteases function over a broader pH range. Therefore, to gain insight into the nature of the protease(s) that can promote reovirus uncoating in U937 cells, we investigated the requirements for acidic pH. L929 and U937 cells were left untreated or pre-treated with E64 in the presence or absence of bafilomycin A1 (Baf) or NH 4 Cl. These latter agents raise vesicular pH by blocking the vacuolar H + -ATPase pump or by acting as a weak base, respectively [48] [49] [50] . After pre-treatment, cells were infected with Lang virions at an MOI of 3 and viral yields were determined at 3 days post infection (d p.i.). A representative experiment is shown in Fig. 2 . Treatment with either Baf or NH 4 Cl did not inhibit viral replication in U937 cells; yields reached 2.9 and 2.7 logs, respectively. Furthermore, these agents had little effect on viral replication in U937 cells even when the cells were also treated with E64 to inhibit cysteine protease activity. In contrast, Baf or NH 4 Cl alone completely blocked reovirus replication in L929 cells, consistent with the requirement for Cat L/B-mediated σ3 removal in these cells. Given that reovirus uncoating is an essential step in the viral life cycle [10] , these findings revealed that a non-cysteine protease that functions at neutral pH can facilitate this step in U937 cells. Treatment of the promonocytic U937 cells with phorbol ester derivatives results in their differentiation into macrophage-like cells [51, 52] . This differentiation is characterized by several major phenotypic changes, including increases in expression of urokinase plasminogen activator receptors, upregulation of collagenase activity and a significant decrease in the expression of NE and Cat G [51, 52] . We predicted, therefore, that PMA treatment might decrease the capacity of reovirus virions to replicate in U937 cells when cysteine proteases were inhibited. To confirm that there was a significant decrease in NE expression in U937 cells differentiated with PMA, U937 cells were treated with 150 nM PMA for 72 h and expression of NE was analyzed by immunoblotting. As shown in Fig. 3A , NE was expressed in untreated U937 cells, but its expression was dramatically reduced following PMAinduced differentiation. To examine the effect of U937 cell differentiation on reovirus infection, PMA-treated and untreated U937 cells were left untreated or were treated with E64 for 3 h and infected with Lang virions or ISVPs at an MOI of 3. Yields were measured at 3 d p.i. and the results of a typical experiment are shown in Fig. 3B . In the absence of E64, PMAtreated U937 cells were permissive to infection by virions. PMA treatment only decreased yields by ~0.5 log relative to untreated cells. In contrast, when PMA-differentiated U937 cells were treated with E64 to inhibit cysteine protease activity, they no longer supported productive infection by Lang virions. Because these results could be explained if E64 was toxic to PMA-treated U937 cells, we examined the replication of ISVPs. In the presence of E64, ISVPs replicated to high yields in both undifferentiated and differentiated U937 cells. Since PMA-induced differentiation of U937 cells caused a substantial decrease in NE expression, these results are consistent with the hypothesis that NE or another similarly regulated neutral protease facilitates productive reovirus infection in promonocytic (pre-differentiated) U937 cells. Effects of agents that raise vesicular pH on reovirus replication in U937 and L929 cells A. Analysis of reovirus replication in U937 cells differentiated with PMA Figure 3 Analysis of reovirus replication in U937 cells differentiated with PMA. A. Lysates generated from 10 5 U937 cells that were untreated (-) or treated with 150 nM PMA for 72 h were resolved on SDS-12% polyacrylamide gels and electrophoretically transferred to a nitrocellulose filter. The filter was subsequently incubated with a polyclonal goat antibody against human NE (1:400) (Santa Cruz Biotechnology). The filter was washed and incubated with a secondary anti-goat antibody conjugated to horseradish peroxidase (1:5000) (Santa Cruz Biotechnology). Protein bands were detected using reagents that generate a chemiluminescent signal (Amersham). B. U937 cells that were undifferentiated (-; black bars) or differentiated (PMA; grey bars) with 150 nM of PMA for 72 h were left untreated (-) or were treated with 300 µM E64. Following pre-treatment with the protease inhibitor, cells were infected with Lang virions or ISVPs at an MOI of 3. Viral yield was quantified at 3 d p.i. as described in the legend to Fig. 1B. A. B. We directly examined the capacity of NE to facilitate reovirus infection by using the irreversible elastase inhibitor, N-(methoxysuccinyl)-Ala-Ala-Pro-Val-chloromethyl ketone [53] . This inhibitor is highly specific for NE and does not inhibit the activity of the related serine protease, Cat G [53] . First, we established the efficacy and specificity of inhibitor treatment under our experimental conditions. U937 cells were treated with the NE inhibitor, E64, Baf or NH 4 Cl for either 3 h or 2 d and the activity of NE in cell lysates was examined using a colorimetric substrate. As shown in Table 1 , the NE inhibitor was active at both time points. In cells treated with the specific inhibitor, NE activity was less than 9% of that in untreated U937 cells. In contrast, in U937 cells treated with E64, Baf or NH 4 Cl, NE activity was only modestly reduced, remaining above 80% even after 2 d. These results are consistent with the capacity of NE to function at neutral pH. To verify the specificity of the NE inhibitor, we also examined its effect on Cat L/B activity using the fluorogenic substrate Z-Phe-Arg-MCA. As expected, Cat L/B activity was completely inhibited by E64 but largely unaffected by the NE inhibitor. To examine the effect of the NE inhibitor on reovirus replication in U937 cells, we pre-treated them for 3 h with E64 in the presence or absence of the NE inhibitor, infected them with Lang virions or ISVPs at an MOI of 3, and quantified viral yields at 2 d p.i. A representative experiment is shown in Fig. 4 . Consistent with the results shown in Fig 1, virion replication was not blocked in E64treated U937 cells. However, in the presence of both E64 and the NE inhibitor, yields were significantly reduced. ISVPs replicated to high yields in treated cells, indicating that the combination of inhibitors was not toxic to U937 cells. These results demonstrate that NE plays a critical role in reovirus infection of U937 cells when cysteine proteases are inhibited. NE, like many cellular proteases, is expressed as a proenzyme that becomes activated only after its pro-region is removed [54] . We envisioned two models by which NE could facilitate reovirus infection of U937 cells. In the first, NE could directly mediate σ3 degradation, leading to the generation of an ISVP-like particle. In the second, NE could act indirectly by activating another protease. To try to distinguish between these models, we examined the capacity of purified NE to directly mediate σ3 removal from Lang virions in vitro. Purified Lang virions were treated with NE for 1 and 4 h and the treated virus particles were analyzed by SDS-PAGE. As shown in Fig. 5A , NE efficiently removed σ3 from Lang virions; after 1 h very little intact σ3 remained on viral particles. After 4 h of NE treatment, σ3 was completely removed and the underlying µ1C was cleaved to the δ and φ fragments (φ was not retained on the gel). When we assayed the infectivity of the resultant particles by plaque assay we found that NE treatment did not negatively affect the titer of Lang particles (data not shown). To determine if NE-generated SVPs required further proteolytic processing of σ3, L929 cells were pre-treated with E64 to block cysteine protease activity and infected at an MOI of 3 with Lang virions, ISVPs or NE-generated subviral particles (NE-SVPs). Viral yields were determined at 1 d p.i. As expected, E64 blocked infection of L929 cells by virions. In contrast, both ISVPs and NE-SVPs replicated efficiently in the presence of the cysteine protease inhibitor (Fig. 5B) . Because virion disassembly in L929 cells requires acidic pH [10] , we also examined the capacity of NE-SVPs to infect L929 cells treated with Baf, NH 4 Cl or monensin, three agents that raise vesicular pH by distinct mechanisms. Cells were treated with these agents and then infected with virions, ISVPs or NE-SVPs at an MOI of a U937 cells were treated with the indicated inhibitors for 3 h or 2 d. b NE activity was assessed using the colorimetric substrate MeOSuc-Ala-Ala-Pro-Val-ρNA and percent activity relative to untreated cells was calculated. c Cathepsin L and B activity were assessed using the fluorogenic substrate Z-Phe-Arg-MCA and percent activity relative to untreated cells was calculated. 10. At 18 hours post infection (h p.i.), cell lysates were harvested and expression of the reovirus non-structural protein µNS was analyzed by immunoblotting (Fig. 5C ). As expected, when treated cells were infected with virions, viral protein expression was blocked. In contrast, µNS expression was evident even in the presence of agents that raise pH when infections were initiated with ISVPs or NE-SVPs (Fig. 5C ). Together, these results demonstrate that NE can directly mediate σ3 removal from virions to generate infectious particles that do not require further proteolytic processing by acid-dependent cysteine proteases in L929 cells. Serine proteases are involved in reovirus infection in the mammalian intestinal tract [31] and in this report we pro-vide evidence that they can mediate uncoating and promote infection in U937 cells. This expands the range of proteases that promote reovirus infection in cell culture to include NE as well as the cysteine proteases Cat L, Cat B, and Cat S. Several lines of evidence now support the notion that protease expression is a cell-specific host factor that can impact reovirus infection. For example, some reovirus strains are inefficiently uncoated by Cat S and thus do not replicate to high yield in P388D macrophages [3] . In this report we demonstrate that PMA-induced differentiation influences the type of protease that mediates reovirus uncoating in U937 cells. In these cells, PMA treatment is reported to increase Cat L expression [55] and decrease expression of the serine proteases NE and Cat G [56, 57] . Accordingly, when we used PMA to induce U937 cell cultures to differentiate, reovirus infection became sensitive to the cysteine protease inhibitor E64. We suspect that Cat L is largely responsible for uncoating in these PMA-differentiated cells, but the acid-independent protease Cat S may also play a role. We are currently addressing this question by analyzing infection in PMAdifferentiated cells treated with either Baf or NH 4 Cl. Our data do not completely resolve this question. Cat G is expressed by U937 cells and, like NE, it is down-regulated by PMA treatment. Furthermore, we found that in vitro treatment of reovirus virions with purified Cat G generates SVPs that behave like NE-SVPs in that they are infectious in the absence of further proteolytic processing (data not shown). Results of our experiment with the NE-specific inhibitor suggest that NE is largely responsible for the E64-resistant infection in U937 cells. While this inhibitor is reported not to inhibit Cat G [53] , we have not independently confirmed this. Another approach to assess the role of Cat G in reovirus infection of U937 cells would be to examine the effect of Cat G-specific inhibitors on infection. We tried one such inhibitor, Cathepsin G Inhibitor I (Calbiochem) [58] , but found that it was cytotoxic to U937 cell cultures. Given that both NE and Cat G can generate infectious reovirus SVPs, more work needs to be done in order to understand the role that these two proteases play in infection in these cells. Previously, we reported that virion uncoating mediated by Cat S does not require acidic pH [3] . These results were consistent with the acid-independence of Cat S activity [37] . Together, the results in Fig. 2 and Fig. 4 reveal that, like Cat S, NE-mediates infection in an acid-independent manner. This finding thus provides further support for a model in which the requirement for acidic pH during reovirus infection of some cell types reflects the requirement for acid-dependent protease activity in those cells rather than some other requisite acid-dependent aspect of cell entry. The small effect of Baf and NH 4 Cl on E64-resistant reovirus growth (Fig. 2 ) may reflect the participation of one or more acid-dependent proteases (such as Cat D) in the activation of NE. Elastase is stored in azurophilic granules that are the major source of acid-dependent hydrolases in neutrophils [59] . Although these granules do not contain LAMP-1 or LAMP-2 [60] they contain the lysosomal markers LAMP-3 [61] and CD68 [62] and are accessible to endocytosed fluid-phase markers under conditions of cellular stimulation [63] . NE can be released from neutrophils during degranulation [64] and its cell surface expression can be induced upon PMA treatment [65] . However, studies in U937 cells have shown that NE is predominantly retained intracellularly and that little if any activity is present in the extracellular medium [45] . Consistent with this, we have been unable to generate ISVP-like particles by treatment of virions with U937 culture supernatants (data not shown). This observation, together with our finding that PMA treatment decreases the capacity of E64-treated U937 cells to support reovirus infection, leads us to favor a model in which NE-mediated virion uncoating in U937 cell cultures occurs intracellularly. In vivo, a number of viruses, including dengue and respiratory syncytial virus, induce the release of IL-8, a cytokine that serves as a chemoattractant for neutrophils and promotes their degranulation [66, 67] . Reovirus replication in the rat lung results in neutrophilic invasion [35, 43] and studies in cell culture indicate that reovirus infection can induce IL-8 expression [68] . Thus, the capacity of reovirus to induce IL-8 secretion in vivo might facilitate the release of neutrophilic lysosomal hydrolases, including NE, into the extracellular milieu. In this report, we have shown that mammalian reovirus can utilize this acid-independent serine protease for uncoating. Our data suggest that, in vivo, one consequence of reovirus-induced IL-8 expression would be the generation of infectious NE-SVPs. Like ISVPs, these particles would be predicted to have an expanded cellular host range because they can infect cells that restrict intracellular uncoating [2] . Thus, inflammation might be predicted to exacerbate reovirus infection by promoting viral spread. Future studies using mice with deletions in the NE gene will be required to elucidate the role this protease plays during reovirus infection in the respiratory tract and other tissues. Finally, given the recent finding that endosomal proteolysis of the Ebola virus glycoprotein is necessary for infection [30] , our results raise the interesting possibility that NE or other neutrophil proteases may play a role in cell entry of other viruses. [Furlong, 1988 #81] . ISVPs were prepared by treating purified virions with chymotrypsin as described elsewhere [Nibert, 1992 #95] . Cysteine protease activity was measured as described previously [23] Samples were frozen and thawed three times and titrated by plaque assay on L929 cells as described elsewhere [69] . Viral yields were calculated according to the following formula: log 10 (PFU/ml) t = x hrlog 10 (PFU/ml) t = 0 +/-standard deviation (SD). To analyze NE expression, cell lysates were generated from U937 cells, either treated or untreated for 48 h with 150 nM PMA as described for the analysis of viral protein expression. Lysate from the equivalent of 1 × 10 6 cells was run on SDS-12% polyacrylamide gels and transferred to nitrocellulose. Membranes were blocked overnight in TBST containing 10% nonfat dry milk. NE expression was analyzed using a polyclonal antibody against NE (1:400 in TBST) (Santa Cruz Biotechnology Inc, Santa Cruz, CA). Membranes were washed with TBST and incubated with a horseradish peroxidase-conjugated anti-goat IgG (1:5000 in TBST). Bound antibody was detected by treating the nitrocellulose filters with enhanced chemiluminescence (ECL) detection reagents (Amersham) and exposing them to Full Speed Blue X-ray film (Henry Schein, Melville, NY). Cells were plated at 10 6 /well in a 6-well plate 18-24 h prior to infection. Virus was allowed to adsorb to cells for 1.5 h at 4°C. At this temperature, virus binds to cells but is not internalized [70] . After adsorption, the cultures were incubated at 37°C in fresh medium. Prior to some infections, cells were pre-treated for 3 h with 300 µM E64, 100 nM Baf, 25 µM monensin (Sigma), or 20 mM NH 4 Cl. In those instances inhibitors were also included in the post-adsorption culture medium. At the indicated times p.i., cells were collected by centrifugation at 179 × g, washed twice in chilled PBS and lysed in TLB. After centrifugation at 179 × g to remove cellular debris, samples were resuspended in sample buffer. Protein samples (representing 1 × 10 5 cells) were analyzed by electrophoresis on SDS-12% polyacrylamide gels and transferred to nitrocellulose membranes for 2 h at 100 V in 25 mM Tris-192 mM glycine-20% methanol. Nitrocellulose membranes (Bio-Rad Laboratories, Hercules, Calif.) were blocked overnight at 4°C in TBST (10 mM Tris [pH 8.0], 150 mM NaCl and 0.05% Tween) containing 5% nonfat dry milk, rinsed with TBST, and incubated with a rabbit anti-µNS polyclonal antiserum [71] (1:12500 in TBST) for 1 h. Membranes were subsequently washed with TBST and incubated for 1 h with horseradish peroxidase-conjugated anti-rabbit immunogloblin G (IgG) (1:7500 in TBST) (Amersham, Arlington Heights, Ill.). Bound antibody was detected by treating the nitrocellulose filters with enhanced chemilumescence (ECL) detection reagents (Amersham) and exposing the filters to Full Speed Blue X-ray film (Eastman Kodak, Rochester, N.Y.). Purified virions (1.4 × 10 11 ) were incubated with 25 µg/ ml of purified neutrophil elastase (Calbiochem) in 40 µL of VDB at 37°C for 3 h. Reactions were terminated by adding 1 mM PMSF and 200 µM NE inhibitor to the reaction mixture. 5.0 × 10 10 particles were run on SDS-12% polyacrylamide gels stained with Coomassie Brilliant Blue to confirm the removal of σ3. Viral infectivity was determined by plaque assay on L929 cell monlayers. Purified Lang virions (1.4 × 10 11 ) were treated with 25 µg/ ml of NE in 40 µL of VDB at 37°C for the times indicated. Reactions were terminated as described above. To verify σ3 removal, the proteins from 5.0 × 10 10 particles were separated on SDS-12% polyacrylamide gels and visualized with Coomassie Brilliant Blue staining. Viral infectivity for each time point was determined by plaque assay on L929 cell monolayers.
36
The influence of locked nucleic acid residues on the thermodynamic properties of 2′-O-methyl RNA/RNA heteroduplexes
The influence of locked nucleic acid (LNA) residues on the thermodynamic properties of 2′-O-methyl RNA/RNA heteroduplexes is reported. Optical melting studies indicate that LNA incorporated into an otherwise 2′-O-methyl RNA oligonucleotide usually, but not always, enhances the stabilities of complementary duplexes formed with RNA. Several trends are apparent, including: (i) a 3′ terminal U LNA and 5′ terminal LNAs are less stabilizing than interior and other 3′ terminal LNAs; (ii) most of the stability enhancement is achieved when LNA nucleotides are separated by at least one 2′-O-methyl nucleotide; and (iii) the effects of LNA substitutions are approximately additive when the LNA nucleotides are separated by at least one 2′-O-methyl nucleotide. An equation is proposed to approximate the stabilities of complementary duplexes formed with RNA when at least one 2′-O-methyl nucleotide separates LNA nucleotides. The sequence dependence of 2′-O-methyl RNA/RNA duplexes appears to be similar to that of RNA/RNA duplexes, and preliminary nearest-neighbor free energy increments at 37°C are presented for 2′-O-methyl RNA/RNA duplexes. Internal mismatches with LNA nucleotides significantly destabilize duplexes with RNA.
Understanding the thermodynamics of nucleic acid duplexes is important for many reasons. For example, such knowledge facilitates design of ribozymes (1), antisense and RNAi oligonucleotides (2) (3) (4) (5) (6) (7) (8) (9) , diagnostic probes including those employed on microarrays (10) (11) (12) (13) (14) (15) (16) (17) (18) (19) (20) (21) (22) (23) and structures useful for nanotechnology (24) (25) (26) (27) . Many modified residues have been developed for such applications. Examples include propynylated bases (28) (29) (30) , peptide nucleic acids (5, (31) (32) (33) , N3 0 -P5 0 phosphoramidates (34-38) and 2 0 -O-alkyl RNA (39) (40) (41) (42) (43) . A modification that is particularly stabilizing in DNA and RNA duplexes (44) (45) (46) (47) (48) (49) (50) (51) is a methyl bridge between the 2 0 oxygen and 4 0 carbon of ribose to form a 'locked nucleic acid' or LNA as shown in Figure 1 . McTigue et al. (48) have shown that the enhanced stability due to a single LNA residue in a DNA duplex can be predicted from a nearestneighbor model. Hybridization of oligonucleotides to RNA is important for applications, such as antisense therapeutics (4, 8, 21, 46, (52) (53) (54) , diagnostics (32, 33, 42, 55) , profiling gene expression with microarrays (18) (19) (20) 56) , identifying bands by Northern blots of gels (57, 58) and probing RNA structure (1, 3, 15, (59) (60) (61) . Oligonucleotides with 2 0 -O-alkyl modifications can be particularly useful for these applications because they are easily synthesized (39, 43) , chemically stable and bind relatively tightly to RNA (39) (40) (41) (42) . However, for many applications, it is desirable to modulate the binding affinity. For example, sequence independent duplex stabilities would benefit applications that involve multiplex detection, such as microarrays. Here, we show that introduction of LNA into 2 0 -O-methyl RNA oligonucleotides can increase stabilities of 2 0 -O-methyl RNA/RNA hybrid duplexes and that the enhancements in stability can usually be predicted with a simple model. High-performance liquid chromatography (HPLC) was performed on a Hewlett Packard series 1100 HPLC with a reverse-phase Supelco RP-18 column (4.6 · 250 mm). Mass spectra were obtained on an LC MS Hewlett Packard series 1100 MSD with API-ES detector or on an AMD 604/402. Thin-layer chromatography (TLC) was carried out on Merck 60 F 254 TLC plates with the mixture 1-propanol/ aqueous ammonia/water ¼ 55:35:10 (v/v/v). Oligoribonucleotides were synthesized on an Applied Biosystems DNA/RNA synthesizer, using b-cyanoethyl phosphoramidite chemistry (62) . For synthesis of standard RNA oligonucleotides, the commercially available phosphoramidites with 2 0 -O-tertbutyldimethylsilyl groups were used (Glen Research). For synthesis of 2 0 -O-methyl RNA oligonucleotides, the 3 0 -O-phosphoramidites of 2 0 -Omethylnucleotides were used (Glen Research and Proligo). The 3 0 -O-phosphoramidites of LNA nucleotides were synthesized according to the published procedures with some minor modifications (44, 47, 63) . The details of deprotection and purification of oligoribonucleotides were described previously (64) . Oligonucleotides were melted in buffer containing 100 mM NaCl, 20 mM sodium cacodylate, 0.5 mM Na 2 EDTA, pH 7.0. The relatively low NaCl concentration kept melting temperatures in the reasonable range even when there were multiple LNA substitutions. Oligonucleotide single-strand concentrations were calculated from absorbencies above 80 C and single-strand extinction coefficients were approximated by a nearest-neighbor model (65, 66) . It was assumed that 2 0 -Omethyl RNA and RNA strands with identical sequences have identical extinction coefficients. Absorbancy versus temperature melting curves were measured at 260 nm with a heating rate of 1 C/min from 0 to 90 C on a Beckman DU 640 spectrophotometer with a water cooled thermoprogrammer. Melting curves were analyzed and thermodynamic parameters were calculated from a two-state model with the program MeltWin 3.5 (67) . For almost all sequences, the DH derived from T m À1 versus ln (C T /4) plots is within 15% of that derived from averaging the fits to individual melting curves, as expected if the two-state model is reasonable. Free energy parameters for predicting stabilities of 2 0 -O-methyl RNA/RNA and 2 0 -O-methyl RNA-LNA/RNA duplexes with the Individual Nearest-Neighbor Hydrogen Bonding (INN-HB) model (64) were obtained by multiple linear regression with the program Analyse-it v.1.71 (Analyse-It Software, Ltd, Leeds, England; www.analyse-it. com) which expands Microsoft Excel. Analyse-It was also used to obtain parameters for enhancement of stabilities of 2 0 -O-methyl RNA/RNA duplexes by substitution of LNA nucleotides internally and/or at the 3 0 end when the LNAs are separated by at least one 2 0 -O-methyl nucleotide. Results from T m À1 versus ln (C T /4) plots were used as the data for the calculations. 3 show typical data from optical melting curves, and Table 1 lists the thermodynamic parameters for the helix to coil transition with either no or one LNA nucleotide in the primarily 2 0 -O-methyl strand of a hybrid with a Watson-Crick complementary RNA strand. Single LNA substitutions at the 5 0 end of heptamer duplexes have little effect on stability The effects of single LNA substitutions at the 5 0 end of the 2 0 -O-methyl strand were studied in duplexes of the form, where superscript M denotes a 2 0 -O-methyl sugar, N is A, C, G, or U with a 2 0 -O-methyl or LNA sugar, r denotes ribose sugars, and Q is the Watson-Crick complement to N. As summarized in Table 1 , 5 0 terminal LNA substitutions make duplex stability more favorable by 0.3-0.6 kcal/mol at 37 C with an average enhancement of 0.45 kcal/mol. Thus, 5 0 terminal LNA substitutions increase the binding constant for duplex formation by $2-fold at 37 C. The effects of single LNA substitutions at the 3 0 ends of heptamer duplexes is idiosyncratic The effects of single LNA substitutions at the 3 0 end of the 2 0 -O-methyl strand was studied in duplexes of the form, Table 1 ). If N is A, C or G, then LNA substitutions have similar effects. On average, an LNA substitution makes duplex stability more favorable by 1.2 kcal/mol at 37 C. In the two sequences with a 3 0 terminal LNA U on the 2 0 -O-methyl strand, duplex stability is, however, affected little, averaging a destabilization of 0.08 kcal/ mol at 37 C. In both cases, the terminal U is preceded by a GC pair, but both orientations of the GC pair give similar destabilization upon LNA substitution at the 3 0 terminal U. Single LNA substitutions in the interior of A M C M U M A M C M C M A M enhance the stability of the duplex formed with its complementary RNA by $1.4 kcal/mol The effect of interior position on the free energy increment for a single LNA substitution for a 2 0 -O-methyl RNA was studied for the duplex 5 0 A M C M U M A M C M C M A M /3 0 r(UGAUGGU). As summarized in Table 1 , a single interior LNA substitution makes duplex stability more favorable by 1.2-1.7 kcal/mol at 37 C, with an average of 1.4 kcal/mol. This corresponds to roughly a 10-fold increase in binding constant. Thus, interior and 3 0 terminal LNA substitutions usually improve binding more than 5 0 terminal LNA substitutions. Table 1 . For 13 of 16 sequences, the LNA substitution makes duplex stability more favorable by 1.0-1.5 kcal/mol at 37 C, with an average enhancement of 1.3 kcal/mol. The enhancement for the other three sequences averages 2.1 kcal/mol at 37 C. The dependence on the 5 0 nearest-neighbor nucleotide of effects from substituting U L for U M was studied in duplexes of the form, neighbor that is preceded by A M and U M , respectively. In both cases, the LNA substitution enhances duplex stability by 1.14 kcal/mol at 37 C. Thus, for seven duplexes, the enhanced stability from an LNA substitution is relatively independent of the nearest-neighbor nucleotide 5 0 to the LNA. The one exception is for the nearest neighbor 5 0 G M U L /3 0 r(CA). Interestingly, this nearest-neighbor combination is also destabilized by LNA substitution at a 3 0 terminal U (Table 1) . Evidently, an LNA substitution in the middle of a 2 0 -O-methyl strand usually affects heteroduplex stability with an RNA strand by about the same amount as an LNA substitution at a 3 0 terminus. The effects of LNA substitutions are approximately additive when LNA nucleotides are spaced by at least one 2 0 -O-methyl nucleotide Table 2 contains thermodynamic parameters measured for duplexes having more than one LNA substitution and Table 3 compares the stabilities at 37 C with those predicted from four simple models. The first model, labeled 'additivity', predicts the DG 37 for duplex formation in the 5 0 ACUACCA/ 3 0 UGAUGGU series by adding the free energy increments measured for single LNA substitutions in the same context to the DG 37 for duplex formation in the absence of LNA nucleotides. The second model predicts the DG 37 (kcal/mol) for duplex formation with the following equation as deduced from fitting the data in Tables 1 and 2 Here, DG 37 (2 0 -O-MeRNA/RNA) is the free energy change at 37 C for duplex formation in the absence of any LNA nucleotides, n 5 0 tL is the number of 5 0 terminal LNAs, n iAL/UL and n iGL/CL are the number of internal LNAs in AU and GC pairs, respectively, n 3 0 tU and n 3 0 tAL/CL/GL are the number of Here, T m À1 is the inverse melting temperature in kelvin, R is the gas constant, 1.987 cal K À1 mol À1 , C T is the total oligonucleotide strand concentration, and both strands have the same concentration. Table 1 . Thermodynamic parameters of duplex formation between RNA and 2 0 -O-methyl oligoribonucleotides with and without a single LNA substitution a Oligonucleotides RNA Average of curve fits 3 0 terminal LNAs that are U or not U, respectively. Both methods that use experimental data for DG 37 (2 0 -O-MeRNA/RNA) provide reasonable predictions that are within 1 kcal/mol of the measured value (Table 3) . Two other methods that use nearest-neighbor models to approximate DG 37 (2 0 -O-MeRNA/RNA) provide somewhat less accurate, but still reasonable predictions as described below. The duplex with the worst prediction, 5 0 G M U L U M C L G M G L /3 0 CAAGCC has a 5 0 G M U L /3 0 CA nearest neighbor, consistent with this motif being unusually unstable by $1.2 kcal/mol. Thus, it is likely that the DG 37 of Equation 1 should be made less favorable by 1.2 kcal/mol for every internal 5 0 G M U L /3 0 CA nearest neighbor in a duplex. Evidently, the effects of multiple LNA substitutions are approximately additive when the LNAs are spaced by at least 1 nt. The data may also be fit to a nearest-neighbor model containing 30 of the LNA enhancement parameters associated with duplexes of RNA strands bound to 2 0 -O-methyl RNA/ LNA chimeras. These parameters are listed in Supplementary Material. The number of occurrences for each nearest neighbor is limited, however, so the values are only roughly determined. Predictions for RNA/RNA duplexes at 1 M NaCl can be used to approximate stabilities of 2 0 -O-methyl RNA/RNA duplexes at 0.1 M NaCl The stabilities of RNA/RNA duplexes at 37 C and 1 M NaCl are predicted well by an Independent Nearest-Neighbor Hydrogen Bonding (INN-HB) model (64) . In this model, the stability of an RNA/RNA duplex is approximated by: Here, DG init is the free energy change for initiating a helix; each DG j NN ð Þ is the free energy increment of the jth type nearest neighbor (see Table 4 ) with n j occurrences in the sequence; m term-AU is the number of terminal AU pairs; DG termÀAU is the free energy increment per terminal AU pair; DG sym is 0.43 kcal/mol at 37 C for self-complementary duplexes and 0 for non-self-complementary duplexes. À0.73 ± 0.26 5 0 AU3 0 À1.10 ± 0.08 Similar sequence dependent parameters may also be applicable to 2 0 -O-methyl RNA/RNA heteroduplexes because they are expected to have A-form conformations similar to those of RNA/RNA homoduplexes (68) . This was tested by comparing the predicted stabilities of RNA/RNA duplexes in 1 M NaCl at 37 C with those measured for 2 0 -O-methyl RNA/RNA duplexes in 0.1 M NaCl at 37 C. The predicted thermodynamics are listed in parentheses in Tables 1 and 2 . On average at 37 C, the RNA/RNA duplexes in 1 M NaCl are 0.12 ± 0.01 kcal/mol of phosphate pairs more stable than the 2 0 -Omethyl RNA/RNA duplexes in 0.1 M NaCl. Presumably, much of this difference is due to a sequence independent effect of salt concentration, which would primarily affect the DS for duplex formation (22, 69) . Thus, a reasonable approximation for the first term on the right hand side of Equation 1 is: Note that DG sym from the RNA/RNA calculation is subtracted because a 2 0 -O-methyl RNA/RNA duplex cannot be selfcomplementary because the backbones differ. For the duplexes studied here, the number of phosphate pairs is one less than the number of base pairs. The effects of LNA substitutions are likely not very dependent on salt concentration. Thus, it is probable that in 1 M NaCl or in the presence of Mg 2+ (70) that DG 37 (2 0 -O-MeRNA/ RNA) can be approximated by DG 37 (RNA/RNA, 1 M NaCl). Table 3 compares measured values for duplexes with more than one LNA to predictions from combining Equation 1-3. The measured DG 37 values average À10.5 kcal/mol and the root-mean-square difference between measured and predicted DG 37 values is 0.6 kcal/mol with the largest difference being 1.7 kcal/mol. Again, the sequence with the largest difference contains a 5 0 G M U L /3 0 CA nearest neighbor so the prediction would be improved if Equation 1 was corrected for the apparent instability of this motif. The results for 2 0 -O-methyl RNA/RNA duplexes provide preliminary nearest-neighbor free energy increments for predicting stabilities of such duplexes The comparison of predicted RNA/RNA stabilities with those measured for 2 0 -O-methyl RNA/RNA duplexes suggests that the INN-HB model will also be applicable to 2 0 -O-methyl RNA/RNA duplexes (71) . The results in Tables 1 and 2 Table 4 ). Three nearest neighbors are only represented once or twice in the database, and these parameters are in parentheses. The parameters for 2 0 -O-methyl RNA/RNA and RNA/RNA duplexes are similar, especially if the RNA/RNA Watson-Crick nearest-neighbor parameters are each made less favorable by 0.12 kcal/mol, which largely accounts for the difference in salt concentration as suggested above. Evidently, the first term on the right hand side of Equation 1 can also be approximated by: Table 3 compares predictions from combining Equations 1 and 4 with measured values for duplexes with more than one LNA. The root-mean-square difference between measured and predicted DG 37 values is 0.6 kcal/mol with the largest difference being the 1.7 kcal/mol associated with the duplex containing a 5 0 G M U L /3 0 CA nearest neighbor. Undoubtedly, this model can be expanded and refined by more measurements, but it appears sufficient to aid sequence design for many applications. Complete LNA substitution is no more stabilizing than substitution at every other nucleotide starting at the second nucleotide from the 5 0 end The effect of complete LNA substitution for a 2 0 -O-methyl RNA backbone was studied for the sequences 5 0 A L C L U L A L C L C L A L /3 0 r(UGAUGGU) and 5 0 G L C L U L A L C L U L G L / 3 0 r(CGAUGAC). As summarized in Table 2 , the stabilities of these duplexes at 37 C are within experimental error of those measured for 5 0 A M C L U M A L C M C L A M /3 0 r(UGAUGGU) and 5 0 G M C L U M A L C M U L G M /3 0 r(CGAUGAC), respectively. Evidently, the most effective use of LNA nucleotides is to space them every other nucleotide with the first LNA placed at the second nucleotide from the 5 0 end. Internal mismatches make duplex formation less favorable Table 5 contains thermodynamic parameters measured for the formation of duplexes containing single mismatches and the difference in stabilities relative to completely Watson-Crick complementary duplexes (Tables 1 and 2 ). All internal mismatches make duplex formation less favorable by at least 2 kcal/mol at 37 C corresponding to at least a 25-fold less favorable equilibrium constant for duplex formation. In general, terminal mismatches destabilize much less than internal mismatches. In fact, when the 3 0 terminal U L of 5 0 A M C M U M A M C M C M U L makes a GU pair, the duplex is stabilized by 0.14 kcal/mol at 37 C relative to a terminal AU pair. For four cases, the effect of a mismatch with an LNA nucleotide was compared with that for the equivalent 2 0 -O-methyl nucleotide. In each case, the mismatch penalty for the LNA was less than that for 2 0 -O-methyl RNA. However, for an A M -G mismatch flanked by LNAs in the context 5 0 A L C M U L A M C L C M A L /3 0 r(UGAGGGU), the LNAs enhanced the mismatch penalty by $1 kcal/mol relative to a completely 2 0 -O-methyl RNA strand. Thus, oligonucleotides containing LNA may discriminate best against mismatches flanked by LNAs. Oligonucleotide hybridization to RNA has many applications, ranging from quantifying gene expression (18) (19) (20) 56) to designing therapeutics (4, 8, 21, 46, (52) (53) (54) . LNA nucleotides have characteristics useful for these purposes. For example, LNA usually stabilizes duplexes (4, 44, 48, 51) and is more resistant than RNA and DNA to nuclease digestion (4, 6, 51) . The results presented here provide insights that are useful for designing 2 0 -O-methyl RNA/LNA chimeric oligonucleotides for various purposes. Some trends may be general for RNA A-form helixes and thus may also be relevant to other chimeras with nucleotides that favor A-form conformations. The results suggest several principles for the design of 2 0 -O-methyl RNA/LNA chimeras for hybridization to RNA The database in Tables 1 and 2 is too small to The magnitude and sequence dependence of the stabilization due to LNAs are surprising. Ribose and therefore probably 2 0 -O-methyl ribose sugars in single strands are typically found in roughly equal fractions in C2 0 -endo and C3 0 -endo conformations. If the methylene bridge of an LNA only locks the sugar into the C3 0 -endo conformation, then the expected stabilization due to preorganization would be: DDG ¼ ÀRT ln 2, which is À0.4 kcal/mol at 37 C (310.15 K). The stabilization observed for a 5 0 terminal LNA is roughly À0.4 kcal/ mol, but the average stabilizations for internal LNAs and 3 0 terminal A L , C L and G L are more favorable at À1.3 and À1.2 kcal/mol, respectively. Moreover, if stabilization was only due to preorganization of an LNA sugar, then the effect would not saturate when alternate sugars are LNA. Evidently, the LNA substitution also affects the 5 0 neighboring base pair in a way that enhances the stabilization beyond that expected from preorganization of a single sugar. Interestingly, NMR structures of DNA/LNA chimeras bound to RNA show that only the DNA sugar 3 0 of the LNA is driven to a C3 0endo conformation for the sequence d(5 0 CTGAT L ATGC)/ 3 0 GACUAUACG, but all non-terminal DNA sugars are C3 0 -endo when all three Ts are LNAs (76) . The free energy increments at 37 C for LNA substitutions ranged from +0.83 to À1.90 kcal/mol with an average of À0.55 kcal/mol. This compares with a range from +0.18 to À2.17 kcal/mol and an average of À1.32 kcal/mol for the single internal LNA substitutions in Table 1 . The comparision suggests that single LNA substitutions are on average more stabilizing to 2 0 -O-methyl RNA/RNA duplexes than to DNA/DNA duplexes. This may reflect the expectation that LNA substitutions do not have a large effect on the conformations of 2 0 -O-methyl RNA/RNA duplexes, but alter the conformations of DNA/DNA duplexes. LNA substitutions should be useful for probing RNA with short 2 0 -O-methyl RNA oligonucleotides RNA structure can be probed with short oligonucleotides on microarrays (3) . To optimize such methods, it is necessary to have tight binding that is sequence independent and that discriminates against mismatches. It appears that LNA nucleotides can be used to achieve this. For example, free energy increments for 2 0 -O-methyl RNA/RNA nearest neighbors range from À0.7 to À3.5 kcal/mol, corresponding to 5 0 A M U M /3 0 UA and 5 0 G M C M /3 0 CG, respectively ( Table 4 ). The average increment of À1.3 kcal/mol of internal and 3 0 terminal LNA can help compensate for such less favorable stability of AU relative to GC pairs. The stability enhancement from LNA can also allow the use of shorter oligonucleotides. The potential disadvantage to LNA substitutions in 2 0 -Omethyl RNA oligonucleotides is that discrimination against mismatches containing an LNA may be less than with a complete 2 0 -O-methyl RNA backbone. This was clearly true for three of the four cases where such direct comparisons were made. Nevertheless, internal mismatches with LNA nucleotides are considerably destabilizing, averaging a penalty of 4.1 kcal/mol at 37 C (Table 5) , which translates to almost a 1000-fold weaker binding due to a single mismatch. When LNAs flanked an A M -G mismatch, the mismatch penalty at 37 C was 4.4 kcal/mol compared with 3.3 kcal/mol in the absence of LNAs. Such an effect may reflect enhanced rigidity due to LNA, which thereby prevents a mismatch from adopting a favorable conformation. Thus, it may be advantageous to use LNAs to flank nucleotides likely to give small mismatch penalties.
37
Draft versus finished sequence data for DNA and protein diagnostic signature development
Sequencing pathogen genomes is costly, demanding careful allocation of limited sequencing resources. We built a computational Sequencing Analysis Pipeline (SAP) to guide decisions regarding the amount of genomic sequencing necessary to develop high-quality diagnostic DNA and protein signatures. SAP uses simulations to estimate the number of target genomes and close phylogenetic relatives (near neighbors or NNs) to sequence. We use SAP to assess whether draft data are sufficient or finished sequencing is required using Marburg and variola virus sequences. Simulations indicate that intermediate to high-quality draft with error rates of 10(−3)–10(−5) (∼8× coverage) of target organisms is suitable for DNA signature prediction. Low-quality draft with error rates of ∼1% (3× to 6× coverage) of target isolates is inadequate for DNA signature prediction, although low-quality draft of NNs is sufficient, as long as the target genomes are of high quality. For protein signature prediction, sequencing errors in target genomes substantially reduce the detection of amino acid sequence conservation, even if the draft is of high quality. In summary, high-quality draft of target and low-quality draft of NNs appears to be a cost-effective investment for DNA signature prediction, but may lead to underestimation of predicted protein signatures.
Draft sequencing requires that the order of base pairs in cloned fragments of a genome be determined usually at least four times (4· depth of coverage) at each position for a minimum degree of draft accuracy. This information is assembled into contigs, or fragments of the genome that cannot be joined further due to lack of sequence information across gaps between the contigs. To generate high-quality draft, usually $8· coverage is optimal (1). Finished sequence, without gaps or ambiguous base calls, usually requires 8· to 10· coverage, along with additional analyses, often manual, to orient the contigs relative to one another and to close the gaps between them in a process called finishing. In fact, it has been stated that 'the defining distinction of draft sequencing is the avoidance of significant human intervention' (1) , although there are computational tools that may also be capable of automated finishing in some circumstances (2) . While some tabulate the cost differential between highquality draft versus finished sequences to be 3-to 4-fold, and the speed differential to be >10-fold (1), others state that the cost differential is a more modest 1.3-to 1.5-fold (3) . In either case, draft sequencing is cheaper and faster. Experts have debated whether finished sequencing is always necessary, considering the higher costs (1, 3, 4) . Thus, here we set out to determine whether draft sequence data are adequate for the computational prediction of DNA and protein diagnostic signatures. By a 'signature' we mean a short region of sequence that is sufficient to uniquely identify an organism down to the species level, without false negatives due to strain variation or false positives due to cross reaction with close phylogenetic relatives. In addition, for DNA signatures, we require that the signature be suitable for a TaqMan reaction (e.g. composed of two primers and a probe of the desired T m s). Limited funds and facilities in which to sequence biothreat pathogens mean that decision makers must choose wisely which and how many organisms to sequence. Money and time saved as a result of draft rather than finished sequencing enables more target organisms, more isolates of the target and more near neighbors (NNs) of the target to be sequenced. However, if draft data do not facilitate the generation of highquality signatures for detection, the tradeoff of quantity over quality will not be worth it. We used the Sequencing Analysis Pipeline (SAP) (5, 6) to compare the value of finished sequence, real draft sequence and simulated draft sequence of different qualities for the computational prediction of DNA and protein signatures for pathogen detection/diagnostics. Marburg and variola viruses were used as model organisms for these analyses, due to the availability of multiple genomes for these organisms. We hope that variola may serve as a guide for making predictions about The online version of this article has been published under an open access model. Users are entitled to use, reproduce, disseminate, or display the open access version of this article for non-commercial purposes provided that: the original authorship is properly and fully attributed; the Journal and Oxford University Press are attributed as the original place of publication with the correct citation details given; if an article is subsequently reproduced or disseminated not in its entirety but only in part or as a derivative work this must be clearly indicated. For commercial re-use, please contact journals.permissions@oxfordjournals.org bacteria, in which the genomes are substantially larger, and thus the cost of sequencing is much higher than for viruses. Variola was selected as the best available surrogate for bacteria at the time we began these analyses because: i. it is double-stranded DNA; ii. it has a relatively low mutation rate, more like bacteria than like the RNA or shorter DNA viruses that have higher mutation rates and thus higher levels of variation; iii. it is very long for a virus, albeit shorter than a bacterial genome; iv. we have access to many genomes, which were sequenced by our collaborators at the US Centers for Disease Control and Prevention in Atlanta, GA; v. there are finished genomes available, so we can compare actual finished data with simulated draft data. Only recently have a fairly large number of Bacillus anthracis genomes become available to us. However, since only some of these are finished, currently we cannot compare finished with draft results for this bacterial genome. The sequencing analysis pipeline uses the DNA and protein signature pipelines The draft SAP simulations are nearly identical to those using finished genomes, described previously (5, 6) . The SAP ( Figure 1 ) performs stochastic (Monte Carlo) simulations and includes our DNA and Protein Signature Pipelines as components, which will be summarized briefly below. It is necessary to describe what the signature pipelines do before the SAP can be clearly described, so signature pipelines will be discussed first. As a step within the DNA and Protein Signature Pipelines, DNA sequence alignments of multiple draft genomes are required. For this we use the WGASA software, also summarized below. Once each of these components of the SAP has been presented, the SAP itself will be described. The DNA Signature Prediction Pipeline, described in detail elsewhere (7) (8) (9) (10) , finds sequence regions that are conserved among target genomes by creating a consensus based on a multiple sequence alignment. WGASA is the software used in the analyses here to create an alignment and will be discussed below. Next, the DNA Signature Pipeline identifies regions that are unique in the target sequence consensus relative to all other non-target bacterial and viral sequences that we have in a >1 Gb database, which is frequently updated from the NCBI GenBank sequence database (11) and other sources (e.g. our collaborators at the CDC, USDA and other public sources, such as TIGR, Sanger Institute and the Joint Genome Institute). From the conserved, unique regions, signatures are selected based on the requirements of a particular technology, in this case, TaqMan PCR. These signature candidates may then proceed for further in silico screening (BLAST analyses to look for undesired inexact matches) before undergoing laboratory screening. For an SAP run, first a pool of target genome and a pool of NN genomes are collected. Then many random subsamples of target and NN genomes are selected from the pool, and each subsample is run through either the DNA signature pipeline or the protein signature pipeline, which identify regions conserved among target genomes and unique relative to non-target genomes, where unique regions are evaluated by comparing to a large sequence database of all currently available bacterial and viral complete genomes or the non-redundant protein database, excluding NNs from the NN pool that are not in that random subsample. Thus, each run of the SAP requires many runs of the DNA or protein signature pipelines with different random samples, generating a range of outcomes that are plotted on range plots. Protein signature prediction and SAP methods have previously been described in detail (6) . The following briefly describes the procedure. First, target genomes are aligned using WGASA. A set of gene (start, end) pairs for both the plus and minus strands relative to the reference genome is required. This implies that coding frames for the translation of nucleic acid codons into amino acids for each protein of the target organism's genome have been correctly determined. From the aligned genomes, nucleotide codons are translated into amino acid sequence based on the gene locations, and conserved strings of six or more amino acids among all the target genomes are recorded. These conserved fragments are then compared with the NCBI GenBank non-redundant (nr) database of amino acid sequences, unveiling peptides that are unique to the target species. For our computations, we require that if a peptide signature is longer than six amino acids, then every sub-string of length six amino acids is also conserved and unique. There may be many conserved and unique peptide signatures on the same and on different proteins. The resulting conserved, unique peptides that are at least six amino acids long from open reading frames are considered to be protein signature candidates. Signature peptides may be used as targets for antibody or ligand binding and may be developed for use in detection, therapeutics or vaccines (12, 13) . Since the signature regions are highly conserved within a species, it is likely that they are functionally important to the organism's survival or reproduction. Those signatures that land on or near protein active sites may be developed into therapeutics, since antibody or ligand binding may interfere with protein function. Signature regions may even be considered as vaccine targets, since these unique peptides may evoke a specific response in the host (14, 15) . For draft genomes, WGASA, or Whole Genome Analysis through Scalable Algorithms, is used to align multiple sequences. This is the only available tool that enables multiple sequence alignment of draft genomes and that is capable of aligning large or many genomes. WGASA requires at least one finished reference genome and the others may be draft. Only recently has it become possible to use the DNA Signature Pipeline to predict signature candidates for draft genomes. This capability is due to the invention of software for multiple sequence alignment of draft genomes with at least one completed full genome. WGASA was developed by David Hysom, Chuck Baldwin and Scott Kohn in the Computations directorate at Lawrence Livermore National Laboratory. They designed the software in close communication with members of our bioinformatics team, and it is tailored for our needs of generating diagnostic and forensic pathogen signatures. WGASA can efficiently align large (e.g. bacterial) genomes. In addition, the developers have created a parallel version that runs in minutes, allowing the SAP simulations, involving thousands of calls to WGASA, to complete in a feasible time frame. In addition to the SAP analyses, this tool has enabled us to revisit signature predictions for several important organisms, such as the food-borne pathogen Listeria, that were previously problematic because some of the sequences were available only in draft format. The tool requires that there be one or more complete, finished genome, and any number of draft sequences. It is based on suffix-tree algorithms (16) . It requires that anchors, identical sequence fragments of user-specified length, be found in each of the genomes to be aligned. Thus, there must be some level of sequence conservation among the genomes in order to discover anchors of sufficient length (e.g. 35-60 bp) that are present in all the genomes. Then the regions between the anchors are aligned using a tool, such as clustalw or HMMer. The algorithm functions most efficiently if anchors are frequent and dispersed across the genomes to provide even coverage. If substitutions, deletions, insertions or gaps in sequence information (e.g. between contigs) result in an anchor's absence in one or more of the genomes, then those regions must be aligned using clustalw, which is slower and more memory intensive for large amounts of sequence data. Similar to all anchor-based alignment algorithms, WGASA is dependent upon a high degree of co-linearity across all input genomes. The SAP for DNA signature analyses operates as follows. First, all available complete genomes of target were gathered into a pool with the total genome count called T. A second pool was created of all available NN complete genomes, with the total count of sequences called N. Next, we selected 10 random samples of size t targets and n NNs, for all t ranging from 1 to T and all n ranging between 1 and minimum(10,N). We ran the DNA Signature Prediction Pipeline for each sample, with signature prediction based on conservation among the t target strains and uniqueness relative to a >1 Gb database minus those NNs in the NN pool that were not chosen in that sample. Thus, for each sample, signature candidates were predicted as though we had only t target and n NN sequences, as well as the rest of the less-closely related organisms in our database that are not considered NNs. In addition to the number of TaqMan signature candidates, the fraction of the genome that is conserved among the t target sequences was also calculated. Based on the combined results of the many signature pipeline runs using random target samples of size t and n, we assessed how much sequence data, that is, the values of t and n, was required to approximate the number of signature candidates c that were predicted when the full data set (all target and NN sequences, t ¼ T and n ¼ N) was analyzed with the signature pipeline. Using the full data set will yield the fewest signatures, because lack of conservation or uniqueness will winnow away all unsuitable candidates. Thus, the SAP performs Monte Carlo sampling from the target and NN genomes, runs each sample through the signature pipeline and summarizes the results of the hundreds of signature pipeline runs in a single plot. On our 24 CPU Sun server, up to seven signature pipeline simulations may be run in parallel, each requiring $15-22 min for viral genomes. All of the SAP analyses of dozens of bacteria and viruses to date have used a total run time of 6.26 years (operating in parallel), with an average pipeline run time of 0.522 h, and a process time span of 2.32 years. The span of predictions generated by different random samples of genomes is illustrated using range plots (Figures 2-8 ). Along the y-axis, whole numbers represent the number of target strains t and the incremental values between the integers represent the number n of NN genomes. Only Figures 3, 5 and 6 have the incremental n values, because for the other plots of target sequence conservation only, the number of NNs was not relevant, and for the protein analyses NN comparisons were not made (described below). Outcomes of the number of signature candidates or the fraction of the target genome that is conserved are plotted along the x-axis as horizontal lines spanning the range (of predicted numbers of signatures or fraction conserved) for the s random samples of size (t,n) with the median and quantiles of the range indicated by colored, short vertical lines. If a random sample of t target strains and n NN strains were sequenced, there would be a 90% chance that the number of signature candidates for that sample would be less than or equal to the 90% quantile mark. The expected outcome is a reduction in the number of signature candidates or the fraction of the genome that is conserved as the number of target and NN sequences used in the simulations increases, due to a reduction in conservation from additional targets and a reduction in uniqueness from additional NNs. The SAP analyses for proteins proceed much like that for DNA signatures. Random samples of size t target sequences are generated, where t ranges from 1 to T, the total number of target sequences in the pool. Either finished data, actual draft data, or draft data simulated as described below are aligned using WGASA. The protein signature prediction pipeline is run on each random sample, and the range, median, 75th and 90th quantiles of the number of protein signature candidates for the samples of a given target size t is plotted in range plots as described above. Our DNA SAP analyses examined the effects of both the number of target as well as the number of NN sequences, but To discriminate samples in which zero NNs were used, the range is drawn as a horizontal gray line, and when n > 0, the range is drawn as a black line. The best estimate of the true value is the quality measure determined using the entire target and NN pools, and is represented by a vertical black line. This best estimate plus a constant c ¼ 20 is at the location of the vertical dashed line and was selected to indicate a reasonable distance from the true answer. The 75% quantile for each range is shown with a black, vertical tick mark. our protein SAP analyses investigated the effects of only the number of target sequences. This is because composing the lists of NN proteins for random, temporary exclusion from the protein nr database (to estimate the value of that NN sequence data) would be difficult to automate for rapid, high-throughput computations. Thus, we compared the target proteins with all the proteins in nr, regardless of their phylogenetic relationship to the target. This was comparable with DNA SAP results using all available NN data. We had sequence data for four strains of Marburg virus, both the actual draft and the finished versions of those same isolates, provided for these analyses by a colleague working at Lawrence Livermore National Laboratory. The draft sequence was of $3· to 6· coverage, which enabled us to compare SAP results using the same strains in finished form. The identities of these sequences are provided in the Appendix. For the draft Marburg analyses, we selected one finished strain, the reference strain from GenBank (gi|13489275|ref|NC_001608.2| Marburg virus, complete genome), as the WGASA reference genome, and then used random sub-samples from the four draft genomes. Marburg was the only organism for which we could obtain a sufficient number of draft genomes for the SAP Monte Carlo simulations. A total of 814 simulations for DNA signatures, i.e. individual runs of the DNA signature pipeline, and 48 simulations for protein signatures were performed using Marburg finished and draft data, requiring an average of 15 min per simulation. We used finished sequence data generously provided by collaborators at the US CDC for 28 variola major genomes and 22 NN genomes from the Orthopox family. The sequence identities are provided in the Appendix. Since we did not have real draft data available, we developed a program to simulate draft sequence from finished sequence, based on guidance from two colleagues who have been involved in sequencing efforts and the finishing process in the Biology and Biotechnology Research Program at Lawrence Livermore National Laboratory. In outline, the draft simulator program randomly cuts a genome into contigs of a size randomly selected from an exponential distribution. Stochastic simulation also determines whether there are gaps or overlaps between contigs, as well as the size of the gap or overlap. Sequencing errors are also simulated. The following paragraphs describe the draft simulation process in greater detail. First, the 5 0 end of the sequence is simulated as missing or present according to a random (Bernoulli) trial based on the probability of there being a gap in the sequence data. If simulations randomly determine that the first part of the sequence is missing, then the size of the missing segment is selected randomly from a uniform distribution ranging from the minimum gap size to the maximum gap size. The length of the first contig is selected randomly from an exponential distribution with a non-zero minimum contig size and a maximum contig size that is a fraction of the mean genome length for the species. The mean of this exponential distribution is also specified as a fraction of the mean genome length. Next, a random Bernoulli trial again determines whether there is a gap or overlap between the first and second contigs, and the size of the gap or overlap is chosen from the appropriate uniform distribution (range for gap size ¼ 1-2000 bases, range for overlap size ¼ 20-40 bases). The size of the contig is selected from the exponential distribution as described above. Additional contigs are simulated in a similar manner. Within each contig, sequencing errors are simulated based on the size of the contig, and whether the base position is at an end (first or last 100 bases) or in the middle of the contig. For long, double-stranded DNA viruses (e.g. variola) and bacteria, the sequencing error rates are larger at the beginning and end of a contig than in the middle, and small contigs are more likely to contain sequencing errors than are large contigs. In contrast, due to differences in generating the products for Sanger sequencing that are employed for smaller RNA and DNA viruses, there are often more sequencing errors in the middle of contigs for such smaller viral draft genomes. Although we did not specifically simulate draft for RNA and short DNA viruses, our simulator should work with minor modification to a few parameters. Thus, there are four parameters that must be specified for simulating sequencing errors: (i) the size cutoff for small versus large contigs, (ii) the probability of errors in the middle portion of small versus large contigs, (iii) the length of the contig ends where sequencing is either less accurate (bacteria and long doublestranded DNA viruses) or more accurate (small viruses, RNA viruses) and (iv) the probability of sequencing errors at the contig ends. If there is a sequencing error at a particular base, we assumed that that base is randomly changed to one of the other three bases with equal probability. Although additional features could be added to the draft simulation tool, the stochastic features that we have incorporated capture the main features of draft sequence and produce data that are suitable for SAP analyses. We performed six sets of analyses using simulated variola draft. Three sets of simulated variola draft runs of the SAP used the following parameters: probability of a gap between contigs ¼ 0.95; probability of overlap between contigs ¼ 0.05; minimum gap size if there is a gap (uniform distribution) ¼ 1 bp; maximum gap size ¼ 2000 bp; minimum overlap if there is overlap (uniform distribution) ¼ 20 bp; maximum overlap ¼ 40 bp; minimum contig size (exponential distribution) ¼ 2000 bp; maximum contig size ¼ 0.5 · (mean genome length) bp mean contig size ¼ 0.05 · (mean genome length) bp cutoff size for small versus large contigs ¼ 10 000 bp; probability of sequence errors inside large contigs ¼ 0.01; probability of sequence errors inside small contigs ¼ 0.05; We will refer to the above set of simulations as those with a high probability of sequencing errors, or low-quality draft. The other three simulated variola draft runs used all the same parameters as above, except that the sequencing error rates were dramatically lower, more in line with error rates of 10 À5 / base that the US Centers for Disease Control and Prevention (CDC) has indicated for their draft variola genomes: probability of sequence errors inside large contigs ¼ 10 À5 ; probability of sequence errors inside small contigs ¼ 10 À4 ; probability of sequence errors in the contig ends ¼ 10 À3 ; These runs were referred to as low error rate, or highquality draft. Finally, we performed SAP runs using high error rate (low quality) simulated draft of the NN sequences and intermediate quality simulated draft of target genomes, using the following probabilities of sequencing errors: probability of sequence errors inside large contigs ¼ 10 À3 ; probability of sequence errors inside small contigs ¼ 10 À3 ; probability of sequence errors in the contig ends ¼ 10 À3 ; The intermediate quality simulated draft is consistent with error rates for draft sequencing cited in the literature (1,3) . For the parameter values specified above, three SAP experiments were simulated. In the first, only the target sequences were simulated into draft, and the NN sequences remained as finished sequences. In the second, the NN sequences were converted to simulated draft and the target sequences remained as finished. In the third, both target and NN sequences were simulated into draft. In the second and third cases, all the NNs were run through the draft simulator each time they were chosen, so that the draft sequences (i.e. location and extent of gaps and sequence errors) differ for the same genome among samples. In the first and third cases, the target sequences must be aligned, and WGASA requires that one of the sequences be a finished genome for reference. Thus, for each random sample from the pool of target genomes, one genome was randomly selected to be the finished genome, and so was left as finished sequence, and the other genomes in the sample were replaced with simulated draft sequence (by running through the draft simulator) before alignment. As with NNs, target draft sequences differ for the same genome among samples due to the randomness of the draft simulation each time it is run. In addition, the target genome that is chosen to be the finished, reference genome differs between samples, and the other target genomes in the sample simulation are replaced with simulated draft versions of the actual finished sequences. Then these sequences were aligned using WGASA and the SAP process was run as described above. A total of 1101 stochastic simulations per 'experiment' were performed, requiring $18 min per simulation. Each simulation involved randomly selecting the subset of target and NN sequences to be included, simulating the draft data based on the finished genomes, aligning the target sample, and finally running the DNA Signature Pipeline. There were four combinations examined: (i) finished variola and finished NN, (ii) draft variola and finished NN, (iii) finished variola and draft NN and (iv) draft variola and draft NN, with each of the draft runs repeated for both low and high sequencing error rates. The combination (iv) was also run with intermediate quality simulated draft variola and low-quality simulated draft NNs. In total, there were eight computational experiments for the finished and simulated draft variola data. We have used the following function to estimate viral sequencing costs, based on discussion with our laboratory colleagues involved in sequencing and finishing. This is merely a rough estimate, and the actual costs of sequencing any given organism may differ substantially from this rule-of-thumb calculation. In Equation 1, it is assumed that the cost of sequencing viruses does not decline for sequencing second and subsequent isolates. While this may be a false assumption in cases where isolates are similar to one another, in other cases where the new sequences are divergent, as isolates from different outbreaks or for viruses with rapid mutation rates, the cost is especially unlikely to decline. In addition, the $0.40/bp figure for draft of 6· to 8· coverage could range from $0.30 to $0.50/bp using shotgun sequencing, but may be as low as $0.10-$0.20/bp if primer walking works well (i.e. known primer sites are found in new isolates). Finishing could be 1-3 times more than the cost of draft, so we used a factor of 2 times more (draft $0.4/bp, finished $0.4 + 0.8/bp) in the equation above as a reasonable estimate. With rapidly evolving sequencing technologies and costs, these are only rough guides that may quickly become outdated. It may be substantially less expensive, on the order of 3-fold, to generate draft compared with finished sequence data for an organism like Marburg virus, according to estimates using Equation 1. For example, for $45K, one could sequence either two finished genomes or one finished and three draft. However, draft sequencing of this low quality (3· to 6·) for Marburg causes a dramatic decline in the ability to computationally eliminate regions of poor conservation, and thus to exclude poor signature regions ( Figure 2 ). This occurs because gaps in the draft data of some of the sequences mask sequence variation among strains. Using the best available data, all six finished genomes, there is 75.2% sequence conservation. The deficiencies of draft data give a false impression that there is 92.6% sequence conservation (Table 1) . Each additional finished genome reduces the conserved fraction by $5%, compared with a reduction of only 2% per genome for the draft data. The overestimation of conservation using draft Marburg data also results in overestimation of the number of signature candidates (Figure 3 ). Samples of four draft targets plus one finished reference yield 43 signature candidates. A smaller sample size of only two draft targets and one finished reference generate upwards of 80 candidates. These results differ from those using finished genomes, where the lack of sequence conservation is more evident and there are zero TaqMan signatures conserved among all strains. Most combinations of four finished genomes are sufficient to eliminate nonconserved signatures ( Figure 3A ). Although predictions that there are 0 signature candidates shared among all finished strains may seem to argue against TaqMan methods, in fact this information provides important guidance for the development of TaqMan signatures with degenerate bases or a set of signatures that will, in combination, pick up all sequenced strains. Other analyses indicate that there are TaqMan signatures conserved among five of the six strains, so that two signatures would form a minimal set that could detect both the one divergent and the other five strains. Estimated sequencing costs of draft variola and draft NNs indicate that draft may require only one-quarter to one-half the costs of finished sequencing. Simulations of high-quality draft data indicated that it is as good as finished data for diagnostic signature prediction. The conservation range plots ( Figure 4A and B) are virtually identical for finished and high-quality draft, and indicate that $98% of the genome is conserved among sequenced isolates. For intermediate quality draft ( Figure 4C ) the conservation range plot is also similar to that for finished sequence, showing that $97% of the genome appears to be conserved. The range plots for the number of TaqMan signature candidates are very similar for finished sequence data, high or intermediate quality draft target, and high or low-quality draft NNs ( Figure 5A-D) . In contrast to the results using high-quality simulated draft or actual Marburg draft, simulations of low-quality variola draft target illustrate that sequence conservation may be underestimated compared with results with finished sequence data, due to sequencing errors (Table 1 and Figure 4D ). With lowquality draft target, it appears that only 58% of the genome is conserved among isolates. Low-quality (high error rate) draft NN data, however, yield results that are very similar to those with high-quality draft or finished NN data, as long as the target sequence information is of intermediate to high quality ( Figure 5C -D andFigure 6): At least four NN sequences are necessary to ensure that signature regions are unique, whether the NN data are low-or highquality draft or finished. That is, our simulations indicate that low-quality draft NN data are adequate for predicting DNA signatures, as long as there is good quality target sequence data. This results because errors in the NN sequences occur at random locations that differ in each NN sequence. As long as at least one of the NNs has enough correct sequence to eliminate each of the non-unique target regions, then the unique regions of the target can be determined. The results illustrated in the figures are emphasized by the data in Table 1 . This table shows the fraction of the target genome that is conserved and conserved+unique, the number of conserved+unique regions that are at least 18 contiguous base pairs long and the number of base pairs in the largest of these regions, since these are the sections that are of sufficient length for one or possibly more primers to be located. The number of these regions is similar for finished data and for draft with a low error rate. Low-quality draft (with a high rate of sequencing errors) for the target data, however, gives the false impression that there are fewer and shorter regions that are conserved and suitable as signature regions than is actually the case. There is an artifact in some of our results that is a consequence of the order in which we calculate conservation and then uniqueness, although this does not affect the signatures that are predicted. First, a conservation gestalt is generated from the sequence alignment, in which non-conserved bases are replaced by a dot ('.'). Then uniqueness is calculated based upon perfect matches of at least 18 bp long between the conservation gestalt and a large sequence database of non-target The percent of the target genome that is conserved varies slightly among the runs using finished target sequences because different genomes were randomly selected to be the reference strain in each multiple sequence alignment. sequences. Non-conserved bases in the conservation gestalt may break up a region into conserved fragments of <18 bases long, and as a result these short fragments are not tested for uniqueness. Consequently, if there is a low level of conservation, then we may overestimate the fraction of the genome that is unique. For example, in Table 1 the conserved+unique fraction is 4% with finished variola target data, but is overestimated at 58% with low-quality draft. This artifact does not, however, affect TaqMan signature prediction, since the regions suitable for primers and probes must have at least 18 contiguous, conserved bases, and all of these are tested for uniqueness, i.e. there is no underestimation of uniqueness in conserved fragments that are at least 18 bp in length, and thus no underestimate of uniqueness in the predicted signatures. We are working to eliminate this issue in future versions of the software. Protein results show a large disparity between finished and draft data. There are 113 protein signature candidates for finished Marburg data compared with only two protein signature candidates for Marburg draft (Figure 7 ). For variola, using all available target data, 97, 14, 6 and 0 protein signatures are predicted using finished, low error, intermediate error and high error draft target data, respectively ( Figure 8 ). Thus, sequencing errors substantially reduce the detection of amino acid sequence conservation, even if sequencing errors occur at the low rate of 10 À4 -10 À5 across most of the genome. The pattern of how additional sequences reduce the number of protein signature candidates also differs for draft compared with finished sequence data. With finished data, there is a large range in the number of peptide signature candidates predicted with 17 or fewer variola genomes, and this range narrows around the lower bound with >17 genomes. With 16 genomes, the 75% quantile mark approaches the final predicted number of 97 signatures ( Figure 8A ). This pattern indicates that there is a set of 97 peptides that are highly conserved among all currently sequenced variolas, which are unlikely to be eroded even as more sequence data are obtained. In other words, additional sequence data are probably not needed at this time in order to computationally predict good peptide signature targets, and as few as 16 finished target sequences would most likely have been adequate to generate this same list of $100 peptide signatures. Draft data, in contrast, whether they are of low or high quality, mask the above pattern ( Figure 8B and C): the range and 75th quantile of the number of peptide signatures gradually decline with each additional target sequence (rather than a sudden, sharp drop as is seen with the finished data), suggesting that additional target sequences would continue to erode the number of peptide signatures. This occurs since sequencing errors occur at random, in different locations in each of the draft target genomes, and obscure the truly conserved peptides. One might falsely infer from peptide SAP results based on the draft data that additional sequencing (beyond the 28 variola major genomes used here) would be useful in generating peptide signature candidates. In actuality, however, SAP analyses using the finished sequence data indicate that there are already ample sequence data for peptide signature prediction. The failure of draft sequencing for Marburg at 3· to 6· coverage or of simulated variola draft with a high error rate to facilitate the prediction of detection signatures highlights a need for finished viral sequences, or at least for draft of high quality such as 8·. Otherwise, a large number of signature candidates either will fail in screening because they are incorrectly designated as conserved among strains (as observed with the Marburg results), or too few regions will be classified as conserved (as observed with variola), and thus not be considered for signatures. The variola simulations with intermediate to high-quality draft (that is, a low error rate, approximating what one might observe with 8· coverage) target and/or NN genomes deliver virtually the same results as finished genomes. Considering that it costs approximately three times as much to generate finished sequence as it does draft, our analyses indicate that investing in more high-quality draft target genomes is better than investing in fewer finished genomes. For our analyses, only one target strain must be finished, and the remaining target sequences and all the NNs may be provided as draft. Our results indicate that NN sequencing may be of low coverage, and thus of low quality, without serious detriment to signature prediction, as long as there are at least four NN draft genome sequences. If high-quality draft sequence is used, and it appears that there is too little sequence conservation among target strains, one might relax specifications for 100% conservation among strains for diagnostic signature prediction. Calculations indicate that it is often possible to generate signatures if one allows a base to be considered 'conserved' if it is present in only a fraction of the genomes (e.g. 75%) rather than the standard requirement for 100% conservation when finished sequence data are used. We have used this 'ratio-to-win' option to generate signature candidates for some highly divergent RNA viruses (for which we have finished sequence), although usually our preference is to include degenerate bases, especially when there are only a few bases with heterogeneity among strains in a given signature candidate. Using a ratio-to-win approach may be particularly important for the generation of protein signature candidates, since draft target data severely compromises the ability to detect conserved strings of amino acids. In summary, intermediate to high-quality draft sequencing of target genomes, combined with low-quality draft sequencing of close phylogenetic relatives, is sufficient for the prediction of DNA diagnostic signatures. Prediction of peptide/ protein signature candidates, in contrast, requires finished sequencing to avoid substantial underestimation of conserved peptide regions.
38
An ontology for immune epitopes: application to the design of a broad scope database of immune reactivities
BACKGROUND: Epitopes can be defined as the molecular structures bound by specific receptors, which are recognized during immune responses. The Immune Epitope Database and Analysis Resource (IEDB) project will catalog and organize information regarding antibody and T cell epitopes from infectious pathogens, experimental antigens and self-antigens, with a priority on NIAID Category A-C pathogens () and emerging/re-emerging infectious diseases. Both intrinsic structural and phylogenetic features, as well as information relating to the interactions of the epitopes with the host's immune system will be catalogued. DESCRIPTION: To effectively represent and communicate the information related to immune epitopes, a formal ontology was developed. The semantics of the epitope domain and related concepts were captured as a hierarchy of classes, which represent the general and specialized relationships between the various concepts. A complete listing of classes and their properties can be found at . CONCLUSION: The IEDB's ontology is the first ontology specifically designed to capture both intrinsic chemical and biochemical information relating to immune epitopes with information relating to the interaction of these structures with molecules derived from the host immune system. We anticipate that the development of this type of ontology and associated databases will facilitate rigorous description of data related to immune epitopes, and might ultimately lead to completely new methods for describing and modeling immune responses.
An epitope can be defined as the molecular structure recognized by the products of immune responses. According to this definition, epitopes are the specific molecular entities engaged in binding to antibody molecules or specific T cell receptors. An extended definition also includes the specific molecules binding in the peptide binding sites of MHC receptors. We have previously described [1] the general design of the Immune Epitope Database and Analysis Resource (IEDB), a broad program recently initiated by National Institute of Allergy and Infectious Diseases (NIAID). The overall goal of the IEDB is to catalog and organize a large body of information regarding antibody and T cell epitopes from infectious pathogens and other sources [2] . Priority will be placed on NIAID Category A-C pathogens (http://www2.niaid.nih.gov/Biodefense/ bandc_priority.htm) and emerging/re-emerging infectious diseases. Epitopes of human and non-human primates, rodents, and other species for which detailed information is available will be included. It is envisioned that this new effort will catalyze the development of new methods to predict and model immune responses, will aid in the discovery and development of new vaccines and diagnostics, and will assist in basic immunological investigations. The IEDB will catalog structural and phylogenetic information about epitopes, information about their capacity to bind to specific receptors (i.e. MHC, TCR, BCR, Antibodies), as well as the type of immune response observed following engagement of the receptors (RFP-NIH-NIAID-DAIT-03/31: http://www.niaid.nih.gov/contract/archive/ rfp0331.pdf). In broad terms, the database will contain two general categories of data and information associated with immune epitopes -intrinsic and extrinsic (context-dependent data). Intrinsic features of an epitope are those characteristics that can be unequivocally defined and are specified within the epitope sequence/structure itself. Examples of intrinsic features are the epitope's sequence, structural features, and binding interactions with other immune system molecules. To describe an immune response associated with a specific epitope, context information also needs to be taken into account. Contextual information includes, for example, the species of the host, the route and dose of immunization, the health status and genetic makeup of the host, and the presence of adjuvants. In this respect, the IEDB project transcends the strict boundaries of database development and reaches into a systems biology application, attempting for the first time to integrate structural information about epitopes with comprehensive details describing their complex interaction with the immune system of the host, be it an infected organism or a vaccine recipient [1] [2] [3] . For these reasons, it was apparent at the outset of the project that it was crucial to develop a rigorous conceptual framework to represent the knowledge related to the epitopes. Such a framework was key to sharing information and ideas among developers, scientists, and potential users, and to allowing the design of an effective logical structure of the database itself. Accordingly, we decided to develop a formal ontology. Over the years, the term "ontology" has been defined and utilized in many ways by the knowledge engineering community [4] . We will adopt the definition of "ontology" as "the explicit formal specifications of the terms in a domain and the relationships among them" [5] . According to Noy and McGuinness [6] , "ontology defines a common vocabulary for researchers who need to share information in a domain and helps separate domain knowledge from operational knowledge". Thus, availability of a formal ontology is relevant in designing a database, in cataloging the information, in communicating the database structure to researchers, developers and users, and in integrating multiple database schema designs and applications. Several existing databases catalog epitope related data. We gratefully acknowledge that we have been able use these previous experiences in the design and implementation of the IEDB. MHCPEP [7] , SYFPEITHI [8] , FIMM [9] , HLA Ligand Database [10] , HIV Immunology Database [11] , JenPep [12] , AntiJen [13] , and MHCBN [14] are all publicly available epitope related databases. In general, these databases provide information relating to epitopes, but do not catalog in-depth information relating to their interactions with the host's immune system. It should also be noted that none of these databases has published a formal ontology, but all of them rely on informal or implicit ontologies. We have taken into account as much as possible these ontologies, inferring their structure by informal communications with database developers or perusal of the databases websites. The ontology developed for IEDB and described herein complements two explicit ontologies that are presently available: the IMGT-Ontology and the Gene Ontology (GO). The IMGT-Ontology [15] was created for the international ImMunoGeneTics Database (IMGT), which is an integrated database specializing in antigen receptors (immunoglobulin and T Cell receptors) and MHC molecules of all vertebrate species. This is, to the best of our knowledge, the first ontology in the domain of immunogenetics and immunoinformatics. The GO project [16] provides structured, controlled vocabularies that cover several domains of molecular and cellular biology. GO provides an excellent framework for genes, gene products, and their sequences, but it does not address the specific epitope substructure of the gene products. The IMGT provides an excellent ontological framework for the immune receptors but lacks information relating to the epitopes themselves. Therefore it was necessary to expand the available ontologies and to create an ontology specifically designed to represent the information of immune receptor interaction with immune epitopes. Wherever possible, the IEDB ontology conforms to standard vocabularies for capturing values for certain fields. For capturing disease names, IEDB uses the International Classification for Diseases (ICD-10) [17]. The NCBI Taxonomy database nomenclature [18, 19] is used to capture species and strain names, and HLA Allele names are consistent with the HLA nomenclature reports [20] . The IEDB is being developed as a web-accessible database using Oracle 10g and Enterprise Java (J2EE). Industry standard software design has been followed and it is expected that IEDB will be available for public users by the end of 2005. Protégé http://protege.stanford.edu was used to design and document the IEDB ontology. Protégé is a free, open source ontology editor and knowledge-base framework, written in Java. It provides an environment for creating ontologies and the terms used in those ontologies. Protégé supports class, slot, and instance creation, allowing users to specify relationships between appropriate entities. Two features that IEDB ontology effort used extensively were Protégé's support for creating ontology terms and for viewing the term hierarchies and the definitions. The support for a central repository on ontologies, along with browsing support, is key in reviewing and reusing ontologies. While there are several open source tools available [21] for developing ontologies, we selected Protégé because of its extensibility to a variety of plug-ins that are readily available for integration. It also has the ability to export to different formats including the Ontology Web Language (OWL) (http://www.w3.org/TR/owl-features/), which allows interoperability with other ontologies. We have previously described some of the general concepts relating to the IEDB design [1, 2] . More information relating to various aspects of the project can be accessed at http://www.immuneepitope.org/. Herein, we report a detailed description of the novel aspects of the IEDB ontology. In designing our application architecture, we have followed the common system engineering practice of first determining the scope and nature of the data involved. A first essential step is to understand the semantics of the domain and to capture that knowledge in an agreed-upon format. Arranging the domain concepts in a taxonomy is one of the initial organizing steps in the ontology design process. The class hierarchy represents the generalization and specialization relationships between the various classes of objects in a domain [6] . Briefly, classes describe concepts in the domain. Subclasses represent concepts (classes) that are more specific than the superclass and these subclasses can have their own unique properties. Slots represent properties of the classes. For example, in Figure 1 , we see that there is a class named Reference and three more specific subclasses of Reference: Journal Article, Patent Application, and Direct Submission. Figure 1 also shows that the class Epitope has a number of properties (slots) associated with it such as "has Epitope Structure" and "has Epitope Source". Our approach for creating the class hierarchy was a topdown development process where we defined each class in a domain and then identified its properties before building the hierarchy. The main classes identified for IEDB are Reference, Epitope Structure, Epitope Source, MHC Binding, Naturally Processed Ligand, T Cell Response, and B Cell Response (Figure 1 ). The Epitope class is the main class that encompasses all the individual concepts that were identified. The individual concepts are related to other classes. The primary relationships use the sub-class relationship or use a property (shown in the figures by the arcs labeled "has") that has a restriction on the type of the value that may fill that slot. "Reference" is the class encompassing information related to the data source from which an epitope and its related information are extracted into the IEDB. We have identified three broad subclasses of References that describe where epitope information will be obtained. They are Journal Article, Patent Application, and Direct Submission. The complete listing of slots (fields) encompassed by the Journal Article, Patent Application, and Submission classes are provided in Figure 2 . The Journal Article class refers to manuscripts published in peer-reviewed journals. The Patent Application class captures all the reference fields for a patent application that contain epitope information. The Submission class captures information about sources that contribute data to the IEDB directly. Data deposited by the Large Scale Antibody and T Cell Epitope Discovery contracts [3] and those transferred from other websites fall into this class. The Epitope Structure and Epitope Source classes capture intrinsic features of an epitope. The Epitope Structure class captures the physical and chemical features of an epitope. Virtually any molecular structure may provoke an immune response, such as proteins, carbohydrates, DNA, and lipids. In the Epitope Structure class, structural information relating to linear sequences and 2-D structures of the epitope, if available, are catalogued. The Epitope Source class captures the phylogenetic source of an epitope, including species of origin, gene name, protein name, and links to other databases for more detailed information about proteins and genes. Figures 3A and 3B show the listings of properties (slots/fields) encompassing the Epitope Structure and Epitope Source classes. The experimental data and information about specific experiments and the methodology utilized are captured in the Assay Information class. The name of the assay used, the type of response measured in the assay, and the readout of the assay are examples of information captured in the Assay Information class. This important class is used as a superclass of several other classes (and thus its properties are inherited by those classes). A complete listing the properties (slots/fields) in the Assay Information class is shown in Figure 3C . As with Assay Information, the classes Immunization, Antigen, and Antigen Presenting Cell are used in multiple other class descriptions. Features relating to the induction of the immune response are captured in the Immunization class ( Figure 4A ). It has relationships to other classes like Immunized Species, Immunogen, In vivo Immunization, and In vitro Immunization. Immunized Species contains information relating to the host that is being immunized. The Immunogen class describes the molecules that induce the immune response and an associated carrier molecule, if present. Features relating to how the immunogen was introduced to the immunized species are captured under the In vivo and In vitro Immunization classes. Similarly, antigens are defined as the whole molecules that react with the products of an immune response (as opposed to the epitopes which are the specific structures, contained within the antigen that engages the immune receptor). Information relating to the antigen and any associated carrier molecule is captured in the Antigen class ( Figure 4B ). During immune responses, antigen-presenting cells process antigens and present peptide epitopes complexed with MHC molecules. This information is captured in the Antigen Presenting Cells class, which has a relationship to the MHC Molecules and the Source Species classes ( Figure 4C ). The Source Species class describes the species information from which the antigen presenting cells are derived. The MHC Binding class captures the details relating to the interaction of the epitope with specific MHC molecules and information relating to the MHC molecule along with any available Epitope-MHC complex structure details. This class also has a slot that is restricted to be an instance of the Assay Information class ( Figure 5A ). Overview of IEDB Class Hierarchy Figure 1 Overview of IEDB Class Hierarchy. classes. Extrinsic features are context-dependent attributes, being dependent upon specific experimental conditions. The Naturally Processed Ligand class captures data related to epitopes that are naturally processed and presented on the cell surface. This class has properties that are instances of classes including Antigen Presenting Cell, Antigen, and Assay Information ( Figure 5B ). The Naturally Processed Ligand class differs from the MHC Binding class in that information related to the antigen that was processed and the cell types in which the processing occurred is represented. MHC Binding class captures data relating to in vitro MHC binding assays, which assess the epitope's binding capacity to the MHC molecule. Hence the MHC Binding class does requires neither the Antigen class not the Antigen Presenting Cells class. In general, naturally processed ligands are assessed in the absence of a T cell response, for example, identified by direct elution from MHC molecules extracted from infected cells or antigen presenting cells. Thus, the Immunization class is not used as a value restriction by the Naturally Processed Ligand class. The T Cell Response class captures all of the T cell mediated immunity-related information ( Figure 6A ). It has properties that are of type: Immunization, Effector Cells, Antigen Presenting Cell, Antigen, Assay Information, and Epitope-MHC-TCR Complex. The Effector Cell class describes the cells that are elicited upon immunization and that acquire measurable functions as a result. The B Cell Response class describes antibody responses that are related to the epitope ( Figure 6B ). This class has properties that are of type: Immunization, Antibody Molecule, Antigen, Assay Information, and Antigen-Antibody Complex. Because B cell responses do not require MHC binding and There are three classes that capture information about the 3D structure of complexes: Epitope-MHC Complex, Epitope-MHC-TCR Complex, and Antigen-Antibody Complex. The Epitope-MHC Complex, Epitope-MHC-TCR Complex, and Antigen-Antibody Complex classes are used as restrictions on properties of the MHC Binding, T Cell Response, and B Cell Response classes respectively ( Figures 5A, 6A, and 6B) . These Complex classes capture the Protein Data Bank (PDB) Identifier, which provides detailed information about 3D structures. The Protein Data Bank [22, 23] contains approximately 1600 3D structures that are of immunological interest. Other information that is not available in PDB, such as the atom pairs that are involved in the interactions between molecules, the specific residues, the contact area of the molecules, and allosteric effect, is also captured here. Each class has numerous slots that capture detailed information associated with epitopes. As mentioned above, a complete list of all the classes, their properties, and relationship, can be found at http://www.immuneep itope.org/ontology/index.html. One of the files provided as supplementary material contains two examples of how two literature references [24, 25] containing epitope information are extracted into the IEDB ontology (additional file 1). Along with the class hierarchy, the IEDB's data dictionary (additional file 2) provides more detailed information about the fields that are defined for the IEDB. The data dictionary contains a textual overview description and a listing of fields that are required to be completed for IEDB entries. The data dictionary also allows database users to provide comments and suggestions to IEDB team to enhance the formal ontology. The IEDB will be a comprehensive resource pertaining to epitopes of the immune system. By extensively curating both intrinsic and extrinsic features associated with epitopes, the IEDB is expected to provide substantially greater detail about specific epitopes than any other databases presently available. The IEDB will be populated with data derived from three main sources, namely the peer-reviewed literature, patent applications, and direct submission. Epitope data published in the literature and patent applications are curated manually by the IEDB's curation team. Data from already existing epitope databases, whose authors have agreed to share their data, will also be imported into the IEDB. Apart from these, a main data source will be the direct submission of data from the Large-scale Antibody and T Cell Epitope Discovery programs [3] that are funded by NIAID. Presently, fourteen contracts have been awarded under this program, and all of them will submit their data to the IEDB. Direct antibody and T cell epitope submissions will also be sought from the broader research community, with an emphasis on antibody epitopes to NIAID Category A-C pathogens. Because of the large scale of the IEDB project, a formal ontology is critical to ensure consistency in the representation of data. Communication between database developers, researchers, analysis tool developers, and team members is crucial, and can be performed in harmony only when a common vocabulary is established. An ontology, which is an explicit formal specification of the terms in the domain and relationships among them, is an effective way to share the knowledge contained in that domain. Accordingly, since the IEDB's domain is epitope-related data, we have created an ontology that captures detailed conceptual structure related to these data. The development of this ontology has relevance for the expansion and modification of the epitope knowledge base. Our ontology design defines individual concepts as separate classes and then defined relationships between these classes and other objects in the domain. These classes serve to restrict the values that will describe properties of objects in the database. For example, the species is a separate concept defined in its own class. Depending on the context, this can refer to an immunized species or the species from which antigen presenting cells are derived. Similarly MHC Molecules is defined as a separate class, and it is used as a value restriction by concepts like MHC Binding and Antigen Presenting Cells. Defining concepts as separate classes and using them to restrict the values of properties in other classes facilitates the expansion and modification of our ontology. Adding properties (slots) to concepts is a task easily accomplished when there are well-defined class descriptions that may serve as value restrictions on the properties, and providing that High-level classification of Immunization (A), Antigen (B), and Antigen Presenting Cells (C) class these class descriptions are general enough to apply in all instances. We have ensured in our design that each concept is atomic and that it can be re-used by various classes. The development of a formal ontology is valuable to database users and in particular to scientists contributing data to, and downloading data from, the IEDB. We anticipate that the availability of a formal ontology will ensure that a common language and shared understanding of concepts will inspire this type of communication, thus ensuring maximum efficiency and accuracy. The formal ontology developed will most likely require refinement and fine tuning when users provide suggestions and new technologies for performing experiments are discovered. The IEDB website will provide mechanisms for the users to provide suggestions and participate in the enhance- ment of the ontology. The IEDB Data Dictionary has a separate column for the users to provide comments on specific data fields. The IEDB website will also host web forms that will guide users to conform to the ontology definitions when submitting data. Apart from the web forms, an XML schema definition (XSD) will be available on the website for users to inspect and use in their data submission. Users will also be able to download epitope records from the website. In the process of developing new ontologies, it is good practice to leverage existing community standards. In our initial analysis, we confirmed that there were no explicit ontologies that efficiently captured epitope details as per the scope of the IEDB program. As mentioned above, we have utilized, as much as possible, inferred ontologies from existing epitope databases. Among the ontologies that we analyzed, IMGT-Ontology and Gene Ontology were the only two formal ontologies that were related to the epitope domain. The IMGT-Ontology was designed for the ImMunoGeneTics database. IMGT is an integrated database specializing in antigen receptors (immunoglobulins and T-cell receptors) and the major histocompatibility complex of all vertebrate species. The ontology developed for this database has specific immunological content, describing the classification and specification of terms needed for immunogenetics. The IEDB does conform to IMGT's standards about receptors and MHC molecule chains in the sense that all the chain names follow IMGT's controlled vocabulary. GO provides structured controlled vocabularies for genes, gene products, and sequences annotated for many organisms. The IEDB complements GO in terms of epitopes of immunological interest since GO is incomplete in this area. Antigens, which are primary sources of epitopes, are annotated in GO. Thus, in essence, the IEDB could be utilized to provide an extension of GO for antigens that contain epitope-related information. Perhaps the most important element in the development of the IEDB ontology is that, to the best of our knowledge, this represents the first immunological ontology specifically designed to capture both intrinsic biochemical and extrinsic context dependent information. In this respect, it is similar in spirit, but different in approach, from other knowledge resources relating to systems biology. We anticipate that the development of this type of ontology and associated databases might lead to completely new methods for describing and modeling immune responses. Accordingly, this new program might represent a novel tool to assist in the design, testing, and development of new ways to combat infectious diseases and other immune related pathologies such as cancer and autoimmune diseases. A complete listing of IEDB's class hierarchy and its properties is available at http://www.immuneepitope.org/ ontology/index.html
39
Evaluation of potential reference genes in real-time RT-PCR studies of Atlantic salmon
BACKGROUND: Salmonid fishes are among the most widely studied model fish species but reports on systematic evaluation of reference genes in qRT-PCR studies is lacking. RESULTS: The stability of six potential reference genes was examined in eight tissues of Atlantic salmon (Salmo salar), to determine the most suitable genes to be used in quantitative real-time RT-PCR analyses. The relative transcription levels of genes encoding 18S rRNA, S20 ribosomal protein, β-actin, glyceraldehyde-3P-dehydrogenase (GAPDH), and two paralog genes encoding elongation factor 1A (EF1A(A )and EF1A(B)) were quantified in gills, liver, head kidney, spleen, thymus, brain, muscle, and posterior intestine in six untreated adult fish, in addition to a group of individuals that went through smoltification. Based on calculations performed with the geNorm VBA applet, which determines the most stable genes from a set of tested genes in a given cDNA sample, the ranking of the examined genes in adult Atlantic salmon was EF1A(B)>EF1A(A)>β-actin>18S rRNA>S20>GAPDH. When the same calculations were done on a total of 24 individuals from four stages in the smoltification process (presmolt, smolt, smoltified seawater and desmoltified freshwater), the gene ranking was EF1A(B)>EF1A(A)>S20>β-actin>18S rRNA>GAPDH. CONCLUSION: Overall, this work suggests that the EF1A(A )and EF1A(B )genes can be useful as reference genes in qRT-PCR examination of gene expression in the Atlantic salmon.
In real-time RT-PCR, the expression levels of the target genes of interest are estimated on the basis of endogenous controls. Various housekeeping genes, ribosomal RNA (rRNA) and total RNA are most commonly used as references in gene expression analysis today. The purpose of these controls is to remove or reduce differences due to sampling, i.e. differences in RNA quantity and quality. The ideal endogenous control should be expressed at a constant level among different tissues of an organism, at all stages of development and should be unaffected by the experimental treatment. It should also be expressed at roughly the same level as the RNA under study [1] . However, data normalization in real-time RT-PCR remains a real problem, especially for absolute quantification [1] . Numerous studies have revealed that no single universal gene has a constant expression level under all developmental or experimental situations. The best choice of reference gene to use as an endogenous control varies, depending on the tissues of interest in the experiment. A large number of genes have for this reason been selected for normalization of mRNA expression data [2, 3] . If the selected reference gene fluctuates randomly between samples, small differences in expression between the genes of interest will be missed. Gene expression coefficient of variation (CV) between different groups of individuals should ideally be as low as possible [4] . In general, the stability of several potential reference genes should be tested in every examined tissue or cell, and under different experimental design [5, 6] . An increasing number of papers are discussing the selection of reference genes in real-time RT-PCR analyses [3, 7] . Two of the most commonly used reference genes are those encoding glyceraldehyde-3P-dehydrogenase (GAPDH) and β-actin. Recently, the use of these two genes as endogenous controls has been scrutinized, and several studies have documented that the GAPDH and β-actin genes should be used with caution as controls [2, 8, 9] . GAPDH in mammals is known to play a role in a broad range of cellular mechanisms (for review see Sirover [10] ), including being a key enzyme in glycolysis. Overall, GAPDH mRNA levels might be regulated under a large number of physiological states, and its use as a reference is inappropriate for most experimental conditions. Actin is a major component of the protein scaffold that supports the cell and determines its shape, and is the most abundant intracellular protein in eukaryotic cells. Even though commonly used as a reference, the application of the β-actin gene has recently been characterized as a historical carryover from northern blots and conventional RT-PCR (for a general discussion on the use of 'classic' reference genes like GAPDH and β-actin, see Huggett et al. [7] ). Eukaryotic elongation factor 1A (eEF1A, formerly elongation factor 1 alpha) plays an important role in translation by catalyzing GTP-dependent binding of aminoacyl-tRNA to the acceptor site of the ribosome. However, the protein is involved in a broad diversity of functions and constitutes 1-3% of the total cytoplasmic protein content of the cell. In human, cDNAs of two actively transcribed isoforms have been cloned (eEF1A-1 and eEF1A-2) (for review see Thornton et al. [11] ). Two paralog EF1A genes (A and B) have recently been applied as references in real-time qRT-PCR of Atlantic salmon [12] . It is plausible to assume that the presence of these highly similar genes is a result of a tetraploidization event that occurred in a salmonid ancestor in the comparatively recent past [13, 14] . Previously, the 18S rRNA gene was considered to be an ideal internal control in qRT-PCR analysis (Ambion [15] ). Ribosomal RNA constitutes up to 80-90% of total cellular RNA, and several studies have shown that rRNA varies less under conditions that affect the expression of mRNAs (discussed in Bustin & Nolan [16] ). However, questions have been raised against the use of ribosomal RNA genes as references. Vandesompele et al. [5] have stressed the fact that there sometimes might be imbalances in rRNA and mRNA fractions between different samples, making genes encoding ribosomal RNAs unsuitable as references. To meet these challenges of accurate interpretation of realtime qRT-PCR data, the authors suggested that an index of the most stable housekeeping genes should be used for normalization, and developed the geNorm VBA applet for Microsoft Excel in this regard [5] . A similar software tool, the BestKeeper, has been developed by Pfaffl et al. [6] . These tools can be used to find the most stable reference genes under different experimental conditions. We used the geNorm software which determines the individual stability of a gene within a pool of genes [5] . The stability is calculated according to the similarity of their expression profile by a pair-wise comparison, using their geometric mean as a normalizing factor. The gene with the highest M, i.e. the least stable gene, is then suggested excluded in a stepwise fashion until the most stable genes are determined, and an index suggested, based on the best genes. geNorm has been used to select the most stable reference genes in several recent studies (e.g. [4, 17, 18] ). The aim of this work was to evaluate the usefulness of six potential reference genes in the Atlantic salmon. Salmonid fish are among the most widely studied model fish species in general, and extensive basic information on many different aspects of their biology has been collected [19] . Large-scale DNA-sequencing projects on salmon have been initiated in several laboratories http://www.sal mongenome.no/cgi-bin/sgp.cgi; http://web.uvic.ca/cbr/ grasp/; http://www.abdn.ac.uk/sfirc/salmon/; http:// www.bcgsc.ca/gc/salmon. In this work we selected the two 'classic' reference genes encoding GAPDH and β-actin, two genes encoding 18S rRNA and S20 ribosomal protein and two paralog genes encoding elongation factor 1A (EF1A A and EF1A B ). To evaluate their usefulness as reference genes, RNA from eight tissues of six adult salmon were subjected to real time PCR. The relative transcription levels of the genes were also estimated in four phases of young salmon going through smoltification, in order to check their stabilities under physiological stressful conditions. The ranking of the six examined genes analyzed by geNorm is shown in Table 3 . In six tissues (muscle, liver, gills, head kidney, spleen and thymus), the EF1A B gene emerged as the most stable, whereas the EF1A A gene was ranked number one in brain and the β-actin gene was ranked number one in intestine. The 18S rRNA and S20 genes were ranked among the worse genes in all tissues. Not surprisingly, the GAPDH gene was ranked worse in five tissues (liver, head kidney, spleen, brain and thymus), confirming the general skepticism against the use of this gene as reference [7, 16, 20] . Combined, the total ranking reads EF1A B >EF1A A >β-actin>18S rRNA>S20>GAPDH. We did not analyze our data with the Bestkeeper software. Analyzing reference genes in virus infected cells, Radonic et al. [4] concluded that the Bestkeeper tool gave results that slightly deviated from, but nevertheless corresponded to, those obtained using geNorm. To be able to evaluate gene stability under stressful conditions, mRNA expression of the selected genes was examined in gills of salmon going through smoltification. Prior to seawater entry, juvenile anadromous salmon undergo a parr-smolt transformation, characterized by behavioral, morphological and physiological changes, known to be challenging for the fish. Physiological alterations include increased seawater tolerance, olfactory sensitivity, metabolic rate, scope for growth and changed hemoglobin and visual pigments [21] . We selected to examine the gills during smoltification, because this tissue plays a major role in ionic and osmotic regulation during adaptation to hyperosmotic seawater. Figure 1 shows the raw Ct values of the studied genes in gills before, during and after smoltification (smoltified in seawater and desmoltified in freshwater). In Figure 2 the same data are presented, but now normalized against an index calculated by geNorm of the three most stable genes (β-actin, EF1A A and EF1A B ). Based on the M values, geNorm ranks the stability of the six genes from 24 fish going through smoltification in the following order: EF1A B >EF1A A >β-actin>S20>18S rRNA>GAPDH ( Figure 3 ). In Figure 1 it can be seen that the 18S gene had the lowest individual raw Ct variation. Most individual raw Ct variation of the studied genes is seen in the presmolt and smolt groups. A characteristic drop in expression can be seen for all genes in the smolt group, compared to the presmolt group. After transfer to seawater, the individual raw Ct variation decreased for all genes. Overall, the raw Ct data suggest that the physiological challenging smoltification process affected the expression of all six genes. When the same data were normalized against an index of the three most stable genes, β-actin, EF1A A and EF1A B , the relative expression levels were altered for all genes. Now the ribosomal 18S gene emerges as the second worse, whereas the two paralog EF1A genes became the most stable. This might have to do with the fact that geNorm will top-rank co-expressed genes [22] , a weakness that has to be considered when evaluating paralog genes likely to be co-regulated. Even though the eEF1A-2 gene has been identified as an important oncogene and has been shown to be differently expressed in human tissues [11] , Hamalainen et al. [23] found the eEF1A-1 gene to be a good reference gene in real-time RT-PCR examinations. A similar finding was reported by Frost and Nilsen [24] in salmon louse, where they showed that the eEF1A and S20 genes were valid candidate references, whereas the 18S rRNA and GAPDH genes were unsuitable. The current findings based on geNorm evaluation question the recommended application of ribosomal genes as references (as suggested for example by Ambion (see reference [15] ), and are in line with earlier warnings against the use of rRNA genes as references [5, 6] . To avoid the normalization of the genes for β-actin, EF1A A and EF1A B against an index partly based on their own expression, the S20 gene was included in the index instead, and the mean normalized expression for these three genes calculated with the new index. The patterns of expression, however, were approximately the same for the three genes as seen in Figure 2 , suggesting that the gene-stability measure M can be used to find the most appropriate reference genes. We see a correlation between the A260/230 absorbance on the NanoDrop and the PCR efficiency (data not shown). We tend to get PCR efficiencies that are too high in some samples with low A260/230 ratios. When the samples are treated with DNase solution, the A260/230 ratio usually drops. After DNase treatment, the A260/280 ratio increased from 1.8 to 2.1 (n = 45 samples). At the Table 3 : Evaluation of the usefulness of six potential reference genes in eight tissues of Atlantic salmon ranked by the geNorm software. 1 = best, 6 = worst. Six individuals were analyzed for six genes in eight tissues. qRT-PCR analysis of six genes in gills of six Atlantic salmon going through smoltification; presmolt (before smoltification), smolt (during smoltification), smoltified (finished smoltified in seawater) and desmolt (desmoltification in freshwater) Figure 1 qRT-PCR analysis of six genes in gills of six Atlantic salmon going through smoltification; presmolt (before smoltification), smolt (during smoltification), smoltified (finished smoltified in seawater) and desmolt (desmoltification in freshwater). Numbers indicate raw Ct values. qRT-PCR analysis of six genes in gills of six smoltifying Atlantic salmon Figure 2 qRT-PCR analysis of six genes in gills of six smoltifying Atlantic salmon. The same data as in Figure 1 , but now normalized against an index of the three best genes (β-actin, EF1A A and EF1A B ) calculated with the geNorm software. The four groups were analyzed with Kruskal-Wallis test, and if significant, the overall p-value is given in the graphs. For β-actin, there were significant differences between the presmolt and the smoltified group (p < 0.05), the presmolt and the desmolt groups (n<0.01) and between the smolt and desmolt groups (p < 0.05). For EF1A A , there was a significant difference between the presmolt and the desmolt groups (p < 0.01). For EF1A B , there was a significant difference between the smolt and the smoltified groups (p < 0.05). An asterisk denotes significant differences between the groups. same time, the A260/230 ratio dropped from 2.4 to 2.1. The DNase treatment therefore adds substances to the RNA solution that increases the absorbance at 230 nm more than it decrease the 260 nm absorbance. The added substance (salt or some other component) may inhibit the RT reaction or the PCR reaction, sometimes called PCR poisoning. We have seen that the A260/230 ratios are quite low in samples that give inadequate PCR efficiency slopes, especially with RNA from head kidney, thymus and intestine tissues, in which the gradient of the standard curve is less than -3.3 ( Table 2 ). The reason one obtain better amplification rate efficiencies with the more diluted samples is because the inhibitor has been diluted below its effective level. The obvious way around this problem is to dilute the amount of cDNA put into the PCR reaction. Alternatively, cleanup columns can be used to purify and concentrate the RNA. Transcription levels of the examined genes and the coefficient of variance (CV) in different tissues varied considerably. mRNA levels in tissues are regulated by numerous endogenous and exogenous stimuli [16] . Transcription rates in metabolic active tissues might be up-regulated compared to those of less active tissues, whereas inter-tissue variation in degradation rates of mRNAs, for example, might affect mRNA stability [25] . The results revealed that muscle had the lowest CVs of the studied genes, compared to higher CVs in more active tissues like thymus, head kidney and spleen. In thymus, intestine, head kidney, gills, brain, liver and spleen, the 18S and S20 genes had the lowest CVs, based on raw Ct values. In all tissues, except intestine, the GAPDH gene had the highest CV. Except for thymus, the two elongation factor genes had relatively similar expression in all eight examined tissues. Their expression are most likely co-expressed in the examined tissues, and therefore favored in geNorm calculations [22] . The results also demonstrated that assays optimized for one tissue of an organism do not necessarily work equally well in other tissues. Of the tissues studied in this work, intestine, head kidney and spleen were the most troublesome. Our data, based on geNorm calculations, suggest that the Atlantic salmon EF1A genes that have been tested in the present study may be good candidate reference genes. The GAPDH gene seems unsuitable as a reference in quantitative real-time RT-PCR. With regard to the 18S rRNA gene, this must be applied with caution. Tools like the geNorm applet for Microsoft Excel can be useful to help select the most stable genes in various experiments. Tissues from 15 individuals were collected (852 ± 702 g, ranging from 254 to 1898 g). These individuals were not separated based on sex, size or sampling time, but treated as one heterogeneous group to examine the width of mRNA expression of the studied genes in eight different organs. This group of fish was handled and fed according to normal aquacultural management, and none of these individuals were exposed to any particular treatment. To examine if physiological stress may alter the gene expression in the gills, a total of 24 individuals were collected before (termed presmolt, 18.3 ± 0.9 g), during (termed smolt, 28.7 ± 3.7 g) and after smoltification. After smoltification, one group was kept and desmoltified in freshwater (termed desmolt FW, 30.0 ± 3.8 g), while the other group was transferred to seawater (termed smoltified SW, 30.2 ± 4.3 g) (n = 6 in each group). The Atlantic salmon examined during smoltification were from the anadromous population "Vosso" of the river Vosso in Southwestern Norway (see Nilsen et al. [27] for details on how these fish were treated). All fish were treated and euthanized according to Norwegian national legislation for laboratory animals. Samples from eight organs, i.e. gills, liver, brain, head kidney, spleen, thymus, white muscle and posterior intestine, were dissected out and immediately frozen in cryo tubes in liquid N and stored at -80°C before RNA extraction. The RNA extracted from three spleen and four head kidney tissue samples were of low quality, and we had to redo the sampling from four individuals, These tissue samples Stability of six genes in gills of Atlantic salmon during smoltifi-cation calculated with the geNorm software were stored on RNA later (Ambion) at -20°C before further processing. RNA was isolated with phenol-chloroform extraction as described by Chomczynski and Sacchi [28] , and stored in 100 µl RNase-free MilliQ H 2 O. Total RNA was extracted using Trizol reagent (Invitrogen, Life Technologies), according to the manufacturer's instructions. Genomic DNA was eliminated from the samples by DNase treatment according to the manufacturer's description (Ambion). The RNA was then stored at -80°C before further processing. The quality of the RNA was assessed with the NanoDrop ® ND-1000 UV-Vis Spectrophotometer (NanoDrop Technologies, Wilmington, DE, USA) and the Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto, CA, USA). A 260/280 nm absorbance ratio of 1.8 -2.0 indicates a pure RNA sample. The RNA 6000 Nano LabChip ® kit (Agilent Technologies, Palo Alto, CA, USA) was used to evaluate the integrity of the RNA. We used the RNeasy MinElute Cleanup kit from Qiagen to purify our most troublesome samples. With this kit the A260/230 ratio increased on average by 5 % (n = 10). The PCR primer and TaqMan MGB probe sequences used for quantification of the genes encoding 18S rRNA, S20 ribosomal protein, β-actin, GAPDH, EF1A A and EF1A B , are shown in Table 1 . Four of these genes, 18S, β-actin, EF1A A and EF1A B , have also been used as references in real-time RT-PCR analyses of Atlantic salmon in other recent studies [12, 28] . The primers amplify PCR products between 57-98 basepairs (bp) long, which is within the range of 50-150 bp as suggested by Applied Biosystems for their Taq-Man assays. qPCR assays were designed using Primer Express 2.0 software (Applied Biosystems, Foster City, CA, USA) to select appropriate primer and probe sequences from known Atlantic salmon genes. The mRNA sequences encoding S20 ribosomal protein and GAPDH were obtained from GenBank accession numbers BG936672 and BU693999, respectively (exon-exon borders were not considered). The EF1A A assay was based on a cDNA clone that we reported to the GenBank previously (AF321836), whereas the EF1A B assay was based on the EST BG933853. An alignment with zebrafish indicated the exon-exon borders [29] . The chosen primers were subsequently used to confirm that the salmon genes contained an intron between the same sites as deduced from the alignment with zebrafish. The PCR products containing the introns were cloned into TOPO vector (Invitrogen) and sequenced (sequences can be provided upon request). PCR primers for β-actin were based on Atlantic salmon BG933897 and designed to span exon-exon borders of this gene, as deduced from corresponding genes in human and zebrafish (NW633959). For 18S rRNA the PCR primers and probe were designed from the Atlantic salmon sequence AJ427629, and placed in a conserved region of the gene based on comparison with the human gene. RNA samples were subjected to DNase treatment to avoid genomic DNA contamination. Amplified PCR products of all actual cDNAs were sequenced to ensure that the correct mRNA sequences were quantified. The fragments were sequenced with BigDye version 3.1 fluorescent chemistry (Applied Biosystems) and run on an ABI PRISM ® 377 DNA apparatus at the University of Bergen Sequencing Facility. A two-step real-time RT-PCR protocol was developed to measure the mRNA levels of the studied genes in eight tissues of Atlantic salmon. For evaluation of the potential reference genes, raw Ct values are presented. The geNorm VBA applet for Microsoft Excel was used to determine the most stable genes from the set of tested genes [5] . The Ct values were transformed to quantities using standard curves, according to the geNorm manual. The gene expression stability (M) was calculated with the geNorm applet, and the genes were ranked from best to worst, based on the M value. The GraphPad Prism 4.0 software (GraphPad Software, Inc.) was used for the statistical analyses in this work. Linear regression was used to determine PCR efficiency based on dilution curves. Non-parametric Kruskal-Wallis test was used to compare differences among four groups of salmon going through smoltification. PAO was responsible for the experiment, data analysis and drafted the manuscript. KKL conducted the real-time RT-PCR analysis, and contributed throughout the experimental process. AEOJ constructed the qPCR assays for two of the genes. TON provided the cDNA from the smoltification experiment. IH participated as a supervisor in the study design, analyses and writing.
40
Relevance of human metapneumovirus in exacerbations of COPD
BACKGROUND AND METHODS: Human metapneumovirus (hMPV) is a recently discovered respiratory virus associated with bronchiolitis, pneumonia, croup and exacerbations of asthma. Since respiratory viruses are frequently detected in patients with acute exacerbations of COPD (AE-COPD) it was our aim to investigate the frequency of hMPV detection in a prospective cohort of hospitalized patients with AE-COPD compared to patients with stable COPD and to smokers without by means of quantitative real-time RT-PCR. RESULTS: We analysed nasal lavage and induced sputum of 130 patients with AE-COPD, 65 patients with stable COPD and 34 smokers without COPD. HMPV was detected in 3/130 (2.3%) AE-COPD patients with a mean of 6.5 × 10(5 )viral copies/ml in nasal lavage and 1.88 × 10(5 )viral copies/ml in induced sputum. It was not found in patients with stable COPD or smokers without COPD. CONCLUSION: HMPV is only found in a very small number of patients with AE-COPD. However it should be considered as a further possible viral trigger of AE-COPD because asymptomatic carriage is unlikely.
Respiratory viruses play an important role in exacerbations of COPD and this has been increasingly recognised since the application of molecular detection methods [1, 2] . The most prevalent viruses detected by polymerase chain reaction so far were respiratory syncytial virus (RSV), Influenza A, Rhinovirus and Parainfluenza 3. Human metapneumovirus (hMPV) is a recently discovered respiratory virus first isolated from a dutch child with lower respiratory tract infection (LRTI) [3] . World wide distribution is probable since it has been isolated in North HMPV has been recognized as a member of the Paramyxoviridae like RSV and it is not only associated with bronchiolitis in most cases, but also with pneumonia, croup and exacerbations of asthma [14, 15] , diseases which share some features with COPD. Up to date reports about hMPV in adults are scarce. In a general Canadian population 14.8% of patients of all age groups with acute respiratory tract infections were hMPV positive. Thirty-three percent of hMPV-infected patients were hospitalized and the hospitalization rates were significantly higher among patients below 5 years and those over 50 years of age [16] . In another prospective cohort of adults, hMPV was detected in 4.5% of all illnesses but also in 4.1% of asymptomatic subjects. HMPV was most prevalent in young adults with children and in frail elderly [17] . HMPV infection can be severe since the virus was isolated from the lungs from a previously healthy man who died from acute pneumonia [18] . The role of hMPV in acute exacerbations of COPD (AE-COPD) has been studied recently in outpatients and only low frequencies have been observed [17, 19] . Up to now the prevalence of hMPV in patients hospitalized with AE-COPD is unknown. Our aim was therefore to investigate the frequency of detection of hMPV in a prospective cohort of hospitalized patients with AE-COPD and to compare these results to patients with stable COPD and to smokers without COPD. Three different groups were studied. The first group consisted of hospitalized patients with an acute exacerbation of COPD (AE-COPD), the second group were subjects with stable COPD and the third group was composed of smokers without COPD. The groups were defined as previously published [20] . Briefly AE-COPD patients suffered from COPD as defined by GOLD [21] . Acute exacerbation was characterized by worsening in dyspnea, cough, and expectoration. A routine posterior-anterior chest radiograph was evaluated on admission by expert radiologists to exclude other other reasons for increased symptoms as pneumonia, tuberculosis, pulmonary fibrosis, bronchiectasis, bronchial carcinoma or congestive heart failure.Stable COPD patients did not have an exacerbation within the last 30 days prior to hospital admission and had no changes in therapy within the last 14 days (including inhaled and oral medication) and had been admitted for other medical reasons into departments of internal medicine other than pulmonary care. COPD subjects were recruited in a 2:1 ratio each month in order to prevent seasonal selection bias. Smokers have been smoking more than 10 pack-years, could have chronic symptoms like cough and phlegm but did not report dyspnea and did not have bronchial obstruction (FEV 1 /FVC>70%, FEV 1 >80% predicted). None of the smokers had a history of COPD or asthma, nor was using systemic or topic pulmonary medication. The smokers were recruited either from our smoking cessation initiative or by newspaper advertisement. The study was approved by the ethical committee of the Ruhr-University of Bochum, Germany. Written informed consent was obtained from all patients and control subjects before inclusion in the study. Clinical evaluation, spirometric tests, nasal lavage, induced sputum, specimen processing and viral ribonucleic acid (RNA) extraction were carried out as described by Rohde et al [2] . Elution volume was 100 µl. cDNA was generated with random-hexamer primers as previously published[2]. A hMPV-specific real-time RT-PCR designed and evaluated by Maertzdorf et al was used [22] . Primers and probe are localized within the nucleoprotein gene (NL-N) and the presence of a degenerate base within the probe allows detection of all four genetic lineages of hMPV. The assays were performed using the TaqMan ® PCR Core Kit. The final volume was 25 µl containing 500 nM of the forward primer (NL-N-forward (5'-CATATAAGCAT-GCTATATTAAAAGAGTCTC-3')), 250 nM of the reverse primer (NL-N-reverse (5'-CCTATTTCTGCAGCATATTTG-TAATCAG-3')) and 500 nM of the probe (NL-N-probe (5'-FAM-TGYAATGATGAGGGTGTCACTGCGGTTG-TAMRA-3', in which Y is either a C or a T residue). Nuclease-free water was used as negative control and a plasmid containing the N gene of hMPV (kindly provided by James Simon, VIRONOVATIVE, EUR Holding, Erasmus University Rotterdam) was used as a positive control in all PCR runs. Cycling parameters were as follows: 5 min at 95°C, 45 cycles of 30 s at 95°C and 1 min at 60°C. Amplification and detection of RNA from virus isolates or clinical specimens were performed using the GeneAmp ® 5700 Sequence Detection System (Applied Biosystems). The real-time PCR product was cloned with the QIAGEN ® PCR cloning kit (QIAGEN, Hilden, Germany) and this standard plasmid DNA was used for absolute quantification of hMPV viral load. Calculations were performed as previously described for absolute quantification of RSV viral load [23] . The primary objective of this study was to compare the frequency of hMPV detection in respiratory specimens between COPD patients with or without an acute exacerbation and smokers without COPD. Continuous data were checked for normal distribution using the Kolmogorov-Smirnov test. The data were of non-parametrical distribution and results were expressed as median and range. Differences between groups were assessed by Kruskal-Wallis test. To further analyse significant differences between two individual groups a pair wise comparison by two-sided Mann-Whitney U-test was performed. All significance levels were set to 5%. Data were analysed and processed using SPSS Version 12.0 on a Windows XP operating system. A total of 229 subjects were investigated between October 1999 and June 2004: 130 patients with AE-COPD, 65 patients with stable COPD and 34 smokers without COPD. The clinical characteristics and lung function measurements are summarized in table 1. FEV 1 , FEV 1 in % of predicted value and FEV 1 /FVC were normal in smokers, significantly decreased in stable COPD patients (all) and further significantly decreased in AE-COPD patients (all p < 0.05 compared to stable COPD and all p < 0.001 compared to smokers). HMPV could be detected in three subjects. All these subjects were AE-COPD patients. The prevalence of hMPV in AE-COPD patients was 2.3%. The virus was simultaneously detected in nasal lavage and induced sputum in one patient only. The viral load was about 100 times higher in nasal lavage than in induced sputum in this patient. Overall the viral load in nasal lavage was about 3.5 times higher compared to induced sputum (for details see table 2). The hMPV positive patients did not differ significantly from other AE-COPD patients when clinical parameters and lung function were analysed. hMPV was detected in the winter season only. The main finding of this controlled study investigating the incidence of hMPV in subjects with COPD and smokers without COPD is that this recently discovered respiratory virus was detectable only during exacerbation of COPD. The frequency of detection was very low but in positive cases the viral load was considerable. There was no detection in patients with stable COPD or smokers without COPD. Recently Vicente et al [19] reported about the incidence of hMPV in 89 COPD patients. Five patients (5.5%) were hMPV positive. Two of these patients had to be transferred to hospital. Although this was not a controlled study and not all details of the study are available due to the fact that the data were published in form of a letter, these results support our findings. The incidence of hMPV in this and in our study is low compared to other respiratory viruses. In a similar previous study we found that Picornaviruses were detectable in 36% of AE-COPD patients, Influenza A in 25% and Respiratory syncytial virus in 22%[2]. There is another prospective cohort study of adults in which hMPV was detected in 4.5% of all illnesses. HMPV was most prevalent in young adults with children and in frail elderly from long term care facilities [17] . Unfortunately this report does not specify how many of the elderly patients suffered from COPD. In our asymptomatic smokers without COPD hMPV could not be detected. A recent study investigating nasal secretions from adults with and without respiratory illnesses found hMPV in 5 of 146 ill patient and in none of 158 control subjects, strongly supporting our data [24] . A further recent study found hMPV in two out of 111 adult patients (1.8%) who presented to the emergency department for AE-COPD during 2 winter/ spring seasons in Quebec, Canada, also in support of our findings [25] . In a US American study investigating clinical samples collected between 1991 and 1995, hMPV could not be detected at all in 196 patients indicating important geographical and seasonal differences in hMPV prevalence [26] . Taken together the results presented here are in keeping with other studies in adults and add important information on the prevalence of hMPV in hospitalized AE-COPD. To our knowledge this is the first study analysing the viral load of hMPV in COPD patients. We found a mean of 6.5 × 10 5 viral copies/ml in nasal lavage and 1.88 × 10 5 viral copies/ml in induced sputum. These values indicate that hMPV may have been the infectious agent triggering exacerbation in these patients. Viral load cut-off values for infectivity in COPD exacerbations have not been studied in detail yet and need further investigation. However, viral loads between 1120 copies/ml in Cytomegalovirus infection in lung-transplant patients [27] and 5.8log 10 copies/ ml in SARS [28] have been considered to indicate infectious disease. Moreover hMPV was only found in acute exacerbation and not in stable disease or in smokers without COPD supporting a triggering role in AE-COPD. HMPV infection can be severe since it was isolated from the lungs from a previously healthy man who died from acute pneumonia [18] . Our hMPV positive patients did not differ in their clinical characteristics or lung function from the other AE-COPD patients which does not indicate a more severe course of AE-COPD in these patients. Taken together this is the first controlled study on the relevance of hMPV in hospitalized AE-COPD. HMPV was detected in a very low frequency but with noticeable viral load in AE-COPD patients. Given that asymptomatic carriage of hMPV is very unlikely it should be considered as another possible trigger of AE-COPD. Since every AE-COPD has considerable impact on the course of the disease and regional outbreaks of hMPV are possible it should be included into future diagnostic and therapeutic considerations. The author(s) declare that they have no competing interests.
41
Bioethical Implications of Globalization: An International Consortium Project of the European Commission
The BIG project looks at some of the ethical concerns surrounding globalization and health.
T he term "globalization" was popularized by Marshall McLuhan in War and Peace in the Global Village . In the book, McLuhan described how the global media shaped current events surrounding the Vietnam War [1] and also predicted how modern information and communication technologies would accelerate world progress through trade and knowledge development. Globalization now refers to a broad range of issues regarding the movement of goods and services through trade liberalization, and the movement of people through migration. Much has also been written on the global effects of environmental degradation, population growth, and economic disparities. In addition, the pace of scientifi c development has accelerated, with both negative and positive implications for global health. Concerns for national health transcend borders, with a need for shared human security and an enhanced role for international cooperation and development [2] . These issues have signifi cant bioethical implications, and thus a renewed academic focus on the ethical dimensions of public health is needed. Future developments in science and health policy also require a fi rm grounding in bioethical principles. These core principles include benefi cence; nonmalefi cence (to do no harm); respect for persons and human dignity (autonomy); and attention to equity and social justice. According to the World Health Organization [3] , global ethical approaches should (1) monitor and update ethical norms for research, as necessary; (2) anticipate ethical implications of advances in science and technology for health; (3) apply internationally accepted codes of ethics; (4) ensure that agreed standards guide future work on the human genome; and (5) ensure that quality in health systems and services is assessed and promoted. The Bioethical Implications of Globalization (BIG) Project is a 42-month dialogue funded by the European Commission that involves a series of expert panel discussions on four specifi c globalization and health subject areas: (1) mobility of people; (2) technological globalization; (3) liberalization of trade; and (4) new global health threats (bioterrorism). In addition, BIG includes a multipleround Delphi Process (Box 1) to solicit input on these issues from a broad, interdisciplinary audience. The project's purpose is both to raise short-term, practical considerations about globalization and health and Delphi is a group communication technique designed to obtain opinions from a panel of selected experts on specifi c issues through the sending of questionnaires to be completed within a specifi ed time. The experts are contacted individually and they do not know other group participants and their opinions-the aim is to submit the group participants to the same conditions. Participants do not meet personally, thereby avoiding undue infl uence. The process foresees the following points: The process is repeated a number of times, until a convergence of all group members is obtained. The process ends with analysis of the answers and formulation of the fi nal report. (Adapted from http://www.bigproject. org/dephi.htm) Thomas E. Novotny*, Emilio Mordini, Ruth Chadwick, J. Martin Pedersen, Fabrizio Fabbri, Reidar Lie, Natapong Thanachaiboot, to provide a longer-term, strategic perspective on the four selected public health-related issues. The fi nal conclusions will be presented to a high-level meeting of European Union (EU) policy makers in 2006; these conclusions may then inform future research directions and stimulate additional critical thinking about globalization and its bioethical implications for health policy. This article presents preliminary results from the BIG Project. Mobility results from the increasing ease of domestic and international travel as well as from instantaneous access to information through the Internet and other electronic resources. Mobility may involve the pursuit of a better quality of life, development of markets for traded goods and services, return of resources to home countries, and improvement of professional and business networks. However, migration may also affect psychological and physical health as a result of confl ict, famine, poverty, and the insuffi cient cultural or economic integration of migrants within their new home society. It may contribute to the spread of infectious diseases across borders ( Figure 1 ). The recent epidemic of SARS was a classic example of an infectious disease propagated through the movement of people across borders; it required attention from the original site to control migration (quarantine) as well as vigilance by secondary sites to protect their populations ( Figure 2 ) [4] . For these and other reasons, the International Organization for Migration is increasingly concerned with migratory patterns and their health consequences in a globalizing world (for an illustration of the emerging confl ict of ideas, see http:⁄⁄www.iom. int and http:⁄⁄www.noborder.org/ iom/index.php). Cross-border health commerce is related to mobility. In Europe, this commerce is likely to increase as the EU enlarges to include Eastern and Central European nations. Such commerce may include the movement of health providers from East to West as well as "medical tourism" in pursuit of less costly or more accessible high-quality health care. In addition, international trade in illegal health products and inconsistent regulatory and safety standards for exports may threaten public health, especially in unregulated pharmaceutical markets. Ethical concerns may also result from the vast growth in international tourist travel. Such travel now accounts for a twelfth of world trade, supporting an economy the size of a middle-income country [5] . Tourism may provide substantial economic benefi ts to many developing countries, and it may improve cultural understanding among travelers. However, these benefi ts require an ethical concern for the environment and for persons employed in the tourist industry. The rights of nations to protect against infectious disease and unsafe medical practices, as well as the rights of human beings displaced by war, traffi cking, and economic and cultural disruption, are critical concerns for health policy makers. Poverty and social disparities are key factors in the growth of global migration. Therefore, it is timely to consider whether mobility is a human right, and whether those who migrate have rights to health care in their new country. These questions should be considered by health policy makers within the ethical contexts of individual autonomy, social justice, nonmalefi cence, and benefi cence. Technology drives globalization and in turn is driven by globalization. However, there is considerable ambiguity as to the value of technological globalization, especially for health in low-income countries. The "digital divide" may be important in improving health or income disparities as the electronic revolution provides scientists and health workers in both the developed and developing world with unprecedented access to information. Much could be done to reduce information inequities for the developing world through collective international action, but new global governance mechanisms may be needed to achieve this goal for information technology [6] . Interestingly, the Internet is a structural necessity for fi nancial and corporate globalization, but the same technology is used by nongovernmental organizations, political groups, and cultural movements to support grassroots social justice and human rights campaigns against these globalizing corporations. Neither side in this struggle would advocate limitations to the expansion of Internet technology, but both sides need to consider the bioethical implications of increased information access. On the other hand, the ethical issues surrounding genomics (with both environmental and human concerns) are quite ambiguous. While there may be signifi cant benefi ts to identifying genetically benefi cial products or genetic determinants of disease, there are also concerns about altering natural environments and about collecting routine genetic information from general populations [7] . For example, some experts assert that genetically modifi ed (GM) crops will signifi cantly increase crop yields without requiring any additional farmland, thus preserving valuable rain forests and animal habitats. herbicides, and pesticides. Farmers are not allowed to trade or save GM seed from one harvest to the next, and "terminator technology" (producing grains that are genetically modifi ed so that they cannot be used to generate new crops) is under development. (See http:⁄⁄www.globalissues.org/ EnvIssues/GEFood/Terminator.asp for more information on this technology.) Thus, ethical considerations of distributive justice and benefi cence must be considered in the debate about the global applicability of GM crops. For the pharmaceutical and health care industries, genetic testing could provide information about the shape of future markets and the possible tailoring of specifi c pharmacotherapy to genomic susceptibility. For governments, genetic testing may provide predictive information on a population basis that could aid future health care planning. Genetic information might also be similarly used by the insurance industry, but the identifi cation of genetically "high-risk" individuals would likely interfere with their autonomy, in that they may not be able to purchase health insurance. For example, the Apolipoprotein E test may indicate that an individual has two copies of one form (allele e4) of the gene that leads to Alzheimer disease [9] . Could this information be used by insurance companies or possible employers to deny insurability, despite no current adverse health effects? In the post-genomic era there is potential to both reduce and increase health inequities, but much will depend on how ethical issues are addressed. If interventions to increase the life span for those with access to high quality health care must compete with expensive investments in genetic research on infectious diseases (which affect the poor most of all), health inequalities may be amplifi ed between those with access and those without access to health care. If research participants or patients in low-income countries have unequal access to information, they may not be properly informed about genetic testing and the counseling needed if adverse genetic information is found. Population-based genomic research may characterize groups of people in such a way that encourages discrimination. Such research may also lead to disputes about ownership of genetic resources from participant populations. Health professionals must have a solid grounding in bioethical issues as they make clinical decisions based on genetic information. However, health policy makers and global governance structures must also be accountable for the potential adverse consequences such decisions might engender. One may ask: will genomic science really help developing nations? To what extent can benefi ts be shared? Will pharmaceutical and biotechnology companies invest in poor countries if they can make more money working on therapeutics for high-income countries? Thus, concern for the bioethical issues of social justice and benefi cence arises. Genomics has the potential to be a global public good, but there is considerable uncertainty as to its bioethical justifi cation in all cultures [10] . In general, globalization helps liberalize trade through removal of import restrictions and tariffs, through removal of restrictions on trade in services, and through linkage of trade sanctions to the protection of intellectual property rights. All these activities may have an impact on population health. Defenders of trade liberalization claim that this process is one of the most effective means of increasing a country's wealth and, by extension, population health. Even if this were always true, there may be specifi c policies that have particularly detrimental effects on health (such as opening markets to trade in manufactured tobacco products). Further, there may be an ethical argument based on social justice against some trade liberalization policies. If, for example, trade liberalization between rich and poor countries produces proportionally more wealth in rich countries compared with poor countries, this may suggest a socially unjust result of liberalization; poor countries' economies may not grow as fast as rich countries' economies in this situation. The relationship between wealth and health is actually somewhat controversial: the so-called Preston curves demonstrate a dramatic relationship between health and economic prosperity up to about a Purchasing Power Parity of US$3,000 per capita per year [11] . However, there are cheap, cost-effective approaches to population health (such as vaccination, clean water, and sewage disposal) that may not be affected by the increase in Purchasing Power Parity. These approaches were relatively more important than economic development per se in early 20thcentury interventions in developed countries, and they are likely to be more important for infl uencing health among developing country populations today than simple economic growth. On the other hand, high-intensity technological improvement rather than economic growth may be more important to health in rich countries compared with developing countries. The concern for intellectual property rights in trade has been an extraordinarily contentious issue in recent years. Newer drugs that are effective against diseases in resourcepoor but highly impacted countries, such as antiretroviral drugs against HIV, have been prohibitively expensive in these countries, in part because of patent protections. With the Trade-Related Intellectual Property agreement, patent protection became linked to trade policy; if countries in need of cheaper essential drugs did not conform to patent rules, trade retaliation from exporting countries might ensue. However, restrictions on poor countries' responses to legitimate public health emergencies may be unethical on the basis of distributive justice, nonmalefi cence, and benefi cence. Exceptions for public health emergencies (such as HIV/AIDS) under the Trade-Related Intellectual Property agreement include the right to compulsory licensing (local companies produce patented medicines in exchange for a royalty payment to the patent holder) or parallel importing (importing patented drugs sold more cheaply elsewhere) that will make essential medicines more available to highly impacted countries without fear of trade retaliation from the originating country [12] . The General Agreement on Trade in Services is a relatively new treaty that covers trade in health services [13] . The agreement has been severely criticized by some, who claim that it increases privatization of health care services and undermines public health care systems. However, given its ambiguities, the actual impact of the agreement on the health sector will be largely determined by the way in which the agreement is further specifi ed in multinational commitments [14] . Social justice, equity, benefi cence, and nonmalefi cence will all come into play in the implementation of this treaty. Concerns for security against biological weapons have recently arisen among both poor and wealthy nations. Some, however, question the enormous sums now being spent to address the perceived threats due to bioterrorism even without strong evidence for actual threats. Even without such evidence, global bioethical principles at least suggest the need for a framework for consideration of distributive justice in this arena. For example, should a nation with a limited supply of a vaccine against weaponized smallpox offer its stockpiles to a neighboring country that is under direct attack? This case is complicated by the fact that the infection could spread to its own territory. In the case of widespread biological attacks, which global governing agency, country, or other entity would be responsible for global resource allocation? Clearly, risks from bioweapons are trans-border, but resources may be unevenly and inequitably distributed, requiring a bioethically based policy determination on a global basis [15] . A further concern with respect to biomedical research is the issue of dual-use technology development for health benefi ts as well as for possible bioweapons. Governments must balance the secrecy necessary for security with the need for disclosure of information that is essential for research and development in health. It is very diffi cult to sequester new knowledge that might be applied to building biological weapons without simultaneously impeding research on defense against those bioweapons and on other benefi cial biomedical advances. Most BIG Project scientists agree that the benefi ts of releasing scientifi c information in general outweigh the risk of its misuse. However, the scientifi c community needs to consider whether new codes of conduct are necessary or whether existing governance is suffi cient to support a bioethical approach to research on possible dualuse technologies. Global bioethical challenges require careful theoretical deliberation and practical considerations for international health policies [16] . The BIG Project seeks to guide these processes in four selected areas of interest to the EU, so that the project results may be helpful to policy makers at local, national, and international levels. The BIG Project has found that bioethical principles are important in considerations of migration, trade, information technology, genomics, and bioweapons threats. Globalization in these arenas is neither a right nor a wrong process, but it demands careful consideration of bioethical principles including social justice, benefi cence, nonmalefi cence, and individual autonomy. These concerns may not be immediately obvious to health policy makers, and thus the BIG Project results may help clarify the larger goals and purposes of bioethically based health policy development within the EU and elsewhere. More information about the BIG Project can be found at http:⁄⁄www. bigproject.org/project.htm.
42
Public awareness of risk factors for cancer among the Japanese general population: A population-based survey
BACKGROUND: The present study aimed to provide information on awareness of the attributable fraction of cancer causes among the Japanese general population. METHODS: A nationwide representative sample of 2,000 Japanese aged 20 or older was asked about their perception and level of concern about various environmental and genetic risk factors in relation to cancer prevention, as a part of an Omnibus Survey. Interviews were conducted with 1,355 subjects (609 men and 746 women). RESULTS: Among 12 risk factor candidates, the attributable fraction of cancer-causing viral and bacterial infection was considered highest (51%), followed by that of tobacco smoking (43%), stress (39%), and endocrine-disrupting chemicals (37%). On the other hand, the attributable fractions of cancer by charred fish and meat (21%) and alcohol drinking (22%) were considered low compared with other risk factor candidates. For most risk factors, attributable fraction responses were higher in women than in men. As a whole, the subjects tended to respond with higher values than those estimated by epidemiologic evidence in the West. The attributable fraction of cancer speculated to be genetically determined was 32%, while 36% of cancer was considered preventable by improving lifestyle. CONCLUSION: Our results suggest that awareness of the attributable fraction of cancer causes in the Japanese general population tends to be dominated by cancer-causing infection, occupational exposure, air pollution and food additives rather than major lifestyle factors such as diet.
In Japan, cancer has been recognized as a major component of the overall pattern of disease for decades. Thus, the importance of cancer prevention by lifestyle modification should now be strongly acknowledged. Internationally, several studies have estimated the proportion of total cancer deaths attributable to various risk factors based on epidemiologic evidence [1, 2] , and various international guidelines and recommendations derived from these have appeared [3] [4] [5] [6] . Not surprisingly, domestic guidelines and recommendations for cancer prevention in Japan such as the 'Twelve recommendations for cancer prevention [7]' and 'Healthy People Japan 21 [8] ' have been significantly influenced by these reports. Public awareness of risk factors in relation to cancer prevention has been surveyed in only a few countries [9, 10] , and results have demonstrated poor awareness. Other studies focusing on specific cancers only have also appeared [11] [12] [13] [14] . However, none of these studies quantitatively evaluated public awareness of the attributable fraction of individual risk factors. In Japan, it appears that most people are aware of the major risk factors of cancer. Although we are unaware of any published evidence, however, public knowledge and information on cancer prevention now seems influenced largely by the mass media and other sources, rather than by information provided directly by health professionals, resulting in a distorted picture of causation. Cancer control policy therefore urgently requires a clarification of the discrepancies which now exist between ideal levels of public concern about risk factors and the current reality, particularly public health policy makers in their formulation of cancer control measures. To address this need, the present study was designed to provide information on awareness of the attributable fraction of cancer causes among the Japanese general population. Since we are interested in quantitatively estimating the awareness of preventability, we placed special emphasis on gauging awareness by attributable fraction of cancer. The study was conducted as a part of an omnibus survey in December, 2003, by commission to a polling agency. The omnibus survey is a monthly multipurpose cross-sectional survey which includes public opinion research, social research, scientific research, market research, and others. Using a stratified two-stage sampling method, a total of 2,000 people aged 20 or older were randomly selected as study subjects, from 160 districts selected from area units representing 12 geographical blocks (Hokkaido, Tohoku, Kanto, Keihin, Koshinetu, Hokuriku, Tokai, Kinki, Hanshin, Chugoku, Shikoku, Kyushu) and 3 types of city scale (14 metropolises, other cities, towns and villages) in proportion to the population distribution as at March 2002. After an initial visit to obtain oral informed consent and schedule a visit for the interview, the survey was conducted by face-to-face interview using trained interviewers in each district. The omnibus survey does not collect any personally identifiable information such as name, date of birth or address details at interview. For the present report, we obtained the electronic data file for the relevant interview component, with no personal identifiers. Ethical approval was not applicable to the present study under the Japanese ethical guidelines for epidemiologic studies, which comply with the declaration of Helsinki. Among the 2,000 people selected for survey (977 men, 1,023 women), interviews were successfully obtained The questionnaire of this survey comprised questions on the awareness of various environmental and genetic risk factors in relation to cancer prevention by enquiring about the attributable fraction of cancer. Fractions were: 1) 12 risk factor candidates, namely alcoholic beverages, unbalanced diet, use of food additives and pesticide chemicals, charred fish and meat, tobacco smoking, obesity, physical inactivity, endocrine-disrupting chemicals, air pollution such as diesel emissions, occupational exposure, cancer-causing viral and bacterial infection, and stress; 2) genetic factors in general; and 3) the preventable fraction of cancer occurrence by lifestyle modification [see Additional file 1]. The first question asked about the preventable fraction of cancer which would result in Japan if each factor were completely and totally eliminated, using the fine categories of <5%, 5 to <10%, 10 to <15%, 15 to <20%, 20 to <25%, 25 to <30%, 30 to <40%, 40 to <50%, 50 to <60%, 60 to <70%, 70 to <80%, 80 to <90%, and 90 to 100%. These categories were exhibited together on a pie chart. These risk factor candidates were selected with reference to previous international and domestic recommendations and guidelines [1] [2] [3] [4] [5] [6] [7] [8] . The second question asked about the fraction of cancer genetically predetermined using the same categories as the first, while the third asked about the preventable fraction of cancer by modification of lifestyle using estimation of an actual percent value. In addition to these questions, subjects were also asked about their smoking and drinking practices, and occupational and educational status. Mean values of the attributable fractions were calculated for each risk factor of cancer and compared by demographic and habitual smoking and drinking status. For analyses, the mid-values of each category were assigned for categorical variables. All analyses were performed using Stata statistical software, S/E Version 8 [15] . A total of 1,355 (67.8%) subjects responded to the survey, with a higher response rate in women (72.9%) than in men (62.3%). Response rate was lower in the 20s age strata than in the other age groups, but no trend to an increase in response rate with increasing age was observed. Overall, no significant difference in area and age distribution was seen between the sampled population and survey respondents. Response rate tended to be lower among subjects who reside in the Kanto region and in cities other than the 14 metropolises than among other subjects ( Table 1) . Characteristics of the 1,355 respondents (609 men, 746 women) are presented in Table 2 . The proportion of current smokers was 44% in men and 15% in women, and decreased with age in both genders. In female subjects aged in their 20s, 26% currently smoke and 49% drink alcohol beverages at least 4 times a week. Awareness of the attributable fraction of cancer causes among the Japanese general population is presented in Table 3 . Among the 12 risk factor candidates, the attributable fraction was considered highest for cancer-causing viral and bacterial infection (51.3%), followed by tobacco smoking (43.0%), stress (39.0%), and endocrine-disrupting chemicals (37.1%). In contrast, the attributable fraction of charred fish and meat (21.4%) and alcohol drinking (21.7%) were considered low compared with other risk factor candidates. The attributable fraction of other risk factor candidates such as occupational exposure, air pollution, food additives and pesticides, unbalanced diet, obesity and physical activity ranked between the high and low fractions. The attributable fraction responses tended to be higher in women than in men, and were increased among inhabitants of larger cities and in homemakers and decreased in those engaged in agriculture, forestry and fisheries. In contrast, risk factor candidate rankings were similar by gender, age group, city scale, and educational and occupational status. In men, those who neither smoke nor drink tended to consider the preventive fraction of the risk factors higher than those who both smoke and drink, whereas in women, the former subjects considered the values lower than the latter. The speculated fraction of cancer which is genetically determined was 31.5% as an average (Table 3 ). This fraction was higher in current heavy smokers and former drinkers, and lower in homemakers and students. On the other hand, an average 35.5% of cancer were considered preventable by lifestyle improvement, with this ratio being higher in homemakers, former smokers, and never and former drinkers. The present survey, targeted at the Japanese general population, showed that the attributable fraction of cancer among Japanese tended to be higher for cancer-causing infection, occupational exposure, air pollution and food additives than major lifestyle factors such as dietary factors. In addition, the attributable fraction of cancer estimated by the Japanese general population was higher than that derived from epidemiologic evidence in the West, which is frequently quoted as 30% for tobacco smoking and 30% for food as a whole [1, 2] . Some of the major cancers in Japan, including gastric and liver cancers, are known to be related to cancer-causing viral and bacterial infection, and a higher level of concern about such infection among Japanese than in Western populations would therefore be understandable [9] . Notwithstanding the validity of such concern, however, the high level of concern for infection, as well as for endocrine-disrupting chemicals, identified in the present survey was most likely due to the severe acute respiratory Likewise, a high level of concern for tobacco smoking was also observed, in spite of a relatively dull reduction in the rate of male current smokers in past decades compared with the U.S. This was probably due to recent enactment of the Health Promotion Law, which curbs passive smoking in public spaces. Respondent estimates for attributable fractions were generally high. This may be in part due to anchoring and adjustment effects of the response categories used and the tendency of people to respond near the middle of the scale. Given that responses tended to be generally high, concern over the present results should probably be focused on rankings rather than absolute values per se. Although tobacco smoking ranked among the top factors, risk factor candidates whose actual contribution is considered to be low, such as endocrine-disrupting chemicals, occupational exposure, air pollution such as diesel emissions and the use of food additives and pesticide chemicals ranked higher than previous estimates of the attributable fraction of cancer causes [1, 2] . In contrast, this should be compared with the results for unbalanced diet, which ranked at only 8th among the 12 risk factor candidates despite an actual ranking which is estimated to be as high as that for tobacco smoking. Particularly in light of findings on long-term exposure to common lifestyle factors such as diet as a cause of cancer, these results suggest that public awareness of cancer prevention is still insufficient. We are unaware of any previous studies aimed at determining public awareness of the attributable fraction of cancer as a whole or at gauging the level the awareness of cancer prevention by attributable fraction. Accordingly, to our knowledge, this is the first attempt to discover the level of awareness for each risk factor candidate, and the questionnaire used has hence not been fully validated. In addition, as indicated above, responses to this type of cross sectional survey are subject to social conditions such as information from the mass media and other sources on disease epidemics and other putative risk factors. Thus, the results might not necessarily reflect actual public awareness. However, the study subjects were recruited from among a nationally representative random sample, and the response rate was similar to that of recent omnibus surveys in other countries [16] [17] [18] [19] . Nevertheless, the exclusion of non-respondents may have distorted the results. In conclusion, awareness of the attributable fraction of cancer causes among the Japanese general population tended to be dominated by infection, occupational exposure, air pollution and food additives rather than dietary factors. The results of the present survey provide valuable clues and perspectives toward the formulation of relevant cancer prevention strategies in Japan.
43
Sequence specific visual detection of LAMP reactions by addition of cationic polymers
BACKGROUND: Development of a practical gene point-of-care testing device (g-POCT device) requires innovative detection methods for demonstrating the results of the gene amplification reaction without the use of expensive equipment. We have studied a new method for the sequence-specific visual detection of minute amounts of nucleic acids using precipitation reaction by addition of cationic polymers to amplicons of Loop mediated isothermal Amplification (LAMP). RESULTS: Oligo DNA probes labeled with different fluorescent dyes were prepared for multiple nucleic acid templates, and the templates were amplified by the LAMP reactions under the existence of the probes. At completion of the LAMP reaction, an optimal amount of low molecular weight polyethylenimine (PEI) was added, resulting in the precipitation of the insoluble LAMP amplicon-PEI complex. The fluorescently labeled Oligo DNA probes hybridized to the LAMP product were incorporated into the precipitation, and the precipitate emitted fluorescence corresponding to the amplified nucleic acid templates. The color of emitted fluorescence can be detected easily by naked eye on a conventional UV illuminator. CONCLUSION: The presence or absence of minute amount of nucleic acid templates could be detected in a simple manner through visual assessment for the color of the LAMP amplicon-PEI complex precipitate. We conclude that this detection method may facilitate development of small and simple g-POCT device.
Loop-mediated isothermal amplification (LAMP) is a unique gene amplification method in which DNA can be isothermally amplified using only one enzyme [1] [2] [3] . Since the advent of the LAMP method, many researchers have been engaged in basic research from a variety of perspectives. As a result, it is currently being put to practical use in the reagents for detecting various pathogens such as SARS [4] and the West Nile virus [5] and reagents for identifying the sex of fertilized eggs in cow in vitro fertilization [6] . Furthermore, LAMP is a gene amplification method with a variety of characteristics and applications in a wide range of fields, including SNP typing [7] and quantification of template DNA [8] . In particular, LAMP is considered to be effective as a gene amplification method for use in gene point-of-care testing (g-POCT) devices, which are used for simple genetic testing whenever and wherever necessary. First, since LAMP can amplify genes isothermally, the amplification reaction can be carried out with a simple heater. There is no need for the special device used Pattern diagram of LAMP reaction and hybridization of fluorescently labeled oligo DNA probes Figure 1 Pattern diagram of LAMP reaction and hybridization of fluorescently labeled oligo DNA probes. The LAMP reaction takes place in three steps (starting material production step, cycling amplification step, and elongation and recycling step) by the primers depicted in the enclosure. In the starting material production step, the starting material (6) is generated by primers (forward inner primer (FIP) and backward inner primer (BIP)). A complementary strand (11) of the starting material (6) is synthesized from the starting material (6) by a reaction that uses itself as a template and by a reaction from an FIP annealed to the loop segment, thus making up the cycle amplification step. During this step, probes (probe F and probe B, respectively) designed for the region between the F1 and F2 region or the B1 and B2 region can hybridize to the loop segment. As the cycle reaction progresses, an elongation and recycling step takes place, during which elongated products (8, 13, etc.) with an inverted repeat structure are generated. Numbers 14 and 15, which have a cauliflower structure, are also generated. They have many loop structures to which probes can hybridize. for polymerase chain reaction (PCR) to rapidly control the temperature [3] . Next, a large amount of DNA (10-30 µg/25 µl) can be synthesized in a short time (15-60 min) while maintaining high specificity. This characteristic greatly facilitates detection of the LAMP reaction [9] . Moreover, since the LAMP reaction progresses by generating a characteristic stem-loop structure, LAMP products have a single-stranded segment in the molecule (loop segment; see Figure 1 and reference No. 1). By using oligo DNA probes designed to recognize the sequence of the single-stranded segment, it is possible to carry out hybridization assay without performing heat denaturation after amplification. This means that all processes, from the amplification reaction to the detection reaction, can be carried out completely isothermally. If these characteristics of the LAMP method are used effectively, we believe it will be possible to develop simple genetic testing devices that have not been realized yet despite a strong awareness of their necessity, in a wide range of fields, including infectious disease testing, food inspection, and environmental testing. The key to developing such simple devices will be figuring out how to simply and clearly present the final amplification results. The objective of this research is to establish new techniques for sequence-specific visual detection of amplification results by means of the LAMP method. To that end, we made use of a reaction that has been known for a long time, i.e., cationic polymers like polyamines form an insoluble complex with DNA [10] . It is well known that one of such the polyamine, polyethylenimine (PEI), strongly interacts with DNA. PEI is widely used as a nucleic acid precipitant for nucleic acid purification [11] and as an in vivo and in vitro non-viral vector [12, 13] . We discovered that an insoluble PEI-LAMP product complex was generated under certain optimized conditions when PEI was added to LAMP reaction solution. Using this precipitation titration, we investigated whether it was possible to perform bound/ free separation of fluorescently labeled probes. In this paper, we report the novel visual detection methods of the presence of sequences of HBV and HCV cloned to plasmid as a model experiment using this method and the results of an investigation of the basic reaction conditions required to achieve this. The primers used for LAMP reaction is schematically depicted in enclosure of Figure 1 . Forward Inner Primer (FIP) consists of F2 and the complementary sequence of F1, and Backward Inner Primer (BIP) contains of B2 and the complementary sequence of B1 when each sequences (F1, F2, B1, and B2) are defined on the template sequence as shown in Figure 1 . In some references such as Notomi et al. [1] or Parida et al. [5] , a spacer of few thymidines was inserted between F1c or B1c and F2 or B2 in the inner primer (FIP or BIP) so that one and two thymidine spacers were inserted in FIP and BIP of HBV, respectively. However, the spacer was not used in this study because the LAMP reaction can progress with the use of inner primers without the spacer as shown by Hong et al. [4] , Hirayama et al. [6] , and Iwasaki et al. [7] . The LAMP reaction takes place isothermally in the three steps shown in Figure 1 , i.e., starting material production step, cycling amplification step, and elongation and recycling step, by using of polymerase with strand displacement activity. First, the starting structure (structure 6) is generated from the template nucleic acid in the starting material production step. Next, the starting structure becomes structure 7 with a stem-loop structure by self-primed DNA synthesis. When the forward inner primer is hybridized to the loop segment and strand displacement synthesis takes place, structure 11, which is a complementary strand of structure 6, is generated. This means that an auto cycle reaction was established between structure 6 and structure 11. In addition, products bound by an inverted repeat with two amplified regions, like structures 8 and 13, are generated in association with the auto cycle reaction (cycling amplification step). Then, with these structures as starting points, products elongated to a length of several kbp and products with complex structures with cauliflower-like structures (14, 15) are ultimately generated. Since the LAMP products have a loop structure, oligo DNA probes (green arcs in the figure) in the reaction solution can sequentially hybridize to products as the LAMP reaction proceeds. The cauliflower structures, in particular, contain two or more loops to which the probes can hybridize. This characteristic of the LAMP reaction plays an important role in the detection method described here. Plasmid DNA cloned with HBV or HCV sequence was added to a LAMP reaction solution containing both HBV primers and the probe (FITC-labeled) and HCV primers and the probe (ROX-labeled) and amplified, after which low molecular weight PEI (Mw = 600; 0.2 µmol as a monomer) was added. As shown in Figure 2A , precipitate emitting green fluorescence characteristic of FITC was obtained in LAMP reaction solution containing the HBV template, precipitate emitting red fluorescence characteristic of ROX was obtained in LAMP reaction solution containing the HCV template, and precipitate emitting a color (orange) that was a combination of FITC green and ROX red fluorescence was obtained in LAMP reaction solution containing both templates. They could be observed using an ordinary UV illuminator or UV-LED (light emitted diode). When LAMP products (lambda DNA) not related to HBV and HCV were present, precipitate with no fluores-Sequence-specific visual detection method that utilizes precipitation titration of LAMP products by adding PEI Figure 2 Sequence-specific visual detection method that utilizes precipitation titration of LAMP products by adding PEI. (A) Results of sequence-specific visual detection after adding PEI to LAMP reaction solution. After LAMP reaction in the presence of both FITC-labeled HBV probes and ROX-labeled HCV probes followed by addition of the prescribed amount (0.2 µmol as monomer) of PEI (Mw = 600), it was centrifuged for several seconds using a desk-top, low-speed centrifuge. The tube was then visually observed as is on a UV illuminator (365 nm). It was possible to differentiate the LAMP reaction by visualizing the presence of precipitate fluorescence and the color of the fluorescence. 1, LAMP reaction negative. 2, When LAMP reaction with PSA amplification (unrelated LAMP reaction) occurred. 3, When it contained HBV template nucleic acid. 4, When it contained HCV template nucleic acid. 5, When it contained both HBV and HCV template nucleic acids. (B) Diagram of principle of sequence-specific visual detection method that utilizes precipitation titration of LAMP products by adding PEI. First, a LAMP reaction is carried out using a LAMP primer set for two types of template nucleic acid and fluorescently labeled probes, which can hybridize to loop segments of each LAMP products. When a LAMP reaction corresponding to a certain fluorescently labeled probe progresses, the probe will sequentially hybridize to the loop segment generated during the reaction. On the other hand, an unrelated probe remains free in the solution. When an optimized amount of PEI is added after reaction for a set length of time, the positive charge of PEI neutralizes the negative charge of the DNA to form an insoluble LAMP product-PEI complex. At this stage, fluorescently labeled probes hybridized to LAMP products are taken up by the LAMP product-PEI complex together with the LAMP products. Since most of the PEI added is used for formation of the LAMP product-PEI complex, free oligo DNA probes cannot form a complex with PEI. When the generated insoluble complex is pelletized by centrifugation and the pellet is irradiated with excitation light, the labeled fluorescent dye hybridized to LAMP products produces fluorescence. Figure 3 Precipitation titration of LAMP products by addition of PEI. (A) Effect of amount of PEI on sequence-specific incorporation of ROX-labeled lambda DNA recognition probes by DNA-PEI complex (Mw of PEI is 600). When 0.2 µmol to 1.0 µmol of PEI was added as a monomer, almost 100% of labeled probes hybridized to the LAMP products for lambda DNA was taken up by the DNA-PEI complex. On the other hand, when 0.4 µmol to 0.8 µmol of PEI was added as a monomer to a reaction solution in which an amplification reaction did not take place, a small amount (<20%) of labeled probes precipitated. Precipitation of this nonspecific probe did not take place when unrelated LAMP products (PSA) were present. (B) Effect of the Mw of PEI on sequence-specific incorporation of ROX-labeled lambda DNA recognition probes by DNA-PEI complex. When PEI with Mw 600 was used, the fluorescence intensity of supernatant decreased only when LAMP product for lambda DNA was present. As the Mw of the PEI used increased from 1,800 to 10,000, the fluorescence intensity of the supernatant decreased in the absence of a LAMP reaction since the formation of insoluble PEI-oligo DNA probe complex occurred. (C) Effect of amount of KCI on sequence-specific incorporation of ROX-labeled lambda DNA recognition probes by LAMP product-PEI (Mw = 600) complex. Normal LAMP reaction solution (control) contains 10 mM of KCI. Since formation of the PEI-LAMP product complex was inhibited by an increase in the amount of KCI added, the fluorescence intensity of supernatant increased regardless of the sequence of the LAMP product. (D) Effect of amount of LAMP product for lambda DNA on sequence-specific incorporation of ROX-labeled lambda DNA recognition probe by LAMP product-PEI complex (Mw of PEI = 600). Even when the amount of LAMP product was 1 µg per 25 µL, almost 100% of labeled probes (1 pmol) was taken up by PEI-DNA complex. cence was obtained, and visible precipitate was not generated in a sample not containing a template (LAMP reaction negative). This means that it was possible to assess whether the HBV template nucleic acid was in the reaction solution, HCV template nucleic acid was in the reaction solution, or both were in the reaction solution by visualizing the fluorescent color of the precipitate. The LAMP method is a nucleic acid amplification method that is so sensitive that it is possible to create an amplification reaction from only six copies of template nucleic acid [1]. Therefore, a combination of LAMP amplification and this detection method makes possible sequence-specific visual presentation of the presence of trace amounts of nucleic acid, i.e., only several copies, found in a sample. A model that represents the principle of this method is shown in Figure 2B . Oligo DNA probes labeled with fluorescent dye are hybridized to a specific LAMP product. When an optimized amount of low-molecular-weight PEI (Mw = 600) is then added, the positive charge of PEI neutralizes the negative charge of the DNA, which results in formation of an insoluble DNA-PEI complex. When this solution is left to stand for a few minutes or it is centrifuged with a small, desk-top centrifuge for a few seconds, the complex is deposited at the bottom of the tube. When the precipitate is observed on an illuminator (365 nm), the fluorescence of the dye from the probe taken up by the precipitate is visualized. On the other hand, probes unrelated to the sequence of the amplified LAMP product are not taken up by the LAMP-PEI complex because they are not hybridized to the LAMP product. When the molecular weight of the PEI used is small, oligo DNA probes and PEI cannot interact sufficiently to form an insoluble complex. Therefore, unrelated probes remain in the supernatant. Since the fluorescent probes in the supernatant are dispersed, the fluorescence cannot be visualized. Consequentially, bound/free separation of labeled oligo DNA is achieved as a result of insolubilization by PEI of LAMP products. The following experiments were conducted to confirm this principle. First, we investigated the effect of the added amount of PEI on this precipitation titration ( Figure 3A ). Oligo DNA probes for lambda DNA were captured in the precipitate when 0.2 to 1 µmol of PEI was added to 25 µL of LAMP reaction solution for lambda DNA. If the amount of PEI added is too high or too low relative to the optimal range, the PEI-DNA complex precipitate is not formed, resulting in oligo DNA probes remaining in the solution. This phenomenon is characteristic in ionic interaction between cationic polymers and anionic polymers [14] . Namely, when the amount of the cationic polymer PEI is too low, the PEI-DNA complex becomes anionic. In contrast, when the amount of PEI is too high, the PEI-DNA complex becomes cationic. In both cases, the PEI-DNA complex is solubilized as a result. This characteristic, which is shown in Figure 3A , indicates that this precipitation titration is based on neutralization of the negative electric charge of the DNA by the cationic polymer PEI. In the case of LAMP negative, a small amount of free probe was deposited under certain conditions when 0.4 µmol to 0.8 µmol of PEI was added as a monomer (< 20%). However, the amount of precipitate in this case was so small that it was impossible to confirm it visually. On the other hand, when LAMP products (PSA) unrelated to the probe sequence were present, no free probe at all was deposited. This is because almost all of the PEI molecules added were consumed in the precipitation of unrelated LAMP products, an excess of which was present relative to the amount of probe. In other words, more reliable detection is achieved by first confirming whether the LAMP reaction has occurred by checking whether white DNA-PEI complex precipitate is generated as a result of addition of PEI and then determining for which nucleic acid template the LAMP reaction occurred based on the fluorescent color of the precipitate. Because of the above results, the amount of PEI added was fixed at 0.2 µmol for the following experiments in order to avoid generation of free probe precipitate as much as possible. We investigated the effect of the Mw of PEI on this detection system ( Figure 3B ). When PEI with different Mw was added to the LAMP solution so that the amount per monomer of each was the same, we found that sufficient BF separation occurred if LAMP amplicons were present, even if the Mw of PEI increased up to 10,000. That is, the fluorescence intensity of the supernatant of LAMP solution with specific amplification (lambda DNA) is lower than that with unrelated amplification (PSA). As the Mw of PEI increased, however, almost all probes in LAMP reaction negative solution formed an insoluble PEI-oligo DNA complex. This result means that the LAMP reaction negative solution cannot be distinguished from the LAMP reaction positive solution, which successfully amplifies the targeting sequence if PEI with a high molecular weight is used. As was also observed in Figure 3A , when LAMP amplicons are present, almost all of the added PEI reacts with LAMP amplicons, so there is little opportunity for interaction with free oligo DNA probes, but when LAMP amplicons are not present, all of the added PEI interacts with oligo DNA probes. Under conditions where a large amount of PEI with a high molecular weight can strongly interact with oligo DNA, an undesirable insoluble PEIoligo DNA probe complex forms because the molecular weight of PEI is high. Therefore, the average molecular weight of PEI in this detection system should be about 600. We conducted a similar experiment using spermine, which is a polyamine with a lower molecular weight. In that experiment, an adequate amount of insoluble complex was not generated compared with PEI of an average molecular weight of 600 under conditions used for the present research (data not shown). It is well known that spermine can also make DNA insoluble, as indicated in many other reports [10] . Therefore, optimal conditions in the case of spermine as a precipitant might exist, but since further investigation would have gone beyond the scope of this paper, no further investigation was carried out. Insolubilization of LAMP products by PEI was inhibited by addition of an excessive amount of KCl to the LAMP reaction solution after amplification ( Figure 3C ). This was because excessive amounts of potassium ions and chloride ions inhibited the electrostatic interaction between DNA and PEI. This finding is further indication that this detection method is based on neutralization of the negative electric charge of DNA by the positive electric charge of PEI. The effect of ionic strength on the LAMP reaction has been investigated and found that the presence of 200 mM or more of KCl markedly delayed the LAMP reaction (data not shown). Therefore, it can be said that this detection method will function without trouble if the solution used has a composition that is optimized for the LAMP reaction. We investigated the sensitivity of this detection method ( Figure 3D ). Almost all probes were taken up by the precipitate in the case of up to 1 µg of LAMP product. Moreover, we were able to visualize the fluorescence in the precipitate even if the specific LAMP product was 0.2 µg. We found that this detection system was sensitive enough as a detection system for visual assessment to be used in simple g-POCT devices. We can see from the results shown in Figure 3C that all of the 1 pmol of probe added hybridized to 1 µg (= 3 nmol nucleotide) of LAMP product. In other words, one molecule of probe bound to every 1,500 base pairs of LAMP products. The LAMP product is a mixture of products of several different sizes, with an average molecular size of 2 kbp [1]. This finding that one molecule of probe binds to every 1,500 base pairs reflects well the fundamental characteristic of this LAMP reaction. The new detection method described above utilizes the unique nature of low-molecular-weight PEI, i.e., it cannot form an insoluble complex with a single-stranded anionic polymer with a low molecular weight such as an oligo DNA probe, but it can form an insoluble complex with DNA with a high molecular weight such as LAMP product. Until now, not much attention appears to have been paid to the complex consisting of oligo DNA and PEI with a molecular weight of 1,000 or lower, which was used in this study. This is probably because the interaction between low-molecular-weight DNA and oligo DNA is very weak and because its practical utility as a vector or nucleic acid precipitant is low. However, Kunath et al. reported that the complex formed by PEI (5 kDa) and plasmid DNA was more unstable than that formed by PEI (25 kDa) [15] . Osland and Kleppe reported that spemidine, which has a structure similar to that of PEI, formed an aggregate with double-stranded oligo DNA, but it did not form an aggregate with single-stranded DNA [10] . Drawing upon these results, the low-molecular-weight PEI used in this study is thought to interact with DNA with high selectivity to differences in DNA concentration and structure (number of strands and molecular weight). The fact that we took advantage of this nature of lowmolecular-weight PEI as a nucleic acid precipitant for detection led to the establishment of this detection method. This detection method effectively utilizes the characteristics of the LAMP method, i.e., a large amount of amplification product can be synthesized in a short time while maintaining high specificity. Since a large amount of amplification product is created by the LAMP reaction, precipitate of a size that can be easily confirmed with the eyes is generated when PEI is added to the LAMP reaction solution. Moreover, the fact that the amplification is highly efficient means that the amount of labeled probe for detection that can be added is large. As a result of these characteristics, the LAMP reaction followed by addition of PEI yields precipitate with a clear color and in a size that can be identified visually, as shown in the photograph in Figure 2 . Furthermore, the fact that the amount of the amplification product is large means that the range of the optimum amount of PEI necessary to generate the insoluble complex is wide. Thus, a precise system for adding the PEI solution is not necessary. This will contribute to simplification of g-POCT devices that use this detection method. In the case of PCR, the most widely used gene amplification method, the amount of amplification product is usually 1/20 or less that of LAMP [9] . Consequently, in order to apply this method to PCR, a device for rapidly cycling the temperature to perform PCR, a high-luminance fluorogenic reagent, a system for washing unreacted probes, and a system for accurately dispensing PEI would be necessary. This would likely be an obstacle to putting to practical use a simple, inexpensive g-POCT device for genetic testing. There are several ways to make LAMP products insoluble besides addition of a cationic polymer like PEI. We tried carrying out the BF separation using chilled ethanol. Since the addition of chilled ethanol lowered the temperature of the LAMP solution, there was a strong tendency for nonspecific probes to weakly hybridize to LAMP products and precipitate (data not shown). Moreover, isopropanol and PEG are believed to be inferior to a PEI solution because it is necessary to add several times the amount of LAMP reaction solution and they are not as easy to handle. Therefore, we believe that low-molecular-weight PEI, which is efficiently insolubilizes amplicons with only a small amount, is the best option for high BF separation, as was shown in this study. If the 5' end of the inner primer is fluorescently labelled, the LAMP product should be visible as in the current study. However, visualization using inner primers fluorescently labelled at the 5' end is not preferred, because the possibility of false positives from self-extension of the labeled primer cannot be excluded. There is no risk of false positives with the oligo DNA probes fluorescently labeled at the 3' end as in this study, so that highly accurate genetic testing can be established. It is necessary to add PEI to the LAMP reaction solution after the LAMP reaction takes place since PEI strongly inhibits the LAMP reaction. However, opening the reaction tube after amplification is generally avoided to prevent carry-over contamination. Therefore, development of a technique for adding PEI in a closed system is needed to put this method to practical use. Some possible solutions could be to apply PEI to the lid of the reaction tube beforehand and turn the reaction tube upside down after LAMP amplification or use wax that responds to heat. Furthermore, we should be able to use technology like micro-Total Analysis system or Lab-on-a-Chip, for which much research is being conducted in recent years, to solve this problem. We established an extremely simple method for visually detecting LAMP products in a sequence-specific manner by simply adding a small amount of low-molecularweight PEI to the LAMP reaction solution. The biggest feature of this technique is the ability to visually present sequence information of amplicons without using an expensive source of light or a detector. In contrast, conventional genetic tests require expensive reagents, complex and skillful manipulation, and large devices equipped with an expensive optical system. These are the main reasons why genetic testing is kept within the walls of specific institutions or university laboratories with special equipment and trained engineers. The combination of the LAMP method and the new detection method described here can overcome several factors that have been preventing true practical application of super-simple g-POCT devices for genetic testing. If a simple, inexpensive g-POCT device that is small and light enough to be held in one hand and whose main components are disposable can be developed, the day when parents will be able to genetically test for a pathogen while sitting next to the bed of their child who has a fever will not be far off [16] . Viral DNA of the hepatitis B virus (GenBank accession number: V00867) digested with BamHI was cloned to pBR322. Reverse transcriptase PCR was conducted for HCV viral RNA (GenBank accession number: AB031663) from patient serum purified using a QIAmp viral RNA Mini kit (Qiagen K.K.) with the PCR primers designed to the sequence of 5' non-coding region and core. The PCR product was cloned to pBR322 according to the established method. Concentrations of the respective plasmids obtained were determined by 260-nm spectrophotometer (Ultrospec 20000, Pharmacia Biotech). The template DNA solutions for LAMP reactions were prepared by serial dilution of the plasmid solutions and purchased lambda DNA solution (Takara Bio, Inc.) with Tris-HCl buffer (10 mM, pH8.0) until it contained 1,000 copies/5 µl. The LAMP reaction was carried out on a scale of 25 µl based on reference 1. Briefly, a forward inner primer (FIP) and a backward inner primer (BIP) at a final concentration of 1. The LAMP primers (Sigma Genosis Japan K.K., HPLC purification grade) were as follows. The hyphens were added between F1c or B1c and F2 or B2. The LAMP reaction for Lambda DNA was used as a LAMP reaction for a basic investigation to clarify the fundamental characteristics of the reaction that occurs between the LAMP product and PEI. A well-established LAMP product [1] for mRNA of prostate-specific antigen (PSA) was used as a LAMP reaction product unrelated to the three LAMP reaction products above. A solution was prepared by serially diluting a LAMP reaction solution with a known concentration (28 µg DNA/25 µl) using LAMP reaction buffer in order to systematically investigate the reaction between the LAMP product and PEI. The concentration of DNA synthesized by the LAMP reaction was determined according to the attached protocol using a PicoGreen dsDNA Quantitation Kit (Molecular Probes, Inc.). The spectrofluorophotometer used was the RF-5000 spectrofluorophotometer (Shimadzu Scientific Instruments, Inc.). The loop segments within LAMP products, i.e., sequences complementary to sequences between either F1 and F2 or B1 and B2, were used as probes for detection, as shown in Fig. 1 . The probes were designed such that the sequences between F1c or B1c and F2c or B2c had a melting temperature (Tm) of 1 to 5°C lower than LAMP reaction temperature (62°C). All fluorescently labeled oligo DNA probes, which were HPLC purification grade, were purchased from Sigma Genosis Japan KK. The sequences of the probe used were as follows. The numbers in parentheses indicate the length and melting temperature (Tm) from supplier's information. Commercially available PEI (Wako Pure Chemical Industries, Ltd.) was used without further purification. The average molecular weight of PEI used was 600, 1,800, and 10,000. In this study, the concentration of the PEI aqueous solution is expressed as the concentration of the monomer unit (-C 2 H 5 N-, 46 g/monomer). The PEI stock solution (2.0 mol/l) was prepared by dissolving 4.6 g of PEI in 50 mL of deionized distilled water in a graduated cylinder. At the time of use, the stock solution was diluted with water to the desired concentration. White precipitate of magnesium pyrophosphate forms in the LAMP reaction solution as a by-product of the amplification reaction [5] . Therefore, after amplification, we first precipitated the magnesium pyrophosphate to the bottom of the reaction tube by centrifugating the LAMP reaction solution for 10 seconds using a small desk-top centrifuge (6,000 rpm). Then, 4 µl of PEI solution adjusted to the desired concentration was added to the supernatant to form an insoluble DNA-cationic polymeric polymer complex at room temperature. A pellet was formed by immediately centrifuging it for 10 seconds using the same centrifuge. A fluorescent image of the pellet was photographed with a fluorescence microscope (VB-G05, Keyence Corporation) equipped with a Handy UV Lamp (wavelength: 365 nm; Vilber Lourmat) and a UV blocking filter (NEO Dynamic L-400, MARUMI Optical Co., Ltd.). Since the magnesium pyrophosphate precipitate was nonfluorescent, it had no effect at all on visualization of pellet fluorescence. To investigate the effect of ionic strength on precipitation titration, aqueous KCl was added to the reaction solution after the LAMP reaction so that it was the prescribed concentration. After centrifugation, 20 µl of supernatant of the PEI-DNA complex solution was aliquoted to a 384-well assay plate (Corning, Inc.) for fluorometry using a fluorescence plate leader (Polarion; Tecan Japan Co., Ltd.). The percentage of probes captured in the precipitate of the PEI-LAMP product complex was calculated according to the following formula based on the obtained results.
44
Injection drug use and HIV/AIDS in China: Review of current situation, prevention and policy implications
Illicit drug abuse and HIV/AIDS have increased rapidly in the past 10 to 20 years in China. This paper reviews drug abuse in China, the HIV/AIDS epidemic and its association with injection drug use (IDU), and Chinese policies on illicit drug abuse and prevention of HIV/AIDS based on published literature and unpublished official data. As a major drug trans-shipment country with source drugs from the "Golden Triangle" and "Gold Crescent" areas in Asia, China has also become an increasingly important drug consuming market. About half of China's 1.14 million documented drug users inject, and many share needles. IDU has contributed to 42% of cumulatively reported HIV/AIDS cases thus far. Drug trafficking is illegal in China and can lead to the death penalty. The public security departments adopt "zero tolerance" approach to drug use, which conflict with harm reduction policies of the public health departments. Past experience in China suggests that cracking down on drug smuggling and prohibiting drug use alone can not prevent or solve all illicit drug related problems in the era of globalization. In recent years, the central government has outlined a series of pragmatic policies to encourage harm reduction programs; meanwhile, some local governments have not fully mobilized to deal with drug abuse and HIV/AIDS problems seriously. Strengthening government leadership at both central and local levels; scaling up methadone substitution and needle exchange programs; making HIV voluntary counseling and testing available and affordable to both urban and rural drug users; and increasing utilization of outreach and nongovernmental organizations are offered as additional strategies to help cope with China's HIV and drug abuse problem.
Illicit drug abuse has become an increasing public health and social concern in the past decades worldwide. Drug abuse causes many problems both to individuals and to societies, including loss of productivity, transmission of infectious diseases, crime, family and social disorder, and excessive health care expenditures [1] . Human immuno-deficiency virus (HIV)/AIDS, associated with injection drug use (IDU) and needle sharing to a large extent, has become one of most stunning tragedies in human history. It has caused more than 20 million deaths, and about 40 million people are living with HIV worldwide thus far, with Africa as the most afflicted continent [2] . As the most populous country in the world, China has also observed rapidly increasing drug abuse and HIV/AIDS occurrence in the past 10 to 20 years. China can still shape the course of its epidemic, but it needs to move swiftly and with great resolve [2] . This paper reviews global illicit drug trafficking, drug abuse and its association with HIV/AIDS epidemic in China, and Chinese policies on illicit drug abuse and prevention of HIV/AIDS, and offers additional strategies to governmental crack down on drug smuggling and drug use prohibition to help cope with China's HIV and drug abuse problem. We searched English and Chinese language literature via Medline and the China National Knowledge Infrastructure and reviewed unpublished official data, including national reports on illicit drug control and HIV/AIDS sentinel surveillance data. More than 100 papers and reports were reviewed. Key databases included: Drug abuse in China can be traced to the late Qing Dynasty (1644-1911 A. D.), when British colonists forcefully brought Indian opium into China for exchange of silk, tea, and cash. Opium was then locally planned. By the founding of new China in 1949, more than 20 million Chinese people were opium addicts, representing 5% of the total population [3] . After a short nationwide antidrug campaign, drug abuse was reported to be eliminated from the mainland in the early 1950s, and for the next three decades China was believed to be a drug-free nation [3] . Illicit drugs reemerged in China in the 1980s as China adopted an open-door policy, and the reemergence was mainly connected with global drug trafficking activities. In Asia, there are two major opiate-producing regions: the "Golden Triangle," comprising three Southeast Asian countries of Myanmar (Burma), Laos and Thailand, and the "Golden Crescent" that includes the three Southwest Asian nations of Afghanistan, Iran and Pakistan [4, 5] . The majority of heroin and opium in the current Chinese market is brought from Myanmar into Yunnan Province or from Viet Nam into Guangxi Province, and then it is transshipped along inland trafficking routes to Sichuan, Guizhou, Gansu and Xinjiang or to Guangdong, Shanghai and Beijing [6] [7] [8] . A small potion of heroin/opium is trafficked into Xinjiang from the "Golden Crescent" [8] . These drugs further penetrate into other provinces. Since the late 1990s, increasing amounts of amphetamine-typestimulants (ATS) and other chemically related synthetic drugs including amphetamine, methamphetamine and ecstasy have been locally manufactured and consumed in China [4] . The number of drug users documented officially by Chinese public security departments increased from 70,000 in 1990 to 1.14 million by 2004 [9] while the estimated number is 3.5 million [1] . The lifetime prevalence rates of illicit drug use among residents age 15 years or older in high-prevalence Chinese cities increased from 1.1% in 1993 to 1.6% in 1996, and the 1-year prevalence rate increased from 0.9% to 1.2% during this period [10, 11] . The main drug of choice in China is heroin. According to a report of the National Narcotic Control Commission (NNCC), 87.6% of drug users abused heroin in 2002 [12] . The abuse of ATS and MDMA (methylenedioxymethamphetamine or ecstasy) has become popular in city night clubs in recent years [1] . In the 1980s, farmers living in rural bordering areas in Yunnan and Guangxi provinces constituted a large fraction of drug users. Since the early 1990s, more and more urban residents use illicit drugs. The majority of drug abusers are young people with a low education level and limited job skills, although a small proportion of urban users regard using drugs as an indicator of "high social class." NNCC data showed that 74% were aged 17-35 years in 2002 [12] . Experimentation (90%), peer pressure (44%), and relaxation (42%) are commonly cited reasons for beginning to use drugs. Initially, drugs were taken primarily through sniffing/snorting (55%) or smoking cigarettes mixed with drugs (38%) [13] . Injection of drugs became increasingly common among drug users, probably as a result of increasing prices of illicit drugs and greater cost effectiveness of injecting to achieve the desired effect. National behavioral surveillance data showed that the median prevalence of IDU among drug users increased from 35% in April 1995 to 49% in April 2004, and median prevalence of needle sharing among IDUs also increased from 26% to 43% during this period [14, 15] . "Buying from drug or grocery stores" and "buying from hospitals" were the two most common routes for obtaining injection needles and syringes [16] . Reasons for sharing a needle included: "do not care about anything else when hooked," "thought it had been cleaned up," "difficult to buy or obtain," "only sharing with selected persons," "no money to buy," and "following other drug users" [16] . Besides IDU and needle sharing, other risk behaviors including unprotected commercial sex put drug users and ultimately non drug users at high risk of HIV [17, 18] . It is believed that many female drug users sell sex for drugs, and limited published research also provides supporting evidence [18] [19] [20] . However, the impact of the interaction between IDU and commercial sex on HIV risk is not clear. Considering the dramatic increase in the commercial sex industry in China, and the potential bridging role of those with dual risk behaviors in transmitting HIV from high risk groups to the general population, research on the interaction of IDU and commercial sex should be given high priority. Since the first AIDS case was detected in 1985 [21] , China has cumulatively reported about 100,000 HIV infections by 2004 [22] . Thus far, the HIV/AIDS epidemic is mainly concentrated in specific geographical areas and sub-group populations [23] . IDU has been the largest contributor to the reported epidemic since 1989, when the first HIV outbreak among IDUs was observed in Ruili City of Yunnan Province bordering with Myanmar [24] . All 31 provinces, autonomous regions and municipalities on mainland China reported HIV infections among IDUs by 2002, and 42% of infections in China were estimated to be due to IDU by 2004 [22] . In 1995, only one out of 8 national sentinel sites for IDU detected HIV infections. The detected prevalence was 0.02% [14] . By 1997, 3 out of 22 sites detected positives, and the average prevalence increased to 6.6%. After that, national average HIV prevalence among IDUs varied between 5.4% and 8.2% [15, 22] . However, dramatic geographic differences in HIV prevalence have been observed. Yunnan and Xinjiang provinces have the most severe HIV infection rates among IDUs. Other provinces along or close to drug trafficking roads, such as Guangxi, Sichuan, Guizhou, Hunan, and Jiangxi, have moderate epidemics [25] . HIV began to spread rapidly in Yunnan province in the late 1980s and early 1990s [26] . High prevalence among IDUs was first detected in 1998 in Urumqi City (28.8%) and in Yining City (82.2%) of Xinjiang Province and later on in selected sentinel sites for IDU in other provinces [27] . These include Guangxi (>10%, 1998 sentinel data), Jiangxi (14.5%, 2000), Sichuan (16-20%, 2002) , Guizhou (17-19%, 2002) , and Hunan (15-20%, 2003) [23] . In 2004, 75% of HIV national sentinel sites for IDU detected HIV infection. More than 20% of HIV prevalence was reported in selected sites in Xinjiang, Sichuan, Guizhou, Guangxi and Hunan [22] . Official HIV prevalence rates in other geographic areas remain <10%; these reports may reflect reality or may be due to failure in detection. At least 7 HIV-1 subtypes (A, B/B', C, D, E, F, and G), and 3 major circulating recombinant forms (CRF01_AE, CRF07_BC, and CRF08_BC) have spread in China [28] [29] [30] [31] . However, the most prevalent HIV-1 strains are subtype B' (44%), C (29%, usually CRF08_BC) and CRF01_AE (13%) [30] . Subtype C and B'/B are the main subtypes among IDUs [30, 32] . Co-circulating of multiple subtypes of HIV-1 in China implies the possibility of interclade recombination. Characterization of genetic variability of HIV-1 may help track the epidemic, generate subtype-specific immunological reagents, develop vaccines, and choose antiretroviral therapy regimens. Drug trafficking and abuse are illegal in China. Offenders may be sentenced to prison if smuggling 10 grams or more of heroin and could receive the death penalty for smuggling more than 50 gram of heroin. In 1990, the "Regulations on Prohibition Against Narcotics" were enacted with three levels of penalty to be applied to drug users. First-time offenders may be fined and/or allowed to go to voluntary detoxification centers, where they receive 10-day methadone treatment managed by the Ministry of Public Health. The cost of treatment is about 2,000 to 5,000 Chinese yuan -a cost considered expensive for many drug users and their families. If drug users who have gone through a voluntary detoxification program are caught again using drugs, they are sent to compulsory rehabilitation centers (CRC) administered by the Ministry of Public Security for 6 to 12 months. Drug users who relapse users after going through CRC are sent to reeducation-through-labor-centers (RELC) administered by the Justice Department for two to three years [33]. In Chinese history, drug abuse and prostitution have been considered "social evils." The Chinese government typically takes "crackdown" measures and tries to eradicate these phenomena. China did achieve a success story in the 1950s. Illicit drug abuse and prostitution were eradicated through national anti-drug and anti-prostitution campaigns [3, 34 ]. However, this success has not been repeated in the past two decades. One possible explanation is that China has expanded its market economy and opened further to the outside world. Under this situation, it seems impossible to completely stem manufacture of highly profitable illicit drugs and import of drugs along China's long porous land borders with drug-producing countries. Currently, the Chinese government adopts more pragmatic policies and takes measures targeting both the root and surface of the drug abuse problem. The measures targeting the root include continuously cracking down on drug smuggling activities and discouraging new users through anti-drug education campaigns. The Chinese government actively seeks to collaborate with neighboring countries to prevent drug smuggling across borders and to help Myanmar, for example, to reduce opium poppy cultivation by replacing with crop plantation. Chinese mass media have increased anti-drug education to the general population. Anti-drug education has been included in the curricula for primary and secondary school students [9] . The measures targeting the surface include providing drug detoxification and harm reduction services to drug users. China has about 300 voluntary detoxification clinics, 700 CRC, and 200 RELC. Each can accommodate 100 to 3,000 patients [9] . Both western medications and traditional Chinese medicines are used for detoxification, and community therapy has been provided in some heavily affected areas since the late 1990s [9] . The National Working Group for Community-based Methadone Maintenance Therapy was established collectively by the Ministry of Health, the Ministry of Public Security, and the State Whether or not China can shun a generalized epidemic of HIV/AIDS may be largely dependent on how China deals with IDU risk factors and breaks the bridge between IDU and heterosexual transmission. Past experience in China suggests that solely cracking down on drug smuggling and prohibiting drug use can not prevent or solve all illicit drug related problems in the era of globalization. Governmental support, harm reduction programs, voluntary counseling and testing, and utilization of non-governmental organizations are recommended. China has a strong central government. Without government support, it would be not imaginable to achieve success in the campaigns against drug use and the spread of HIV/AIDS. Since the late 1990s, the Chinese central government has stepped up HIV/AIDS control efforts, including setting out national policy framework for responding to HIV/AIDS, increasing funding inputs, expanding collaborations with international organizations. The Chinese National Medium-and-Long-Term Strategic Plan for AIDS Prevention and Control (1998-2010) was formulated in 1998 and set one goal that "by 2002, health education on preventing HIV/AIDS and STDs should be carried out at all detoxification centers and re-education centers as well as in 80% of jails..." [37] . The Action Plan (2001-2005) calls for creating "drug-free communities" through drug prohibition education and drug detoxification activities, together with active promotion of healthy life styles and behaviors and harm reduction for drug users. On the other hand, there is substantial autonomy at provincial level in some areas. Responses to drug use and HIV/AIDS epidemic vary significantly at provincial and lower administrative levels. For example, Yunnan and Guangxi provinces have done far more than other provinces in supporting, implementing, and advocating for harm reduction interventions for IDUs. Some local governments are not fully motivated to confront drug abuse and HIV/AIDS problems. Some government leaders still ignore or even cover up these problems. They are far more interested in economic growth than HIV/AIDS control and wish for their administrative areas to become "economy provinces" or "economy cities" rather than "AIDS provinces" or "AIDS cities," which is believed to be helpful for their career promotion. Advocacy to and support from government leaders at all administrative levels for harm reduction and community-based prevention are needed. Harm reduction includes many strategies, such as methadone maintenance, needle exchange, dispensing other drugs, and outreach services [38] . Harm reduction has been a controversial issue as compared to abstinencebased philosophies [39, 40] . Harm reduction seems to encourage tolerance of social phenomena that are undesirable and hazardous and that may result in social turpitude. Instead, abstinence is considered to be the proper way to address drug problems [39] . In the United States, law and policy restrict the use of federal funds in supporting needle and injection equipment distribution projects. However, many studies refute the concern that access to sterile syringes is an endorsement of IDU and is likely to result in increases in injection and initiation of injection [41, 42] . Harm reduction projects in China are in the pilot phases. Some scholars and health officials still have similar concerns as in some Western countries. They believe that needle exchange services may send a wrong signal of encouraging drug abuse to drug users and the public, and they consider methadone substitution unethical because it uses one drug to replace another drug [43] . As drug use and HIV/AIDS spread rapidly across the country, it is urgent to find supporting domestic evidence that harm reduction will reduce HIV transmission by evaluating harm reduction projects and ultimately scaling up harm reduction efforts. It is easy to access sterile needles and syringes in urban areas of China because they are legally sold and available at pharmacies and medical clinics [16, 44] . However, many drug users live in rural areas; in addition, drug users may share needles and syringes because they can not buy them during night time or do not have money to buy them [16] . Methadone is orally administered, and metha-done substitution can reduce injection and needle sharing of opiate drug addicts. But methadone maintenance therapy is costly and requires drug users to attend clinics on a regular basis. Therefore, methadone substitution and needle exchange services should be made available and affordable at convenient times and in both urban and rural settings, especially in the communities with heavy drug use. In addition to cost and availability, other factors might also affect acceptability of methadone maintenance therapy, such as concerns about the safety and efficacy of the therapy [45] . Greater retention in treatment has been found to result in greater decreases in drug use, criminal activity, and unemployment [46] . The length of drug treatment has a positive association with better post-treatment outcome [46] . However, limited experience with methadone maintenance therapy (MMT) in China shows a high rate of dropouts. International studies have shown that motivational enhancement therapy or motivational interviewing enhances treatment initiation, retention and outcomes in MMT program [47, 48] , and adding behavioral intervention components into MMT programs increases abstinence and reduces HIV risk behaviors [49] . Policy-oriented operational research is needed in China to better understand how to increase the effectiveness of MMT and other harm reduction interventions in the Chinese context. There are still persistent conflicts in the policy and legal landscape. The central government has given explicit support to harm reduction, for example, as stated in the Medium-and Long-Term Strategic Plan and the Action Plan. Some programs have been implemented successfully [36,50,51]. However, in China, as in many other countries, public health and public security authorities frequently approach drug abuse from different perspectives, leading to conflicting approaches at local levels. The crackdown philosophy and detention of drug users in China reflect inconsistent interpretations of "harm reduction" and present a challenge to public health officials in implementing methadone substitution and needleexchange programs [52] . Drug users may be reluctant to participate in these programs due to fear of being caught by police officers [50] . It might be impossible to completely solve the dilemma in the near future, but this conflict is expected to gradually reduce for the following reasons. First, Chinese national policies for HIV prevention and control have become much more pragmatic in the past years. MMT and needle exchange programs were almost unimaginable several years ago, but now they are ready to be expanded across the country. We expect that the open policy trend will continue as the Chinese economy is increasingly merged with international markets, and this trend will favor harm reduction programs. Fur-thermore, China's centralized government may achieve an advantage in promoting public health policies if these policies are believed to be correct. Second, inter-agency coordination on public health crisis has been enhanced at both central and local governmental levels since SARS outbreak in 2003, which reduces potential conflict of public health policies. Public health workers should provide policy advocacy to public security authorities and help them change their traditional norms about illicit drug control and obtain their supports for harm reduction. Third, operational research is needed to provide evidence on the benefits of harm reduction programs and convince policy enforcers and lead to revision of unfavorable policy components. HIV voluntary counseling and testing (VCT) is often considered the first step for initiating prevention and/or therapy. One of the strategies for addressing the AIDS epidemic is to give people an opportunity to know their HIV status so that they can take precautions to avoid further spread and receive early therapy if they are infected [53, 54] . However, even in developed countries, many atrisk people do not take VCT. A national British survey in 2000 showed that only one-third of IDUs had VCT in the past 5 years [55] . About one-fourth of the 0.8 to 0.9 million infected people in the United States remain unaware that they are HIV positive [54] . In China, there is a large discrepancy between reported (about 100,000) and estimated (about 1 million) cumulative HIV/AIDS cases thus far. Many at-risk individuals do not seek out standard HIV counseling and testing services. The stigma associated with drug use and HIV/AIDS and fear of arrest or knowing a positive result can be major barriers to access to VCT. A survey among 840 pregnant women and 780 health professionals in Yunnan Province -an epicenter of the HIV/ AIDS epidemic in China -found prevalent negative attitudes toward HIV/AIDS. Twenty-three percent of health professionals and 45% of pregnant women thought HIV was a disease of "low class and illegal" people; 48% of health professionals and 59% of pregnant women thought that HIV positive individuals should not be allowed to get married; and 30% of the health professionals were not willing to treat an HIV-positive individual [56] . Cost of traditional VCT, low awareness of risk factors for HIV infection, distance, and inconvenience in time also may prevent access to VCT. Possible solutions include development of outreach programs to offer anonymous testing and counseling to those at heightened risk of HIV infection and adoption of new technologies such as rapid saliva testing and counseling strategies to improve the outreach and efficacy of programs. Non-governmental organizations (NGO) can play a critical role in the delivery of HIV prevention services and other assistance to persons living with AIDS. The flexibility of NGOs enables them to respond quickly to fill in gaps in health care and social services. NGOs can do what government agencies cannot do or are not willing to dofor example, reaching out without perceived threat to IDUs and other marginalized sub-groups whose behaviors are often stigmatized and also put them at higher risk of HIV/AIDS. A recent survey of 29 NGOs in Central and Eastern Europe showed that most NGOs targeted injection drug users; provided needle exchange and HIV prevention peer education; and delivered AIDS presentations and distributed educational materials [57] . In Africa, where the main transmission occurs via heterosexual activity, NGOs are most likely to direct their attention to the general public and to youth; they provide peer-education or community outreach [58] . In both Thailand and Brazil, where success has been observed in controlling the HIV/AIDS epidemic or reducing AIDS mortality, NGO are believed to play a key role, but their programs lack rigorous and systemic evaluation [59, 60] . NGOs often face several difficulties: lack of financial resources [57, 58] ; lack of communications with governmental organizations [61] ; governmental indifference or opposition; and AIDSrelated stigma [57] . China has large number of government organized NGOs (GONGOs), including Family Planning Associations, Women's Federation, Red Cross, Youth League, trade unions, and diverse academic associations. The members in these GONGOs have formal positions in governmental organizations while they volunteer at GONGOs. The "true" NGO that has no relationship to the government is just emerging in China. More and more of existing Chinese GONGOs are getting involved in sexually transmitted disease/AIDS prevention. Since the 1990s, the Chinese government has encouraged them to participate in HIV/ AIDS control. These government-sponsored NGOs support HIV/AIDS education and academic publications and participate in AIDS research and education with foreign governmental and non-governmental partners, and they can serve as a powerful aid to the Chinese government to achieve the goal of stopping further spread of drug use and HIV/AIDS epidemic. In spite of its potential for greatly contributing HIV prevention in China, there is little literature in this area. An exception is a recent study (Chen and Liao, 2005 ) of a Women Federation's HIV prevention program in south China. The study showed that the Women Federation was able to deliver a culturally oriented, multi-level intervention program targeted at female drug users. The data also indicated that the program was successfully in increasing knowledge about HIV/AIDS, increasing condom use, and decreasing needle and syringe sharing among the female drug users in the project. Studies which systematically evaluate the implementation and effectiveness of NGO based intervention programs are greatly needed in the future. The author(s) declare that they have no competing interests. HZQ and JES conceived of the study and wrote the first draft of the manuscript, HZQ, HTC and YHR collected the data, and all authors participated in the data interpretation and manuscript revisions.
45
Development of a humanized monoclonal antibody with therapeutic potential against West Nile virus
Neutralization of West Nile virus (WNV) in vivo correlates with the development of an antibody response against the viral envelope (E) protein. Using random mutagenesis and yeast surface display, we defined individual contact residues of 14 newly generated monoclonal antibodies against domain III of the WNV E protein. Monoclonal antibodies that strongly neutralized WNV localized to a surface patch on the lateral face of domain III. Convalescent antibodies from individuals who had recovered from WNV infection also detected this epitope. One monoclonal antibody, E16, neutralized 10 different strains in vitro, and showed therapeutic efficacy in mice, even when administered as a single dose 5 d after infection. A humanized version of E16 was generated that retained antigen specificity, avidity and neutralizing activity. In postexposure therapeutic trials in mice, a single dose of humanized E16 protected mice against WNV-induced mortality, and may therefore be a viable treatment option against WNV infection in humans. SUPPLEMENTARY INFORMATION: The online version of this article (doi:10.1038/nm1240) contains supplementary material, which is available to authorized users.
Development of a humanized monoclonal antibody with therapeutic potential against West Nile virus E protein by nickel-affinity chromatography (data not shown). After immunization and screening 2,000 hybridomas, we isolated 46 new monoclonal antibodies that recognized WNV E protein (Supplementary Table 1 online). We evaluated the antibodies for their ability to block WNV infection in BHK21 cells using a standard plaque-reduction assay 23 . Twelve had strong neutralizing activity that greatly exceeded the potency of immune human γ-globulin, with 50% plaque reduction neutralization titers (PRNT 50 ) below 2 µg, whereas immune human γ-globulin had a PRNT 50 value of 500 µg 5 . The inhibitory activity of two neutralizing antibodies, E16 and E24, was reproduced in J774.2 mouse macrophages and SW13 human adrenal carcinoma cells ( Supplementary Fig. 1 online) and thus was not specific to fibroblasts. One of the potent neutralizing monoclonal antibodies, E16, inhibited infection of genetically diverse WNV lineage I strains that were isolated from mosquitoes, birds and horses in New York. E16 neutralized all WNV strains with PRNT 50 values of 4-18 ng and PRNT 90 values of 53-297 ng (Supplementary Table 2 online) . Notably, Fab fragments of E16 inhibited WNV (PRNT 50 , 23 ng), suggesting that neutralization does not require bivalent E protein binding. E16 potently blocked infection with strain 956, the original lineage II strain isolated in 1937 (ref. 24 ), yet was virus specific, as it neither recognized nor neutralized other flaviviruses including distantly related dengue and yellow fever viruses (data not shown) and closely related Japanese and St. Louis encephalitis viruses (Supplementary Table 2 Figure 1 Mapping of monoclonal antibodies to DIII with yeast. (a) Surface display of WNV E protein on yeast. A fusion protein is composed of the ectodomain or DIII of WNV E protein and the yeast Aga2 protein, which becomes attached to the Aga1 protein on the yeast cell wall. Yeast were transformed with the vector alone (pYD1; top row), the entire WNV E ectodomain (amino acids 1-415; middle row), or DIII alone (amino acids 296-415; bottom row). 24 h after induction, yeast were stained with the indicated monoclonal antibodies (negative control, anti-SARS ORF7a) and processed by flow cytometry. Data for a representative neutralizing (E16) and non-neutralizing (E18) antibody are shown. (b) Flow cytometric enrichment for DIII-expressing yeast variants that lose binding of E24. After each round, an increased percentage of DIII expressing yeast are recognized by the polyclonal WNV E-specific antibody but not by E24. After the final round, DIII-expressing variants (boxed region) were harvested. a b Figure 2 Fine epitope mapping of DIII neutralizing and non-neutralizing monoclonal antibodies. (a) Flow cytometry profiles for immunoreactivity by nonneutralizing (E2), weakly neutralizing (E1) and strongly neutralizing (E34) monoclonal antibodies with yeast expressing wild-type and variant DIII. The red, yellow and green arrows point to mutations that abolish yeast surface binding of individual monoclonal antibodies and correspond to distinct regions of DIII (see b). (b) Mapping of neutralizing and non-neutralizing monoclonal antibodies on the surface of WNV DIII. A molecular surface representation is depicted based on the nuclear magnetic resonance structure of WNV DIII 16 . The indicated amino acid residues associated with binding of neutralizing, weakly neutralizing and non-neutralizing monoclonal antibodies are shown in red, yellow and green, respectively. To map our strongly neutralizing monoclonal antibodies, we developed a strategy using yeast surface display 25 . The ectodomain (amino acids 1-415) or DIII (amino acids 296-415) of WNV E protein were expressed as fusion proteins on the yeast cell surface (Fig. 1a) . Monoclonal antibodies that recognize DIII alone are considered DIII specific. Monoclonal antibodies that recognized the E ectodomain but not DIII alone may contact residues that map to domain I or II or both, although cooperative contacts with DIII cannot be ruled out. Most of our 46 monoclonal antibodies recognized yeast that displayed the entire ectodomain of E ( Fig. 1a and Supplementary Table 1 online) . Sixteen antibodies recognized yeast that displayed DIII alone, and 10 of 12 strongly neutralizing antibodies localized to DIII. Only two neutralizing antibodies (E53 and E60) recognized the E ectodomain but not DIII alone. We used error-prone PCR mutagenesis of DIII of WNV E protein and yeast surface expression to map antibody contact residues in a highthroughput manner. We performed individual screens to identify DIII mutants that lost binding selectively to strongly neutralizing (E16, E24 and E34), weakly neutralizing (E1), and non-neutralizing (E2 and E22) monoclonal antibodies. To eliminate mutants that abolished surface expression of DIII, yeast were stained sequentially with an Alexa Fluor 647-conjugated individual monoclonal antibodies and an Alexa Fluor 488-conjugated oligoclonal antibody derived from a pool of individual monoclonal antibodies. After cell sorting, we identified yeast that selectively lost expression of an individual monoclonal antibody epitope but retained surface expression of DIII (Fig. 1b) . Multiple independent yeast that lost binding of individual monoclonal antibodies were subjected to plasmid recovery and sequencing. Monoclonal antibodies that localized to DIII and strongly neutralized WNV had reduced binding when residues S306, K307, T330 or T332 were altered ( Fig. 2a and Table 1 ). These are located on adjacent loops and form a contiguous patch 16 on the solvent-exposed surface at the lateral tip of the DIII (Fig. 2b) . Only two other mutations caused considerable loss of binding of any of the ten neutralizing monoclonal antibodies tested: K310E or P315R reduced binding only of E49. No two neutralizing monoclonal antibodies had identical loss-of-binding patterns. For example, S306L reduced binding of E16, E27, E40, E43 and E49 but not E24, E33, E34, E47 and E58. K307R abolished binding of E34, E40, E43, E47, E49 and E58 but affected E16, E24, E27 and E33 less strongly. In contrast, K307E decreased binding of all neutralizing monoclonal antibodies yet did not affect non-neutralizing or poorly neutralizing monoclonal antibodies. Changes in residues T330 and T332 also abolished binding of neutralizing but not non-neutralizing monoclonal antibodies. T330I or T332M strongly reduced binding of all neutralizing monoclonal antibodies with the exception of E27, and T332V or T332A weakened binding of only E24, E27, E33 and E58. Six monoclonal antibodies that recognized DIII were either poorly neutralizing or non-neutralizing, and none engaged the dominant neutralizing epitope defined by S306, K307, T330 or T332. E2 and E9 were abolished or reduced by mutation of D381 and H396, and binding of E22 was weakened by a change in P315. D381 and H396 are proximal to one another but physically distinct from the four residues that affect binding of neutralizing monoclonal antibodies (Fig. 2b) . None of the mutations identified by loss-of-binding sorts for E2 or E22 had any effect on two other non-neutralizing monoclonal Individual WNV-specific monoclonal antibodies (25 µg/ml) were mixed with yeast that displayed wild-type or mutant DIII on their surface. After washing, and incubation with an Alexa-647 goat-anti mouse IgG secondary antibody, yeast were analyzed for antibody binding by flow cytometry. The two values represent the percentage of yeast that were positive for DIII expression with a given monoclonal antibody, and in parentheses, the mean linear fluorescence intensity of the positive cells. Yeast were analyzed at 24 h after induction with galactose, which gives a baseline surface expression of DIII of 60-80% positive cells (Fig. 1a) . Bold values indicate an almost complete (>80%) loss of binding, whereas underlined values show a marked (50-79%) reduction in either the percentage or the mean fluorescence intensity of the positive cells. Results are representative of at least two independent experiments for each antibody and DIII variant. antibodies, E21 and E23. E1, a monoclonal antibody with weak neutralizing activity, mapped to a site between the non-neutralizing and neutralizing monoclonal antibodies, as mutation of K310 and N394 strongly inhibited binding. To determine whether human antibodies specific for WNV recognize the neutralizing epitope on DIII during infection, plasma was obtained from WNV-positive individuals. Samples from convalescent individuals were negative for WNV RNA but positive for neutralizing WNV-specific antibodies. The individuals reported mild systemic illness, though none progressed to severe disease. To determine whether these samples contained antibodies that localized to the neutralizing epitope on DIII, we tested whether E16 Fab or IgG could compete binding to recombinant, wildtype and mutant N394K and K307N forms of DIII ( Supplementary Fig. 2 online); N394K retains wild-type binding to E16, whereas K307N has markedly reduced binding. E16 equivalently inhibited binding of patient WNV-specific antibodies to wild-type (Fab, 35% ± 8; IgG, 40% ± 12) or N394K DIII (Fab, 32% ± 9; IgG, 34% ± 11), whereas E53 IgG, a WNV-specific monoclonal antibody that recognizes an epitope outside DIII, did not compete binding to wild-type (1% ± 4) or N394K (-2% ± 5) DIII. As expected, E16, which only weakly recognizes K307N, poorly competed (Fab 6% ± 3; IgG 9% ± 4) binding to K307N DIII. These data suggest that humans, who clear WNV infection, develop antibodies that recognize an epitope in close proximity to that defined by E16. To evaluate the correlation between neutralization, epitope localization and in vivo protection, we assessed the therapeutic activity of different neutralizing monoclonal antibodies in an established mouse model 5 . Studies were performed with 5-week-old wild-type C57BL/6 mice, which have a ∼10% survival rate 5 . Mice were inoculated subcutaneously with 10 2 plaque-forming units (PFU) of WNV and administered a single dose of monoclonal antibody at day 2 after infection. Notably, 500 µg of the non-neutralizing monoclonal antibody E2 provided no protection (data not shown). In contrast, 100 µg of any of three different neutralizing monoclonal antibodies that map to K307 (E16, E24 or E34) protected greater than 90% of mice from lethal infection ( Fig. 3a-c) . Even a single 4 µg treatment of E16 or E34 on day 2 after infection prevented mortality. Given that humans can present with WNV infection of the CNS, we evaluated the therapeutic efficacy of monoclonal antibodies at later , or (c) E34 monoclonal antibodies. As controls, mice were independently administered saline (PBS) or a negative control monoclonal antibody (anti-SARS ORF7a, 500 µg). The survival curves were constructed using data from two independent experiments. The number of animals for each antibody dose ranged from 20 to 30. The difference in survival curves was statistically significant for all WNV-specific monoclonal antibody doses shown (P < 0.0001). (d) WNV burden in the brain of 5-week-old wild-type mice. At days 4, 5 and 6 after WNV infection, brains were harvested and viral burdens were determined by plaque assay. The following percentage of mice had viral burdens below detection (<20 PFU/g): day 4, 33%; day 5, 22%; day 6, 17%. (e,f) Efficacy of WNV-specific monoclonal antibody therapy at days 4 (e) and 5 (f) after infection. A single dose (0.5 mg at day 4 or 2 mg at day 5) of monoclonal antibody (E16, E24, E34 or anti-SARS ORF7a) was administered either 4 or 5 d after WNV infection. Data reflect approximately 20 mice per condition. The difference in survival curves was statistically significant for all WNV-specific monoclonal antibody doses shown at day 4 (P < 0.0001) and day 5 (E16, P = 0.0009; E24, P = 0.027). (g) Effect of E16 therapy on viral burden. Mice were treated with 2 mg of E16 or PBS on day 5 after WNV infection. On day 9, brains were recovered, homogenized and subjected to plaque assay. For a subset (6) that received PBS treatment, brains were harvested at days 7 and 8 from moribund mice. The data is expressed as PFU/g. Of 16 mice treated with E16 68% (11) had undetectable viral loads in the brain at day 9 whereas all mice (14 of 14) treated with PBS had detectable virus at the time of harvest. The dotted line represents the limit of sensitivity of the assay and the dark bars represent the mean of the log values. The differences were statistically different (P < 0.0001). time points. At days 4, 5 and 6 after WNV infection, we detected virus in the brains of 67%, 78% and 83% of mice, respectively (Fig. 3d) . A single 500 µg dose of E16 or E34 at day 4 resulted in an 80-90% survival rate (Fig. 3e) . A single 2 mg dose of E16 at day 5 resulted in 90% survival (Fig. 3f) and complete clearance of WNV from the brain in 68% of mice by day 9 (Fig. 3g) . Thus, administration of neutralizing monoclonal antibodies to mice with active CNS infection improved survival and induced a virologic cure. As expected, lengthening the interval before treatment was associated with decreased benefit. Administration of E16 at day 6 did not enhance survival (data not shown), although average survival time was increased (9.5 d ± 0.4 to 11.5 d ± 0.4; P = 0.003). We considered humanizing E16 or E34 as a possible therapeutic measure. Humanized monoclonal antibodies have substantially longer half-lives in humans than their mouse counterparts 26, 27 . Sequencing studies indicated that E16 had greater homology to human framework regions, making it simpler to construct a humanized version of E16. We amplified the cDNA encoding the heavy (VH) and light (VL) variable domains from the hybridoma cellular RNA by a 5′ rapid amplification of cDNA ends (RACE) procedure. The VH belongs to mouse heavy chain subgroup II (J558 family) and the VL belongs to mouse κ chain subgroup V. The complementarity-determining regions of E16 were grafted onto the human VH1-18 backbone (Fig. 4a) and human Vκ-B3 backbone (Fig. 4b) to create Hm-E16.1. One (VL-Y49S) and two framework back-mutations (VH-T71A and VL-Y49S) were introduced to create two variants, Hm-E16.1 and Hm-E16.3, respectively. The resulting humanized VH and VL were combined with human γ1 and κ constant regions, fused to an IgG signal sequence and inserted into expression plasmids. To construct the chimerized antibodies, the mouse VH and VL sequences were combined with the human γ1 and κ constant regions. We expressed humanized (Hm-E16) and chimerized (Ch-E16) E16 in HEK-293 cells, and purified them from supernatants by affinity and size-exclusion chromatography (data not shown). The affinity was analyzed by surface plasmon resonance using purified antibody in the solid phase. Mouse E16 binds DIII with an affinity of 3.4 nM and a half-life of 3.9 min. The affinity of the Ch-E16 and Hm-E16 was similar with K D ranging from 7.1 to 21 nM (Fig. 4c) . Hm-E16, Ch-E16, and the parent E16 all had similar PRNT 50 values (Fig. 4d) . We hypothesized that E16 could also control infection in mice through effector functions including antibody-dependent complement fixation and cytotoxicity. To test this, we generated a Ch-E16 N297Q aglycosyl variant that neutralizes WNV (Fig. 4d) but does not efficiently fix complement or bind Fc γ receptors 28 . Mice were administered Ch-E16 or Ch-E16 N297Q at day 2 after WNV infection. Although high doses of Ch-E16 and Ch-E16 N297Q provided virtually complete protection, lower doses of the aglycosyl variant afforded less protection (Fig. 4e) . Administration of 4 µg of Ch-E16 resulted in 84% survival, whereas 4 µg of Ch-E16 N297Q provided only 31% protection. To test which effector function enhanced the activity of E16, we performed studies with 8-week-old C1qa, C4, or Fcgr1 and Fcgr3-deficient C57BL/6 mice (Fig. 5) . These mice all show increased susceptibility to lethal WNV infection compared to wild-type controls. In C1qa or C4-deficient mice, which cannot activate complement by the antibody-dependent classical pathway, E16 had a similar potency compared to wild-type mice. In contrast, in Fcgr1-and Fcgr3-deficient mice, although high doses afforded complete protection, lower doses resulted in higher mortality rates. A dose of 20 µg at day 2, which strongly protected wild-type, C1qa or C4-deficient mice, did not improve the survival rate of Fcgr1-and Fcgr3-deficient mice (P = 0.4). Thus, the Fc region enhances the potency of E16 in mice, by virtue of its ability to bind to Fc γ receptors. To confirm the efficacy of humanized E16, wild-type mice were administered three different versions of purified Hm-E16 at day 2 after infection. Although Hm-E16 variants protected mice against lethal infection (Fig. 4f) , at a dose of 4 µg, the variant (16.3) that showed the highest affinity for DIII (Fig. 4c) was more protective (67% versus 46% survival; mean survival time of 14 ± 1 d versus 11 ± 2 d, P = 0.04). We generated a panel of 46 monoclonal antibodies against WNV E protein, and applied a new high-throughput epitope-mapping strategy to identify a dominant epitope that was recognized by the majority of neutralizing monoclonal antibodies in DIII. This epitope was also detected by convalescent antibodies from individuals who had recovered from WNV infection without clinical consequence. Three neutralizing monoclonal antibodies protected against WNV mortality in a postexposure therapy model. One of these, E16, was humanized and confirmed as therapeutically effective in mice. Previous studies have mapped amino acid contact residues of neutralizing monoclonal antibodies by sequencing in vitro neutralization escape variants, through site-specific substitution of specific charged or polar residues, and by performing binding assays with overlapping peptide libraries. We used error-prone PCR mutagenesis and yeast surface expression to identify contact residues in a high-throughput manner. Although this technique has been used previously 29 , this is the first highthroughput epitope-mapping application. By having a large panel of DIII monoclonal antibodies and selecting only variants that abolished or markedly reduced binding of a few monoclonal antibodies, we minimized the possibility that mutations altered protein folding. We have recently confirmed the recognition sites on DIII for E16 and the validity of the yeast display strategy by solving the crystal structure of the E16 Fab-DIII complex (G.N., T.O., M.D, & D.F., unpublished data). Of 12 neutralizing monoclonal antibodies, 10 localized to the distal lateral surface of DIII, results that are consistent with prior studies that mapped three neutralizing monoclonal antibodies against WNV using in vitro escape variants 15, 30 . Our weakly neutralizing and non-neutralizing monoclonal antibodies did not recognize this epitope but localized to distinct regions. E16 recognized the dominant epitope and neutralized all strains that were tested. Sequence analysis of 124 WNV strains in public databases showed almost complete (98.4-100%) conservation of the contact residues S306, K307, T330 and T332. Only two clinically attenuated lineage II isolates had mutations at these residues. Because of the structural homology among flaviviruses 7, 8, 16 , the analogous amino acids that map to the distal lateral surface of DIII can be readily identified. Although this region is highly variable among flaviviruses, most mutations that abolish binding of virus-specific neutralizing antibodies map here 15, 18, 30, 31 , suggesting the existence of an analogous dominant neutralizing epitope for other flaviviruses. We speculate that successful vaccines against WNV or other flaviviruses should induce potent humoral responses against this neutralizing epitope. Although passive administration of immune human γ-globulin after WNV infection improved survival in mice 4,5 , it may be limited by its low-titer neutralizing activity, variability and risk of transmission of infectious agents. Only two prior studies have shown postexposure therapy of neutralizing monoclonal antibodies with flaviviruses: 6B5A-2 reduced mortality 3-4 d after infection with St. Louis encephalitis virus 6 ; and 503 reduced mortality resulting from infection with Japanese encephalitis virus 5 d 32 . Here, we show that three different neutralizing monoclonal antibodies improved survival even when administered 4 and 5 d after WNV infection. Moreover, therapy with E16 at day 5 completely cleared WNV from the brain at day 9 in 68% of mice. Thus, inhibitory WNV monoclonal antibodies improve clinical and virologic outcome even after viral spread through the CNS, results that agree with studies showing that antibody can mediate viral clearance from infected neurons 33, 34 . Our experiments are consistent with a model in which the therapeutic efficacy of monoclonal antibodies is determined by properties in addition to neutralization: (i) the monoclonal antibody (E24) with the strongest neutralizing activity in vitro did not have the greatest efficacy in vivo; (ii) an aglycosyl version of E16 that lacked the ability to fix complement or bind to Fc γ receptors had equivalent neutralizing but reduced therapeutic activity; (iii) E16 was less potent in mice that lacked Fc γ receptors. E16 was humanized as a possible therapeutic measure for humans. Hm-E16 bound DIII with similar affinity and showed efficacy as postexposure therapy. Moreover, it may be possible to improve Hm-E16 by a b c introducing mutations that enhance affinity, creating forms of E16 that more readily cross the blood-brain barrier, and combining monoclonal antibodies that neutralize WNV infection through independent mechanisms. Our results are the first successful demonstration of a humanized monoclonal antibody as postexposure therapy against a viral disease, and suggest that antibody-based therapeutics may have more broad utility than previously appreciated, especially in the treatment of CNS infections in which an effective antibody response is important for limiting virus dissemination and injury to neurons. We cultured BHK-21, Vero and C6/36 Aedes albopictus cells as previously described 35 24, 36) . We also performed neutralization experiments with prototype strains of St. Louis (59268 (Parton)) and Japanese encephalitis (Nakayama) viruses 37 . For in vivo experiments, viruses were diluted and injected into mice as described 23 . Purified WNV E protein expression. WNV E protein ectodomain was generated using a baculovirus expression system according to previously described methods for related flaviviruses 38 . The last 45 nucleotides of prM (endogenous signal sequence) and the first 1,290 nucleotides of WNV E protein from the New York 1999 strain 39 were fused downstream of the polyhedrin promoter and upstream of a histidine repeat in a baculovirus shuttle vector (pFastBac, Invitrogen) by PCR using a high-fidelity Taq polymerase (Platinum Taq, Invitrogen). Three days after baculovirus infection of Hi-5 insect cells at a multiplicity of infection (MOI) of 1, supernatants were harvested, filtered, buffer-exchanged and purified by nickelaffinity chromatography according to the manufacturer's instructions. The purified WNV E ectodomain lacks the C-terminal 71 amino acids that are associated with the membrane proximal, transmembrane and cytoplasmic domains. Purified WNV DIII. The construction, expression, purification and refolding of DIII of WNV E protein is described in greater detail elsewhere (G.N., T.O., M.D. & D.F., unpublished data). Briefly, wild-type, N394K and K307N DIII were generated from an infectious cDNA clone of the New York 1999 strain of WNV (gift of R. Kinney, Centers for Disease Control and Prevention, Fort Collins, CO) using PCR and Quik-change mutagenesis (Stratagene). After cloning into a PET21 vector (Novagen) and sequence confirmation of the mutations, we transformed plasmids into BL21 Codon Plus E. coli cells (Stratagene). Bacteria were grown in Luria broth, induced with 0.5 mM isopropyl thiogalactoside (IPTG) and pelleted. Subsequently, we lysed bacteria after the addition of lysozyme, sonicated them and recovered DIII as insoluble aggregate from the inclusion bodies. DIII was denatured in the presence of guanidine hydrochloride and β-mercaptoethanol and refolded by slowly diluting out the denaturing reagents in the presence of L-arginine, EDTA, PMSF, reduced glutathione and oxidized glutathione. We separated refolded DIII from aggregates on a Superdex 75 16/60 size-exclusion column (Amersham Bioscience) and concentrated it using a centricon-10 spin column into 20 mM Hepes pH 7.4, 150 mM NaCl and 0.01% NaN 3 . After refolding, wild-type DIII reacted with all domain III-specific monoclonal antibodies including those that recognized conformationally sensitive epitopes. Generation and purification of monoclonal antibodies. BALB/c mice were primed and boosted at 3-week intervals with insect cell-generated, purified, recombinant WNV E protein (25 µg) that was complexed with adjuvant (RIBI Immunochemical). Approximately 1 month after the last boost, we harvested serum and tested it for immunoreactivity against solid-phase purified E. Mice with high titers (>1/10,000) were boosted intravenously with purified E protein (5 µg) in PBS. We harvested splenocytes 3 d later and fused them to P3X63Ag8.653 myeloma cells to generate hybridomas according to published procedures 40 . We purified monoclonal antibodies against WNV or other control antigens by stan-dard protein A or protein G chromatography according to the manufacturer's instructions (Pharmacia). For WNV infection experiments, all wild-type C57BL/6J mice were purchased from a commercial source (Jackson Laboratories). We obtained the congenic C1qa-deficient and C4-deficient mice from G. Stahl (Beth Israel Deaconess Medical Center, Boston, MA) and M. Carroll (The CBR Institute for Biomedical Research, Harvard Medical School, Boston, MA), respectively. We obtained the congenic Fcgr1-and Fcgr3-deficient mice commercially (Taconic). We used mice between 5 and 8 weeks of age depending on the particular experiment and inoculated them subcutaneously with WNV by footpad injection after anesthetization with xylazine and ketamine. Mouse experiments were approved and performed according to the guidelines of the Washington University School of Medicine Animal Safety Committee. For passive-transfer experiments, we administered to mice a single dose of purified monoclonal antibody by intraperitoneal injection at a given time point (day 2, 4 or 5) after infection. To analyze virus production in the brain, infected mice were killed on a given day after inoculation. After cardiac perfusion with PBS, we removed the brains, weighed and homogenized them, and performed plaque assays as previously described 23 . Expression of WNV E protein on yeast. The ectodomain or DIII of WNV E protein was expressed on the surface of yeast using a modification of a previously described protocol for surface expression of T cell receptors 29 . Amino acid residues 1-415 (ectodomain) or 296-415 (DIII) of WNV E protein were amplified with BamHI and XhoI sites at their 5´ and 3´ ends, respectively, by PCR from the New York 1999 infectious cDNA clone (R. Kinney, Centers for Disease Control and Prevention, Fort Collins, CO). The resulting products were digested with BamHI and XhoI, and cloned as downstream fusions to the yeast Aga2 and Xpress epitope tag genes in the yeast surface display vector pYD1 (Invitrogen). An upstream GAL1 promoter controls fusion protein expression. These constructs were transfected into the S. cerevesiae yeast strain EBY100 (refs. 25,41) resulting in yeast that expressed the WNV E ectodomain or DIII. Yeast that only expressed the Xpress epitope tag linked to Aga2 were prepared in parallel by transfecting EBY100 cells with the parent vector pYD1. Individual yeast colonies were grown to log phase overnight in tryptophan-free media containing 2% glucose at 30 °C and harvested in log phase. Fusion protein expression was induced on the yeast surface by growing yeast for an additional 24 h in tryptophan-free media containing 2% galactose at 25 °C. We harvested yeast, washed them with PBS supplemented with 1 mg/ml BSA and immunostained them with 50 µl of monoclonal antibody (25 µg/ml) against the Xpress tag or WNV E protein. After 30 min, we washed yeast three times and stained them with a goat anti-mouse secondary antibody conjugated to Alexa Fluor 647 (Molecular Probes). Subsequently, the yeast cells were analyzed on a Becton Dickinson FACSCaliber flow cytometer. Library construction and screening. We mutated DIII of the WNV E protein using an error-prone PCR protocol 25 that included Mn 2+ and Mg 2+ at concentrations of 0.3 and 2.0 mM respectively. Subsequently, the cDNA library was ligated into pYD1 and transformed into XL2-blue ultracompetent cells (Strategene). The colonies were pooled and the plasmid DNA was recovered using the Qiagen HiSpeed Maxi kit. For each individual antibody, we screened the yeast library of DIII mutants according to the following protocol. To identify yeast that selectively lost binding to a given monoclonal antibody epitope, the library was initially stained with an Alexa Fluor 647-conjugated WNV-specific monoclonal antibody for 30 min at 4 °C. To control for the surface expression of DIII, after washing, yeast were subsequently stained for 30 min at 4 °C with an Alexa Fluor 488-conjugated oligoclonal antibody that was derived from a pool of individual monoclonal antibodies (E1, E2, E9, E16, E24 and E34). After immunostaining, we subjected yeast to flow cytometry and identified the population that was single monoclonal antibody negative but pooled oligoclonal antibody positive. The yeast cells were sorted at an event rate of ∼4,000 cells/s and this population (monoclonal antibody-negative and oligoclonal antibody-positive) was enriched after three rounds of sorting. After the final enrichment sort, we plated yeast and selected individual colonies and tested them for binding to individual monoclonal antibodies. For individual clones that had lost only the desired monoclonal antibody epitope, the DIII-pYD1 plasmid was recovered using the Zymoprep Yeast Miniprep kit (Zymo Research). The plasmid was then transformed into DH5α cells, purified using the Qiaprep Spin Miniprep kit (Qiagen) and sequenced. In some cases, DIII variants with two independent mutations were isolated. To determine which mutation conferred the loss-of-binding phenotype, single independent mutations were engineered by site-directed mutagenesis of DIII-pYD1 using mutant oligonucleotides and the Quik Change II mutagenesis kit (Strategene). All mutations were confirmed by sequencing. We determined the titer of neutralizing antibodies by a standard plaque reduction neutralization titer (PRNT) assay using either BHK21 or SW13 cells 23 . Results were plotted and the titers for 50% (PRNT 50 ) and 90% inhibition (PRNT 90 ) were calculated. The inhibition assay with J774.2 mouse macrophages was performed as follows: we mixed medium and E16 or E24 (2.5 µg of monoclonal antibody) with 5 × 10 2 PFU of WNV, incubated the mixture for 1 h at 4 °C, and then added to 5 × 10 4 J774.2 mouse macrophages in individual wells of a 24-well plate. After 1 h, cells were washed four times with PBS to remove free virus and monoclonal antibody, DMEM with 10% FBS was added, and the cells were incubated for an additional 24 h. We subsequently harvested supernatants for a viral plaque assay on Vero cells. Competition ELISA with human anti-WNV antibodies. After purification and refolding, wild-type, K307N and N394K DIII were diluted (5 µg/ml) in 0.1 M Na carbonate buffer (pH 9.3) and adsorbed to 96-well plates overnight at 4 °C. After blocking with PBS, 2% BSA and 0.05% Tween 20 (PBS-BT), wells were preincubated for 1 h at 23 °C with PBS-BT containing no antibody, E16 IgG (50 µg/ml), E16 Fab (50 µg/ml) or E53 IgG (50 µg/ml). E53 serves as a negative control as it recognizes an epitope in domain I and II of WNV E protein. Subsequently, human plasma (1/40 dilution in PBS-BT, heat-inactivated) was directly added for an additional 1 h at 23 °C. We obtained the human samples with informed consent from seven different WNV-infected patients (gift of M. Busch and L. Tobler, Blood Systems Research Institute, San Francisco, CA). Because the samples were sequentially numbered and not linkable back to the original subjects, they satisfied the criteria for exemption from approval from the Human Studies Committee at Washington University. After six washes with PBS-BT, plates were serially incubated with biotin-conjugated goat anti-human IgG (1 µg/ml), streptavidin-horseradish peroxidase (2 µg/ml) and tetramethylbenzidine developing substrate (DAKO). We determined optical densities at 450 nm with an automatic ELISA plate reader (Tecan) and adjusted them after subtraction of the value obtained from nonimmune human plasma. Surface plasmon resonance. Antibody affinity analysis for DIII of WNV was performed by surface plasmon resonance (BIAcore 3000, Biacore, Inc). Binding curves and kinetic parameters were obtained as follows: we captured E16 antibodies by flowing (300 nM, rate of 5 µl/min for 2 min) them over immobilized F(ab)'2 fragment specific for goat anti-human or mouse IgG with Fc region specificity. Subsequently, DIII of the New York 1999 strain of WNV E protein (amino acids 296-415), which was generated in E. coli, was injected (6.25-200 nM, flow rate 70 µl/min for 1.5 min and then allowed to dissociate over 5 min). The F(ab)'2 surface was regenerated by pulse injection of 10 mM glycine (pH 1.5) and 100 mM NaOH before each E16 injection. We analyzed curves with a global fit 1:1 binding algorithm with drifting baseline. Cloning and humanization of E16. E16 heavy-and light-chain RNA was isolated from hybridoma cells after guanidinium thiocyanate and phenol-chloroform extraction, and converted to cDNA by reverse transcription. The VH and VL segments were amplified by PCR using the 5′ rapid amplification of cDNA ends (RACE) system (Invitrogen). Gene-specific primers (GSP) for VH and VL were as follows: VH-GSP1: 5´-GGTCACTGTCACTGGCTCAGGG-3´; VH-GSP2: 5´-AGGCGGATCCAGGGGCCAGTGGATAGAC-3´; VL-GSP1: 5´-GCACACG ACTGAGGCACCTCCAGATG-3´; and VL-GSP2: 5´ CGGATCCGATGGATAC AGTTGGTGCAGCATC-3´. The PCR products were inserted into the plasmid pCR2.1-TOPO using the TopoTA kit (Invitrogen). We then subjected the resulting plasmids to DNA sequencing to determine the VH and VL sequences for E16. The cDNA sequences were translated and the predicted amino acid sequence determined. From these sequences the framework and CDR regions were identified as previously defined 42 . We joined the mouse VH to a human C-γ1 constant region and an Ig leader sequence, and inserted it into pCI-neo for mammalian expression. We joined the mouse VL to a human Cκ segment and an Ig leader sequence and also cloned it into pCI-neo for mammalian expression of chimeric E16 (Ch-E16). For Ch-E16, site-directed mutagenesis was also performed to change residue 297 from asparagine to glutamine of the heavy chain to eliminate the single glycosylation site on the γ1 Fc. Humanized E16 VH consists of the framework segments from the human germline VH1-18 VH segment and JH6 segment 43, 44 , and the CDR regions of the E16 VH, respectively. The humanized E16 VL consists of the framework segments of the human germline VK-B3 VL segment and JK2, (refs. 45-47) segment and the CDR regions of E16 VL. The humanized VH segments were assembled de novo from oligonucleotides and amplified by PCR. The humanized VL segments were assembled by PCR and overlapping PCR. The resulting VH and VL segments were subsequently combined by overlapping PCR with a leader sequence and the appropriate constant region segment and cloned into the expression vector pCI-neo as NheI-EcoRI fragments. We confirmed the DNA sequence of the resulting plasmids by sequence analysis. Site-directed mutagenesis was then performed to substitute mouse for human residues at key framework positions VH-71 (T71A) and VL-49 (Y49S). The resulting plasmids were cotransfected into human 293 cells using lipofectamine-2000 and humanized antibody was recovered from the resulting conditioned medium and purified by protein A and size-exclusion chromatography. Statistical analysis. All data were analyzed with Prism software (GraphPad Software). For survival analysis, Kaplan-Meier survival curves were analyzed by the log-rank and Mantel-Haenszel test. For viral burden experiments, we determined statistical significance using the Mann-Whitney test. Note: Supplementary information is available on the Nature Medicine website.
46
Local public health workers' perceptions toward responding to an influenza pandemic
BACKGROUND: Current national preparedness plans require local health departments to play an integral role in responding to an influenza pandemic, a major public health threat that the World Health Organization has described as "inevitable and possibly imminent". To understand local public health workers' perceptions toward pandemic influenza response, we surveyed 308 employees at three health departments in Maryland from March – July 2005, on factors that may influence their ability and willingness to report to duty in such an event. RESULTS: The data suggest that nearly half of the local health department workers are likely not to report to duty during a pandemic. The stated likelihood of reporting to duty was significantly greater for clinical (Multivariate OR: 2.5; CI 1.3–4.7) than technical and support staff, and perception of the importance of one's role in the agency's overall response was the single most influential factor associated with willingness to report (Multivariate OR: 9.5; CI 4.6–19.9). CONCLUSION: The perceived risk among public health workers was shown to be associated with several factors peripheral to the actual hazard of this event. These risk perception modifiers and the knowledge gaps identified serve as barriers to pandemic influenza response and must be specifically addressed to enable effective local public health response to this significant threat.
Local health departments are considered the backbone of public health response plans for any and all infectious disease outbreaks. An influenza pandemic is considered increasingly likely, and is now considered one of the most significant and urgent threats to the nation's public health preparedness infrastructure. It has been argued that of the 12 disaster scenarios recently assessed by the U.S. Department of Homeland Security, pandemic influenza is the most likely and perhaps the most deadly [1] . The United States pandemic influenza plan released in November 2005, lays out a critical role for local and state public health agencies during a pandemic, including: providing regular situational updates for providers; providing guidance on infection control measures for healthcare and non-healthcare settings; conducting or facilitating testing and investigation of pandemic influenza cases; and investigating and reporting special pandemic situations [2] . These specified activities would require an extensive prompt response by local health departments. Current contingency plans account for possible personnel shortages due to influenza morbidity, but previous studies have shown that during extreme scenarios, a varying proportion of healthcare workers may be unable or unwilling to report to duty [3] [4] [5] . This may be even truer for health departments, where unlike more "traditional" first responder agencies (such as law enforcement, fire services, and emergency medical services), the capacity and willingness to respond 24/7 to crises is not historically ingrained in the workforces' professional cultures and training. Even in the post-9/11 environment, recent data indicate inconsistent and sometimes slow after-hours response by health departments to urgent events involving communicable disease [6] . Risk perception theory provides a revealing framework for better understanding response limitations and needs of the public health workforce. The perceived risk, according to this theory, is a multifactorial phenomenon, involving the summation of actual risk and other peripheral influences independent of the actual risk, such as perceived authority, trust, and situational control; these peripheral influences have been termed "outrage" [7] or "dread." [8] . Based on these models, it was previously suggested that contributing factors peripheral to the actual risk will have a considerable practical impact on how public health employees would respond in a crisis [9] . Aside from physical and circumstantial barriers such as availability of transportation or dependency of family members, we have identified specific risk perception issues whose impact may be markedly high and of unique importance for the public health workforce's response to a crisis. These factors, or modifiers, stem from a number of features previously suggested to have been associated with elevated risk perception, including manageability of the threat; risk to future generations; direct personal impact; and sense of control over events. Based on these modifiers, several major barriers to effective public health workforce emergency response were suggested; these include uncertainty regarding working environment safety, unclear expectations of role-specific emergency response requirements, safety and well being of family members, inadequate emphasis on the critical value of each employee to the agency response efforts, and insufficient emphasis on stress management techniquesall of which may heighten employees' sense of dread due to a lack of personal control [9] . In light of the projected impact of an influenza pandemic, health departments must optimize the response rate of their employees in this crisis scenario. Based on the emer-gency response principle that all disasters are "local" [10] , we have set out to assess local public health employees' risk perception and likelihood of reporting to duty during a local outbreak of pandemic influenza, and to uncover the variables that affect these outcomes, thus providing a needed evidence base for health departments' planning and training efforts. We conducted the study in Carroll, Dorchester, and Harford county health departments between March 2005 and July 2005. All three health departments are located in Maryland, and range in size from 132 employees to 225 employees. We selected these health departments because of their location in communities ranging from 30,000 on Maryland's Eastern Shore (Dorchester County) to 235,000 in the greater Baltimore/Towson metropolitan area (Harford County) [11] , thus reflecting the 96% of the nation's local public health agencies serving communities with populations of 500,000 or fewer [12] . Self-administered anonymous surveys were sent to all health department personnel by their respective health departments. Completed surveys were directly mailed to investigators at the Johns Hopkins Center for Public Health Preparedness. The survey included questions on personal characteristics such as job classification, gender, and age. The respondents used a 5-point Likert scale for questions pertaining to a possible flu pandemic: probability of them reporting to work ("very likely" to "not at all likely"); possibility of being asked by their health department to respond to an emergency ("very likely" to "not at all likely"); how knowledgeable they thought they were about the potential public health impact of pandemic influenza ("very knowledgeable" to "not at all knowledgeable"); how confident they were about being safe in their work roles ("very confident" to "not at all confident"); how likely was their family prepared to function in their absence ("very likely" to "not at all likely"); how likely they felt their health department would provide them with timely updates ("very likely" to "not at all likely"); how familiar they were with their role specific response requirements ("very familiar" to "not at all familiar"); how well they thought they could address the questions of a concerned member of the public ("very well" to "not at all"); how significant a role they thought they would play in the agency's overall response ("very significant" to "not at all significant"); how important would be pre-event preparation and training ("very important" to "not at all important"); how important it was for them to have psychological support available during the event ("very important" to "not at all important"); and how important it was for them to have psychological support available after the event ("very important" to "not at all important"). The job classification variable was collapsed into technical/support staff (such as computer entry staff, clerical staff (e.g. receptionists), computer specialists, health information systems data analysts etc.), and professional staff. The latter included public health officials, clinical staff (e.g., nurse, dentist, physician), public health communicable disease staff, environmental health staff, public information staff, and other public health professional staff (e.g., health educator, legal professional, financial officer, other). We dichotomized the responses to the job classification question into professional and technical/support categories. Questions about likelihood of reporting to work and pandemic influenza-related attitudes and beliefs were dichotomized into responses with a score two or less, and all other responses. We used logistic regression to compute Odds Ratios to evaluate the association of demographic variables and attitudes and beliefs with selfdescribed likelihood of reporting to work. We used multivariate logistic regression to explore associations between attitudes and beliefs related to pandemic influenza preparedness and self-described likelihood of reporting to work. The model included adjustment for age, gender, and job classification. Similarly, we used bivariate and multivariate (adjusted for age, gender, and job classification) logistic regression models to evaluate the association between the various attitudes and beliefs. In order to assess non-response bias, we compared age, gender, and job classification distributions for the respondents and for all health department personnel. We used TeleForm Version 8 (Cardiff, Vista, CA) and Stata Version 9 (Stata Corporation, College Station, TX) for data capturing and analysis respectively. We received 118 out of 205 (57.6%), 74 out of 128 (57.8%), and 116 out of 198 (58.6%) surveys fromCarroll, Dorchester, and Harford county health departments respectively, resulting in an overall response rate of 58.0% (n = 308). We did not find a statistically significant difference in age and gender distribution between the respondents and all health department personnel. A small yet statistically significant difference in the proportion of technical/support staff (vs. professional staff) was detected (22.4% vs. 32% in the study group and all personnel respectively, p = 0.003), yet no significant difference in the proportions of professional staff subgroups was detected. Of the 303 who responded to the question about their likelihood of reporting during a pandemic influenza related emergency, 163 (53.8%) indicated they would likely report to work during such an emergency. Age and gender did not have an association with likelihood of reporting. Clinical staff indicated a higher likelihood of reporting (Multivariate OR: 2.5; CI 1.3-4.7) than technical/support staff (Table 1) . Only 40% of all respondents-45.1% professional staff and 26.1% technical/support staff -felt it was likely they would be asked by their health department to respond to a pandemic influenza related emergency. Perception of Table 2) . Perception of one's existing knowledge about pandemic influenza, and perception of having an important role in the agency's overall response were significantly higher among professional staff compared to technical/support staff (Figure 1 ), In multivariate analysis, increased self-described likelihood of reporting to work during an influenza pandemic emergency was significantly associated with agreement with several constructs, most notably perception of the capacity to communicate risk effectively, perception of the importance of one's role in the agency's overall response, and familiarity with one's role-specific response requirements in a pandemic influenza related emergency. ( Table 2 ). The vast majority (83%) of the respondents felt they would benefit from additional training activities. A lower perceived level of familiarity with one's role was not significantly associated with a higher perceived need for additional training ( (Figure 2 ). The World Health Organization has urged all countries to prepare for the next influenza pandemic, which it termed . The federally adopted U.S. model of all-hazards emergency readiness has presented local health departments with new organizational challenges and learning curves. The all-hazards approach entails an ability and willingness to respond to a broad spectrum of disasters, ranging from the intentional (e.g., chemical, biological, or radiological terror) to the naturally occurring (e.g., weather-related crises or non-bioterrorism related infectious disease) [14] . Current national contingency plans account for possible personnel shortages within the healthcare and public health settings, mainly due to the expected influenza morbidity among workers. Yet our data suggest that regardless of the expected morbidity among personnel during an influenza pandemic, nearly half of the local health department workers are likely not to report to duty during such an extreme public health crisis. In fact, most of the workers (and nearly three out of four technical/support workers) do not believe they will even be asked to report to work. We have found that the willingness to report to duty during a pandemic varies considerably according to the individual's job classification. Clinical staff state they are significantly more likely to report to duty, compared with all other workers. This difference correlates well with the single most influential construct associated with willingness to report to duty -the perception of the importance of one's role in the agency's overall response. Less than a third of the respondents believed they will have an important role in the agency's response to local outbreaks of pandemic influenza, but within this subgroup, willingness to report to duty was as high as 86.8%. Belief in the importance of one's role was lowest among technical/support staff, environmental health staff, and other non-clinical professional staff (15.1%, 18.4% and 18.8% respectively), groups in which willingness to report was shown to be lowest. We therefore believe further efforts must be directed at ensuring that all local public health workers, but most notably non-clinical professional staff, understand in advance the importance of their role during an influenza pandemic -otherwise they will fail to show up when they are most needed. Our findings fit well in the theoretical framework emphasizing risk communication needs of public health workers, who themselves serve as risk communicators [9] . Several factors, previously suggested to be risk perception "modifiers" [9] of substantial impact on public health Proportion of individuals who agreed with each of the attitude and belief constructs by staff type Figure 1 Proportion of individuals who agreed with each of the attitude and belief constructs by staff type. Technical/Support Staff Professional Staff * * workforce's response to a crisis, indeed proved to be important in this context. Lack of knowledge, ambiguity regarding one's exact tasks, and questionable ability in performing one's role as risk communicator were all significantly associated with a higher perceived personal risk and a two-to ten-fold decrease in willingness to report to duty; these factors proved to be more influential even than the perceived level of family preparedness to function in one's absence. It is therefore important to recognize that public health employees, who are intended to serve as purveyors of risk communication for their communities, themselves represent a community with specific perceptions that must be addressed in the context of emergency readiness training. The threat of an impending influenza pandemic is not a new one -pandemics have been taking place once every several decades for over 300 years. Yet it was only in the last couple of years, as highly pathogenic H5N1 strain became increasingly endemic in southeast Asia and as lethal infections with the virus occurred in an alarmingly increasing rate among humans, that the urgency of the situation was openly declared by national and international health authorities. The rapidity of this evolving situation may serve to explain why only one third of the respondents felt they were adequately knowledgeable on pan-demic influenza, and why only one in five respondents felt capable in effectively communicating pandemic risks. This finding is especially noteworthy, in that members of the public health support staff may become frontline telephone risk communicators in a crisis, serving as the first points of interface for concerned callers contacting a health department. Only one of the 35 technical/support staff workers who felt incapable of effective risk communication was willing to report to duty, even though most of them believed the health department will have the ability to provide timely information. The study has some relevant limitations that must be factored into the overall analysis. First, the sample was limited to three non-randomly selected health departments, none of which serves a community larger than 250,000 residents, and all of which have staff sizes under 250. The sample size of 308 survey employees limited this study's power. As the study includes Maryland health departments only, it does not account for potential jurisdictional or regional variations nationwide in response capacity or risk perceptions toward pandemic influenza response. Furthermore, the job classifications -based on those used to develop the CDC-adopted emergency preparedness competencies [15] -do not necessarily map neatly onto functional responsibilities in disaster Odds Ratios of reporting to work in case of a pandemic-influenza-related emergency by staff and attitude or belief construct Figure 2 Odds Ratios of reporting to work in case of a pandemic-influenza-related emergency by staff and attitude or belief construct. Technical/Support Staff Professional Staff response. For example, health educators may play as frontline a role as clinical staff, in terms of their degree of interface with the public in a disaster. Our job categories therefore do not necessarily reflect the relative impacts of job-specific cohorts on disaster response in the event that they do not report to work. We assessed the presence and the direction of nonresponse bias by comparing the distribution of personal characteristics for the respondents and for all health department personnel. The lack of significant difference in age and gender distribution, as well as the lack of significant difference in job classification other than technical/ support staff indicates that the extent of such a bias in the study is probably limited. The small yet statistically significant over-representation of technical/support staff in our study group may potentially have caused a slight underestimation of overall willingness to report. However, as the internal associations between the various variables were also studied separately for the technical/support staff and professional staff (Figure 2) , this over-representation should not impact the general conclusions presented above. Having accounted for these limitations, it is important to note that the findings were internally consistent among the three surveyed health departments. Although none of the health departments served large metropolitan areas and all had fewer than 250 employees, it must also be recognized that only 4% of the nation's local health departments serve populations of 500,000 or more, and that local public health agencies tend to have small staff sizes (with a median of 13 full time employees) [12] . Interestingly, our findings show similar patterns to data on the willingness of urban healthcare workers from nonpublic health settings to respond to emergencies: a survey of 6248 employees from 47 healthcare facilities in the New York City area revealed that these workers were least willing (48%) to report to duty during an untreatable naturally-occurring infectious disease outbreak affecting their facility (SARS), compared to other disaster scenarios [5] . In our study we have detected similar rates of likelihood to report to duty, although lower rates could have been expected in our study population since the New York City survey focused on healthcare workers whose organizational cultures are historically much more accustomed than that of local public health workers to emergency response, in a city with a heightened awareness of disaster preparedness in the wake of the World Trade Center attacks and subsequent anthrax attacks [5] . In the face of a pandemic influenza threat, local health department employees' unwillingness to report to duty may pose a threat to the nation's emergency response infrastructure. Addressing the specific factors associated with this unwillingness is necessary to help ensure that existing local health department preparedness competencies [15] will translate into the scope of response described in the nation's pandemic influenza plans [2] . Interventions suggested to enhance the willingness of healthcare workers in non-public health department settings to report to duty in disasters include workforce preparedness education [5] , provision of appropriate personal protective equipment, [4, 14] crisis counseling, family preparedness and social support [5, 16] . These recommendations fit well within the framework of our findings, and we further recommend that such education programs include specialized training emphasizing the specific nature of, and guidelines for, one's role in response to pandemic influenza; the relevance of each worker's role in the effectiveness of an overall public health response; and the workers' ability to provide effective risk communication. Additional research must further focus on best practice models for addressing the above described gaps in local public health response to this urgent public health threat. These data offer a current, evidence-based window into the needs of public health workers who would serve as a backbone of locally-driven emergency response in an influenza pandemic setting. We found that most of these workers feel they will work under significant personal risk, in a scenario they are not adequately knowledgeable about, performing a role they are not sufficiently trained for, and believing this role does not have a significant impact on the agency's overall response. These specific perceptions and needs must be attended, and specific intervention programs must be initiated. In order to reduce the perceived risk associated with the worker's role in an influenza pandemic, each worker must have better understanding of the scenario and importance of his or her personal role within these settings, confidence that the agency will provide adequate protective equipment for its employees, psychological support and timely information, and a belief of being well-trained to cope with emergency responsibilities including the ability to communicate risk to others. In view of what is currently considered to be an impending influenza pandemic, a wide gap between these desired targets and current status exists, that may lead to significant hindrance in the ability of local health departments to function adequately.
47
On pandemics and the duty to care: whose duty? who cares?
BACKGROUND: As a number of commentators have noted, SARS exposed the vulnerabilities of our health care systems and governance structures. Health care professionals (HCPs) and hospital systems that bore the brunt of the SARS outbreak continue to struggle with the aftermath of the crisis. Indeed, HCPs – both in clinical care and in public health – were severely tested by SARS. Unprecedented demands were placed on their skills and expertise, and their personal commitment to their profession was severely tried. Many were exposed to serious risk of morbidity and mortality, as evidenced by the World Health Organization figures showing that approximately 30% of reported cases were among HCPs, some of whom died from the infection. Despite this challenge, professional codes of ethics are silent on the issue of duty to care during communicable disease outbreaks, thus providing no guidance on what is expected of HCPs or how they ought to approach their duty to care in the face of risk. DISCUSSION: In the aftermath of SARS and with the spectre of a pandemic avian influenza, it is imperative that we (re)consider the obligations of HCPs for patients with severe infectious diseases, particularly diseases that pose risks to those providing care. It is of pressing importance that organizations representing HCPs give clear indication of what standard of care is expected of their members in the event of a pandemic. In this paper, we address the issue of special obligations of HCPs during an infectious disease outbreak. We argue that there is a pressing need to clarify the rights and responsibilities of HCPs in the current context of pandemic flu preparedness, and that these rights and responsibilities ought to be codified in professional codes of ethics. Finally, we present a brief historical accounting of the treatment of the duty to care in professional health care codes of ethics. SUMMARY: An honest and critical examination of the role of HCPs during communicable disease outbreaks is needed in order to provide guidelines regarding professional rights and responsibilities, as well as ethical duties and obligations. With this paper, we hope to open the social dialogue and advance the public debate on this increasingly urgent issue.
In 2003, the world witnessed the spread of a novel and deadly virus, namely SARS CoV. The health care workers (HCWs) and hospital systems that bore the brunt of the SARS outbreak continue to struggle with the aftermath of the crisis. Indeed, HCWs -both in clinical care and in public health -were severely tested by SARS. Unprecedented demands were placed on their skills and expertise, and their personal commitment to their profession was severely tried. Many were exposed to serious risk of morbidity and mortality; indeed, approximately 30% of reported cases were among HCWs, some of whom died from the infection [1] . As a number of commentators have noted, SARS exposed the vulnerabilities of our current health care systems and governance structures [2] [3] [4] . The aftermath of SARS and the spectre of pandemic avian influenza make imperative the need to consider the obligations of HCWs for patients with severe infectious diseases, particularly diseases that pose risks to those providing care. It is of pressing importance that organizations representing HCWs -professionals and non-professionals alike -give clear indication of what standard of care is expected of their members in the event of a pandemic. Many experts believe that the SARS outbreak was merely a preview of the next flu pandemic that is soon to arrive, possibly from an avian influenza virus [5] . Quite clearly, avian flu threatens to be more widespread than SARS, with the potential to become a truly global pandemic. An honest and critical examination of the role of HCWs during such a crisis is needed in order to provide guidelines regarding professional rights and responsibilities, as well as ethical duties and obligations [6] . In this paper, we address the issue of special obligations of health care professionals (HCPs) during an infectious disease outbreak. We contend that there is a pressing need to clarify the rights and responsibilities of HCPs, especially in the current context of pandemic flu preparedness. Moreover, we argue that these rights and responsibilities would best be codified in professional codes of ethics. Finally, we present a brief historical account of the treatment of the duty to care in professional health care codes of ethics with the intention of opening the social dialogue and advancing the public debate on this increasingly relevant issue. Given that the response by HCPs to the SARS crisis was generally regarded as exemplary, one might ask whether an ethical problem truly exists. There is little doubt that the vast majority of HCPs performed their jobs admirably under considerable stress and significant personal risk. Many HCPs provided exemplary care, and still others behaved in truly heroic fashion. So why, then, formally problematize something that is not a problem? As noted, many HCPs acted in a supererogatory manner during the SARS outbreak [7], none more so than Dr. Carlo Urbani of the World Health Organization, who himself died of SARS after being exposed to the yet unknown virus in the course of carrying out his professional duties. Likewise, scores of nurses, doctors, respiratory technicians, and other professional and nonprofessional health workers laboured extremely long hours at personal risk. This demonstration of going above and beyond the call of duty, which proved necessary to control the disease, was highly morally commendable. At the same time, however, serious concerns did surface during SARS about the extent to which HCPs would tolerate risks of infection [8, 9] . Some baulked at providing care to those infected with the unknown virus. In some circumstances, staffing became an issue in SARS wards and assessment centres; indeed, failure to report for duty during the outbreak resulted in the permanent dismissal of some hospital staff. As a consequence, the risk that was faced during SARS was not distributed equitably, and those HCPs who volunteered to provide care faced the greatest exposure. Following the outbreak, many of those who treated SARS patients raised concerns about the protections that were provided to safeguard their own health and that of their family members. Conflicting obligations were another significant concern. HCPs are bound by an ethic of care, therefore, obligations to the patient's well-being should be primary. At the same time, however, HCPs have competing obligations to their families and friends, whom they feared infecting, in addition to obligations to themselves and to their own health (particularly those with special vulnerabilities, such as a co-morbid condition). During SARS, some HCPs questioned their choice of career; indeed, some decided to leave their profession and pursue new ventures, indicating an unwillingness or inability to care for patients in the face of risk. Recent survey data from the U.S. indicate that there exists mixed views on the duty to care for patients during infectious disease outbreaks [10] . What is clear is this: the issue of duty to care has emerged as a matter of paramount concern among health care professionals, hospital administrators, public policy makers, and bioethicists [11] [12] [13] [14] . The ethical foundations of the duty to provide care are grounded in several longstanding ethical principles. Foremost among these is the principle of beneficience, which recognizes and defines the special moral obligation on the part of HCPs to further the welfare of patients and to advance patients' well-being. In modern health care, it is commonly understood and generally accepted that the principle of beneficence constitutes a foundational principle of the patient-provider relationship [15] . For the HCP in general, and for the physician in particular, there are a number of compelling reasons to provide care in the context of an infectious disease outbreak. Clark [12] has recently outlined three such reasons: 1. The ability of physicians and health care professionals to provide care is greater than that of the public, thus increasing the obligation to provide care Although self-care and self-protection, as well as the care and protection of friends and family members, are acknowledged in pandemic plans, it is evident that the expertise of HCPs is an integral and principal component of the response to a pandemic. There is no other sector of society that can be legitimately expected to fulfil this role and to assume this level of risk. Arguably, HCPs have consented to greater than average risk by their very choice of profession. While it may be granted that the risk of contracting an infectious disease was likely not a concern for a generation of prospective health care workers, any informed reading of the medical literature in the last 20 years has shown that infectious diseases remain ubiquitous and problematic -notwithstanding overly-optimistic statements regarding the future threat of infectious diseases. It is therefore not unreasonable to argue that HCPs were aware of the greater than average risks posed by their choice of profession. In publicly-funded health care systems, such as those found in many Western societies, there is a strong claim for a social contract between the HCP and society. It is a reasonable and legitimate expectation by the public that HCPs will respond in an infectious disease emergency. Society has granted and permits professions to be self-regulating on the understanding that such a response would occur. One of the characteristics of a self-regulating profession is the development of standards of practice, sometimes referred to as best practice guidelines. These standards are articulated in professional codes of ethics, which are developed on the basis of the fundamental principles and values of the particular profession, as is the case, for instance, with respect to the codes of ethics that were developed long ago in medicine and nursing. Indeed, the code of ethics has a long and respected tradition in the health professions and today most, if not all, the various health and social care professions have codes of ethics in place to provide guidance to their members. The code of ethics is sometimes referred to as an instrument of "soft law," owing to its non-legislative nature [16] . As such, in the health care professions, codes of ethics should be interpreted as guides for ethical reasoning and frameworks for the treatment of individual patients, rather than as substitutes for such reasoning or as an absolute mandate [17] . At the same time, a code that is too vague can render it ineffectual and irrelevant. In an era in which health care and technology are evolving at a rapid pace, efforts are necessary to ensure that codes of ethics remain current, practical, and concordant with public expectations. An informative and comprehensible code of ethics has numerous tangible benefits. Perhaps the greatest benefit would be to dispel confusion and uncertainty for HCPs concerning their professional rights and responsibilities as regards the duty to care. Of course, a detailed treatment of the issue in professional codes of ethics would also serve to increase awareness and comfort levels, perhaps resulting in increased willingness to provide care in uncertain and risky conditions [18] . Additionally, codes guiding professional conduct may effectively serve as norms of standards recognizable and enforceable by law, acting as the foundation of legal obligations and decisions [16] . Finally, codes of ethics also serve as potent forms of symbolic communication to the public that is served by the professions. By making explicit the values that health care professions represent, professional codes of ethics can reassure the public that the trust invested in the professions is justified and legitimate, as is properly noted in the following excerpt from the College of Nurses of Ontario Practice Standard on Ethics: "Nurses have a commitment to the nursing profession. Being a member of the profession brings with it the respect and trust of the public. To continue to deserve this respect, nurses have a duty to uphold the standards of the profession.... As members of a self-regulating profession, nurses also have a commitment to help regulate nursing to protect the public's right to quality nursing services. It is in the public's interest that the profession continue to regulate itself by developing and changing the methods of self-regulation to meet the changes in health care and society. Nurses have an obligation to participate in the effective evolution of self-regulation. Self-regulation is a privilege and each nurse is accountable for the responsibilities that accompany this privilege." [19] What do current codes of ethics say regarding duty to care during epidemics? It is of no small concern that many current professional codes of ethics fail to provide explicit guidance sufficient to set policy or assure the public in the event of an infectious disease outbreak. The Canadian Medical Association, for instance, released a revised Code of Ethics in 2004, one year after the SARS pandemic in which Canada was particularly affected [20] . Despite the seemingly fortuitous timing of the publication, the revised Code is, quite astoundingly, altogether silent on physicians' duty to care, which might be described as the first among equals of the myriad ethical dilemmas that emerged during the global outbreak. The key revision in the 2004 edition of the CMA Code was the addition of the following item to the 'Fundamental Responsibilities' section: "Consider the well-being of society in matters affecting health" [20] . This addition, however, does little to address, in any substantively meaningful way, the duty to care obligations of HCPs in the context of an infectious disease outbreak. Does the addition of this responsibility obligate physicians to provide treatment even when doing so would put their own health in peril? The wording is too vague to be of any significant guidance in clinical practice. In contrast, the American Medical Association (AMA) appears to have recognized the present need to address the issue of duty to care. In the wake of the 9/11 terrorist attack, the AMA has adopted several new ethics policies that focus specifically on the medical profession's obligations and responsibilities in the context of a public health emergency. The following passage is from the AMA policy document "Physician Obligation in Disaster Preparedness and Response" that was adopted in June 2004: "National, regional, and local responses to epidemics, terrorist attacks, and other disasters require extensive involvement of physicians. Because of their commitment to care for the sick and injured, individual physicians have an obligation to provide urgent medical care during disasters. This ethical obligation holds even in the face of greater than usual risks to their own safety, health or life. The physician workforce, however, is not an unlimited resource; therefore, when participating in disaster responses, physicians should balance immediate benefits to individual patients with ability to care for patients in the future." [21] While the AMA has taken a step in the right direction by stating the obligations of its members, and it is to its great credit for initiating this process, it remains to be seen whether other national medical associations and other health care professions will follow suit and redress the silence of codes of ethics on the duty to provide care. To some extent, codes of ethics can be seen as reflections of enduring professional values. At the same time, they are also clearly influenced by and are the product of historical circumstances. The CMA, for instance, previously included in its Code of Ethics a strongly worded statement explicitly addressing the obligations of physicians in infectious disease outbreaks. The 1922 version reads as follows: "When pestilence prevails, it is their [physicians'] duty to face the danger, and to continue their labours for the alleviation of suffering, even at the jeopardy of their own lives" [22] . This is the only appearance in the CMA Code of this type of strong categorical language regarding the professional duty to care. Interestingly, the specific text cited above appears for the first time in the revision following the 1919 influenza pandemic and then, conspicuously, disappears from the next revision released in 1926. The AMA included the very same provision in its Code of Ethics from 1846 through until the 1970s when it was likewise excised. This marked professional retrenchment from a strong obligation to provide care -as reflected in current codes of ethics -is attracting increased interest of late. A number of explanations have been proposed by academic commentators. For instance, the retrenchment has been linked to the rise of government and corporate intrusions into medical practice [23] . Others, including Clark [12] , have pointed to an increasing general belief originating circa 1950 that infectious diseases had been vanquished. It is most likely the case that both these factors played a significant role in the observed retrenchment over time. Irrespective of the reasons underlying the current silence, there can be little doubt that infectious diseases are an increasing clinical reality in the developed world, and have long been a tremendous challenge in the developing world. For this reason alone, the continuing silence of codes of ethics is greatly problematic, both clinically and normatively. There is no current consensus as to how explicitly and stringently the requirements for the duty to care should be stated [14] . In a 2003 survey of 1000 American physicians, respondents reported decidely mixed views on whether they would continue to care for patients in the event of an outbreak. Given that only a narrow majority of the surveyed physicians reported believing in a professional duty to treat patients in epidemics, the authors of the study concluded that there should be a reinforcement of "the [medical] profession's ethical duty to treat" in the event of a public health crisis [10] . This call for the reinforcement of the duty to care echoes the 1922 CMA Code of Ethics that clearly stipulates that physicians have a duty to provide care, even at the jeopardy of their own lives. This statement may indeed be considered too strong and too categorical by many today. To require the provision of care even when doing so entails significant risk to the provider would appear to be demanding that all HCPs behave like "supreme Samaritans" [12] . Is this reasonable? Is this ethical? These remain open questions, but it is doubtful that all HCPs would adhere to such stringent obligations when faced with a SARS-like crisis. As Emmanuel [24] has instructively noted, the historical record of physicians is decidedly mixed in this regard; indeed, it was in response to the vocal and mounting opposition to treating seropositive patients at the height of the HIV/AIDS epidemic in the 1980s that the medical profession reconsidered the ongoing retrenchment of the duty to care. With the threat of a new epidemic, another round of full and open discourse is required as to whether the acceptable standard of professional engagement should occur at the level of "supreme", "good", or "merely decent" Samaritan [12] . There is presently a need to address the professional duties of HCPs to their patients with the risks to the well-being of society, which may include family, friends, co-workers, and other patients, in addition to the population at large. This is a matter of balance. The content of current professional codes of ethics offers little guidance or reflection of consensus in the health care community. In the wake of SARS and with the current threat of avian influenza, it is clearly time for medicine, nursing, and other self-regulating health care professions to address the issue head on. A number of options are open to the professions. One option, which we have already argued is unacceptable, would be to remain silent. On the other extreme, codes of ethics could be revised with a strong emphasis on the professional obligation and duty to provide care during infectious disease emergencies; that is, assume a position leaning towards supererogation, or performing acts that are 'above the call' of duty [7] . Several other options exist between these two extremes. For instance, the codes could reflect a strong but limited duty to care, with the limits clearly specified. Alternatively, there could be a weak emphasis on duty to care -an option more sympathetic to the self-regarding concerns of HCPs [8] -although this may run the risk of dissolution of the generally high regard for the health care professions that exists in society today. We maintain that, with respect to the duty to care, it is not acceptable for codes of ethics to be vague, ambiguous, or otherwise avoid explicit statements of position. This is particularly true in light of calls for additional protections for HCPs during infectious disease outbreaks, including a position statement to that effect issued by the Ontario Medical Association [25] . Such calls for danger pay and/or enhanced disability insurance could be justified if the professions expressed a strong commitment to the ethical duty to provide care during public health emergencies. In the absence of such commitment, however, any additional measures to protect and safeguard the well-being of HCPs would appear self-serving and misplaced. In the current context of pandemic influenza planning, as with other public health emergencies, there is an acute need for strategies to encourage greater discussion and dialogue among all interested parties and stakeholders [26] . A first step would be for the professional colleges to create a forum to engage their memberships and encourage the exchange of views on the issue. Such an exchange could then inform the development of formal position statements on the duty to care during communicable disease outbreaks, as well as the development of clear and unambiguous guidelines regarding the professional rights and responsibilities and the ethical duties and obligations of HCPs during such outbreaks. Such statements ought to be made publicly available (e.g., prominently posted on the websites of professional health care colleges and associations) in order to encourage sustained dialogue on the issues raised by professional colleges. A next step would be to foster public debate and dialogue on the positions taken by the various health care professions. To promote this public debate, a working group at the University of Toronto Joint Centre for Bioethics has produced an ethical framework to guide preparedness planning for pandemic influenza, based in part on experiences and study of the SARS crisis [27] . The framework presents a 15-point, value-based ethical guide for pandemic planning, including the value of duty to care. This report has been made publicly available via the internet and the use of webcasting and electronic town hall meetings are being planned to facilitate an open exchange. It is of utmost importance to promote a public discourse on these issues and, most importantly, to give a voice to all those who would be directly affected by a communicable disease outbreak. Together, health care professionals and the general public should participate in discussions to determine whether and when it is legitimate for HCPs to eschew the duty to care in the face of personal risk. In light of the recent experience of Canadian physicians, nurses, and other HCWs on the frontlines of the SARS outbreak, we submit that the Canadian health care community should lead the charge to address issues of duty to care and ethical obligations in times of public health emergencies. In place of open and honest discussion, we currently have vagueness and ambiguity. In our view, health care codes of ethics should speak specifically to this issue in order to guide professional behaviour during infectious disease outbreaks. Indeed, the time to address the ethical duty to provide care is at hand -before the arrival of the next public health emergency.
48
Epicatechins Purified from Green Tea (Camellia sinensis) Differentially Suppress Growth of Gender-Dependent Human Cancer Cell Lines
The anticancer potential of catechins derived from green tea is not well understood, in part because catechin-related growth suppression and/or apoptosis appears to vary with the type and stage of malignancy as well as with the type of catechin. This in vitro study examined the biological effects of epicatechin (EC), epigallocatechin (EGC), EC 3-gallate (ECG) and EGC 3-gallate (EGCG) in cell lines from human gender-specific cancers. Cell lines developed from organ-confined (HH870) and metastatic (DU145) prostate cancer, and from moderately (HH450) and poorly differentiated (HH639) epithelial ovarian cancer were grown with or without EC, EGC, ECG or EGCG. When untreated cells reached confluency, viability and doubling time were measured for treated and untreated cells. Whereas EC treatment reduced proliferation of HH639 cells by 50%, EGCG suppressed proliferation of all cell lines by 50%. ECG was even more potent: it inhibited DU145, HH870, HH450 and HH639 cells at concentrations of 24, 27, 29 and 30 µM, whereas EGCG inhibited DU145, HH870, HH450 and HH639 cells at concentrations 89, 45, 62 and 42 µM. When compared with EGCG, ECG more effectively suppresses the growth of prostate cancer and epithelial ovarian cancer cell lines derived from tumors of patients with different stages of disease.
There is accruing evidence that green tea may have anticancer activity (1) , but the mechanisms for this action are poorly understood. Green tea is produced from the shrub Camellia sinensis (Fig. 1) ; leaves are dried but not fermented so that the green coloration attributed to polyphenols is retained. Commercially prepared green tea extracts contain $60% polyphenols (1) . These polyphenols are the source of bioflavonoids, which have strong antioxidant activity. The major bioflavonoids in green tea are epicatechins. Like all bioflavonoids, the tea catechins have three hydrocarbon rings; hydroxyl molecules are found at the 3, 5, and 7 positions (Fig. 2) . The four major tea catechins are epicatechin (EC), EC 3-gallate (ECG), epigallocatechin (EGC) and EGC 3-gallate (EGCG). The relative proportions of EC, ECG, EGC and EGCG in non-decaffeinated green tea are 792 ± 3, 1702 ± 16, 1695 ± 1 and 8295 ± 92 mg 100 g À1 dry wt, respectively; corresponding proportions in non-decaffeinated black tea are 240 ± 1, 761 ± 4, 1116 ± 24 and 1199 ± 0.12 mg 100 g À1 dry wt (1) . Epicatechins have apparent activity against human cancer: they reportedly may promote apoptosis (2) (3) (4) (5) (6) , arrest metastasis by inhibiting metalloproteinases (7, 8) , impair angiogenesis (9, 10) and reverse multidrug resistance (11, 12) . Although all epicatechins except EC can potentially suppress cell proliferation (13) (14) (15) (16) (17) (18) , EGCG appears the most promising and is therefore under clinical investigation in chemoprevention trials (19) . However, given the wide range in physiologic potency of the different catechins, an exclusive focus on EGCG is probably short-sighted. EGCG is reportedly more effective than EGC in decreasing the intestinal absorption of cholesterol (20) and it is the most potent catechin inhibitor of HIV-1 reverse transcriptase (21) , but ECG has the strongest collagenase inhibitory effect (22) and the highest antioxidant potential (23) . By contrast, only EGC is a potent mediator of oxidative modification and an inhibitor of xanthine oxidase during hepatic catabolism of purines (24) . We hypothesized that the in vitro anticancer action of the various catechins varies with the type and stage of malignancy. We tested this hypothesis by examining proliferation of catechin-treated cell lines derived from organ-confined or metastatic prostate cancer (CaP) and from moderately or poorly differentiated epithelial ovarian cancer (EOC). The goal was to obtain data that would be useful for developing chemopreventive and therapeutic clinical trials in patients with gender-specific and non-specific solid tumors. Four gender-specific human cancer cell lines were used. The HH870 androgen-receptor-negative CaP cell line was developed at Hoag Cancer Center, Newport Beach, CA, from an organ-confined primary tumor that had been resected from a 56-year-old, previously untreated Caucasian (25) . This tumor was Gleason Grade 3/4, with no evidence of vascular or perineural invasion or extracapsular extension (stage T2b). The DU145 metastatic CaP cell line (American Type Culture Collection line HTB-81) was derived from a brain lesion of 69-year-old male Caucasian. It is androgen insensitive and does not express prostate-specific antigen. Two EOC cell lines developed at Hoag Cancer Center were also used: HH639 was from a poorly differentiated clear cell, Grade 3 carcinoma in the omentum and left ovary of a 56year-old Caucasian female; HH450 was from moderately differentiated metastatic cells recovered from the abdominal fluid of a 52-year-old Asian female. All four cell lines were cryopreserved in liquid nitrogen freezer at À70 o C. For recovery of cryopreserved cells, the vials were transferred to a 37 o C water bath for 15-30 s, further thawed at room temperature and then transferred to a 15 ml polypropylene tube with a Pasteur pipette. An aliquot of 9 ml of RPMI-9% fetal bovine serum (FBS) was added in drops. The cells were allowed to settle for 5 min and then centrifuged at 4 o C for 10 min at 300 g. Supernatant was removed, and cells were suspended in fresh RPMI, gently tapped and vortexed. Cell viability was monitored by 0.2% trypan blue dye exclusion, and cell count was determined using a hemocytometer. Cells recovered from cryovials were grown in RPMI-1640 with glutamine (Invitrogen, Carlsbad, CA) supplemented with 9% FBS, HEPES buffer, gentamycin (5 mg%) and fungizone (0.5 mg%), at 37 C in a humidified atmosphere of 5% CO 2 . Upon confluency, cells were detached with sterile EDTA-dextrose (137 mM sodium chloride, 5.4 mM potassium chloride, 5.6 mM dextrose, 0.54 mM ethylene diamine tetra acetate (EDTA), 7.1 mM sodium bicarbonate) at 37 C for 5-15 min (or $45 min for HH639), recovered with cold RPMI-1640-9% FBS and resuspended in the same medium. Use of trypsin was avoided for harvesting the cells. Cell viability and cell count were reassessed before cells were seeded in culture flasks. All epicatechins used in this study (Fig. 2) with 50, 60 or 100 mM of each epicatechin or with no epicatechin (control) in RPMI-1640 with glutamine (Invitrogen), 9% FBS, 0.54% HEPES buffer, gentamycin (5 mg%) and fungizone (0.5 mg%). All experiments used 25 ml sterile polystyrene tissue culture flasks with a vented blue plug seal cap (Beckton Dickinson, Franklin Lakes, NJ, Cat. No. 353107). Each flask contained stock solution with or without epicatechin in concentrations of 50 mM (five flasks for each epicatechin and five flasks for control) and 25, 75 and 100 mM (three flasks for each epicatechin and three flasks for control). Cells (0.25 · 10 6 ) suspended in 10 ml of the RPMI-1640-FBS solution described above were transferred to each flask and allowed to grow until control cells reached confluency. The cells were detached with sterile EDTA-dextrose at 37 C for 5 min, recovered with cold RPMI-1640-FBS medium and resuspended in the same medium. Cells were counted using a hemocytometer; trypan blue dye exclusion was used to determine the number of viable versus dead cells. The interval between seeding and confluent growth of control cells was used to calculate the doubling time and the number of cell cycles. The 50% inhibitory concentration (IC50) of each catechin in each cell line was calculated using a software program (Microcal Origin Corp, OriginLab Corporation, Northampton, MA). The cells were photographed directly from the flask using light microscopy (Olympus IX-70, Japan). Analyses of variance and Fisher's least significant difference (LSD) method were used for pairwise comparisons of values significant at the 0.05 level. Organ-confined prostate cancer cell line HH870 and primary and metastatic epithelial ovarian cancer cell lines (HH450 and HH639) seeded (2.5 · 10 5 cells) in flasks with or without various concentrations (25, 50, 75 or 100 mM) of ECG or EGCG were photographed under a light microscope after the untreated control cells reached confluency (Fig. 3 ). Both ECG and EGCG significantly affected the density of each cell line at or above 75 mM. The decrease in cell density at higher concentrations is much pronounced for ECG than for EGCG, a finding significant considering recommendations of clinical trials with EGCG (19) . The mean density or viable cell number (in millions) (n ¼5 per treatment) of different cell lines was examined with or without catechins (50 mM) (Fig. 4) . The cell density was measured when growth of untreated cells reached confluency. Statistical analysis by ANOVA as well as by pairwise comparison showed that both ECG and EGCG significantly affected the cell density. ECG decreased the cell density of prostate cancer cells DU145, HH870 and ovarian cancer cell line HH639 more potently than EGCG. But EGCG inhibited the growth of ovarian cancer cell line HH450 better than ECG, suggesting the need to determine relative efficacy of ECG and EGCG in clinical trials for different cancers. Tumor Cell Doubling Time: ECG versus EGCG Figure 5 shows the influence of the four epicatechins on cell doubling time. ECG and/or EGCG prolonged the doubling (14) EGC and ECG inhibited the growth of a human lung cancer cell line, PC-9 cells as potently as did EGCG, but EC did not show significant growth inhibition. The mechanism of growth inhibition by EGCG was studied in relation to cell-cycle regulation. EGCG (50 and 100 mM) increased the percentages of cells in the G The relative cytotoxicity (CTX) of ECG to carcinoma HSC-2 cells and normal HGF-2 fibroblasts cells from the human oral cavity, as compared with other polyphenols in tea, was evaluated. For the HSC-2 carcinoma cells, ECG, CG and EGCG grouped as highly toxic, EGC as moderately toxic, and C and EC as least toxic. For the HGF-2 fibroblasts, ECG and CG grouped as highly toxic, EGCG as moderately toxic, and EGC, C and EC as least toxic. The CTX effects of the polyphenols were more pronounced to the carcinoma, than to the normal, cells. Table 1 summarizes the effects of EC, ECG, EGC and EGCG on viability, doubling time and cycling of the four cell lines. Untreated cells from each line reached confluency in about 2.5 cell cycles. EC did not affect the proliferation of DU145, HH870 or HH450 cells but it reduced the proliferation of HH639 cells by half (P < 0.05) and prevented their confluent growth (Table 1) . EGC did not affect the proliferation of any cell line (Table 1) , whereas EGCG arrested proliferation of all four lines. ECG, followed by EGCG, was the most potent inhibitor of cell growth and cycling. Proliferation of each cell line (n ¼ 3 per treatment) was monitored with or without ECG or EGCG at concentrations of 0, 25, 50, 75 and 100 mM. The dosimetric results plotted in Fig. 6 shows concentration-dependent suppression of cell growth by ECG and EGCG. The suppressive effect on cell density was striking at higher concentrations of ECG and EGCG. ECG was a more potent inhibitor of cell growth than EGCG. At 25 mM of EGCG, cell numbers for HH870 and DU145 were significantly higher than control values. Based on the results plotted in Fig. 6 , IC50 values were calculated. The IC50 values are 24-30 mM for ECG, versus 42-89 mM for EGCG (Table 2 ). ECG suppressed growth at all higher concentrations tested (Fig. 6) , whereas EGCG significantly (P < 0.05) enhanced proliferation of CaP cells at 25 mM, a finding relevant to chemoprevention trials with EGCG only. Green tea is widely consumed in Japan and China and its polyphenolic components have a chemopreventive effect against cancer in vitro and in vivo (39) . A cup of green tea contains 100-150 mg catechins, of which 8% are EC, 15% are EGC, 15% are ECG and 50% are EGCG (40) . Although numerous investigations have shown the role of EGCG in cancer chemoprevention, only a few studies have attempted to compare the relative antitumor efficacy of all four catechins (Table 3) . When we used a systematic approach to assess the effect of various catechins on cell lines derived from gender-based cancers, we found that each catechin's antitumor activity depended on the type of tumor. EGCG was not always the most potent chemopreventive agent. Most of the earlier literature (Table 3) indicates that EGCG is the most potent growth inhibitor of cell lines from glioblastoma, melanoma and cancers of the breast, colon, lung, prostate (androgen-receptor-positive), pancreas, liver and mouth. EGCG prevents proliferation of DU145 cells by arresting the cell cycle at G 0 /G 1 -phase (19) . Gupta and others (26) have documented that G 0 /G 1 -phase arrest is independent of p53 mutation, and EGCG treatment of DU145 induces the cyclin kinase inhibitor WAF1/p21. These observations suggest that EGCG imposes a cell-cycle checkpoint (19) . However, our results showed that ECG may be more potent than EGCG for inhibition of primary and metastatic CaP and EOC cells (Fig. 4, Tables 1 and 2) . ECG significantly reduced cell proliferation (Table 1, Figs 2 and 3) and increased mean doubling time (Table 1, Fig. 4) . The in vitro effect of chemopreventive agents can be studied when tumor cells are in a matrix (1, 4, 27) or in a suspension (28, 29) . We used the suspension method because it exposes the entire cell surface to the chemotherapeutic agent. Our findings confirm an earlier report that used the matrix method to show that ECG is more potent than EGCG in suppressing the proliferation of DU145 CaP cells (4) . Thus reported differences in the relative efficacy of different catechins may not be due to differences in methodology. Not all tumor cells are killed by catechins. In our study, ECG (50 mM) induced death of most but not all HH639 cells. Doubling ECG's IC50 concentration might increase the tumor kill rate if ECG does not epimerize to CG. Our in vitro dose of 100 mM is equivalent to 29 mg (EC/EGC) to 45 mg (EGCG/ ECG), far less than the 100-150 mg (50% of which is EGCG) in one cup of green tea. However, Lee et al. (41) reported that plasma levels of EGCG and EGC in healthy volunteers increased to 78 and 223 ng ml À1 , respectively, 20 min after drinking brewed green tea (1.2 g of tea solids in 200 ml hot water). This suggests that drinking more than 10 cups of green tea may be necessary to maintain a plasma concentration of EGCG equivalent to that used in vitro by a dose of 50 mM or 22.5 mg. Kaegi (42) suggested a daily intake of 13 cups of green tea as a chemopreventive measure. Because this level of tea consumption is impractically high, chemoprevention of cancer with catechins may require administration of the appropriate catechin in a purified form. In conclusion it may be stated that both green and black tea polyphenols are important components of antitumor aspect of complementary and alternative medicine (CAM), which play a significant role in the American health care system and in patients who suffer from chronic problems (43) . While green tea catechin gallates such as EGCG and ECG possess potent antitumor activities, their epimers, commonly found in black tea, act as potent inhibitor of proteases involved in replication of viruses, including coronoviruses (44) . There is a need to understand preventive and therapeutic potential of catechin gallates from both green and black teas. We are currently designing a phase I chemopreventive study to examine the effects of purified EGCG and ECG in patients who have been chosen observational management of organ-confined prostate cancer.
49
Markers of exacerbation severity in chronic obstructive pulmonary disease
BACKGROUND: Patients with chronic obstructive pulmonary disease (COPD) can experience 'exacerbations' of their conditions. An exacerbation is an event defined in terms of subjective descriptors or symptoms, namely dyspnoea, cough and sputum that worsen sufficiently to warrant a change in medical management. There is a need for reliable markers that reflect the pathological mechanisms that underlie exacerbation severity and that can be used as a surrogate to assess treatment effects in clinical studies. Little is known as to how existing study variables and suggested markers change in both the stable and exacerbation phases of COPD. In an attempt to find the best surrogates for exacerbations, we have reviewed the literature to identify which of these markers change in a consistent manner with the severity of the exacerbation event. METHODS: We have searched standard databases between 1966 to July 2004 using major keywords and terms. Studies that provided demographics, spirometry, potential markers, and clear eligibility criteria were included in this study. Central tendencies and dispersions for all the variables and markers reported and collected by us were first tabulated according to sample size and ATS/ERS 2004 Exacerbation Severity Levels I to III criteria. Due to the possible similarity of patients in Levels II and III, the data was also redefined into categories of exacerbations, namely out-patient (Level I) and in-patient (Levels II & III combined). For both approaches, we performed a fixed effect meta-analysis on each of the reported variables. RESULTS: We included a total of 268 studies reported between 1979 to July 2004. These studies investigated 142,407 patients with COPD. Arterial carbon dioxide tension and breathing rate were statistically different between all levels of exacerbation severity and between in out- and in-patient settings. Most other measures showed weak relationships with either level or setting, or they had insufficient data to permit meta-analysis. CONCLUSION: Arterial carbon dioxide and breathing rate varied in a consistent manner with exacerbation severity and patient setting. Many other measures showed weak correlations that should be further explored in future longitudinal studies or assessed using suggested mathematical modelling techniques.
Chronic obstructive pulmonary disease (COPD) is a respiratory disease characterized by an airflow limitation and inflammation of the lower airways [1] . As the disease worsens, some patients experience 'exacerbations' of their principal symptoms of dyspnoea, cough and sputum. These exacerbations frequently result in a visit to a general practitioner's office or to a local hospital for treatment. Exacerbations occur in COPD patients at a median of three times a year with half of them being unreported [2] [3] [4] . The heterogeneity of COPD exacerbations make them difficult to define, classify and manage due to their range of symptoms, varied treatment requirements, seasonal occurrence, and ambiguous aetiology [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] . To address this problem, attempts have been made to develop a consensus definition for COPD exacerbations [15] . Recently, the American Thoracic Society (ATS) and European Respiratory Society (ERS) adopted the following definition: 'an event in the natural course of the disease that is characterised by a change in the patient's baseline dyspnea, cough and sputum beyond day-to-day variability sufficient to warrant a change in management' [1] . The severity of an exacerbation has been also difficult to classify despite the various schemes that have been proposed to deal with this issue [4, [15] [16] [17] . The ATS and ERS have also jointly suggested a classification based upon severity and the type of medical management used, i.e., Exacerbation Level I is home treatment, Level II is hospitalization, and Level III is specialised care [1] . The aim of this scheme is to improve the existing management of exacerbations and to serve as an aid in the assessment of treatment efficacy. Different operational definitions for COPD exacerbations have been proposed in the past and these have helped determine their relative importance, in particular their relationship to COPD progression [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] . However, these definitions have relied primarily on symptoms, and this along with the absence of a standard classification for the degree of symptom severity, has delayed the development of new therapies for this condition. The current therapies for exacerbations have been evaluated based on their ability to reduce symptoms, and to improve a patient's forced expiratory volume in one second (FEV 1 ) since the latter is strongly correlated with COPD mortality. However, FEV 1 does not discriminate well between the stable and exacerbative states of COPD, particularly during the later stages of this disease. Hence, the development of biological markers, or biomarkers that are more sensitive and specific to the severity of COPD exacerbations would provide investigators with new insights and directions for further research. At this time, only a few clinical variables or inflammatory mediators have been shown to be associated with COPD exacerbations and their related morbidity and mortality. Some of those include: age [18] [19] [20] ; FEV 1 , forced vital capacity and peak expired flow [19, 21, 22] ; body mass index [20] ; albumin [20, 22, 23] ; sodium [23] ; pH [24, 25] ; eosinophils [26] [27] [28] [29] ; interleukins 6 and 8 [29] [30] [31] [32] ; fibrinogen [31] ; and C-reactive protein [33] . Significant clinical events such as the number of exacerbations per year, the number of hospital admissions per year, time to relapses, and days in hospital have been regarded as useful measures in clinical studies designed to assess drug efficacy and cost-effectiveness as well as to standardize existing hospital support programs for COPD [34] [35] [36] [37] [38] [39] . However, it is not known how these measures change with increasing severity of COPD exacerbations. Therefore, we have surveyed the medical literature to identify which of the commonly accepted variables and suggested markers for COPD exacerbations change according to the ATS/ERS' levels of exacerbation severity. The long-term aim of our work is to assess the sensitivity and specificity of potential markers for use in future COPD studies as well as to determine how such markers can be further studied and fully integrated into the development of new drugs for COPD. We searched standard databases since 1966 using medical search headings and related terms as obtained from major consensus documents related to COPD exacerbations. The major keywords were 'exacerbation', 'unstable', 'acute', 'bronchitis', and variants of the term 'COPD'. This phase of our search retrieved a total of 843 citations. For these citations, we read the title and abstract of each citation so as to exclude citations that concerned exacerbations of coronary artery disease, myocardial infarction, cystic fibrosis, asthma, pulmonary emboli, and community pneumonia. Citations for case studies, letters, reviews, meta-analyses, and animal studies were also excluded. After this initial screening, we identified 387 citations to papers that were of possible interest. We retrieved the original articles in electronic and hard copy forms, and then critically read each article. As a result of this step, we arrived at a total of 268 studies in our final review and analyses. We selected these studies based on the availability of demographics, spirometry, clear study eligibility criteria, and the potential markers being used to assess exacerbations. The objectives of this literature review and data analyses were to determine which of the baseline measures com-monly used in COPD exacerbation studies change with the extent of the exacerbation and disease severity, and to determine whether COPD exacerbations can be modelled as 'events' or 'time-to-event' in future investigations. Initially, we considered various exacerbation definitions and classification schemes, in particular, those suggested by Rodriguez-Roisin [15] as well as those described by Pauwels and colleagues [17] . However, we determined that the ATS/ERS' operational classification of exacerbation severity [1] was the most sensible and feasible system for systematically assessing the patient baseline characteristics and biomarker information from the majority of published studies. We therefore used this classification scheme and the related clinical history, physical findings and diagnostic procedures for managing exacerbations to perform our data abstraction. From each study, we retrieved the reported demographics, spirometry, smoking status, clinical, cytological and biochemical variables as well as suggested markers of the severity of the exacerbation at baseline conditions, i.e., immediately prior to, or during the exacerbation event but before the time in which the intervention of interest was investigated (Table 1 ). Whenever such variables were measured in stable conditions, we also abstracted this information. For each study, we noted the type of definition used to define an exacerbation such as symptom-or event-based as well as the research question asked, the experimental design used, any sponsorship, and the presence or absence of data from individual study patients. Data was then further organized according to sample size and smoking status when available. Cytological and biochemical data were also classified according to their collection methods. These included sputum induction, bronchial biopsy, bronchoalveolar lavage (BAL), exhaled breath sampling, and blood sampling. We were also aware of the possibility that for some study groups in severity Levels II and III (as per the ATS/ERS criteria) included in this review may have experienced a similar quality of care or medical management that was not reported adequately in the original publication. In attempt to correct for this problem, we combined the exacerbation data from Levels II and III into an 'in-patient' category and then compared it to Level I that we regarded as the 'out-patient' category. We collected and calculated study means, medians, standard errors, standard deviations, 95% confidence interval, and inter-quartile ranges using the statistical algorithms in Microsoft Excel 2002. We then conducted fixed effect meta-analyses to obtain mean point estimates, 95% confidence intervals, and two standard deviations for each exacerbation level [40] . Exacerbation Severity Levels I and II, II and III, and I and III were each compared using a twotailed Z-test. The alpha level of p < 0.05 was adjusted for multiple testing according to the Bonferroni correction procedure [41] . In the event that a specific exacerbation severity level had a large number of studies in which only median data were available, the data were considered to be normally distributed and medians were treated as means. Since many studies did not publish data for individual patients, we were limited in addressing non-normality in the data by using a log 10 -transformation. We again performed a fixed effect meta-analysis to obtain mean point estimates, 95% confidence intervals, and two standard deviations for in-patient and out-patient categories of each measure. We then compared each category using a two-tailed Z-test and a p-value of 0.05. Our search strategy yielded 268 suitable studies that met our selection criteria. These studies were published between 1979 and July 2004 -Week 2. (The references for these studies can be found at the LACDR Division of Pharmacology website [42] ). The total number of study subjects included in this review was 142,407. Of this group, 18% fell in Exacerbation Severity Level I, 78% in Level II, and 4% in Level III. When we re-analysed the data according to out-or in-patient settings, 18% were out-patients and 82% in-patients. Meta-analyses of typical study demographics showed that there was significant overlap in 95% confidence intervals and study data distributions for the three exacerbation severity levels except for age where study patients in Level II had a mean age of 64.2 years (95% confidence interval (CI): 62.9 to 65.5 years) compared to 68.0 years (95% CI: 65.9 to 70.1 years) for patients in Level III (p = 0.002) ( Table 2 ). When the demographics were re-analyzed according to patient settings, we determined that only body mass index was statistically different between the out-patient setting (mean point estimate: 26.2 kg/m 2 ; 95% CI: 23.8 to 28.7 kg/m 2 ) and the in-patient setting (mean point estimate: 23.4 kg/m 2 ; 95% CI: 22.5 to 24.3 kg/m 2 ) (p = 0.038) ( Table 3) . The spirometry measures Forced Expired Volume in 1 Second (FEV 1 ) and Forced Vital Capacity (FVC), both in percent predicted, decreased from Exacerbation Levels I to II (p < 0.017) but remained unchanged from Levels II to III ( Figure 1A and 1C, respectively). However, when Levels II and III were combined to create an 'in-patient' category for each of these variables, there was a statistically signifi-Respiratory Research 2006, 7:74 http://respiratory-research.com/content/7/1/74 cant decrease for the in-patients versus the out-patients (p < 0.05) ( Figure 1B and 1D, respectively). We also observed the same trend for FEV 1 /FVC (Figure 2A and 2B ). For all other spirometry measures, there were too few studies available in Level III for meta-analysis. We found for smoking that pack years increased with exacerbation severity, but only Levels I and II were statistically different (p = 0.015) ( Figure 2C ). When we compared pack years between patient settings, it was statistically higher for the in-patients than the out-patients (p = 0.010) ( Figure 2D ). In terms of the hemodynamic measures, only heart rate showed a statistically significant difference being higher in Level II than Level I (p = 0.014) with no difference Many study variables were measured at or around the time of the exacerbation. If these variables were measured in the stable condition of these COPD patients, i.e., measurements were taken weeks or months prior to the exacerbation, then these were also obtained. between Levels II and III ( Figure 3A ). Heart rates were also higher for in-patients than out-patients (p = 0.011) (Figure 3B ). The clinical measures of dyspnoea, i.e., the breathing rate ( Figure 3C ) and Borg dyspnoea score, tended to increase from Levels I to II and then decrease from Levels II to III. However, only breathing rate demonstrated clear statistical differences between the three levels (p < 0.017). Only Levels II and III of the Borg Dyspnoea Score were statistically different (p < 0.001); a statistical comparison of these levels with Level I was not possible due to lack of data. When patient settings were compared, only breathing rate showed a clear statistical difference being statistically lower for in-patients than out-patients (p = 0.003) ( Figure 3D ). Exacerbation Levels II and III were statistically different with respect to pH (p = 0.003) and bicarbonate (p = 0.002) in that pH decreased from Level II to III whereas bicarbonate increased. However, there was insufficient Level I data for each variable to allow for statistical comparisons with the other Levels. There was also insufficient data available to compare out-patients with in-patients. In terms of blood gas measures studied, only arterial carbon dioxide tension (PaCO 2 ) showed a statistically significant increase with increasing exacerbation severity (p < 0.017) ( Figure 4A ) as well as out-versus in-patients (p < 0.05) ( Figure 4B ). In the case of oxygen saturation, it gradually decreased with increasing exacerbation severity with statistically significant differences between Levels I and II (p < 0.001) as well as Levels I and III (p = 0.011) ( Figure 4C ). It also decreased going from an out-patient to an inpatient setting (P < 0.001) ( Figure 4D ). The six minute walking distance challenge test seemed to show a decreasing trend with increasing exacerbation severity but such changes did not reach statistical significance. This was also the case when the out-and in-patients were compared. Many other variables related to spirometry, respiratory status, exacerbation and hospital event categories also did not change significantly with exacerbation severity or out-and in-patients (See additional file 1). There was not enough data in the bacteriology and virology categories to permit any meta-analyses. Of the 268 studies sampled, only half contained data about the biochemical variables. We conducted this review of the COPD exacerbation literature to determine which commonly-accepted baseline variables and suggested markers changed in a consistent manner with the severity of COPD exacerbations. As our index of COPD severity, we used the recently published ATS/ERS operational classification of exacerbation severity for medical management. This is because most of the published literature rarely provides sufficient details to characterise the severity of a patient's exacerbation. In addition, we also analyzed the same data according to out-and in-patient settings so as to account for possible overlaps in medical management between Levels II and III but were not reported in the original publication. The long-term aim of our work is to improve the quality and applicability of exacerbation management through the identification of sensitive and specific markers that can be used for the assessment of treatment effects. This review identified a few potential markers of exacerbation severity. When we assessed the spirometry measures FEV 1 and FVC in % predicted, as well as FEV 1 /FVC, we observed statistically significant differences with exacerbation severity, and between out-and in-patients ( Figures 1A-D and 2A -B). One draw-back was the paucity of such information in Level III studies. This confirms the clinical situation that as exacerbations worsen and more specialised care is required, spirometry measurements are less likely under baseline conditions or during an exacerbation [14] . Thus, such data is rare in many published studies. The number of smoking-related pack years increased with exacerbation severity and showed a clear difference between out-and in-patient settings ( Figures 2C and 2D) , a finding that is consistent with the idea that the more a COPD patient smokes, and for longer, the higher the likelihood that COPD exacerbations will be more severe. According to the mean point estimates obtained in this In-patient (N=81;n=5768) PaCO2 (mmHg) * B study, COPD patients with 40 to 60 pack-years of smoking will experience an increase in the severity of COPD exacerbations. However, our conclusion regarding this finding is limited by there being data from only two studies at Level III. Although heart rate varied little between Exacerbation Levels II to III, it is important to note that it was substantially elevated in patients ( Figure 3A ) with the clearest difference being between in-and out-patients. This is possibly associated with the anxiety and dyspnea that experienced when an exacerbation occurs. The increase in heart rate of course increases the oxygen requirements of the heart. The increased heart rate may also be the result of underlying cardiovascular disease that is more prominent in severe COPD patients [43] . The relationship of pH and bicarbonate to exacerbation severity are consistent with the signs of respiratory acidosis evident in COPD patients with exacerbations [1, 24, 25] . However, due to the shortage of data in Level I, proper statistical conclusions about each of these variables are difficult to make. In relation to this, breathing rate significantly increased from Levels I to II and then decreased from Levels II to III ( Figure 3C ). The first observation may reflect components of the exacerbation episode (i.e., anxiety and dyspnea) as well as the physiological need to breathe more to maintain adequate blood gas levels. The reduction at Level III possibly reflects the results of the specialized care where patients are given ventilatory support so as to return the breathing rate to normal. The Borg Dyspnea Score showed the same trend as breathing rate, although insufficient data in Level I did not allow for further comparisons. When out-and inpatient data were compared for each of these variables, only breathing rate demonstrated a clear statistical difference ( Figure 3D ). The Borg Dyspnoea Score on the other hand did not have enough studies in the out-patient category to perform any statistical test. Overall, the observed trends were consistent with the fact that management of dyspnoea is one of the main factors generating the high hospital costs associated with COPD exacerbations [44] . In keeping with the direct measures of dyspnoea, arterial carbon dioxide tension showed a clear relationship with exacerbation severity and patient management settings ( Figures 4A and 4B ) that is consistent with the conclusions reported in the medical literature [20, [45] [46] [47] . Arterial oxygen tension in contrast did not change with exacerbation severity or patient setting. Possibly this lack of correlation reflects the immediate administration of supplemental oxygen given to hypoxaemic patients in a hospital setting. There was however a decreasing trend in oxygen saturation with increasing exacerbation severity and clear differences between out-and in-patient settings ( Figures 4C and 4D ) that are consistent with the present thinking on blood gas changes. Most of the other commonly accepted measures and suggested biomarkers poorly reflected exacerbation severity, or the fact that there was not sufficient data to undertake a meta-analysis (See additional file 1). This finding recalls a 2001 US Department of Health and Human Services report on exacerbation treatment outcomes from over 200 randomised controlled trials [14] . The aim of that study was to create new guidelines to improve the management of COPD exacerbations. That study also concluded that the current literature was limited in terms of the number of studies and the amount of detail available as well as the reliability and accuracy of the clinical assessments used to discriminate between COPD exacerbations and other causes of worsening respiratory status. Thus, our observations agree with previous observations regarding the assessment of the unstable COPD literature. As previously discussed, most of the studies used for this review were predominately with hospitalized patients (Level II). However, most COPD occurs in an out-patient setting (Level I) [48] [49] [50] [51] [52] [53] . This has implications for our study since the latter population was poorly represented. Our basic categorisation was according to the ATS/ERS' operational scheme for classifying the severity of COPD exacerbations as well as to out-and in-patient categories. To our knowledge, we are the first to undertake this type of literature review and thus we were faced with a lack of consistency in the definition of exacerbations as used in the various studies. We tried to overcome this difficulty by selecting and ranking clinical studies so as to improve the comparability of subjects between studies. We were also aware that the clinical studies we analysed differed with respect to which comorbidities or identifiable causes for exacerbations were reported. Most patients were elderly and therefore were more likely to be suffering from one or more co-existing diseases such as asthma or cardiovascular disease. Such co-morbidity makes interpretation of our findings more difficult with respect to the true causes of exacerbations. If their aetiology could be determined, then susceptible patients such as those in Level I could be identified and new treatments developed to help prevent their onset and related hospital costs. Finally, the compatibility between the studies of COPD exacerbation that we analysed may have been limited by substantial variations in the time and location of studies. Exacerbations are more likely in summer [5] but many studies failed to report the time of year or the time period for study implementation. Thus, seasonal effects, combined with the low incidence of exacerbations per patient, could represent an inherent bias. In addition, different institutions probably had different standards with respect to diagnosis and management of COPD exacerbations when these studies were performed. Such variations may also explain any observed inconsistencies in our findings. However, we attempted to overcome this possible bias in Exacerbation Levels II and III by the subsequent re-analysis of this data on the basis of out-patient and in-patient settings. As observed in The additional online file, there was a scarcity of information particularly for biomarkers at different exacerbation levels. It is also unclear to us whether any of the variables that changed with exacerbation severity are causally-related. Hence, longitudinal studies and/or less restrictive eligibility criteria would be needed to address all these questions. One difficulty in tackling such problems is the enormous amount of time and expense involved in implementing such studies. In addition, the current methods for data analysis in clinical studies have limitations imposed by the assessment of the reduction in frequency or total suppression of exacerbation episodes (i.e. rare event or "non-event"). To overcome these drawbacks and obtain more accurate evaluation of treatment effect on COPD exacerbations, alternative analytical methods based, for example, on predictive mathematical models such as hidden Markov chains or Bayesian forecasting should be tried. Such models can characterise and predict rare events without undertaking a full-scale, long-term longitudinal study. This approach to predicting rare events has been used previously in studies of migraine, epilepsy and various cardiovascular diseases where the size of treatment effect is measured in terms of a reduction in the frequency of the repetition of an event within a given probability or within a given time period [54, 55] . One example of a mathematical model development includes the use of a Markov model to predict COPD exacerbation rates in a clinical trial of the inhaled anticholinergic bronchodilator tiotropium [56] . In this example, the model was developed on the basis of prior knowledge of the exacerbation rate as estimated from meta-analyses of randomised controlled trial data. This gave the probabilities for COPD exacerbations for different stages of COPD. In another study, a proportional hazards model was used to identify risk factors for COPD patients hospitalised due to an exacerbation [44] . The current ATS/ERS guidelines for exacerbations do not consider the implications of using probabilistic models as a means of assessing the severity of COPD exacerbations or the effect of treatment [1] . A modelling approach may offer new insights into which variables related to COPD exacerbations should be investigated. From a research planning perspective, our study findings have generated some hypotheses and related considerations that could be evaluated in future clinical trials. One hypothesis is that the combination of variables that we observed to change in our study (i.e., FEV 1 , FVC, FEV 1 / FVC, arterial carbon dioxide, breathing rate, heart rate, pack years, and oxygen saturation) could represent a new definition for a 'severe' exacerbation event. Most definitions in the literature, including the recent ATS/ERS definition, do not indicate any assessment of (patho)physiological variables as signs of an exacerbation. They simply regard the exacerbation as a worsening of the normal day-to-day symptoms and/or an adjustment in medical management [17] . A definition that encompasses a clear set of objective measures would be useful to medical practitioners who predominantly rely on clinical judgement or past experiences for diagnosing an exacerbation and its severity as well as for assessing treatment effect. Another important consideration for future clinical trials is the assessment of treatment effect based on predictions of exacerbation frequency and intensity. In other words, the collection of data such as the rate of onset and resolution of an exacerbation from longitudinal studies could be used to determine probabilities of second, third, fourth, etc., exacerbation events in individual patients [54] . The alteration of such probabilities with an experimental treatment could be a more sensitive and reliable approach for assessing treatment effect in clinical trials than recording daily changes in symptoms or medical management. Lastly, our findings were obtained from COPD patients that had experienced at least one exacerbation during the study assessment period. In the same studies, there were also patients who did not experience an exacerbation. This indicates that a fraction of COPD patients may be regarded as being susceptible to an exacerbation whereas another fraction is 'exacerbation-free'. It would be interesting to determine how the variables we identified in our study change in the latter patient group according to FEV 1 . Some published studies have stratified COPD patients on the basis of exacerbation frequency; this is generally done by categorising patients as having either 'infrequent' or 'frequent' exacerbations if they had less than or greater than a mean of three exacerbations per year, respectively [57] . In our study, we were unable to make this distinction between COPD patients since many of the published studies did not provide individual patient data on exacerbation frequency. We are currently investigating a commercial database of clinical trials that will enable us to look at patients with 'infrequent' or 'frequent' exacerbations. The results of this work could help us better select patients as well as identify potential markers for future longitudinal studies. The current management and treatment of COPD exacerbations is primarily dependent on the evaluation of the symptoms rather than the signs related to the exacerbation event. We found that arterial carbon dioxide tension and breathing rate consistently varied with the severity of COPD exacerbations and with in-versus out-patients. Other commonly-accepted measures and suggested biomarkers for exacerbations failed to show consistent trends or lacked sufficient data to permit any meta-analysis. We recommend the design of longitudinal studies looking at the frequency of exacerbations as well as the use of more advanced modelling techniques to improve the selection of potential markers for the categorization of the severity of COPD exacerbations and the assessment of treatment effect in future studies. COPD -Chronic Obstructive Pulmonary Disease
50
Differentially profiling the low-expression transcriptomes of human hepatoma using a novel SSH/microarray approach
BACKGROUND: The main limitation in performing genome-wide gene-expression profiling is the assay of low-expression genes. Approaches with high throughput and high sensitivity for assaying low-expression transcripts are urgently needed for functional genomic studies. Combination of the suppressive subtractive hybridization (SSH) and cDNA microarray techniques using the subtracted cDNA clones as probes printed on chips has greatly improved the efficiency for fishing out the differentially expressed clones and has been used before. However, it remains tedious and inefficient sequencing works for identifying genes including the great number of redundancy in the subtracted amplicons, and sacrifices the original advantages of high sensitivity of SSH in profiling low-expression transcriptomes. RESULTS: We modified the previous combination of SSH and microarray methods by directly using the subtracted amplicons as targets to hybridize the pre-made cDNA microarrays (named as "SSH/microarray"). mRNA prepared from three pairs of hepatoma and non-hepatoma liver tissues was subjected to the SSH/microarray assays, as well as directly to regular cDNA microarray assays for comparison. As compared to the original SSH and microarray combination assays, the modified SSH/microarray assays allowed for much easier inspection of the subtraction efficiency and identification of genes in the subtracted amplicons without tedious and inefficient sequencing work. On the other hand, 5015 of the 9376 genes originally filtered out by the regular cDNA microarray assays because of low expression became analyzable by the SSH/microarray assays. Moreover, the SSH/microarray assays detected about ten times more (701 vs. 69) HCC differentially expressed genes (at least a two-fold difference and P < 0.01), particularly for those with rare transcripts, than did the regular cDNA microarray assays. The differential expression was validated in 9 randomly selected genes in 18 pairs of hepatoma/non-hepatoma liver tissues using quantitative RT-PCR. The SSH/microarray approaches resulted in identifying many differentially expressed genes implicated in the regulation of cell cycle, cell death, signal transduction and cell morphogenesis, suggesting the involvement of multi-biological processes in hepato-carcinogenesis. CONCLUSION: The modified SSH/microarray approach is a simple but high-sensitive and high-efficient tool for differentially profiling the low-expression transcriptomes. It is most adequate for applying to functional genomic studies.
Microarray is a powerful technique for simultaneously determining the expression of thousands of genes [1] [2] [3] . Such studies can quickly yield a genome-wide description of mRNA expression, called transcriptomes, in a given cell or tissue at a given physiologic or pathologic condition [4] . Even though, one of the main challenges in such genome-wide gene-expression profiling is the difficult inspection of genes with rare transcripts. On the other hand, PCR-based suppressive subtractive hybridization (SSH) techniques are highly sensitive for identifying differences in gene expression [5] [6] [7] [8] [9] . However, the potentials of SSH in assaying dynamic changes of gene expression in minute levels have never been addressed before. In addition, SSH techniques are also restricted in terms of limited specificity and difficulties in identifying enriched genes. Combining the SSH technique with highthroughput screening of the harvested clones through the use of cDNA microarrays could greatly reduce the tedious work for northern blot analysis, as well as the likelihood of false-positive clones enriched via SSH [10] . Such combined approaches by printing the clones obtained from the SSH amplicones on chips have been successfully used for profiling the differentiation of gene expression [10] [11] [12] [13] . Nevertheless, in such approaches, the genes of the subtracted clones remain to be sequenced for identification, and a large portion of redundancy in the enriched amplicons must also be identified. Moreover, since the targets used to hybridize the amplicon clones printed on chips were the un-enriched cDNA pools rather than the subtracted, enriched clones, such approaches would not increase the sensitivity for detection of the low-expression genes. Herein, we report our modifications using the subtracted amplicons as the targets to hybridize the pre-prepared microarray chips for SSH/microarray analysis. Since all of the probes on the microarray chips have been well characterized, the hundreds and thousands of genes in the subtracted amplicons can be determined by a single hybridization. Moreover, the relative expression status between the compared tissues for each gene can be augmented and easily determined by combining the use of the targets prepared both from forward and reverse subtractions of SSH. We named this modified approach "SSH/microarray" and used human hepatocellular carcinoma as a model to demonstrate the feasibility of this approach. The SSH/cDNA microarray versus regular cDNA microarray approaches To conduct the comparative transcriptomic studies on human hepatoma particularly for the genes at low-expres-sion levels, we set out an approach as shown in Figure 1 . We used the PCR-based suppressive subtractive hybridization to enrich the differentially expressed cDNA clones between human hepatoma and non-hepatoma liver tissues. The SSH was conducted both forwards using RNA prepared from hepatoma and non-hepatoma liver tissues as tester and driver, respectively, to yield the hepatoma (T-N) subtracted amplicons, and reversely using the mRNA prepared from non-hepatoma and hepatoma liver tissues as tester and driver respectively to yield the non-hepatoma (N-T) subtracted amplicons as well. Instead of the previously reported combination with cDNA microarray approaches, in which clones in the subtracted cDNA libraries were cloned and then printed on chips, we directly labeled the subtracted T-N and N-T amplicons and then used as targets to hybridize the pre-made microarray chips printed with the known 14,811 cDNA clones. We named this modified approach as "SSH/microarray". The same tissue RNA pairs and cDNA microarray chips were also used for the regular cDNA microarray assays for comparison. Since human hepatoma generally occurs in patients of chronic hepatitis with/without cirrhosis, to identify the genes implicated in the common biological processes leading to hepatocellular carcinogenesis, we selected three hepatoma patients with different underlying liver disease. The first was with chronic hepatitis C and cirrhosis, the second had chronic hepatitis B and cirrhosis, and the third had chronic hepatitis B but no clinical or histological evidence of cirrhosis. Figure 2 demonstrates the representative results obtained from the corresponding regular cDNA microarray and the SSH/microarray assays. Of note, in the SSH/microarray assays the genes in the subtracted T-N and N-T amplicons were readily identified without tremendous sequencing works, which were required for the original SSH and microarray combination assays. In addition, many of the clones, including those of the house-keeping genes, such as β-actin and GAPDH, with unremarkable difference in hybridization intensities by the regular cDNA microarray assays were efficiently excluded from the subtracted amplicons ( Figure 2 , indicated by white circles). This indicated the efficiency of subtraction hybridization. Most importantly, many of the clones with hybridization intensity below the evaluation threshold by the regular cDNA microarray approaches (hybridization intensities lower than or close to the noise) became detectable by the SSH/ microarray method ( Figure 2 , indicated with red circles). These observations suggested a more sensitive detection of the low abundance transcripts using the SSH/microarray assays. Figure 1 Experimental flowchart. Messenger RNA was prepared from hepatoma (T) and the corresponding non-hepatoma liver tissues (N), and then subjected to 1) PCR-based suppressive subtractive hybridization (SSH) followed by using the resulted subtracted cDNA libraries as targets for cDNA microarray analysis (the SSH/microarray), and 2) conventional cDNA microarray analysis. SSH was performed in both the forward (T as tester) and reverse (N as tester) direction to enrich up-regulated (T-N amplicon) as well as down-regulated transcriptomes (N-T amplicon) in human hepatoma, respectively. The two subtracted amplicons were labeled with fluorescent cy-dyes as targets for microarray analysis, in which dye-swapping approaches were used. The results thus obtained were then compared to those obtained from the conventional cDNA microarray assays. The differentially expressed genes were categorized into three groups: I) only detected by conventional cDNA microarray approaches, II) identified by both conventional cDNA microarray and SSH/microarray approaches, and III) only obtained from SSH/microarray assays. The results were further confirmed by qRT-PCR in 6 genes randomly selected from group III differentially expressed genes. Labeled and used to hybridize chips To further address whether the SSH/microarray approach was able to assay the expression of those genes with rare transcripts, we selectively inspected those genes with low expression. Of the total of 14,811 clones on the chips, 9376 clones that were filtered out for further analyses in the regular cDNA microarray studies due to the low hybridization intensity (intensities <500 unit after the noise subtracted) ( Figure 3A ) were selected for examination of their hybridization intensity in the SSH/microarray assays ( Figure 3B ). Of these, 5015 clones were enriched via the SSH/microarray assays and could be analyzed for differential expression (intensities >500 unit after the noise intensity subtracted) ( Figure 3B &3C ). That is, the SSH/microarray approach allowed about 100% more of the genes, most of which were low abundant, for the subsequent analyses. Since SSH specifically amplified the difference of gene expression, the SSH/microarray approach should also be able to identify more differentially expressed genes between hepatoma and non-hepatoma liver tissues, which were originally undetectable by the regular cDNA microarray methods. As shown in Fig 3C, of the 5015 lowexpression genes identified only by the SSH/microarray assays, we identified additional 512 and 1923 genes with at least a 2-fold decrease and increase in hepatoma tissues, respectively. We further examined the sensitivity for detection of differential expression by both approaches. We compared the ratio of Cy5/Cy3 intensity of each clone obtained from the regular cDNA microarray assays to that obtained from the SSH/microarray assays. As shown in Figure 4 , a total of 3028 genes with a ratio lower than 2 folds (-1 < log 2 < 1) using the cDNA microarray assays had a ratio greater than 2 folds (log 2 > 1 or < -1) using the SSH/microarray assays The representative results of the regular cDNA microarray vs Figure 2 The representative results of the regular cDNA microarray vs. SSH/microarray assays. RNA were prepared from the hepatoma and para-hepatoma liver tissues of a patient with HCC. Left panels are the representative results obtained via the regular cDNA microarray assays using the Cy 5 and Cy3 labeled aRNA generated from hepatoma and non-hepatoma liver tissues, respectively. Right panels are the representative results obtained via the SSH/microarray assays using Cy3 and Cy5 to label the forward and reverse subtracted amplicons, respectively. Lower panels are the close views of the upper panels. White circles indicate the clones, which had equal Cy3 and Cy5 hybridization intensities in the conventional cDNA microarray assays but been subtracted away from the subtracted amplicons. Red circles mark the clones, which had low Cy5 and Cy3 hybridization intensities in the conventional cDNA microarray but presented with differential expression in the subtracted amplicons. SSH/microarray ( Figure 4 , 1091 and 1937 genes indicated by the red box, respectively). That is, using the modified SSH/microarray approaches resulted in much more differentially expressed genes, particularly for those with rare transcripts that using the regular cDNA microarray approaches. A total of 69 and 701 genes differentially expressed in human hepatoma were finally identified (based on the data derived from the three tissue pairs in duplicate with log 2 > 1 or < -1, and P < 0.01) by the regular cDNA microarray assays and by the SSH/microarray assays, respec-tively ( Figure 5 & see additional file 1). That is, the SSH/ microarray assays detected 10-fold more differentially expressed genes, particularly for those low-expression genes, than did the conventional cDNA microarray. To confirm the differential expression of genes with low expression identified by the SSH/microarray assays (including 446 and 255 up-and down-regulated genes, respectively), we quantified the relative expression level of nine randomly selected genes, whose differential expression was detected by the SSH/microarray assays, in 18 pairs of HCC and the matched non-HCC liver tissues. We found the results were consistent with those obtained from SSH/microarray assays ( Figure 6 ). Detection of differential low-expression transcriptomes by the SSH/microarray Figure 3 Detection of differential low-expression transcriptomes by the SSH/microarray. A) Of the 14811 distinct genes on the chips, 9376 genes had hybridization intensities lower than 500 units (the defined low intensity threshold for exclusion from further microarray analysis) in a representative set of results of the conventional cDNA microarray assay. B)These low-expression genes were then subjected to re-calibration for their Cy3/Cy5 intensities in the SSH/microarray assay and 5015 genes became detectable in SSH assays (blue spots). C) The distribution of these 5015 low-expression genes identified only in the SSH/microarray assays, including 512 and 1923 genes down-and up-regulated in hepatoma tissues, respectively. The data presented in this figure were based on the results derived from patient 1, a case of chronic hepatitis C with cirrhosis and hepatoma. We compared the HCC up-and down-regulated transcriptomes identified by the SSH/microarray assays in accordance with their potential molecular functions, implicated biological processes and sub-cellular localization. As presented in Table 1 Microarray techniques are limited in the detection of genes with low expression [14] , while subtractive hybridization methods are restricted by their specificity and the tremendous work needed for validating the results, as well as for sequencing to identify genes. There have been many reports using microarray techniques for rapid and high throughput validation of the subtraction specificity of SSH by spotting the enriched clones on the chips or membranes [10] [11] [12] [13] [15] [16] [17] [18] [19] [20] [21] [22] [23] . However, such approaches would not only require tremendously sequencing works to identify genes including a great number of redundant clones in the subtracted amplicons, but also lose the sensitivity of SSH techniques to detect the low abundance transcripts, since the sensitivity of microarray analysis is determined by the targets used to hybridize chips but not by the probes printed on chips. Herein we report our modified approaches. Instead of using the subtracted clones as probes spotted on chips, we directly labeled the enriched amplicons and used them as targets to hybridize the pre-made microarray chips for microarray analysis (named as "SSH/microarray" in this report). This approach allowed us not only to readily identify genes without tremendous sequencing works in the subtracted amplicons regardless of many redundant clones, but also to easily evaluate the subtraction efficiency and specificity. Moreover, the SSH/microarray approach made it possible to conduct a transcriptomewide identification of differentially expressed genes particularly for those with low expression. It has been reported that the absolute expression level is not a crucial determinant for identifying genes, while the relative difference in expression levels does impact on whether or not a gene is recovered by subtractive hybridization [5, 24] . In this report using the SSH/microarray approach, we identified about ten times more of the differentially expressed genes, particularly for those with low expression, in human hepatoma. Our findings successfully demonstrated the transcriptome-wide assays of differentially expressed genes at low abundance. This simple but very powerful approach would greatly facilitate future researches on functional genomics [15, 19, 25] . One of the main concerns about SSH techniques is their specificity. However, as combined with microarray tech- Figure 4 Comparison of the relatively expression by regular cDNA microarray vs. the SSH/microarray. The distribution of the ratios of Cy5/Cy3 intensities obtained by the conventional cDNA microarray assays vs. the SSH/microarray assays is shown. Only the clones with significant Cy3 and Cy5 intensities in both assays were included in the analysis. The clones inside the red-boxes were those with their log 2 ratios < 1 and > -1 via the conventional cDNA microarray assays, while their log 2 ratios > 1 or < -1 by the SSH/microarray assays. The data presented were based on those derived from patient 1, a case of chronic hepatitis C with cirrhosis and hepatoma. Log 2 ( Cy5/Cy3) of SSH Log 2 ( Cy5/Cy3) of cDNA niques, the subtraction efficiency can be readily evaluated. As shown in Figure 2 , the subtraction efficiency was determined by the rare presence of clones with equal hybridization intensities between the forward and reverse amplicons, and by the consistent exclusion of the housekeeping genes in the subtracted amplicons. The specificity was further confirmed by quantification of the differential gene expression between hepatoma and para-hepatoma The differentially expressed genes Figure 5 The differentially expressed genes. The genes differentially expression in human hepatoma were identified by either only the conventional cDNA microarray assays (26 genes) or SSH/microarray assays (658 genes), or both (43 genes). The results were derived from three pairs of hepatoma liver tissues with at least two-fold difference in gene expression and P < 0.01. APEG1, TTN, HTN1, MEST, EVL3, VLDLR, MYOG, TNNI2, HOXC11, IGF2, AHSG, TTID, BIRC1, PAFAH1B1, SYK, LMNA, TNFRSF11B, FALZ, BMP1, ETS2, NRD1, CXCL1, JAG1, SEMA4G, HEY1, ZNF22, CRMP1, TEAD4, IGFBP1, IGFBP3, IGFBP7, PTGS1 , NEUGRIN, CUGBP1) related to the regulation of cell differentiation and embryonic development, and 59 genes (ASGR2, CAP2, RGS5, ITGB4BP, NCK1, EVL3, PLCG2, EPHA7, IGF2, IFNGR2, AHSG, CAP1, PTPRF, CXCL2, GNAI1, TGFBR1, NCSTN, DOK1, FRBB3, SYK, MC1R, CD79B, CSNKIE, CD69, AVPR1A, GNB2L1, ACVR1, BIRC2, FLT4, CXCL1, FAG1, PNOC, SLC9A3R1, LANCL1, GNB3, ADORA2B, FGR, NCK1, RAB11A, PLCG2, LOC91614, PRKAR1A, FLJ22595, ARF3, TYK2, MAP3K7IP1, MC1R, TNFSF10, ARF4, VAV2, AVPR1A, GNB2L1, MAPK10, STMN1, SNX17, ADORA2B, ECT2, RGS5, IGFBP3 ) associated with signal transduction in response to extra-cellular proliferation and growth stimuli were found. Our findings that differentially expressed genes were related to multi-biological processes suggest the complexity of the molecular mechanisms for hepatocellular carcinogenesis. In this study, we modified the previously method of the combination of the suppressive subtractive hybridization and microarray techniques to differentially profile the low-expression transcriptomes of human hepatocellular carcinoma by directly labeling both of the reciprocal subtracted amplicons as target for cDNA microarray assays. Compared to SSH or previous SSH in conjunction with microarray approaches, this modified approach provided us with three additional advantages: 1) easy inspection of the subtraction efficiency, 2) avoidance of tremendous sequencing work for gene identification, 3) high sensitiv-ity for identifying the low-expression, differentially expressed genes. This approach allowed for the detection of about ten times more of the differentially expressed genes than did the regular cDNA microarray approach in human hepatoma, particularly for those with low expression. Using this approach, we identified many genes potentially implicated in human hepatocarcinogenesis, which were not identified before. For its high efficiency and high sensitivity, this SSH/microarray approach is powerful for the rapid differentially profiling the lowexpression transcriptomes, and most adequate for applying to functional genomic studies. For SSH and cDNA microarray analysis, hepatoma and the corresponding non-cancerous liver tissues were obtained from 3 patients who had liver surgery at the Chang Gung Memorial Hospital. For reverse-transcription real-time PCR, hepatoma and the matched non-hepatoma liver tissues were obtained from additional 18 patients of hepatoma. Diagnoses of HCC and non-hepatoma liver tissues were based on histo-pathologic findings. The Internal Review Board for Medical Ethics of Chang Gung Memorial Hospital approved the specimen collection procedures and informed consent was obtained from each subject or subject's family. Total RNA from hepatoma and the para-hepatoma liver tissues and the poly(A) RNA was prepared as described before [26, 27] . SSH was performed with the Clontech PCR-Select cDNA Subtraction Kit (Clontech Laboratories Inc., Palo Alto, CA) as described by the manufacturer but with the following modifications. Starting material consisted of 2 μg hepatoma mRNA as tester and 2 μg nonhepatoma liver tissue mRNA as driver, and vice versa. Primary and secondary PCR conditions were altered to increase specificity of amplification according to either plan A or B. Both A and B reduced the extension time and the number of cycles of the primary PCR to 2 min and 26 cycles, respectively. The primary PCR products were diluted 1/50 prior to use in the secondary PCR. All other aspects of plan A were as per the instructions of the manufacturer. Plan B diverged from plan A in two ways. First, the initial cycle of primary PCR was performed using annealing and extension times that had been reduced to 15 s and 1.5 min, respectively. Second, for subsequent cycles, the denaturing time was increased to 10 s while the annealing and extension times were reduced to 15 s and 1.5 min, respectively. In this study, we used the GMRCL Human 15 K set, Version 2 chips as previously described [28] , which contained 14,811 sequence-verified, human cDNA clones mapped to 12,530 distinct genes. All of the samples for the regular cDNA microarray or SSH/microarray assays were performed with the dye-swapping microarray design for minimizing labeling bias and statistical variances of data. For the regular cDNA microarray experiment, we used 2 μg of the total RNA for labeling and hybridization using a 3DNA Array 350RP Detection kit (Genisphere, PA, USA). For the SSH experiment, 1 μl subtractive PCR products were labeled with Cy3 and Cy 5-dCTP (NEN, Boston, MA) using random primers. Unincorporated fluorescent nucle-otides were removed using a Qiaquick PCR purification kit (Qiagen). The fluorescent-labeled DNAs were mixed with 30 μg of human cot-1 DNA (Invitrogen) and 100 μg yeast tRNA, precipitated and then resuspended in 30 μl of Microarray Hybridization Buffer Version 2 (Amersham Pharmacia). The hybridization solution was heated to 80°C for 10 min to denature the DNA and was then incubated for 30 min at 37°C, allowing cot-1DNA and yeast tRNA to block the repetitive sequences in genome probes. The probes were hybridized to a human cDNA microarray (GMRCL Human 15 K). We scanned the slides with a confocal scanner ChipReader (Virtek, Canada) and acquired Quantification of relatively gene expression in human hepatoma Figure 6 Quantification of relatively gene expression in human hepatoma. A total of nine differentially expressed genes including six (GBA, PTPRF, GNL3, ERBB3, OGDHL, FMO3 belonging to Class III) identified only by the SSH/microarray assays and three (C9, MT1F, MT1X belonging to class II) identified by both of the SSH/microarray and regular cDNA microarray assays were assayed for their relative expression between HCC and the corresponding non-HCC liver tissues of 18 patients of hepatoma using real-time semi-quantitative RT-PCR. The results are presented with the log 2 values of the gene expression ratio's between HCC vs. the matched non-HCC liver tissues. Left panel is the relatively expression level of these genes initially determined by the SSH/microarray and regular cDNA microarray assays, respectively. Of note, the difference of the expression for the six genes belonging to Class III was originally only identified by the SSH/microarray assays, but not by the regular cDNA microarray assays. the spot and background intensities with the GenePix Pro 4.1 software (Axon Instruments, Inc., CA, USA). The within-slide normalization was done using programs written with MATLAB 6.5 software (The MathWorks, Inc., MA, USA). To validate the results obtained from the SSH/microarray assays, we randomly selected six differentially expressed genes identified only by the SSH/microarray assays and three differentially expressed genes identified by both the regular cDNA microarray and SSH/microarray assays for the comparison of gene expression between hepatoma and the corresponding non-hepatoma liver tissues in eighteen patients of HCC using reverse transcription realtime PCR (qRT-PCR). Total RNA was extracted from tissues with Trizol reagents and reverse transcribed using the SuperScript III first strand synthesis system (Invitrogen, Carlsbad, CA). qRT-PCR was conducted using the ABI PRISM 7000 sequence detection system (Applied Biosystems). Pre-designed Assays on Demand TaqMan probes and primer pairs for these 9 genes were obtained from Applied Biosystems Incoporated (ABI) (Foster City, CA). For each gene, two to four sets of Taq-Man probes and primers were tested. The probes contained a 6-carboxyfluorescein phosphoramidite (FAM dye) label at the 5' end of the gene and a minor groove binder and non-fluorescent quencher at the 3' end. These were designed to hybridize across exon junctions. As a result, no fluorescent signal was generated by these assays when genomic DNA was used as a substrate, which confirmed that the assays measured only mRNA. Equal amounts of RNA were used for all qRT-PCR reactions, which were performed in triplicate, and 18S ribosomal RNAs were used as internal controls. All data (derived from three pairs of HCC and non-HCC liver tissues in a dye-swapping approach) were filtered so that 446 up-regulated and 255 down-regulated genes in HCC with at least two-fold difference and p value less than 0.01 were included in the further studies. After removal of the un-annotated genes, a total of 360 and 202 HCC up-and down-regulated genes were subjected to the subsequent gene ontology analyses. The two lists of the differentially expressed genes were analyzed using the online software FatiGo for comparative gene ontology categories including molecular function, biological process and cellular component [29] . They were also imported into the on-line software KEGG2 for pathway mapping [30, 31] . The statistical significance was defined as P < 0.01 between the HCC up-and down-regulated gene groups using Fisher's exact test.
51
From Functional Genomics to Functional Immunomics: New Challenges, Old Problems, Big Rewards
The development of DNA microarray technology a decade ago led to the establishment of functional genomics as one of the most active and successful scientific disciplines today. With the ongoing development of immunomic microarray technology—a spatially addressable, large-scale technology for measurement of specific immunological response—the new challenge of functional immunomics is emerging, which bears similarities to but is also significantly different from functional genomics. Immunonic data has been successfully used to identify biological markers involved in autoimmune diseases, allergies, viral infections such as human immunodeficiency virus (HIV), influenza, diabetes, and responses to cancer vaccines. This review intends to provide a coherent vision of this nascent scientific field, and speculate on future research directions. We discuss at some length issues such as epitope prediction, immunomic microarray technology and its applications, and computation and statistical challenges related to functional immunomics. Based on the recent discovery of regulation mechanisms in T cell responses, we envision the use of immunomic microarrays as a tool for advances in systems biology of cellular immune responses, by means of immunomic regulatory network models.
During the past decade, the highly successful field of functional genomics experienced huge growth as a result of the development of DNA microarray technology [1] [2] [3] [4] , which made it possible for the first time to measure the RNA expression of thousands of genes in parallel, in a single assay. Immune responses are complex phenomena that supervene on genomics, that is, immune responses ultimately depend on the expression of genes inside a variety of cells, but explaining the function of the immune system only in terms of gene expression in those cells would constitute a reductionist approach. While studying the immune system in terms of genomics is an important goal [5, 6] , the function of the immune system, from antigen processing to epitopespecific immune responses, may be better understood through an integrated approach that takes into account properties of the immune system as a whole. We quote from [7] , ''The immunome is the detailed map of immune reactions of a given host interacting with a foreign antigen, and immunomics is the study of immunomes.'' Whereas functional genomics strives to identify the role of genes in cellular processes via the paradigm of hybridization of mRNA to complementary DNA, functional immunomics aims to identify the roles of chemical/biological targets involved in immunological processes via the paradigm of specific cellular and humoral immune responses elicited by antigens presented to the immune system [8] [9] [10] [11] . This is an effort that promises great rewards, both in terms of our basic understanding of the immune system and in terms of disease diagnosis/prognosis [12] and the design of vaccines [13] [14] [15] to combat a variety of human infirmities ranging from pathogenic infections to allergies and cancer. Enabling technologies. Functional genomics was made possible by the significant advances that had previously been made in sequential genomics, including not only the massive efforts required to identify genome-wide DNA sequences [16] , but also the computational methods used to parse and align those sequences [17] . Sequential genomic data are deposited in large public-access databanks such as GenBank [18] , and researchers or companies who make DNA microarrays use the sequences in these databases as probes. In a similar fashion, the field of functional immunomics has now come of age as a result of advances in sequential immunomics, which consists of methods to catalogue the chemical/biological targets capable of eliciting an immune response, also known as epitopes. Computational and statistical methods are now available for automated large-scale epitope prediction (please see the next subsection), in addition to classical immunoassays such as ELISPOT [19] and tetramer staining by flow cytometry [20] that together enable high throughput identification of epitopes. Recently, a coordinated effort has been initiated by the National Institute of Allergy and Infectious Disease at the US National Institutes of Health, under the auspices of the Large-Scale Antibody and T Cell Epitope Discovery Program (in which the authors of this paper participate), to create an integrated immunome database and resources such as a toolbox of epitope prediction methods. This initiative is designed to identify immune epitopes from selected infectious agents; the information will be made freely available to scientists worldwide through the Immune Epitope Database and Analysis Resource (IEDB) [7, 21, 22] (http://www. immuneepitope.org). The sequencing information produced by this and other epitope mapping efforts being carried out will be essential for the construction of immunomic microarrays (discussed in detail below), leading to an experimental paradigm similar to that employed in functional genomics. Computational epitope prediction methods. The immune system recognizes antigen via binding of antibody (humoral response) or T cell receptors (cellular response) to self or foreign proteins. B cell epitopes correspond in general to the 3-D features on the surface of antigen where recognition by the immune system occurs; a continuous or linear epitope is a sequential fragment from the protein sequence, while a discontinuous or conformational epitope is composed of several fragments scattered along the protein sequence and brought together in spatial proximity when the protein is folded [23] . Humoral response is targeted mainly at conformational epitopes, which may represent up to 90% of the total B cell responses. This makes prediction of B cell epitopes a hard problem [24] , even more so because B cell responses are virtually only restricted by immunoglobulin access to the epitope, B cell receptor activation, and self versus no-self discrimination rules of the immune system. Ideally B cell prediction systems would use 3-D surface models of the protein antigens and measure surface energy interactions of variable regions of the immunoglobulins that correlate with B cell activation. However, so far B cell prediction systems make estimations of the probability of a primary peptide sequence being present at the surface of a protein based on hydrophilicity and secondary structures [25] . Cellular responses, on the other hand, are restricted through the binding of T cell receptors to short linear peptides, which are bound by a specific groove in two main classes of major histocompatibility complex (MHC) molecules, and presented on the surface of cells to the T cell receptors of CD4 þ and CD8 þ cells [26] . Binding affinity between the peptide and the MHC molecule is therefore a necessary requirement for effective cellular immune response. A complicating factor is the highly polymorphic nature of the MHC molecule, which displays large variability in human populations. Using experimental affinity data deposited in public databases as training data, researchers have developed statistical methods to predict the MHC affinity of a given unknown peptide [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] . Typically, such computational epitope prediction methods scan the full length of pathogen or self-immunogenic protein sequences by taking consecutive overlapping peptides. In addition, such methods can predict ''promiscuous'' epitopes, that is, the ones that bind to a large class of different MHC alleles, know as supertypes [43] . Computational MHC-binding prediction methods have become essential for the systematic search for epitopes, in situations where techniques such as ELISPOT and flow cytometry are effectively impractical due to the large number of peptides to be assayed. Early MHC-binding studies identified characteristic amphipathic chemical patterns on the binding peptides [44] , and enhanced versions of these systems continue in use in association with other methods [41, 44, 45] . Today, the most basic MHC-binding prediction methods are based on the identification of specific amino acids commonly found at particular positions, called binding motifs, within peptides that bind to a specific MHC molecule. However, all the amino acids of a peptide bound to an MHC groove (normally 8-10 amino acids for MHC I and 8-14 amino acids for MHC II) can potentially play a positive or negative role in binding, and more complex methods assign positive or negative values for each amino acid at each position of a peptide and combine these values to define scores that predict binding; these ''quantitative-matrix'' approaches have been very successful. One of the limitations of the ''quantitative-matrix'' approach is that it does not take into consideration the influences of interactions between amino acids at different peptide positions of the epitope. The value of these interactions are difficult to measure and have been explored only in a limited fashion by combining pair-wise interactions between two peptide positions [46] . The combination of independent binding calculated by quantitative matrices with coefficients derived from the pair-wise interaction provided better predictions. Moreover, quantitative-matrix scores have been generated for several HLA alleles, and studies using HLA sequence homology have allowed the development of virtual quantitative matrices to be applicable to many more HLA alleles [28, 39] . General methods for MHC-binding prediction systems, such as artificial neural networks, and statistical models such as Hidden Markov, can incorporate nonlinear complex interactions between the MHC molecule and the peptide epitope, and can evolve as more data is included in the training set. These general methods have shown to be potentially superior to the previous ones [27, 37] . The greatest challenge, however, of the general methods is that, to be reliable, they require a larger amount of peptide-binding data. Strategies using query-by-committee approaches to compare predictions from different training sets have been used to identify the most informative peptide-binding data to be determined in the biochemical binding assays in order to more efficiently build a representative training dataset [40] . Recently a collection of more than 48,000 peptide binding affinities of class I molecules were made public [47] . The development of prediction models will be greatly accelerated by this community resource benchmark. We remark that epitope recognition by the immune system involves more than a receptor-ligand problem between the MHC molecule and a peptide, as other biological processes are involved in preparation of the peptide for loading into the MHC molecule. Epitope prediction systems continue to evolve, and many steps in antigen processing and presentation required for development of cognate immune responses are being modeled and combined into rational prediction systems. The greatest current challenge is the development of models incorporating the rules of B cell and T cell receptor engagement and their possible outcomes. The growth ''boom'' of immunomics. While the growth boom in genomics took place in the 1990s and this field has now begun to enter a mature stage of development, a similar growth boom in immunomics is likely to take place over the remaining years of the current decade. We recently searched the PubMed database (with Sente 2.3) using the following query: ''immunomics OR immunomic OR immunome OR (antigen AND microarray AND functional) OR (epitope AND microarray).'' After removing nine irrelevant articles from the output list and adding five articles for the new journal Immunome Research, we obtained a list of 71 articles covering the years from 1999 to the present (please see Figure 1 ). It is clear that interest in this field has accelerated, supporting the expectation of a continuing boom in growth. It is expected that the number of publications will increase at an exponential pace as immunomic microarrays became commercially available for research use. As immunomic array technology evolves, we expect that immunomic arrays with a small number of features will eventually be designed for specific clinical diagnostic purposes and used regularly in medical practice. However, these clinical applications might still be in the distant future. The basic functioning principle behind all microarray technology is the binding, and subsequent measurement, of target biological specimens of interest to complementary probes arrayed in a spatially addressable fashion. Typically, a planar surface, such as a glass slide, is used to support an array of spots containing the probes. As a consequence of using spatially addressable probes, a large number of different targets can be measured in a single experiment. For example, in the case of DNA microarray technology, which provides the basic enabling technology for functional genomics, the targets are fluorescent mRNA molecules (indicators of genomic expression) that are hybridized to gene-specific DNA probes immobilized on a planar surface. In a similar fashion, the enabling technology for functional immunomics is the immunomic microarray. The basic technologies for immunomic microarrays that we consider in detail in this paper are antibody, peptide, and peptide-MHC microarrays (see Table 1 for a summary of these technologies). Other functional immunomic approaches include dissociable antibody microarrays [48] , cell microarrays [49, 50] , serum microarrays [51] , peptide libraries [52, 53] , and serological analysis of cDNA expression libraries (SEREX) [54] [55] [56] [57] [58] . There are significant technological challenges inherent in fabricating immunomic microarrays, including the identification of a workable surface coating for the glass, appropriate probe concentration and target incubation times, and suitable spot size and interdistance [59] [60] [61] [62] [63] [64] . Antibody microarrays consist of antibody probes and antigen targets; thus, they can be used to measure concentrations of antigens for which the antibody probes are specific [65, 66] . As such, antibody microarrays are quite useful in proteomic applications, such as in the proteomic profiling of cancer antigens [67] [68] [69] . Antibody microarrays have also been proposed for post-translational functional genomics [70] . The rationale for directly measuring protein concentration, rather than using a traditional DNA microarray format, is the existence of evidence of poor correlation between concentrations of mRNA and its corresponding protein, which reflects post-translational modification of the protein [71] . Given the possibility of measuring antigens or proteins associated with ''foreign agents,'' antibody microarrays can be employed in functional immunomics applications [72, 73] (The application in [73] actually used cells, which display the target protein markers on their surfaces.) As a general rule, however, using antibody microarrays in data-driven functional immunomic applications may be problematic. One of the main reasons is that this approach requires the production of specific antibody sets for use in defining each of the antigen targets, and the development of large numbers of interrogative features (antibody sets) is a tremendous challenge, since humoral responses are much broader than MHC-restricted T cell responses, are highly conformationally dependent, and can be developed against a great variety of chemical/biological elements present in biological fluids, including small molecules. Peptide microarrays use the opposite technical approach; that is, they use antigen peptides as fixed probes and serum antibodies as targets [74] . This format is promising for functional immunomic applications. Published studies using peptide microarrays include applications to autoimmune disease [10, [75] [76] [77] [78] [79] , allergy [80] [81] [82] [83] [84] [85] , B cell epitope mapping [13, [86] [87] [88] , vaccine studies [89, 90] , detection assays [91, 92] , serum diagnostics [12] , characterization of weak protein interactions [93] , and analysis of antibody specificity [94] . Peptide microarrays essentially correspond to highthroughput parallelized ELISA assays [12, [95] [96] [97] [98] and thus can reveal the repertoire status of antigen-specific B cell antibody responses. However, B cell responses are highly dependent on CD4 þ T cell immune responses, and thus peptide microarrays should ideally be used in parallel with extensive analysis of T cell responses (e.g., by using peptide-MHC microarrays; see below). One of the pioneer studies that best depict the usefulness of peptide immunomic technology was performed using an array of 87 protein antigens to search for specific antibody reactivity patterns in the serum of 20 normal health volunteers; these were compared to the patterns of 20 type-1 diabetes mellitus patients, and simple classifiers were designed to discriminate between healthy and diabetic patients, with an overall sensitivity of 95% and specificity of 90% [78] . In a subsequent study, the 87-feature array was able to identify prognostic signatures that could predict the susceptibility of healthy animals to develop diabetes [10] . Another important report describes the use of a panel of 225 selected peptides of several protein antigens known to be recognized by autoimmune disease patients [79] . The autoimmune peptide array was used to study the profile of the autoantibody reactivity pattern of rheumatoid arthritis (RA) patients. The RA study used serum from 18 RA patients, 38 healthy controls, and 58 recently diagnosed RA patients, and found early clinical prognostic markers able to predict which patients are more likely to develop severe RA, and also markers to identify the group of patients with the milder form of the disease [77] . Moreover, an array developed from a panel of 213 peptides derived from allergenic peanut proteins established that patients responding to a greater diversity of peptide peanut epitopes had the worst allergic reactions [80] . The importance of the breadth of antibody response against the simian-human immunodeficiency virus (SHIV), a experimental nonhuman primate model for human immunodeficiency virus (HIV), was demonstrated through studies performed with an array of 430 peptides derived from simian immunodeficiency virus (SIV) and HIV amino acid sequences [90] . This study indicated that the reduction of the repertoire of the antibody response was associated with development of acquired immunodeficiency syndrome (AIDS). These examples underscore the great potential of peptide microarrays to identify several valuable clinical markers. The most recent technology to be proposed is the peptide-MHC microarray or artificial antigen-presenting chip [9, 11, 99] ; in this case, recombinant peptide-MHC complexes and costimulatory molecules are immobilized on a surface, and populations of T cells are incubated with the microarrays, whose spots effectively act as artificial antigen-presenting cells [100] containing a defined MHC-restricted peptide. Different methods have been proposed for detecting T cells expressing receptors with affinity for specific peptide-MHC complexes on the microarray; these can include simple inspection of T cell clusters bound to a spot [9] or identification of activated cells secreting specific cytokines with cytokine-specific capture antibodies [11, 99] . Peptide-MHC microarrays correspond to high-throughput parallelized ELISPOT assays [19] , particularly when low enough densities of cells are used, in which case a direct counting of activated cells is possible [11] . Quantitation can involve cell counts alone, detected cytokine intensities alone, or a combination of both, as in [99] , which used a cell-count score adjusted by an intensity score. The benefit of using peptide-MHC microarrays is that it can map MHC-restricted T cell epitopes, which are involved in several helper and regulatory functions of the immune system, and can be used in conjunction with peptide-based B cell epitope microarrays to study the adaptive system as whole. Figure 2 illustrates the functioning of peptide-MHC microarrays. In Figure 2a , a peptide-MHC microarray is depicted, with an inset showing the probe molecules that are deposited on a microarray spot. Figure 2b depicts T cells that bind to, and are activated by, specific peptide-MHC complexes, with the help of co-stimulatory antibodies; these T cells secrete cytokines that are captured by specific detection antibodies. Finally, as depicted in Figure 2c , the T cells and excess cytokine are washed away, and the bound cytokine is revealed by fluorescent antibody (other methods can be employed to reveal the cytokine [11] ). Therefore, a peptide-MHC spot is designed taking into account two elements: a peptide-MHC complex, and the detection antibody specific to the particular cytokine one wants to measure. A third element can be the kind of T cell population (e.g., T helper or CTL) that is used as a target (i.e., incubated with the microarray). The choice among these three different elements will lead to a vast number of immunomic responses that can be measured. An exciting feature that distinguishes immunomic from DNA microarray data is the possibility of measuring two or more signals simultaneously, determined by a single feature, the epitope. In the case of DNA microarrays, one response value is obtained for each gene per sample, namely the concentration of mRNA produced by the gene (note that twodye experiments use mRNA from two different samples). In the case of peptide-MHC chips, a single epitope can generate different response values corresponding to different cytokines or different target T cell populations, or even different antibody isotypes, in the case of peptide microarrays. In other words, in the case of genomic microarrays, one parameter is measured, namely the level of transcription of each individual gene, whereas in the case of immunomic microarrays, it is possible to measure several parameters regarding immune responses against a single epitope. For instance, a single B cell epitope can be recognized by different isotypes of immunoglobulins, such as IgE or IgG1. Therefore, in this case it is not only the intensity of the antibody response that can be measured, but also the quality of the antibody response. This aspect can be very relevant since a high IgE titer in relation to IgG1 may be associated with allergy, whereas the opposite, a high IgG1 titer in relation to IgE to the same epitope, is not. This situation is even more significant in the case of the peptide-MHC array, where the same peptide-MHC epitope can induce several different cytokine responses. These ''multicolor'' peptide-MHC microarrays have a counterpart in the multicolor ELISPOT assays currently in use [101] . It is known that the combined effect of multiple cytokines is essential to the control of immune responses; this is described by the suggestive term ''cytokine chord'' in [102] . Thus, given a family of epitopes, one may want to simultaneously measure both inflammatory (effector) and anti-inflammatory (regulatory) T cell responses, which are known to be associated with the concentrations of IFN-c and IL-10, respectively [103] . In this case one would have more than one spot on the microarray containing the same epitope (peptide-MHC complex) but use distinct cytokine antibodies for detection (see Figure 3 ). The result of this analysis is not a real-valued profile, as is obtained from functional genomics microarrays, but rather a vector-valued profile. Such profiles are sometimes called ''multispectral'' profiles (see the section on data analysis below). We will, however, adopt the term ''multicolor'' when referring to immunomic data, due to the fact that this term is already used in the similar setting of ELISPOT assays. The technological challenges mentioned previously in connection with antibody and peptide microarrays are much more complex in the case of peptide-MHC microarrays, which in fact involve elements of the two previous technologies, namely presentation of peptide and antibody detection of secreted cytokine. The technology of peptide-MHC microarrays, though still in its infancy, is viewed as a simple and economical method for screening the T cell repertoire of a host [99] , and thus holds great potential. The first clinical research application of peptide-MHC microarray technology was the study of the correlates of protection regarding the effects of an experimental therapeutic cancer vaccine [104] . Ten patients with melanoma were immunized with a peptide vaccine, and their immune responses were examined with a peptide-MHC microarray, which contained seven types of peptide-MHC epitopes and probed for 26 secreted factors. This peptide-MHC array was shown to have the sensitivity to detect one peptide-MHC specific T cell in 10,000, and 10 6 CD8 þ cells were incubated with the array (so, in theory, this peptide-MHC array could detect up to 100 distinct reactive features that have reached minimum frequency of 1:10,000). Analysis of the peptide-MHC microarray response patterns demonstrated that patients who presented both IFN-c and TNF-a secretory responses against a specific epitope remained free of melanoma. The complexity of the statistical analysis with regard to immunomic microarray data is on a whole different level than that of genomic microarray data. The total number of genes in humans is estimated to be ;30,000; in comparison, the total number of different T cell receptors in humans, generated by somatic recombination, which recognize peptides within the context of a major histocompatibility molecule, is estimated to be on the order of 10 7 to 10 15 , and the number of B cell clonotypes, generated by somatic recombination of V(D)J genes, is estimated to be on the order of 10 12 . DNA is based on a four-letter ''alphabet,'' consisting of the four nucleotide bases A,G,C,T, whereas peptide epitopes are based on a 20-letter alphabet, consisting of the amino acids known to be involved in life processes. Clearly, the combinatorial complexity in the case of functional immunomics is several orders of magnitude higher than that of functional genomics. In addition, for functional genomics the number of interrogative features that need to be built on microarrays is on the order of 10 4 to 10 5 . In functional immunomics, the total number of interrogative features included in microarray analysis can be much larger and may be estimated as follows, in the case of peptide-MHC microarrays: analysis of MHC peptide-binding motifs [26] suggests that a core of nine amino acids within a peptide is sufficient for characterization of a T cell epitope. The total number of interrogative features would thus correspond to the number of possible nine-letter words based on a 20-letter alphabet, which is 20 9 ' 10 11 ; fortunately, only ,1% of these are able to bind to MHC molecules, which makes the number of interrogative features more manageable. In addition, B cells and T cells go through a process of clonal selection, where leukocytes that either do not react or react too strongly are eliminated, and dangerous clones that react with self antigens are deleted or anergized as part of the process of immune tolerance. The number of features is still very large, however, and methods to further reduce this number are essential; such methods include the prediction methods and immunomic databases mentioned previously in connection with epitope mapping, as well as the selection of peptides from specific genomes of pathogens, allergens, and self-antigens involved in human infirmities, such as tumor antigens, diabetes, and autoimmune diseases. One additional procedure that we and other groups have used after screening the genomes of pathogens for putative binding peptides is to compare those candidate epitopes with known host protein sequences, and in some cases we have found peptides that are identical to host peptides. This is very important because several critical diseases are caused by pathogen molecular mimicry, that is, some diseases, such as diabetes, dengue hemorrhagic fever, and Guillian Barret syndrome, are hypothesized to be the result of infections that induce self-reactive pathogenic immune responses. It is important to state that the goal of the immunomic array is not to test all possible naïve T cell or B cell clones that can be generated by the somatic, VJ, or V(D)J recombinations against all possible combination of peptides-there simply would not be enough patient blood to do that, even if it were considered to be a relevant pursuit. The goal of the immunomic array is to identify primed cells that have reached a reasonable level of precursor frequency and are thus expected to have biological relevance. If the frequency of a circulating T cell in the peripheral blood is less than 1:100,000 we can expect that the biological relevance of it is small in contrast to a T cell that has a frequency of 1:1,000. In addition, a T cell has a binding affinity for the peptide-MHC that the T cell is specific for, and it should not bind to the ones with which it has no affinity. The current reported limit of detection of the peptide-MHC immunomic array is 1:10,000 cells when using 10 6 CD8 þ cells. So, in theory, a peptide-MHC array incubated with 10 6 purified CD8 þ cells (;10 ml blood) could detect up to 100 distinct reactive features that have reached minimum frequency of 1:10,000. However, if we expect to be able to detect rare clones, with very low frequencies, more cells would be needed. This requirement may be overcome with larger amounts of blood, and it is not unreasonable to collect 100 ml blood, or by adding a T cell expansion step to grow the cell population ex vivo before it is incubated with the array. This technique of T cell clonal amplification is commonly used to detect rare populations of cells by flow cytometry or by ELISPOT and may as well be applied in immunomic studies. Normally, 250,000 human PBMCs are used in an ELISPOT assay with one peptide, and the limit of detection is estimated to be 4-fold to 5-fold more sensitive than flow cytometry. However, ELISPOT is quite distinct experimentally from immunomic microarrays. In ELISPOT, T cells and APCs are present in the same mixture, and for a T cell to be activated, it has to be in close contact with its APC. In the peptide-MHC array the spot surface is equivalent to a very large defined APC, completely loaded with one specific peptide epitope, for which a reactive T cell has a binding affinity, so that it adheres to its specific peptide-MHC spot and not to the other spots. In [104] , the authors used 10 6 cells and compared the limit of detection of the array with flow cytometry. It turned out that both had similar limits of sensitivity, approximately 1:10,000 cells or 0.01%. It is possible that in the future the peptide-MHC spot surface can be improved and the sensitivity of the peptide-MHC array may become even greater than the current ELISPOT assay. Another great challenge is the polymorphism of HLA genes, in particular HLA class II, and the several combinations of different alpha and beta chains. Some approaches may be useful to limit the number of features, such as the selection of specific alleles most frequently found in a population to be used in broad screening arrays. A second such approach could be the use of supertype prototype HLA molecules compatible with a set of several HLA alleles. A third approach would be through customization of the arrays, by having many different arrays of single alleles and combining them according to the HLA types of the individuals being tested. An issue that sets immunomic microarray data apart is the availability of vector-valued response profiles (in the case of peptide-MHC microarrays). The statistical challenge here is reminiscent of the data analysis problem in the engineering field of remote sensing [105] , where different materials have characteristic vector responses, called spectral signatures. In the case of immunomic data, the analogous notion to the spectral signature is the cytokine profile associated with a given epitope and T cell population; see Figure 4 for an illustration. One simple technique to address the data analysis problem for multicolor immunomic data is to combine the responses into one long feature vector, by juxtaposing the individual cytokine response profiles for each epitope, with the caveat that there may be systematic correlation among the features in the resulting feature vector. The large number of features that can be measured simultaneously with microarray technology also presents a challenge. On one hand, it is likely that a large number of irrelevant features will be present; on the other hand, the scientist would like to work with a small number of strong, relevant features that can be used for diagnostic/prognostic panels, or as the basis for further biochemical validation studies of the mechanisms involved. This problem of feature selection also arises due to a fundamental limitation in statistics, sometimes called the ''curse of dimensionality'', according to which the existence of a large number of features necessitates an even larger (exponentially larger) number of samples to achieve consistent and accurate results. As the number of patients in microarray-based studies is severely restricted by factors such as the cost of the technology and difficulties in patient enrollment, it is almost always the case that only a small number of samples are available. Thus, only a small number of features at a time can be considered. The recommended approach to feature selection is to consider combinations of m features at a time, do classifier design based on the feature set under consideration, and use an estimate of its probability of error as the performance score. Application of this type of analysis to immunomic microarray data would allow the identification of sets of epitope-specific immune responses, associated for example with distinct disease states. However, feature selection presents an explosive combinatorial problem. For example, in an exhaustive selection of sets of three features among 1,000 initial features, the total number of feature sets to be assessed is equal to 166,167,000. If an initial set of 10,000 features is used instead, the number of feature sets of size three to be searched is larger than 10 11 . The complexity of feature selection is especially crucial in functional immunomics applications, since in this case the number of initial features to be considered is huge. The use of high-performance computing architectures, such as large computer clusters, is almost mandatory. Given that a set of features has been selected, one should be able to design a classifier that takes as input microarray data for an unknown sample and generates as output a predicted class label (e.g., clinical outcome, kind of infection, or other conditions). Figure 5 illustrates this approach in a hypothetical functional immunomics application. In this case, there are two class labels, corresponding to control and protected patients, in a situation in which protection is achieved by immunization with an attenuated-virus vaccine for a given infectious disease. The objective is to identify the epitopes that show a discriminatory response between the two groups and are therefore prime targets for rational epitope-based vaccine design. A set of two features, corresponding to epitopes X and Y, have been identified via feature selection among the thousands of microarray probes. For instance, we can imagine a situation in which the protected individuals presented a higher TNF-a response to epitope X, as well as a higher IFN-c response to epitope Y. Based on the response values observed for each patient (note that a patient corresponds to a point in the plane), a linear classifier is designed. This classifier corresponds simply to two decision regions separated by a line. If a future unknown patient has response values that fall in the upper decision region, he/she is likely to be a protected patient, provided that the classifier has a small probability of error. In this case, the responses to epitopes X and Y (in terms of the TNF-a and IFN-c cytokines) characterize immunological memory induced by the attenuated-virus vaccine: large response values to both epitopes X and Y indicate protected patients (note that the response of neither epitope X nor epitope Y by itself is a good discriminator in this example, indicating the need to consider the multivariate, combined effect of both responses). Note, in Figure 5 , that the apparent error rate (i.e., the number of misclassified sample divided by the total number of samples) is 2 4 20 ¼ 10%; the actual probability of classification error on future data typically exceeds the apparent error rate [106] . Systems biology makes use of mathematical modeling in order to provide a theoretical core for biology, analogous to the way that mathematical theories provided that core for physics in the 20th century. The success of engineering and computational methodology in the physical realm is due to the predictive capability of mathematical modeling. We quote from [107] : ''Predictive mathematical models are necessary to move biology in the direction of a predictive science. They are also necessary to the application of engineering methods to translate biological knowledge into therapies with a mathematical and computational basis.'' The large complexity of biological systems, in comparison with most physical systems, makes even more urgent the application of mathematical and computational modeling techniques to them. Dynamical systems provide ''the natural language needed to describe the 'integrated behavior' of systems coordinating the actions of many elements'' [108] , and are also capable of displaying emerging self-order from massively disorganized complexity, which is believed to be a fundamental feature of life. In what follows, we will describe the notion of immunomic regulatory networks, a dynamical system model for immune regulation. An important recent development in immunology has been the discovery of regulatory T cells [103, [109] [110] [111] . These T cells suppress immune responses, helping to stem runaway inflammatory processes and to avoid autoimmune disease. It is thought that this beneficial suppression activity can turn deleterious when it is taken advantage of by pathogens, leading to chronic and abnormal infectious processes. It has been observed that some regulatory T cells are not antigenspecific; these are called natural regulatory T cells [103] . In addition, there exist regulatory T cells, both CD4 þ and CD8 þ , that are antigen-specific and thus epitope-driven [103] . Given the suppressive action of epitope-driven regulatory T cells in conjunction with the promoting activity of epitope-driven helper T cells, it follows that the immunological response to a given epitope may be suppressed or promoted by the immunological response to other epitopes. Thus the notion of regulatory networks arises as a fundamental concept in understanding the functioning of the immune system. In fact, most human diseases are the result of an unbalance in immune system homeostasis. In functional genomics, DNA microarray data is used to infer genomic regulatory networks [112] . For biological and efficiency reasons, gene expression is often quantized to two levels: on and off [113] . The multivariate methods of classification and feature selection discussed in the previous section have proved to be essential in the inference of such Boolean (binary) regulatory networks. By the same token, immunomic microarray data can be used to infer immunomic regulatory networks. As is true for gene expression, one may quantize each epitope response measured with an immunomic microarray at one of two levels-on (immunogenic/responder) and off (nonimmunogenic/ nonresponder)-leading to the inference of Boolean immunomic regulatory networks. In the general case, each node of an immunomic regulatory network represents a combination of the epitope, the cytokine response measured, and the T cell population used as the target; in most cases, each node is in a one-to-one relationship with a single physical spot on an immunomic microarray experiment with a given T cell population. The edges between nodes represent putative regulatory relationships between the cells that respond to the respective epitopes. Figure 6 depicts a simple example with quantized Boolean responses, where peptide-MHC microarrays are used in conjunction with three kinds of T cell targets: CD4 þ helper T cells, CD4 þ regulatory T cells, and CD8 þ cytotoxic T cells. Epitope A is specific to the CD4 þ helper T cells that promote the response to epitope C, which is specific to the CD8 þ effector T cells that produce the actual protective mechanism. In addition, there is an epitope B that activates the CD4 þ regulatory T cells that suppress the effector response to epitope C, thereby producing an antiinflammatory response. In practice, such a model would be derived from microarray data by automatic epitope (feature) selection and determination of predictive relationships (classifier design). In this example, the effector response is activated only in the presence of both help from epitope A and an absence of regulatory response to epitope B. The suppressing response to epitope B is itself promoted by the presence of a response to epitope C, providing a negative feedback mechanism. These relationships can be represented by a wiring diagram and transition rules, depicted in Figure 6a . Since the responses have been quantized to two values, on and off, and there are three epitopes, the total number of possible states of the system is 2 3 ¼ 8. Figure 6b depicts the state transition table obtained from the transition rules in Figure 6a . Using this table, one can determine the attractors of the system, which are the states or sets of states in which the system stays in the long run in the absence of external disruptions. For each attractor, there is associated a basin of attraction, containing attractors and transient states, which are the sets of states that lead to the attractor [108] . In the context of biological systems, attractors are a mathematical model for homeostasis. In our minimalist example, we see that there are two different behaviors, corresponding to the two distinct basins of attraction in Figure 6c ; the respective attractors are indicated by dashed rectangles. As can be seen, which of the two behaviors the system is in depends only upon the state of the response to epitope A. If it is off (there is no help), then the system may pass through some irrelevant transient states but it will always tend toward the resting single-state 000 attractor and thus an absence of activity; see the diagram on the left in Figure 6c . If the response to epitope A is on (there is help), then the activity of the system corresponds to that of a cyclic attractor, with the effector DOI: 10.1371/journal.pcbi.0020081.g006 Figure 6 . Example of a Simple Immunomic Network, Consisting of Three Epitopes Epitope A is a promoter (A is specific to CD4 þ helper T cells), epitope B is a suppressor (B is specific to CD4 þ regulatory T cells), while epitope C produces the effector response (C is specific to CD8 þ cytotoxic T cells), while also promoting the suppressing response of epitope B (negative feedback). response being turned on and off cyclically; see the diagram on the right in Figure 6c . This situation corresponds to modulation of the effector response and regulation of the inflammatory response by means of a negative feedback mechanism. Note that when there ceases to be a response to epitope A, the system jumps to the other basin of attraction, and tends to the resting 000 state. The immune response to epitope A in this example determines the behavior of the system, and thus it functions as the master of the overall immunological response, with the individual immune responses to epitopes B and C being slave to it. The concept of master-slave regulatory units is quite important for the understanding of complex regulatory systems and has in fact been considered as a mechanism for genomic regulation (Michael Bittner, unpublished data). Inference of immunomic regulatory networks from immunomic microarray data constitutes, after proper validation, computational knowledge discovery. There are subtle epistemological issues involved in using data-driven, computer-based methodology to obtain scientific knowledge. As Karl Popper explained in his classic book The Logic of Scientific Discovery [114] , every scientific theory consists of an initial irrational act of creativity (induction) followed by rigorous logical consequences (deduction) and testing of the initial hypothesis. Where is the initial act of creativity, in other words the scientific hypothesis, in computational knowledge discovery? Computers are clearly not capable of irrational creativity. Can this be considered science? We maintain that the answer is yes [115] . In fact, the initial irrational act of creativity is there; it is involved in all the steps of experiment design, selection of patients/samples, and choice of statistical methodology. Once these are settled, the actual data analysis is purely logic deduction via the machinery of mathematical operations, as prescribed by Popper. In this deductive stage the computer plays a critical role, as it facilitates the application of very complex computational methods. Therefore, scientist (and statistician) bias here is in fact unavoidable, as it is in all scientific disciplines. In particular, the term data-driven is really a misnomer in describing computational knowledge discovery. Functional immunomics promises great rewards, both in terms of our basic understanding of the immune system and in disease diagnosis/prognosis and rational epitope-driven vaccine design. Research into the basic biology and statistical methods associated with functional immunomic experiments will lead to the advancement of medical science and public health. Functional immunomics is however still at an early stage of development. In this review we have attempted to provide a coherent vision of this nascent field, and have speculated on future research directions for this technology. In our discussion we have often compared and contrasted immunomics with genomics. Immunomics supervenes on genomics, in the epistemological sense, since immunology ultimately depends on the functioning of genes inside cells, but immunomics has its own independent character and properties. In the same manner that each cell has its own pattern of gene expression that defines its unique cellular properties, each reaction of the cognate immune system to an antigen has its own pattern of epitope-specific responses that define its final outcome. There exists a large collection of mathematical models to describe the immune system [116] . Here, we have proposed Boolean immunomic regulatory networks as a new mathematical model for immune system regulation. This is a dynamical system model, with parameters that can be estimated from immunomic microarray data. A somewhat similar concept was suggested in [102, 117] , where a network model for cytokine action was proposed, albeit without explicit reference to large-scale immunomic technology or regulatory T cell response. In addition, immune system regulatory networks have been previously discussed in the context of genomics [118] . The immunomic regulatory network model may be useful for computational knowledge discovery and simulation of regulatory mechanisms of the immune system in health and disease, which may lead to advances in practical applications (e.g., vaccine design) as well as in the basic scientific understanding of the immune system. "
52
Transmission patterns of smallpox: systematic review of natural outbreaks in Europe and North America since World War II
BACKGROUND: Because smallpox (variola major) may be used as a biological weapon, we reviewed outbreaks in post-World War II Europe and North America in order to understand smallpox transmission patterns. METHODS: A systematic review was used to identify papers from the National Library of Medicine, Embase, Biosis, Cochrane Library, Defense Technical Information Center, WorldCat, and reference lists of included publications. Two authors reviewed selected papers for smallpox outbreaks. RESULTS: 51 relevant outbreaks were identified from 1,389 publications. The median for the effective first generation reproduction rate (initial R) was 2 (range 0–38). The majority outbreaks were small (less than 5 cases) and contained within one generation. Outbreaks with few hospitalized patients had low initial R values (median of 1) and were prolonged if not initially recognized (median of 3 generations); outbreaks with mostly hospitalized patients had higher initial R values (median 12) and were shorter (median of 3 generations). Index cases with an atypical presentation of smallpox were less likely to have been diagnosed with smallpox; outbreaks in which the index case was not correctly diagnosed were larger (median of 27.5 cases) and longer (median of 3 generations) compared to outbreaks in which the index case was correctly diagnosed (median of 3 cases and 1 generation). CONCLUSION: Patterns of spread during Smallpox outbreaks varied with circumstances, but early detection and implementation of control measures is a most important influence on the magnitude of outbreaks. The majority of outbreaks studied in Europe and North America were controlled within a few generations if detected early.
The anthrax attacks that followed the events of September 11th, 2001 focused attention on the threat of terrorism with biological agents including variola major, the causative agent of smallpox [1] . There is concern that an attack with smallpox could result in many deaths because of the susceptibility of the U.S. population [2] [3] [4] [5] [6] [7] [8] . A number of issues, such as the amount of virus released and the number of people exposed, need to be considered in order to estimate the impact of a bioterrorist attack with smallpox. However, one key factor in determining the size of an outbreak resulting from such an attack is the transmissibility of smallpox, which can be difficult to estimate. One measure of transmissibility is the effective reproductive rate (R), which is equal to the expected number of infections produced by an infectious host in a real world setting [5, [9] [10] [11] . R can be determined in a straightforward manner for individual epidemics and is related to the circumstances, such as population vaccination status, under which the individual epidemics occur. When considering intentional attacks, the greatest interest lies in initial spread, much of which often occurs before an outbreak is recognized. For this reason, we focused on the transmission from index cases to the first generation of secondary cases, which we termed the initial R. The data for this study was obtained from a systematic literature review for reports of smallpox outbreaks. We limited our review to reports of smallpox outbreaks occurring in Europe and North America after 1945 because these societies were much more similar to the modern United States in terms of their demographic, social, and physical structure, than were the pre-World War II western and the developing societies from which most smallpox outbreak reports have emanated. Moreover, population immunity was also relatively more similar to a contemporary US population than might be expected a priori. While early post-war European and North American populations were generally vaccinated in childhood, re-vaccination rates were generally low and few adults had more than remote vaccination [12, 13] . Immunity naturally acquired through infection was also similarly low in both populations because outbreaks were rare. Although terrorist use of smallpox in the 21 st century may have substantially different epidemiological characteristics (see discussion below), the outbreaks analyzed in this paper are the closest actual population experience that may help partially guide modern smallpox control efforts. The following databases were searched for literature on smallpox through 2001: National Library of Medicine 1800 to 2001 (keywords: smallpox or variola), Embase 1974 to 2001, Biosis 1969-2001 (keywords: smallpox or variola virus and human and not in-vitro), Cochrane Controlled Clinical Trials Library (keywords: smallpox or variola) and Defense Technical Information Center 1947 to 2000 (keywords: smallpox). The search was later updated by a PubMed search from 2001 to 2004 (keywords: smallpox or variola). Duplicate titles were removed. Secondary searches included more specific searches for outbreaks in specific countries and sites, and a review of reference lists from publications identified in other searches (Figure 1 ). There was no language restriction. Two authors (SB and VB) independently scanned titles for relevance, and publications chosen by either reviewer were reviewed further. One of the authors (VB) scanned the selected articles to identify post-1945 smallpox outbreaks from Europe and North America. Two abstractors (RB or research associate and VB) independently extracted data from the identified outbreaks. Discrepancies between reviewers were resolved by consensus. Non-English publications were abstracted by one author (RB) or by a research assistant under an author's (VB) direction. To be included in the analysis, the following had to be available: the date of occurrence, and the number of index cases, total cases, cases in each generation, and generations. The following outbreak details were also extracted if available: the number of cases acquired within a hospital, a household, or at a distance (cases not acquired by household or hospital contact); the number of missed cases (cases not diagnosed as smallpox at the time of illness) and deaths; the generation at which the outbreak was identified, and the clinical presentation of the index case. Index cases were classified as having had typical smallpox, atypical smallpox (a milder illness as seen in people with prior smallpox vaccination and/or a rash not characteristic of smallpox), or hemorrhagic smallpox (a usually fatal illness characterized by internal hemorrhage and/or petechiae) [14] . The following data regarding the efficacy of control measures were also extracted if available: the use of ring vaccination (the vaccination of people who were in close contact with a patient with smallpox), case isolation (the separation of suspected cases of smallpox from the general public), quarantine (the restriction of movement in or out of a region that has been exposed to smallpox), mass vaccination (the vaccination of an entire region), and/or prophylactic vaccinia immune globulin administration [15] [16] [17] . The number of outbreaks over the period studied was small (total of 45 outbreaks) and the circumstances varied widely. Therefore, a descriptive and qualitative analysis was emphasized. For this analysis, the initial R was meas-ured by dividing the total number of first generation cases by the number of index (zero th generation) cases. In addition, several outbreaks had sufficient details to be classified by setting. Because hospitals were often the centre of outbreaks, setting was defined according to the proportion of secondary cases (non-index cases) acquired within a hospital [12, 13, [18] [19] [20] . After examination of the distribution of case locations across outbreaks, hospital outbreaks were defined those in which 70% or more of secondary cases were hospital acquired, mixed outbreaks as those in which 30% to 70% of secondary cases were hospital-acquired, and community outbreaks as those in which 30% or less of secondary cases were hospitalacquired. Based on the results of the primary searches, 5,976 titles were scanned; 1,389 of these were selected. Despite initial de-duplication, 10 of these were duplicates (Figure 1 ). Secondary searches identified 14 additional publications. Of the resulting 1,393 publications, 1219 (87%) were obtained; 92 journal citations (7% of the selected publica-tions) and 41 government reports (3% of the selected publications) were unobtainable. Forty-nine post-1945 European and North American outbreaks were identified. Thirty-six of these were included in a single World Health Organization (WHO) publication [21] and prior work by T.M. Mack [12] . Several outbreaks were documented in multiple publications (Tables 1 and 2 ). Four outbreaks were excluded. The source of the first outbreak could not be confirmed as the outbreak may have resulted from the importation of smallpox on a rug (Union of Soviet Socialist Republics or U.S.S.R, 1959). The second and third outbreaks occurred under unique circumstances; one occurred aboard a ship (Poland, 1962) and the other may have been the result of testing aerosolized smallpox (U.S.S.R, 1972) [21] [22] [23] . A fourth outbreak was excluded because the number of cases in each generation was not documented; therefore, the initial reproductive rate (R) could not be derived from the data (United States or U.S., 1946) [24] . The remaining 45 original outbreaks occurred over a thirty-year period in different countries and under different circumstances. The majority of reported outbreaks occurred in three coun- (Tables 1 and 2 ). Most resulted from involved index cases imported from endemic countries, but two outbreaks were the result of laboratory exposure [25, 26] . Although published or listed as single outbreaks, four outbreaks included multiple epidemiologically distinct components, such as occurred when a case travelled to a different region resulting in a secondary outbreak in the new area (see reference column and footnotes for Table 2 ). The components of these four outbreaks were treated as 10 separate outbreaks so that an effective total of 51 outbreaks were analyzed [21, [27] [28] [29] . For all 51 outbreaks, available information included the total number of generations and cases, the number of cases in each generation, the duration of the outbreak, and the number of deaths (Tables 1 and 2 ). Thirty of the outbreaks could also be categorized by setting and had detailed information on the generation that the outbreak was identified, control measures, clinical presentation of the index case, and the number of missed cases ( Table 2) . The median initial reproduction rate (R) across all 51 outbreaks was 2 with a range of 0 to 38 (Table 3, Figure 2a ). About half had an initial R of 1 or less, and over two-thirds had an initial R of 3 or less. The median duration, as measured by number of generations, was 1 with a range of 0 to 9 (Table 3, Figure 2b ). About a third did not extend beyond the index generation, and nearly three quarters lasted for 3 or fewer generations. The median outbreak size, as measured by the total number of cases, was 4 with a range of 1 to 134 cases (Table 3, Figure 2c ). About half involved 3 or fewer cases, and two-thirds involved 15 or fewer cases. The median number of deaths was 1 with a range of 0 to 26 (Table 3, Figure 2d ). About two-fifths involved no deaths, and three quarters involved 3 or fewer deaths. Outbreaks with greater initial R-values tended to be larger and longer. For example, in outbreaks with an initial R of 5 or less versus more than 5, median values for the total number of cases and generations were 2 versus 23 and 1 versus 2.5 respectively. There was no temporal trend in the amount of detail reported, but the thirty outbreaks reported in detail typically had a larger R (median 2 vs.0), lasted more generations (median 2 vs. 0), and had more total cases (median 13 vs. 1) and deaths (median 2 vs. 0) compared to lessdetailed outbreaks (Table 3 , Figures 2a-d) . In the detailed outbreaks, most of the index cases had atypical (mild) or hemorrhagic disease. Because these clinical presentations of smallpox are less common, many of these index cases were missed or not diagnosed with smallpox at the time of illness. The majority of index cases were infected The median initial R across the 30 detailed outbreaks was 2 with a range of 0 to 38 (Table 3 , Figure 2a ). About a third had an initial R of 1 or less, and about two thirds had an initial R of 3 or less. Outbreaks that remained in the community had a lower initial R (median 1) than those that were hospital based from the first generation (median 12) ( Table 3 , Figure 3a ). Mixed outbreaks, which generally started in the community and later spread through hospital contacts, had an intermediate initial R (median 2) ( Table 3 , Figure 3a ). The initial R was smaller when the index case(s) had a typical versus atypical or hemorrhagic presentation (medians 1 vs. 4 or 5, respectively), and when the index case was identified as smallpox versus unrecognized (median 1 vs. 5) ( Table 3 , Figures 3b and 3c) . The effect was most pronounced in hospital outbreaks -median initial R's for identified versus unidentified index cases in hospital outbreaks were 3 versus 15, but the same values for community and mixed outbreaks were 1 versus 4, and 2 versus 2, respectively. The median duration for 30 outbreaks was 2 generations with a range is 0 to 9 generations (Table 3, Figure 2b ). About a third did not extend beyond the first generation, Table 2 ) analyzed by components: (1) Outbreak from Marseille, France (1952) broken into two components: Entries 3 and 15 [27] . (2)Outbreak from Cardiff, U.K.(1963)broken into three components: Entries7, 24 and 26 [21, 28] . (3) Outbreak from Poland (1963) broken into three components: Entries 8, 17 and 27 [41] . (4)Outbreak from Kosovo, Yugoslavia (1972) broken into two components: Entries 10 and 19 [21] . Footnotes for Table 2 : a Nature of the index cases was not clear from the text (limited information) [27] . b Total number of cases listed as 74, consistent with text [38] and second reference [27] . c Total number of cases listed as 5 [28, 34] ; 6 is listed in World Health Organization Table 23 .4 [21] . d Proportion of hospital acquired cases listed as 0.44 because three nurses infected as a result of household exposure and were setting excluded as hospital contacts [21, 36] . e Total number of cases listed as 1 [37] ; 2 is listed in World Health Organization Table 23 .4 [21] . and another third lasted for only through the second generation. Outbreaks that remained in the community lasted a median of 1 generation, while those that were hospital-based lasted a median of 2 generations and those that were mixed lasted a median of 3 generations (Table 3) . Outbreaks tended to be shorter when the index case was typical compared to atypical or hemorrhagic (median number of generations 1 vs. 3 vs. 2, respectively) and when the index case was identified (median number of generations 1 vs. 3) ( Table 3 ). The duration of hospital outbreaks was not affected by the identification of the index cases, but the duration of mixed and community outbreaks lasted up to 6 and 9 generations, respectively, when index cases were unidentified. The median outbreak size, as measured by the total number of cases, was 13 with a range of 1 to 134 cases (Table 3, Figure 2c ). There were approximately equal proportions of small (total number of cases less than 5) and large (total over 20) outbreaks ( Table 2 ). Outbreaks that remained in the community involved a median of 2 cases, while those that were hospital based involved a median of 20 cases and those that were mixed had a median of 22 cases (Table 3) . Outbreaks tended to be smaller when the index case was typical versus atypical or hemorrhagic (median number of cases 2 vs. 27 vs. 16.5) and when the index case was identified (median number of cases 3 vs. 27.5) ( Table 3 ). The outbreak size was similarly small across all 3 settings when the index case was identified and similarly large across the settings when it was not. The median number of deaths was 2 with a range of 0 to 26 (Table 3 , Figure 2d ). About one-fifth involved no deaths, and four fifths involved 4 or fewer deaths (Table 2) . Community, hospital based, and mixed outbreaks had medians of 1, 4, and 3 deaths, respectively (Table 3) ; however, case fatality rates were 0.24, 0.20 and 0.17. The number of deaths was smaller if the index case was identified (median equals 1 vs. 6, Table 3 ). The case fatality rate was highest if the index case presented with hemorrhagic-type smallpox. There was a strong association between typical presentation and early identification of index cases: index cases were identified as smallpox in 9 of 9 outbreaks when the presentation was typical, 3 of 13 when it was atypical, and 2 of 6 when it was hemorrhagic. This relationship aids in the understanding of outbreak severity. For example, 93% of outbreaks had greater than 10 cases when the index case was not identified. Moreover, in all of these larger outbreaks, the index cases were atypical or hemorrhagic. When the index case was identified, most outbreaks had less than 10 cases regardless of whether the index case was typical, atypical, or hemorrhagic. Descriptions of smallpox transmission are important to biosecurity planning, especially since some planning efforts have posited outbreaks resulting in numbers of cases -hundreds, thousands, or more -that are very large relative to the clinical experience shortly before eradication of the natural disease [32] . In a systematic review designed to provide an understanding of smallpox transmission patterns, we identified and analyzed what were effectively 51 smallpox outbreaks from post-1945 Europe and North America. In these outbreaks, we found a median of 4 total cases and 1 death. The characteristics of the outbreaks varied greatly: one outbreak with an initial R of 11 involved 134 people and caused 26 deaths over only 3 generations of spread; another outbreak with initial R of only 2 involved 28 cases but lasted 9 generations. However, most outbreaks were small and contained within a few generations: 31 percent involved only one case, 41 percent caused no deaths, and 31 percent had an initial R of 0 (i.e., no first generation cases). Examination of outbreaks within categories of setting, identification of the index case, and clinical characteristics provided additional insights. By reviewing original publications, we were able to categorize and analyze many outbreaks by setting. For example, were able to review the original publications for eight outbreaks documented by the WHO [25, 26, 28, [33] [34] [35] [36] [37] as well as, identify thirteen additional outbreaks [20, 27, [29] [30] [31] [38] [39] [40] [41] . Several of these outbreaks also had epidemiologically distinct components, resulting in 51 effective outbreaks for analysis. Many (six out of eleven) of the hospital outbreaks were only found in the above-mentioned additional outbreaks [20, 27, 29, 30, 38, 39] . Hospital outbreaks had higher initial reproductive rates than either community or mixed outbreaks, and had more cases and deaths than did community types. Moreover, the interaction of identification and setting in the studied outbreaks reveals a pattern similar to that seen in frequently unrecognized infectious diseases, such as tuberculosis and, more recently, SARS [42] . When index cases were identified, control is established early and both the median initial R and the number of generations were similarly low in all 3 settings. When the index cases were unidentified, a large increase in the median initial R was seen only in the hospital setting and a large increase in the number of generations was seen only in community and mixed settings. This is consistent with the observation that, within the close quarters of an institution, the transmission of smallpox is relatively effective before identification and containment measures are quite effective after identification. This impression is reinforced by examination of the two shipboard out-breaks that were excluded from the analysis. In these outbreaks, the initial R's were 6 and 25, respectively, but recognition was followed by control within two generations [22] . In community outbreaks, transmission is intrinsically less effective. The consequence of missed initial identification in the community was not an elevated initial R, but an increase in the length of the outbreaks. The longer outbreaks may be due to a decrease in the effectiveness of containment measures after initial spread in the community, possibly because of the greater difficulty of tracking and managing cases in that setting. Nonetheless, all of the 30 outbreaks reported in detail were brought under control, generally by case isolation and ring vaccination. Other measures were rarely used (mass vaccination was used in the New York City, 1947 outbreak and mass vaccination and quarantine were used in the Kosovo, 1972 outbreak [21, 30] ) and there is limited information on the efficacy of these additional measures. It should also be noted that newer treatment modalities, such as Cidofovir, have been shown to be protective in animal studies against vaccinia and other orthopox viruses (monkeypox and cowpox) after exposure [43, 44] . Outbreaks beginning with an atypical or hemorrhagic versus a typical presentation of the index case had more cases, a higher initial transmission rate, and were somewhat longer. This is related to the ease of initial diagnosis: only non-typical cases were mis-identified, and outbreaks in which the index case was mis-identified were generally larger, deadlier, longer, and had a median initial R of 5 compared to 1 if the case was identified. We speculate that the large difference in initial R is because, when the index case is identified, this parameter is effectively the initial reproductive rate with early institution of control measures. However, it is clear that the types of outbreaks seen in developed countries in the second half of the 20th century were well and quickly controlled. Despite searching several databases, using broad searches, scanning nearly 6,000 titles, and reviewing papers in several languages, our requirement that included epidemics involve post-World War II western populations limited us to identifying only 45 separate outbreaks, some of which were divided for analysis. The majority of the outbreaks were from only three countries, suggesting that outbreaks from certain countries were more likely to have been documented. The 30 outbreaks reported in detail were larger than the 21 that were briefly reported; it is probable that smaller outbreaks were not routinely published in detail because outbreaks were still relatively common and smallpox was still endemic in many parts of the world. This apparent publication bias, however, suggests that there may have been many unpublished small outbreaks and that this analysis may overestimate the transmission potential of smallpox because it is based only on published reports. The recent literature contains several estimates for the reproductive rate of smallpox. It is difficult to directly compare our findings and estimates for initial R, the initial effective reproduction rate for smallpox, to these other studies. Two of the studies do not use contemporary western outbreaks. One estimates the basic reproductive rate, (R 0, a theoretical parameter defined as the expected number of new infected hosts that an infectious host will produce in a large randomly mixed population of susceptible individuals [9] ), rather than effective reproduction rate. Specifically, Gani and Leach used epidemic modelling of data from 19th Century European smallpox outbreaks to estimate a R 0 of 3.5 to 6.0 [19] . Eichner and Dietz obtained an R of 6.9 from data on a 1967 outbreak in Nigeria [5] . Meltzer et al examined post-1961 outbreaks from countries around the world in estimating an R of 3 [8] . Our estimate of the overall initial R is quite comparable to Meltzer et al., but use of any single point estimate for R 0 or initial R in policy analysis ignores a main lesson of this review. This work clearly shows that the pattern, speed, and duration of an outbreak's spread vary widely according to the specific circumstances surrounding an outbreak. Different smallpox release scenarios should be expected to yield different results in part because the parameters and modified natural history of the outbreak itself will vary with the scenarios. For example, our previous work modelled the outcome of terrorist attacks under different scenarios [4] . These scenarios included community based outbreaks resulting from infected cases riding a busy transit system and the intentional release of aerosolized smallpox in airport terminals, and a mixed outbreak resulting from the release of aerosolized smallpox in a large office building. Although adjustments had to be made for a poorly vaccinated population and healthcare workforce, the understanding of smallpox transmission patterns in hospital, mixed and community settings, provided important insights into how smallpox may spread under different attack scenarios. In a contemporary US population, one might expect transmission from introduced smallpox to be higher than we found to be typical. Population density is higher; mobility is greater, and, though more similar that one might think, immunity in the contemporary U.S. population is likely somewhat lower than in the populations studied. The healthcare workforce is also poorly vaccinated and inexperienced with smallpox. This will decrease the human resources available for care and containment, and increase the likelihood that initial cases will be missed. Counterbalancing this is the fact that few index cases will be vaccinated recently so the likelihood of more easily recognized typical cases is much higher. As the outbreak scenarios of greatest concern involve intentional release, outbreaks could be larger and more obvious than those reviewed, or could involve strains selected or engineered to be more virulent or contagious. Finally, the outbreak from Kosovo deserves notice because this region is not as developed or as wealthy as most of the other countries studied. This outbreak was the largest identified for this review (N = 176 for both combined hospital and mixed components). Moreover, the hospital and mixed component of this outbreak had the second highest (R = 38) and highest (R = 11) initial reproductive rate, and the hospital component lasted 9 generations (median number of generations for a hospital outbreak = 2). This suggests that a smallpox outbreak in a less developed country with limited resources for healthcare, disease surveillance, and case isolation could be potentially more devastating than a bioterrorist attack in a Western/industrialized country [44] . It is clear that most post-World War II outbreaks in Western countries were small and had low transmission rates. This was almost uniformly the case when the index case presented with typical smallpox and was recognized early. However, exceptions were common and heterogeneous, particularly when (usually atypical or hemorrhagic) index cases were not initially recognized. Initially unrecognized community outbreaks lasted several generations despite low initial R values. This suggests that control measures were less effective in this setting, perhaps because of the difficulty tracing and vaccinating all cases and contacts. In contrast, unrecognized hospital outbreaks were controlled quickly despite high initial R values, suggesting that control measures worked very well in a closed, contained setting. Early detection, particularly of patients who do not present with typical smallpox, coupled with early implementation of control measures decreases both the duration and size of outbreaks in all settings. Publish with Bio Med Central and every scientist can read your work free of charge
53
Reliability of case definitions for public health surveillance assessed by Round-Robin test methodology
BACKGROUND: Case definitions have been recognized to be important elements of public health surveillance systems. They are to assure comparability and consistency of surveillance data and have crucial impact on the sensitivity and the positive predictive value of a surveillance system. The reliability of case definitions has rarely been investigated systematically. METHODS: We conducted a Round-Robin test by asking all 425 local health departments (LHD) and the 16 state health departments (SHD) in Germany to classify a selection of 68 case examples using case definitions. By multivariate analysis we investigated factors linked to classification agreement with a gold standard, which was defined by an expert panel. RESULTS: A total of 7870 classifications were done by 396 LHD (93%) and all SHD. Reporting sensitivity was 90.0%, positive predictive value 76.6%. Polio case examples had the lowest reporting precision, salmonellosis case examples the highest (OR = 0.008; CI: 0.005–0.013). Case definitions with a check-list format of clinical criteria resulted in higher reporting precision than case definitions with a narrative description (OR = 3.08; CI: 2.47–3.83). Reporting precision was higher among SHD compared to LHD (OR = 1.52; CI: 1.14–2.02). CONCLUSION: Our findings led to a systematic revision of the German case definitions and build the basis for general recommendations for the creation of case definitions. These include, among others, that testable yes/no criteria in a check-list format is likely to improve reliability, and that software used for data transmission should be designed in strict accordance with the case definitions. The findings of this study are largely applicable to case definitions in many other countries or international networks as they share the same structural and editorial characteristics of the case definitions evaluated in this study before their revision.
between countries. One of the first case definitions used for national disease reporting was the case definition for AIDS published by the Centers for Disease Control and Prevention (CDC) in 1982 [2] . In 1985 Sacks published a survey among all 50 US states, Puerto Rico, and Washington, DC, that revealed important variations in the case definitions between the different states, and concluded the necessity to unify case definitions if surveillance data between states are to be compared [3] . In 1990 the CDC in collaboration with the Council of State and Territorial Epidemiologists published an edition of case definitions for public health surveillance [4, 5] . Since then case definitions have become an important tool of other national surveillance systems and international surveillance networks. Koo and colleagues have analyzed surveillance data for Cholera in Latin America and have described the importance of uniform case definitions to make data comparable between countries [6] . In 2003 the European Union (EU) case definitions for the European networks have reached obligatory status for the member states reporting to the EU [7] . During the SARS epidemic the case definition had a major impact on whether and how countries were considered affected or not, resulting in severe political and economic consequences for a number of countries [8] . Coggon and colleagues have demonstrated the difficulties of determining optimal case definitions if a satisfactory diagnostic gold standard is lacking [9] . In sharp contrast to the importance of case definitions hardly any research has been published on the performance of surveillance case definitions. Studies are rare on how local health departments and other health professionals are able to understand case definitions and to what extent case definitions are unambiguous enough to really assure reliability. To our knowledge, the only publication investigating this issue was focused on case definitions for nosocomial infections: Gastmeier and colleagues had investigated how uniform the case definitions of the nosocomial infections surveillance system in Germany had been applied by different investigators using a set of 60 case studies [10] . Due to the general importance of case definition for public health surveillance and the current need for harmonization in international surveillance systems we conducted a systematic evaluation of the national case definitions with the objective to identify general as well as specific criteria and recommendations for improvement of case definitions. Germany is a federal republic with 16 states subdivided into 440 counties. As in many countries the local (county) health departments (total number: 425) are the primary recipients of infectious disease notifications made by physicians and laboratories. Local health departments verify the incoming notifications and assess the need for public health action. Local health departments use one of five software products on the market to classify the case reports according to the national edition of case definitions and to report these cases electronically to the state health department. From there the report is being forwarded to the Robert Koch Institute (RKI), the federal institution in charge of national infectious disease surveillance in Germany [11] . The edition of national case definitions for all notifiable infectious diseases was introduced in Germany in 2001, following the implementation of a new law to control infectious diseases (Infektionsschutzgesetz, IfSG) [12] [13] [14] . The IfSG determines the set of diseases and pathogens to be notified by physicians and laboratories throughout the Federal Republic of Germany. The five eastern states, which formerly belonged to the Democratic Republic of Germany (East Germany) and the State of Berlin have enacted complementary rules that make certain diseases additionally notifiable within the state jurisdiction, that are not notifiable in all of Germany. The case definitions were developed by the RKI, using the delphi method including the expertise of state epidemiologists, national reference laboratories and medical and scientific associations for the specific diseases. The case definitions for infectious conditions under public health surveillance published by the CDC were also taken into account [5, 15] . After having published the IfSG case definitions in the fall of 2000 to be implemented with the beginning of 2001 the RKI also published additional case definitions in January 2002 for some of the diseases exclusively notifiable in the eastern states jurisdictions [11, 16] . From June 2002 to September 2003 we had conducted a systematic evaluation of the case definitions with the purpose to revise them by the end of 2003. The German case definitions are divided into three types of evidence: Clinical picture, laboratory detection, and epidemiological confirmation. The types of evidence are specifically defined for each disease (see table 1 ). Based on whether or not requirements for these three types of evidence are fulfilled a case is classified into five categories. In the revised 2004 edition of case definitions these categories are named: A) clinically diagnosed illness (neither epidemiologically nor laboratory-confirmed), B) clinically and epidemiologically confirmed illness (not laboratory-confirmed), C) clinically and laboratory-confirmed illness, D) laboratory-detected infection not fulfilling clinical criteria, E) laboratory-detected infection with unknown clinical picture. (In the 2001 edition of case definitions these five categories were named slightly differently) For most notifiable diseases only categories B, C, D and E are reportable from the local health department to the next level, requiring at least laboratory detection of the pathogen or epidemiological confirmation. For some exceptions (e.g. tuberculosis, polio, measles, Creutzfeldt-Jakob disease), cases are also reported from the local health department to the next level if category A -the clinical picture alone -is fulfilled. In June 2002 we conducted a Round-Robin test in analogy to the established quality control procedure of laboratories and other testing units [17] . Round-Robin tests are mainly used in proficiency tests in order to determine laboratory performance by means of comparing tests on identical items by two or more laboratories in accordance with predetermined conditions [18] . We asked each local and state health department to classify a selection of 68 written case examples on the basis of the case definitions that were implemented in 2001 (2002 respectively for disease only notifiable in East German States). While proficiency tests generally intent to assess the ability of laboratories in finding identical results, we applied this method to assess to which extend the case definitions were unambiguous enough to assure identical classification by the health departments. We applied four different outcome variables in our analysis: 1) Disease identification: A disease was defined as being correctly identified if the participant of the Round-Robin test was able to identify the correct disease of the case example. 2) Case categorization: A case example was considered correctly categorized if the participant classified the case example with the correct disease and the correct case definition category as defined in the gold standard. 3) Reporting: The decision on reportability was considered correct if a case that should have been reported to the Clinical picture: Clinical picture compatible with salmonellosis, characterized by diarrhea, abdominal pain, malaise, vomiting, fever. Salmonella can also cause infections outside the intestinal tract (for example: arthritis, endocarditis, pyelonephritis, septicaemia). Laboratory diagnosis: Isolation (culture) of pathogen from stool or other clinical material (e.g. blood, urine). The identification of serogroup has to be attempted. Clinical picture • Clinical picture of acute salmonellosis, defined as at least one of the following four symptoms: diarrhea* • cramp-like abdominal pain • vomiting • fever* additional information: Samonella can also cause generalized (septicemia) and localized infections outside the intestinal tract (for example: arthritis, endocarditis, pyelonephritis). These should in case of an acute infection also be reported. The reactive arthritis, which may also be caused by Salmonella infection, is not to be reported. Laboratory diagnosed Positive finding using the following method: • Direct detection of pathogen: isolation of pathogen (culture) Additional information: Results of identified serogroup and lysotype should also be reported. Epidemiological confirmation Epidemiological confirmation, defined as at least one of the following three constellations while taking into account the incubation period (about 6 to 72 hours): • Epidemiological link to another laboratory-diagnosed human infection through ❍ Person-to-person transmission OR ❍ Same source of exposure (e.g. animal contact*, food*) • Consumption of food (including drinking water), for which Salmonella spp. was laboratory-detected in non-consumed food. • Contact to animal (e.g. poultry) with a laboratory-detected infection, or contact to its secretions or consumption of its products (e.g. eggs). * terms marked with an asterix are defined in more detail in a glossary of the case definitions next level would have been forwarded according to the case definition category, given that the correct disease was identified. Inversely decision on reporting was also seen to be correct if a case that should not have been reported to the next level was in fact classified in a way that the case would have been held back. However, cases forwarded with wrong disease identification (see above) were a priori considered incorrect. Thus reporting was based on the question whether the case needed to be forwarded to the state level or not, which is a direct result of the disease identification and the case definition category. Sensitivity of reporting was defined as the number of cases that would have been correctly forwarded divided by the number of cases that should have been forwarded according to the gold standard. The positive predictive value of reporting was defined as the number of cases that should have been forwarded among those that would have been forwarded. Precision of reporting is defined as the number of cases that would have been either correctly forwarded to the state level, or would have been correctly held back at the local health department level, divided by the total number of case examples. Unless stated otherwise, reporting precision was the outcome parameter used in the following analysis. To specifically assess the effect of different styles in formulating case definitions, a fourth outcome variable was used. The clinical classification was considered correct if the part regarding the clinical picture was classified according to the gold standard, regardless whether other parts of the case definition were correctly classified or not. This analysis was done to compare case definitions with narrative description of the clinical pic-ture (as in all former IfSG case definitions) to case definitions with a more explicit check-list format of clinical criteria, that was implemented for diseases additionally notifiable in specific states and for the new IfSG case definitions. The case examples consisted in facsimile excerpts of one or more of the following sources: laboratory report form, physician form, and protocol of the patient interview [see additional file 1]. The case examples were created based on real cases that have come to the attention of the RKI in the quality control process and in the information service hotline that the RKI is offering to the health departments. The case examples were pre-tested among epidemiologists within the RKI and among epidemiologists and public health nurses in the state and local health departments. After the data of the respondents had been analyzed, the classification originally intended while creating the case definition, was challenged with the results of the respondents. Three epidemiologists then reassessed each individual case example and re-examined whether the classification originally intended was still justified. Based on this process the gold standard was defined for each case example. We compared the responses to the established gold standard and stratified by the following variables: health department being in an East German versus a West German state, disease of the case example, whether or not physicians participate in routine quality control of case reports (versus this being done exclusively by public health nurses), institutional level (local health depart-ment, state health department, RKI), acceptance and style of case definitions (check-list vs. text) and software used at local health department. Because of the selection and distribution of case examples described above, we conducted the individual analyses for each group. After univariate analysis we conducted a multivariate analysis using SPSS 13.0 for Windows (Version 13.0.1). The distribution of the classifications was compared to the gold standard, in order to identify common discrepancies. Based on these discrepancies we identified which part of the case definition was affected and identified specific aspects of the case definitions that had repeatedly been interpreted differently by the participants, indicating failure of the case definition to be unambiguous and reliable. These aspects were then summarized in order to deduct commonalities which could then lead to specific recommendations on how to improve this particular case definition and also on how to improve formulation of case definitions in general. In May 2002 -simultaneously with the Round-Robin test -we conducted a written survey addressed to all 425 local health departments in Germany. Among various questions on the structure and equipment of the local health departments, and their experiences with the new IfSG, we also asked about the profession of the person who had actually filled the questionnaire and about his or her attitudes and experiences towards the case definitions. The multivariate analysis was limited to data from the local health departments and without additional case examples for the East German states (n = 5995). Only statistically significant associations are mentioned in the following. The disease of the case examples was for all groups significantly associated with reporting (p < 0.001 in group 1, 2 and 4, p = 0.022 in group 3). Local health departments using the RKI-software showed a higher chance to identify the disease (disease identification) of the case example according to the gold standard compared to health departments using any of the commercially available software programs (group 2: OR = The administrative level at which the respondents worked, was significantly associated with the outcome reporting. For the analysis we used all cases of set A and set E (without the additionally diseases for the East German States, n = 2213). Adjusted for the diseases the chance of correct reporting to the next level was 1.5 times higher in cases done by state level staff compared to those done by local health department staff (OR = 1.52, CI: 1.14 -2.02). The following observations have been made in the qualitative analysis of the responses: • The concept of epidemiological confirmation was not well understood. For example travel in endemic countries was equivocally seen as an epidemiological confirmation (e.g. haemorrhagic fever and travel to Egypt). Re-evaluation of the case definitions showed that in fact there was only a vague definition of the epidemiological confirmation. • Participants appeared to have difficulties in deciding whether all clinical signs and symptoms mentioned in the case definition had to be existent in a case, or whether they were only listed as descriptive examples. • Case examples of diarrheal disease without any evidence of a specific pathogen, were frequently classified as salmonellosis. • Laboratory findings with only one elevated antibody value in serum were repeatedly classified as laboratory detection although the case definition required a rise in antibody level. • In some case definitions detection of the pathogen is only accepted if the detection was done in specific materials (normally sterile material such as blood for detecting N. meningitidis). This limitation was frequently neglected. • Some of the information in the case definition intended to serve as additional background information was mistakenly used as selection criteria (e.g. statement that clinician described rash as "very typical" for measles, but fever was missing). When asked about the availability of the case definitions, 395 (99%) of 398 local health departments responded that the case definition were accessible at the work place. The case definitions were seen as useful by 377 (95%) of 397 health departments who answered this question and not useful by 20 (5%). The clarity of the individual sections of the case definitions was rated differently: The section on the clinical picture of the case definitions was seen as unambiguous in all case definitions by 72 respondents (18%), in the majority of case definitions by 305 (76%), in the minority by 20 (5%), and in none of the case definitions by one (0.3%) of the respondents (n = 398 respondents). The section on the laboratory confirmation of the case definitions was seen as unambiguous in all case definitions by 137 respondents (34%), in the majority of case definitions by 248 (62%), and in the minority by 11 (3%) (n = 396). Three-hundred and three (87%) of 347 health departments stated that case classifications were done exclusively or primarily by public health nurses. With respect to the case examples presented to the participants, 220 (55%) of 396 respondents (from the local health departments) stated that the case examples were realistic. The results of our evaluation have shown that although case definitions may appear to be clearly defined, they may be interpreted quite differently by their users, which may result in severe misclassifications and reduced sensitivity and positive predictive value. This study is believed to be the first to systematically assess these effects Also the complexity of the case definition itself is likely to affect reporting precision. Unfortunately much of the complexity of the case definition is a result of methodological limitations of available laboratory tests and cannot be influenced. The case definition system with its three different types of evidence leading to five different categories may appear very complex and less intuitive that the classical categories of "suspect", "probable" and "confirmed". The detailed differentiation of the German case definitions however enables us to apply computer algorithms in order to translate these to the EU case definitions and thus make the data compatible to the standards of various European surveillance networks and to WHO reports. Reassessment of the gold standard after receipt of the responses resulted in modifications of 5 of the 68 case examples. This procedure took place in an initial review process of gold standard before the actual analysis was done. We believe it was legitimate and necessary in order to correct for biases caused by unforeseen ambiguity of the case examples. The software used at the local health department was significantly associated with the quality of the data in only some subgroups and outcomes. Apparently the software is not a very strong determinant in the given study design, although our experience in implementing the electronic surveillance system in Germany showed that commercially available software products often do not fully implement the standards published by the RKI for data transmission software or they do so with a delay of several years [21] . The other interesting finding is that the administrative level of the participants was significantly associated with the outcome: Participants from state health departments had a significantly higher rate of agreement with the reporting gold standard than the participants from local health departments. This might be explained by the fact that staff at the state level is generally higher trained in epidemiology and infectious diseases than local health department staff and they are routinely involved in quality control of incoming case reports and also training and supervision of local health departments' staff. All the observed quantitative effects and their propagated explanations merge into the one main conclusion: Case definitions must be very carefully formulated in order to assure their unambiguous interpretation by local health department personnel. The detailed evaluation of our study has resulted in a substantially revised edition of the German case definitions [23,24]: • We rephrased the case definitions in a check-list format indicating clearly how many of the symptoms and signs had to be fulfilled in which combination. • Some diseases previously jointly described in one case definition were defined separately (Dengue was separated from other haemorrhagic fever; hemolytic uraemic syn-drome was created new, separately from EHEC and Shigella.) • We rephrased the definitions in a way that for serologic confirmation the necessity for two samples is clearly apparent at the beginning of the phrase. • The material in which the pathogen has to be detected is now highlighted and is only listed if it is relevant for the case definition. • A glossary now defines the expressions that are being used repeatedly in the case definitions • The case definitions are now limited to criteria relevant for the decision process. All additional explanatory information is clearly indicated as such in a separate section of the case definition • The evidence type "epidemiological confirmation" was completely redesigned and replaces the previously used term "epidemiological link". The accepted types of epidemiological links are now specified individually for each case definition. One practical implication, that is supported by this analysis is, that software used at the local health department must be designed with strict accordance to the case definitions using identical terminology and structuring which would have been more easily archived if all local health departments had been equipped with one identical software system developed within or under supervision of one institution. Possibly other countries in the process of developing or implementing new electronic surveillance systems might want to learn form this experience [21,25]. The case example book, which resulted from this study, constitutes a detailed feed back for the participants of the study and is now being used as training material for public health nurses. We have demonstrated that rigorous reduction of case definitions to testable yes/no-criteria in a check-list format is likely to improve their reliability. Reducing the differential diagnostic complexity of a disease to a limited number of yes/no-criteria, is a major challenge, but it also carries the benefit of facilitating computerized testing algorithms for quality control and for case classifications. As the reliability of epidemiologic surveillance largely depends on the reliability of its case definitions, it is essential to create and revise case definitions based on systematic evaluations [9] . Most of the basic principles for the revision of the German case definition edition deducted from this analysis may also be applicable for case definitions in other countries (such as the United States, Ireland, Sweden, Mexico) or international networks (EU, WHO) as they share the same structural and editorial characteristics that we identified to be problematic in the first edition of the German case definitions [4, 7, 8, 26, 27] . We therefore believe that our findings are highly relevant for many national and international surveillance systems.
54
Ventilator associated pneumonia and infection control
Ventilator associated pneumonia (VAP) is the leading cause of morbidity and mortality in intensive care units. The incidence of VAP varies from 7% to 70% in different studies and the mortality rates are 20–75% according to the study population. Aspiration of colonized pathogenic microorganisms on the oropharynx and gastrointestinal tract is the main route for the development of VAP. On the other hand, the major risk factor for VAP is intubation and the duration of mechanical ventilation. Diagnosis remains difficult, and studies showed the importance of early initiation of appropriate antibiotic for prognosis. VAP causes extra length of stay in hospital and intensive care units and increases hospital cost. Consequently, infection control policies are more rational and will save money.
Nosocomial pneumonia (NP) is defined as parenchymal lung infection, occurring after the first 48 hours of hospital admission [1] . It accounts for 13-18% of all hospitalacquired infections, but leading cause of death from nosocomial infections [2] . It is a major threat to patients admitted intensive care units (ICU) and receiving mechanical ventilation (MV). In the recent studies, it was shown that ventilator associated pneumonia (VAP) was the most common infectious complication among patients admitted ICU [3, 4] . It results in high mortality and morbidity, prolonged lengths of hospitalisation, and also increased cost of hospitalisation. The mortality rates for VAP range from 20% to 75% according to the study population [5] [6] [7] [8] [9] [10] [11] [12] [13] . The average excess cost of pneumonia was estimated to be U.S.$1255 per patient in 1982 (14) , U.S.$2863 per patient in 1985 [15] , and in a recent study in 1999, it was >U.S.$40 000 per patient [16] . Despite the clinical experience and major advances in diagnostic techniques and management, VAP remains a significant problem for intensivists. In this review, epidemiology, diagnosis and, mainly, infection control of VAP were discussed. In different studies, the incidence of VAP was reported different, depending on the definition, the type of hospital or ICU, the population studied, and the type of rate calculated and varies from 7% to 70% [17] [18] [19] [20] [21] [22] [23] . In a large database, 1-day point prevalence study, conducted in 1417 European ICUs, pneumonia accounted for 47% of nosocomial infections [4] . In the National Nosocomial Infections Surveillance System (NNIS), NP accounted for 31% of all nosocomial infections in ICU [24] and in another NNIS data in medical ICUs, it was accounted for 27% [25] . The recent studies reported the device-associated incidence rate 13.2-51 per 1000 ventilator days [20, 23, 26, 27] . Generally, the rates of VAP in surgical ICU were higher than in medical ICUs, depending on the differences in the patient population, surgical disorders, the proportion of patients that needed MV and the duration of ventilation. Kollef et al. [28] reported incidences of NP of 21.6% in patients admitted to a cardiothoracic ICU, 14% in other surgical ICU, and 9.3% in a medical ICU. NP may occur by four routes; haematogenous spread from a distant focus of infection, contiguous spread, inhalation of infectious aerosols and aspiration. Aspiration of the pathogenic gram-positive and gram-negative bacteria, colonized on the oropharynx and gastrointestinal tract, is the main route. The role of other routes is very rare [1] . Once microorganisms reach the distal lung, they multiply and cause invasive disease. The host defence, including filtration and humidification of air in the upper airways, epiglottic and cough reflexes, ciliary transport by respiratory epithelium, phagocytes and opsonins in distal lung, and systemic cell mediated and humoral immunity, prevent bacterial invasion [29] . In ICU, the host defence of patients are usually altered because of their underlying diseases, and devices that are used. They can not cough efficiently due to sedation or underlying disease. And also, when they are intubated, the endotracheal tube holds the vocal cords open and facilitates aspiration. The most important risk factor for NP is tracheal intubation; associated with a 3 to 21 fold risk [30] [31] [32] [33] . It increases the risk by; 1) causing sinusitis and trauma to nasopharynx (nasotracheal tube), 2) impairing swallowing of secretions, 3) acting as a reservoir for bacterial proliferation, 4) increasing bacterial adherence and colonization of airways, 5) requiring the presence of a foreign body that traumatizes the oropharyngeal epithelium, 6) causing ischemia secondary to cuff pressure, 7) impairing ciliary clearance and cough, 8) causing leakage of secretions around the cuff, and 9) requiring suctioning to remove secretions [34] . Microorganisms can adhere to the surface of the endotracheal tube and some species exude an exopolysaccharide that acts as a slime-like adhesive. That microbial biofilm on the tube surface provides a reservoir of microorganisms, and they are greatly resistant to the action of antimicrobials and host defence [35] . Also, the patient requiring MV exposes to other devices, such as nebulizers, humidifiers, which can be the source of microorganisms. The duration of MV increases the risk of infection. Cook et al. [22] reported a cumulative increased risk of VAP with time, with 3% per day in the first week of MV, 2% per day in the second week, and 1% per day in the third week. In other studies, similarly, it was shown that the risk of pneumonia increased by the duration of MV and the highest risk was during the first 8-10 days [36] [37] [38] . The need for reintubation, urgent intubation and documented massive aspiration are also associated with high incidence of VAP [1, 21, 29, 39] . The effect of prior antibiotic therapy is still controversial. Sirvent et al. [40] reported that a short course of cephalosporin prophylaxis was associated with a lower rate of VAP in patients with structural coma. Also other investigators showed that antibiotics administered during the first days, reduced the risk of early-onset ventilator associated pneumonia [22, 37] . However, prior antibiotic exposure predisposes patients to subsequent colonization and infection with resistant pathogens [28, 41] . Nasogastric tubes impair the function of the gastroesophageal sphincter and increase the risk of maxillary sinusitis, oropharyngeal colonization and reflux, all of which may lead to migration of bacteria and pneumonia [29, 39] . Enteral nutrition given by nasogastric tube is also associated with increased risk of VAP. Moreover, it may predispose to VAP by elevating gastric pH, leading to gastric colonization and increasing the risk of reflux and aspiration by causing gastric distension [22, 42, 43] . Also patient transportation was found risk factor for VAP by facilitating the aspiration of contaminated secretions from the upper airway and the ventilator circuits in the supine position [44] . The other independent risk factors for VAP are shown in Table 1 [6, 7, 11, 22, 28, 33, 34, 45] . Identifying these risk factors will guide to prevention measures of VAP. The causative organisms vary according to the patients' demographics in the ICU, methods of diagnosis, the durations of hospital and ICU stays, and the antibiotic policy. Gram negative bacteria are the most common pathogens cause VAP in several studies [7, 17, 44] . In NNIS data, although the most frequent reported isolate was Staphylococcus aureus (17%), 59% of all reported isolates were gram-negative. The most common gram-negative species were Pseudomonas aeruginosa (15.6%), Enterobacter species (10.9%), and Klebsiella pneumoniae (7.0%) (24) . In recent years, gram-positive bacteria have become more common in ICU and in the EPIC study, S. aureus accounted for 31% of the 836 cases with identified microorganisms (46) . NNIS data from medical ICU, also, reported high percentage (20%) of S. aureus [25] . Polymicrobial infection rate is usually high in VAP [12, 17, 47, 48] . The duration of MV and the prior exposure to antimicrobials significantly influence the distribution patterns of etiologic agents. In early onset VAP (<5 days), methicillinsensitive S. aureus, Streptococcus pneumoniae, and Haemophilus influenzae are the most common pathogens, whereas methicillin-resistant S. aureus (MRSA), P. aeruginosa, Acinetobacter baumannii and Stenotrophomonas maltophilia are more frequent in late onset VAP (≥ 5 days) [49, 50] . Also, MRSA, P. aeruginosa, A. baumannii and the other multi-resistant gram negative pathogens are the most common pathogens in the patients expose to prior antibiotic. The other special factors that predispose patients to infection with specific microorganisms are summarized in Table 2 [29, [51] [52] [53] [54] [55] [56] . Determining risk factors for microorganisms will help to select appropriate antimicrobial treatment, that improves the outcome. The diagnosis of pneumonia in mechanically ventilated patients is difficult, and still there is no "gold-standard" diagnostic method. It is usually based on the combination of clinical, radiological, and microbiological criteria defined by Centers for Disease and Control (CDC) ( Table 3) . But these criteria have low sensitivity and specificity. The systemic signs (fever, leukocytosis, etc.) of infection can be seen by any condition in ICU (pulmonary edema, pulmonary infarction, after surgery, trauma, devascularized tissue, open wounds, etc.). Investigators reported that the clinical diagnosis of VAP is associated 30-35% falsenegative and 20-25% false-positive results [57] [58] [59] . And also, ICU patients do not always have systemic signs of infection due to their underlying disease (chronic renal failure, immunosuppresion, etc.). Radiological infiltration has limited value, mimicking by cardiogenic pulmonary edema, noncardiogenic pulmonary edema, adult respiratory distress syndrome (ARDS), atelectasis, pulmonary contusion, which are not uncommon in ICU [29] . In an autopsy-proven VAP study, Wunderink et al. [1] reported that no radiographic sign had a diagnostic accuracy greater than 68%. The presence of air bronchograms was the only sign that correlated well with pneumonia, correctly predicting 64% of pneumonias in the entire group. The upper respiratory tract of patients is colonized with potential pulmonary pathogens a few hours after intubation [48, 61] . Consequently, isolation of pathogens from tracheal secretions do not always indicate pulmonary infection. But a positive Gram's stain may guide the initial antibiotic therapy. However prior antibiotic and corticosteroid therapy can reduce the sensitivity of this technique [62, 63] . Pugin et al. [64] proposed to combine the seven variables (temperature, leukocytes, tracheal aspirate volume and purulence of tracheal secretions, chest X-ray, oxygenation-PaO 2 /FiO 2 -and semiquantitative culture of tracheal aspirate) for the diagnosis of VAP, defined as clinical pulmonary infection score (CPIS) ( Table 4 ). The score varied from 0 to 12 points and was reported that a CPIS of more than six was associated with a sensitivity of 93% and a specificity of 100% for the diagnosis of pneumonia. In a post mortem study, Papazian and colleagues [65] reported a high diagnostic accuracy of CPIS at a threshold of 6 (72% sensitivity and 85% specificity). However, the original scoring system has some limitations; that it requires 24-48 hours for the results of tracheal aspirate cultures, and also identifying pulmonary infiltrates progression depends on intensivist experience. Singh et al. [66] used a modified CPIS (calculated at baseline from the first five clinical variables, and CPIS at 72 hours was based on all variables of the score) that antibiotics were stopped in patients with a persistent low score (<6) after 3 days of therapy, avoiding unnecessary use of antibiotics, and all patients who discontinued the therapy improved. In a recent study, Fartoukh et al. [67] reported that the modified CPIS does not perform better when the clinical suspicion of pneumonia is high, so they proposed incorporating the results of specimens gram stain (by adding two more points when gram stains were positive) to modified CPIS to increase the sensitivity of the score and the physicians' diagnostic accuracy. Qualitative cultures of tracheal aspirate (TA) is not a specific diagnostic method because of the lower respiratory tract colonization and a high percentage of false-positive results [48, 68] . However, investigators reported that quantitative cultures of TA have equal diagnostic accuracy to the other invasive techniques [69] [70] [71] [72] [73] . In a recent study, quantitative cultures of TA were compared with plugged telescoping catheter (PTC). The specificity of TA was similar to PTC when a cut-off point of 10 6 cfu/mL or higher was used, although the sensitivity of TA at ≥ 10 6 cfu/mL was lower than PTC. But when a cut-off point of 10 5 cfu/ mL was used, the sensitivity of TA was not statistically different from that of PTC [74] . Although, quantitative cultures of TA is non-invasive, inexpensive and a simple method, it has some risks, that if the cut-off value ≥ 10 6 cfu/mL is used, sensitivity will be low and some patients with VAP may not be identified or when the cut-off value ≥ 10 5 cfu/mL is used, unnecessary antibiotic treatment will be given because of low specificity [75] . In the recent years, many investigators favour invasive techniques for diagnosis of pneumonia (protected-speci-men brush -PSB-or bronchoalveolar lavage-BAL-) that may have more diagnostic accuracy [76] [77] [78] [79] [80] . In PSB, 0.001 mL of secretions are collected and the presence of >10 3 cfu/mL bacteria has 80-90% sensitivity and 95% specificity for the diagnosis of VAP. In BAL, larger proportion of lung can be sampled and the diagnostic threshold is >10 4 cfu/mL. The sensitivity and specificity of BAL are 86-100% and 95-100%, respectively [76, [81] [82] [83] . Heyland and colleagues [84] proposed that PSB or BAL may increase physician confidence in the diagnosis and management of VAP and allows for greater ability to limit or discontinue antibiotic therapy. Also in this study, patients who underwent bronchoscopy with PSB and BAL had a lower mortality rate compared with patients who did not undergo bronchoscopy. However, a recent meta-analysis concluded that regular use of bronchoscopy for the diagnosis of VAP does not alter mortality, because it does not directly affect the initial antibiotic prescription [85] . The disadvantages of these invasive techniques are [71, 86, 87] ; a) prior antibiotic use may decrease the sensitivity and accuracy of these methods. However, in a recent study, Souweine et al. [88] reported that if a current antibiotic treatment prescribed for a prior infectious disease other than VAP, the diagnostic accuracy of protected specimen brush or bronchoalveolar lavage is not changed, b) these techniques are based on quantitative culture and results of these cultures require 24-48 hours, and, therefore miss early cases, and also give no information about appropriate initial antibiotic therapy, c) these invasive tests may worsen the patient's status (cardiac arrhythmias, hypoxemia, bleeding, etc.), d) increase the costs of caring, e) it has not been proven that the use of these invasive techniques lead to a decrease in patients' mortality. The spread of microorganism to blood or pleural space is <10%, so blood and pleural effusion cultures have low sensitivity and specificity. Luna and colleagues [89] demonstrated that the positive predictive value of blood cultures to detect the etiologic microorganism was 73% and the sensitivity of blood cultures was only 26%. They con- Three or more of the following criteria: Rectal temperature >38°C or <35.5°C Blood leucocytosis (>10.10 3 /mm 3 ) and/or left shift or blood leukopenia (<3.10 3 /mm 3 ) More than ten leukocytes in Gram stain of tracheal aspirate (in high power field) Positive culture from endotracheal aspirate AND New, persistent, or progressive radiographical infiltrate cluded that blood cultures in patients with VAP are useful if there is suspicion of another probable infectious condition, but the isolation of a microorganism in the blood does not confirm that microorganism as the pathogen causing VAP. Therefore, two sets of blood samples for culture and tapping pleural effusions >10 mm should be performed in patients suspected VAP [50] . In conclusion, microbiological testing should be always performed to decide the appropriate initial empirical antibiotic therapy. Clinicians can choose optimal diagnostic test for specific patients in their clinical setting. Because of longer duration of mechanical ventilation, longer stay in the ICU, increased use of antibiotics, higher costs for healthcare, and most importantly, increased mortality, the prevention of VAP is the major priority. But, despite the advances in the pathogenesis of VAP, intensivists still struggle with the prevention strategy. Basic hygiene principles of infection control (hand washing/disinfection just before and after each patient contact, the use of glove and sterile equipment) remain important for the prevention of VAP. Healthcare workers (HCW) can spread microorganisms from patient to patient by their hands easily. Although HCWs realize the importance of handwashing/disinfection, their compliance is still low (25-40%) [90] [91] [92] . Especially their compliance rate is lowest in activities that carried higher risk for transmission and in ICU. High workload decreases their compliance [93] . Wrist watches, bangles, and other jewellery act as reservoirs for organisms, and inhibit effective hand cleaning [94] . Therefore, staff have to take off wrist watch and jewellery to achieve effective hand cleaning. They have to use gowns and gloves when appropriate and must change and wash/disinfect their hands between patients [95] . Bedside hand antiseptics (alcohol-based handrub solution), easier access to sinks and availability of washing equipment, decrease in workload, communication and education tools (posters) and feedback improve compliance and decrease the cross-transmission of nosocomial infection [90, 96] . The internal machinery of mechanical ventilators is not an important risk factor for VAP. Therefore, using a filter between the inspiratory phase circuit and the patient is not necessary. Furthermore, the importance of filters on the expiratory limb of the mechanical-ventilator circuit in preventing cross comtamination has not known and needs further investigation [34] . The devices used on the respiratory tract come into contact with mucous membranes, therefore cleaning and highlevel disinfection (at 75°C for 30 minutes) of reusable equipments are required [97] . Resuscitation bags, spirometers, and oxygen analyzers must be cleaned and disinfected between patients to avoid cross-transmission [95] . In several studies, routine changing of the ventilator circuits is not recommended [26, [98] [99] [100] [101] [102] [103] [104] . Replacement is required when there is gross soilage and mechanical malfunction [105] . The condensate fluid in the ventilator circuit, which contains high concentration of pathogenic bacteria and a risk factor for VAP, has to be removed regularly [106] . And accidental drainage of condensate into the patient's airway and contamination of caregivers during ventilator disconnection or during disposal of condensate should be avoided. In-line devices with one-way valves, put in place into disposable circuits and emptied regularly, are recommended for collecting condensate [75] . Humidification of the inspired air is an important care in the ventilator management. Humidification may be achieved by active humidifiers (bubble-through or wick) or passive humidifiers (hygroscopic condenser-artificial nose-or heat-moisture exchanger-HME-). In humidification, formation of condensate in tubing and colonization of this condensate with microorganisms is an important risk factor for VAP. HME recycle heat and moisture that reduces the condensate formation and also bacterial colonization in the circuit. Moreover they have bacterial filtration characteristics [34, 98] . They do not need to be changed daily and can be used for at least 48 hours, sometimes for up to 1 week [107, 108] . Also with other advantages (reduced nurses workload, reduced financial cost, and better safety), HMEs are favourable devices in many ICUs. Indeed several investigators [109] [110] [111] [112] reported lower rates of VAP in HME groups than conventional heated-water humidification systems, the effect of HME on the prevention of VAP is still controversial and a recent study showed no significant difference in VAP rates [113] . Furthermore, HMEs increase dead space and resistance to breathing, cause airway occlusion, and more tenacious secretions [34, 98, 104, 111] . Additional studies are needed to identify the benefits of HMEs on infection control of VAP. Nebulizers are used for medication or humidification of air, and inserted into the inspiratory phase of the mechanical ventilator circuit. They can be contaminated by condensate in the tube or by using contaminated solutions, and inoculate infectious aerosol particles directly into the lung parenchyma and can cause outbreaks in ICUs. Recommendations for the infection control of nebulizers are; a) filling immediately before use, b) using sterile water and drugs, c) never refilling the liquid to be nebulised, d) cleaning and disinfecting the receptacle daily, d) using sterile water for rinsing and allowing to dry, e) using patient-specific large-volume nebulizers, f) using patient specific mask, mouthpiece, connecting pieces and medicine cups [34, 114] . Suctioning the secretions in the trachea is another approach to VAP prevention. Two types of tracheal suction catheters are used on ventilated patients; the open, single-use catheters and the closed, multiple-use catheters. In single-use system, HCWs have to use sterile solutions during rinsing these catheters and have to care aseptic technique when suctioning endotracheal secretions. In closed suctioning systems, secretions can be suctioned without removal of mechanical ventilation support. This may cause less hypoxia, hypotension and arrhythmias, and also less environmental contamination [115, 116] . However similar VAP rates with closed and open system were suggested in the earlier trials [115, 117] , Combes and colleagues [116] reported a 3.5 times greater risk of VAP in open suctioning system than closed suctioning system in a recent study. Indeed, closed suction catheter is an extension of the ventilator circuit, daily change of this catheter is not necessary for infection control, and in one study no significant difference in VAP rate was reported when daily changes were compared with no routine changes, that may decrease the costs [118] . The use of closed suction system is recommended as part of a VAP prevention program [104] . The relationship between the use of invasive devices and nosocomial pneumonia directed the investigators to use noninvasive ventilation to reduce the VAP rates. In several studies, lower risk of VAP, with less antibiotic use, with a shorter length of ICU stay, and with lower mortality were reported in the use of non-invasive ventilation [119] [120] [121] [122] [123] . Therefore, care can be taken to use non-invasive mechanical ventilation more often, and to reduce the frequency of tracheal intubation. Endotracheal tube alters host defences, impairs mechanical clearance from the respiratory tract, causes local trauma and inflammation, and allows pooling of secretions around the cuff. The pressure of the endotracheal tube-cuff should be sufficient to prevent the leakage of colonized subglottic secretions into the lower airway [37] . Also, continuous or intermittent suctioning of oropharyngeal and upper respiratory tract secretions above endotracheal cuffs can prevent aspiration. Endotracheal tubes with a separate dorsal lumen above the cuff, designed to suction the subglottic secretions continuously, were found able to decrease the rates of early-onset VAP [124, 125] . However, in another randomised trial, no benefit with continous subglottic suction on the overall VAP frequency was found [126] . It can reduce but not eliminate the volume of fluid aspirated into the lungs. The lack of effect on prevention of late-onset pneumonia and the high cost of these tubes restrict the usage of them. Microbial biofilm on the endotracheal tube surface is a reservoir for the pathogens and prevent the microorganisms from the action of antibiotics [35] . Adair and colleagues [127] proposed that high concentrations of antibiotic on the endotracheal luminal surface, achieved either by nebuliser or endotracheal surface modification, would be expected to prevent biofilm formation on the endotracheal tube and may have a role in reducing the incidence of VAP, also minimising patient exposure to systemic antibiotics. Nasotracheal intubation increases the risk of nosocomial sinusitis, that may predispose VAP by the aspiration of infected secretions from nasal sinuses [128, 129] . There- fore, endotracheal intubation should be preferred to decrease the risk of VAP. As nasotracheal intubation, nasogastric tube can cause orophanryngeal colonization and nosocomial sinusitis. By impairing the function of the upper oesophagus sphincter, it may facilitate the reflux of bacteria from the gut. As a result, it increases the risk of VAP [130, 131] . In a randomized study, gastroesophageal reflux and microaspiration of gastric contents to the lower airways were not influenced by the size of the nasogastric tube. Because of their potential complications (tracheal malposition by coiling and clogging) and high cost, the small-bore nasogastric tubes are not routinely recommended for the prevention of VAP [132] . Poor nutritional state and hypoalbuminemia contribute the development of VAP. For this reason, early initiation of enteral nutrition may have preventive effect in mechanically ventilated patients. Moreover, it helps to maintain the gastrointestinal epithelium and reduces the need for stress-bleeding prophylaxis. However, because of using nasogastric tubes and alkalinization of the stomach contents by these feeds; gastric colonization, gastrooesophageal reflux, aspiration and pneumonia might be promoted [29] . In a recent study, postpyloric enteral access placement improved tube-feeding tolerance and reduced the rates of VAP [133] . Heyland et al. [134] used acidified feeds in critically ill patients and demonstrated a dramatic reduction in bacterial growth from aspirates of stomach contents and a lower rate of gram-negative bacterial growth in tracheal secretions in patients receiving acidified feeds, but not a significant reduction in nosocomial pneumonia. They cannot be used in patients with active gastro-intestinal bleeding, acidemia, or renal failure. Furthermore, it was a small size study and further research on the effect of prevention is needed, before it is used in practice. In the recent years, selective decontamination of the digestive tract (SDD) is one of the most extensively studied prevention strategy in VAP. In SDD, topical non-absorbed antimicrobials (usually combining polymyxin, aminoglycoside and amphotericin B) are used to prevent gastrointestinal colonisation by pathogenic microorganisms. It selectively eradicates the potential pathogenic microorganisms (gram-negative aerobic intestinal bacteria, S. aureus and fungi) and does not affect anaerobic flora, as elimination of anaerobic flora leads to increased colonization with gram-negative aerobic flora. Although some investigators used only topical antibiotics applied to the oropharynx and through a nasogastric tube, many of them added systemic therapy with a broad spectrum (e.g. cefotaxime) during the first few days, to prevent early infections with S. pneumoniae, H. influenzae and S. aureus [29] . In a recent meta-analysis that searched 33 randomized controlled trials published from 1984 to 1996, significant reductions in the incidence of respiratory tract infections (65%) and in total mortality (20%) were determined [135] . Also in this meta-analysis and in the other recent prospective, randomized studies, it was mentioned that using only topical antibiotics reduced respiratory infections, but did not influence the survival [135] [136] [137] . The threat of SDD is to lead the selection and overgrowth of antibiotic resistant microorganisms. In a recent study from the Netherlands, where the incidence of MRSA and vancomycin resistant enterococcus (VRE) are very low, a reduction in the frequency of colonization with resistant gram-negative bacteria and no effect on the acquisition of MRSA were reported (138) . But the studies from the ICUs where MRSA was endemic, increased incidence of MRSA was reported (139, 140) . Therefore, in ICUs with high incidence of multi-resistant microorganisms, SDD cannot be used. On the other hand, in trauma and surgical patients, SDD seems more effective than in medical patients, may be due to less colonisation [141] . In conclusion the routine use of SDD in ICUs is not recommended, it should be decided according to the patient population studied and the characteristic of ICU. In addition, the colonization of the oral cavity with pathogens is important risk factor for the develoment of VAP, it is unclear if oral care with chlorhexidine reduces VAP. Also the concern over a chlorhexidine-related increase in colonization of gram negative bacteria should be considered [142] . Semirecumbent position (45°) prevents aspiration and the passage of bacteria into the airways, and should be preferred in ICU patients, if there is no contraindication [143] . "Kinetic Beds" or Continous Lateral Rotational Therapy (CLRT) turn continuously and slowly and change the patient's position. Investigators believe that it helps the drainage of pulmonary secretions. However, these beds are so expensive and their effectiveness are not demonstrated. So, the routine use of these beds is not recommended [97] . Also, chest physiotherapy, to improve the clearance of secretions, for the prevention of VAP is not recommended because of its lack of proven benefits and the associated risks (e.g., arterial oxygen desaturation) [105, 144] . Stress ulcer prophylaxis is proposed to be a risk factor due to the alkalinization of gastric content. The effect of stress ulcer prophylaxis with H 2 -antagonists or antacids on VAP is still controversial. In some studies [145, 146] , the use of sucralfate was associated with a decreased incidence of VAP, however the other reports did not support this [147, 148] . Also, H 2 -antagonists are more efficient for anti-ulcer prophylaxis than the sucralfate [148] . Therefore, the choice of agent for prophylaxis should be done according to the patient and cost-effectiveness. To reduce the aspiration of oropharyngeal contents, over use of sedatives should be avoided. Kress et al. [149] reported that for reducing over use of sedatives, daily interruption of sedative-drug infusions until the patients were awake decreased the duration of mechanical ventilation and the length of stay in the ICU.
55
Sentinel surveillance for human enterovirus 71 in Sarawak, Malaysia: lessons from the first 7 years
BACKGROUND: A major outbreak of human enterovirus 71-associated hand, foot and mouth disease in Sarawak in 1997 marked the beginning of a series of outbreaks in the Asia Pacific region. Some of these outbreaks had unusually high numbers of fatalities and this generated much fear and anxiety in the region. METHODS: We established a sentinel surveillance programme for hand, foot and mouth disease in Sarawak, Malaysia, in March 1998, and the observations of the first 7 years are described here. Virus isolation, serotyping and genotyping were performed on throat, rectal, vesicle and other swabs. RESULTS: During this period Sarawak had two outbreaks of human enterovirus 71, in 2000 and 2003. The predominant strains circulating in the outbreaks of 1997, 2000 and 2003 were all from genogroup B, but the strains isolated during each outbreak were genetically distinct from each other. Human enterovirus 71 outbreaks occurred in a cyclical pattern every three years and Coxsackievirus A16 co-circulated with human enterovirus 71. Although vesicles were most likely to yield an isolate, this sample was not generally available from most cases and obtaining throat swabs was thus found to be the most efficient way to obtain virological information. CONCLUSION: Knowledge of the epidemiology of human enterovirus 71 transmission will allow public health personnel to predict when outbreaks might occur and to plan interventions in an effective manner in order to reduce the burden of disease.
Hand, foot and mouth disease (HFMD) is a common acute viral illness that primarily affects infants and young children, and often occurs in clusters or outbreaks. It is characterized by rapid onset of fever and sore throat, accompanied by vesicles and ulcers on the gums, tongue, buccal mucosa and palate. Punctate and usually transient skin lesions appear on the palms, soles and occasionally on the buttocks, knees or other areas. While the fever and rash may subside rapidly, the mouth lesions may last more than a week, and virus may continue to be shed for several weeks [1] . In temperate countries HFMD occurs during the summer but in the tropics HFMD can occur at any time during the year. The major causative agents of HFMD are coxsackievirus A16 (CVA16), human enterovirus 71 (HEV71) and coxsackievirus A10 (CVA10) of the genus Enterovirus in the family Picornaviridae [2] . Other enteroviruses isolated from HFMD cases are the other species A human enteroviruses such as coxsackievirus A (CVA) 4, CVA5, CVA6 and CVA7, and coxsackievirus B (CVB) 1, CVB2, CVB3 and CVB5 [2] [3] [4] . Unlike other aetiological agents of HFMD that normally cause mild disease, HEV71 infection has been reported to cause neurological disease manifesting as aseptic meningitis, encephalitis or poliomyelitis-like acute flaccid paralysis [5] . First isolated from a child suffering from encephalitis in California in 1969, HEV71 was further isolated from 23 cases with severe neurological disease in California during the next three years [3] . Historically, HEV71-associated outbreaks have been reported in Australia in 1972 [6] , Japan in 1973 and 1978 [7] , Bulgaria in 1975 [8] and Hungary in 1978 [9] . In the past decade, countries in the Asia-Pacific region have experienced an increased occurrence of HEV71-associated HFMD outbreaks [10] . HEV71 outbreaks have been reported in Sarawak in 1997, Taiwan in 1998, Perth in 1999, then in Singapore, Korea, Malaysia and Taiwan in 2000 [11] [12] [13] [14] [15] [16] [17] . In an outbreak of HEV71 in 1997 in Sarawak, a state of Malaysia on the island of Borneo, a cluster of unusual paediatric deaths due to encephalitis and cardiac failure was observed [12, 13] . This raised a lot of fear and anxiety and because of the heightened concern about HEV71 in Sarawak, we implemented a sentinel surveillance programme for HFMD beginning in March 1998. This programme was set up as part of the operational functions of the Sarawak Health Department and was approved by the Director of Health. The principles of the Helsinki Declaration were followed throughout the surveillance operation. Our aims were to investigate the epidemiology of this common childhood disease in Sarawak, and to determine if there were any differences in the patterns of transmission of HEV71, CVA16 and other aetiological agents of HFMD. It was also the aim of this programme to provide data of practical value for doctors and public health personnel with a view to efficient and effective virological surveillance of HEV71, in particular, to provide an early warning system for HEV71 outbreaks. This paper describes the preliminary observations from our surveillance programme from March 1998 through June 2005. In early 1998 after discussion with a number of community paediatricians, our team set up a protocol for a sentinel surveillance programme for HFMD. The doctors who had consented to actively participate were provided with a standard reporting and specimen collection form, sterile swabs and virus transport medium, a telephone number for obtaining assistance for transport of specimens to the laboratory and a facsimile number to report cases to the Health Department. All sentinel clinic doctors obtained parental consent before swabs were taken. Sentinel clinic doctors were provided with feedback on viruses isolated from their patients and contact was maintained through both outbreak and inter-outbreak periods to assure doctors that the surveillance programme was active and ongoing. Data obtained in this manner were expected to provide accurate information about disease trends and molecular epidemiology of the relevant viruses. Three specialist paediatric clinics located in the towns of Kuching and Sibu in the state of Sarawak actively participated in this study from March 1998. In 2000 we included a fourth specialist clinic in Sibu. Two government polyclinics in Kuching and Sibu also participated as sentinel clinics. All children presenting to the sentinel clinics with a history of oral or other skin lesions typical of HFMD were enrolled into the surveillance study, and throat and rectal swabs were obtained from each child enrolled in the first 18 months. Where possible, swabs were also obtained from mouth ulcers, vesicles and other skin lesions. After a preliminary analysis of data from the first 18 months, the protocol was modified to require only throat swabs from sentinel clinics. Rectal swabs were optional and doctors were requested to provide vesicle swabs whenever possible. Specimens were to be transported on ice to the laboratory in 2 ml of viral transport medium (VTM) where they were vortexed, freeze-thawed and aliquoted. Since the primary objective of programme was a sentinel surveillance system for HEV71 HFMD, we inoculated specimens into human rhabdomyosarcoma (RD) cells susceptible to both CVA16 and HEV71. It was not a particular objective of this exercise to identify the minor causative agents known to be associated with HFMD and hence we did not include multiple cell lines as part of our virus isolation protocol. RD cell cultures normally showed the characteristic enterovirus CPE in 2 to 10 days and were harvested after the monolayer showed extensive CPE. A blind passage was done with all cultures showing no CPE after 10 to 14 days. RNA was extracted from all culture harvests using Tri Reagent LS (Molecular Research Centre, Cincinnati, OH, USA) according to the manufacturer's instructions. The dry RNA pellet was dissolved in 20 μl of sterile ultra high quality RNase-free water and stored at -80°C until use. The presence of enterovirus RNA in culture fluids was determined by a previously described pan-EV RT-PCR method [18] with some modifications. The duration of all the steps in the PCR was reduced to one minute and the final extension was reduced to 5 minutes. From 1998 through 2002, specimens positive using the pan-EV primers were tested for the presence of HEV71 genome by RT-PCR using the primers 159S and 162A, which anneal to the VP1 gene of HEV71. Dr. Mark Pallansch (Centers for Disease Control, Atlanta) generously made the primer sequences available to us prior to publication [19] . All PCR products were sequenced to confirm the identification. In 2002, we changed our protocol for identification of HEV71 due to problems of misidentification of local strains of CVA16 as HEV71 using the primer set 159S/162A [20] . Currently, HEV71 specific primers designed in-house are used for specific identification of HEV71 [20] . All primers used are listed in Table 1 . Sequencing reactions were performed using the Big Dye Terminator Cycle Sequencing Kit version 3.0 or 3.1 (Applied Biosystems, Foster City, CA, USA). Molecular serotyping of non-HEV71 enteroviruses isolated was carried out using the methods and sequences published by Oberste and colleagues [21] and Chu, Ishiko and colleagues [22, 23] . Prior to 2002, serotyping of selected non-HEV71 enteroviruses was performed exclusively according to Oberste's method. When Ishiko's method [23] was published in 2002, we made a comparison of the methods by serotyping new isolates using both methods. We determined that for human species A enteroviruses circulating in our region, Ishiko's primers gave identical serotype identification to that obtained using Oberste's method [21] . Since there was no discrepancy between the methods, we modified Ishiko's primers to convert the method from a semi-nested to a non-nested method for ease of use. Our modifications were verified and described by Cardosa and colleagues [24] . We then retrospectively retested all non-HEV71 enteroviruses isolated prior to 2002 using the modified method. A subset of isolates identified by sequencing of VP4 were subjected to confirmation by using VP1 specific primers and DNA sequencing of the PCR products, as described previously [24] . DNA sequences of VP1 and VP4 gene products generated by RT-PCR from isolates were used in this analysis essentially as previously described [24] . The software package ClustalX [25] was used for alignment and to generate a bootstrapped phylogenetic tree using the neighbour joining method according to Saitou and Nei [26] .Primers and DNA sequences TTCCAATACCACCCCTTGGATGA Antisense (1195-1173) HEV71 VP4 gene. 159 [19] ACYATGAAAYTGTGCAAGG Sense (2385-2403) HEV71 VP1 gene. 162 [19] CCRGTAGGKGTRCACGCRAC Antisense (2869-2850) HEV71 VP1 gene. 161 [19] CTGGGACATAGAYATAACWGG Sense (2766-2785) HEV71 VP1 gene. NP1A [19] GCICCICAYTGITGICCRAA Antisense (3355-3336) HEV71 VP1 gene. MAS01S [20] ATAATAGCAYTRGCGGCAGCCCA Sense (2355-2377) partial VP1 gene. MAS02A [20] AGAGGGAGRTCTATCTCYCC Antisense (2731-2712) partial VP1 gene. MD91 [18] CCTCCGGCCCCTGAATGCGGCTAAT Sense (450-474) partial 5UTR. MD90 [18] ATTGTCACCATAAGCAGCCA Antisense (603-584) partial 5UTR. *Position relative to the genome of HEV71 strain 7423-MS-87 (GenBank accession number U22522) All primers used in the methods described above are listed in Table 1 . All new DNA sequences used in the phylogenetic analysis but hitherto unpublished have been deposited in GenBank and have the accession numbers AY794032, AY794033, AY794035 to AY794042. All other sequences used are from previous publications by our own as well as other groups [17, 20, [22] [23] [24] 27] . Detailed protocols, sample collection methods and other practical information have been placed in the public domain through our APNET (The Asia-Pacific Enterovirus Surveillance Network) website [28] . Statistical analysis was performed using the software package JMP Statistics version 5.01(SAS Institute Inc., USA) and Prism 4 for Macintosh (Graphpad Software, Inc., USA). The first provisional protocol we provided to the sentinel clinics for the collection of specimens required both throat and rectal swabs and vesicle or ulcer swabs where possible. Results from virus isolation studies of specimens obtained from both sentinel clinics as well as hospitals during this period were used to review the protocol that was originally implemented. A total of 579 specimens from 263 children with a clinical diagnosis of HFMD were received during the 18-month period from March 1998 through August 1999. The age of the children ranged from 6 months to 13 years, with 153 (58.2%) males and 110 (41.8%) females. All specimens received were subjected to virus isolation. Fifty specimens from 44 children, of a total of 579 (8.6%) specimens, were too heavily contaminated with bacteria. Twenty four of the contaminated specimens were rectal swabs, 19 were throat swabs and 7 were from various skin lesions. All remaining uncontaminated cell culture harvests were tested for enteroviruses by using the pan-EV set of primers and 235 of the 529 (44.4%) specimens tested yielded an enterovirus, but only 15 of the 235 (6.4%) enteroviruses were HEV71. These specimens were from 259 children and an enterovirus was isolated from 153 (59.1%) children. Only 6 (3.9%) of the children had HEV71. From this early set of specimens, we were able to isolate an enterovirus from 44% of the throat swabs, 40% of the rectal swabs, 44% of the mouth ulcers and 66% of the vesicle swabs. Clearly vesicle swabs are very useful specimens, but only 18% of the children had had vesicle swabs taken because not all children presented with skin lesions filled with abundant fluid. Since throat swabs provided a reasonably high yield of enterovirus isolates, we made the decision in 2000, to require throat swabs as the primary specimen from the sentinel clinics, with vesicle swabs where possible, while rectal swabs were not required. This served to reduce the laboratory workload during an outbreak. Our laboratory received 4290 specimens from 2950 children from March 1998 through June 2005, with a male to female ratio of 1.35:1. The histogram in the top panel of Figure 1 shows the distribution of HFMD cases seen in our sentinel clinics during this period. There have clearly been two large outbreaks of HFMD in 2000 and 2003 (bottom panel of Figure 1 ), with some sporadic activity between these peaks. The dominant enterovirus serotype isolated during both the outbreaks was HEV71 as shown in the middle panel of Figure 1 . CVA16 was always isolated during HEV71 outbreaks as well but was also isolated in interoutbreak periods. Other species A human enteroviruses such as CVA2, CVA4, CVA5, CVA10 and CVA12 were also isolated in inter-outbreak periods. Phylogenetic analysis of the HEV71 strains isolated in Sarawak from 1998 to 2005 show that both genogroup B and genogroup C strains circulated in Sarawak during this period (see Figures 2 and 3) . We have used both VP4 and VP1 genes in the phylogenetic analysis in order to be certain that there is no major discrepancy in genotyping associated with using VP4 and VP1 gene regions. Furthermore, we wish to provide both options to other groups who may wish to compare their data with ours since it is known that many groups may still use VP4 sequencing as a first step in molecular identification of human enteroviruses. Although both genogroup B and genogroup C HEV71 strains co-circulated in Sarawak, the predominant genogroup in both the HEV71 outbreaks of 2000 and 2003 was genogroup B. Besides co-circulating with genogroup B strains during outbreaks, genogroup C viruses also appeared sporadically between outbreaks along with other species A human enteroviruses. We never isolated a genogroup B HEV71 in non-outbreak periods. The distribution of genogroup B and genogroup C HEV71 strains during the surveillance period is shown in the bottom panel of Figure 1 . The phylogenetic trees in Figures 2 and 3 The HFMD epidemiological curves for the outbreak years 2000 and 2003 were plotted according to epidemiological week ( Figure 4 ) and show clearly that the first HFMD cases began to be seen early in the year. By week 7 a clear rise in the number of cases was seen. This early rise in cases differs from the summer outbreaks seen in temperate countries, and we suggest that the HEV71 outbreaks in Sarawak preceed the summer outbreaks of countries in the northern hemisphere in each year. In 2000 the HFMD outbreak stretched to the end of the year, peaking between Phylogenetic tree generated from the VP1 gene, showing relationships between HEV71 isolated in different years Figure 1 ). In 2003, the number of HEV71 cases declined sharply by the end of April (by week 18) and were no longer detected by the end of June, coinciding with the last HFMD cases seen that year. Interestingly this outbreak coincided with the SARS outbreak in the region and the public health measures put into place during this time evidently served to control the transmission of enteroviruses as well. A detailed analysis was done on data collected from two sentinel clinics, coded S1 and S2, which had sent samples to our laboratory consistently and reliably throughout the seven-year study period. A total of 2570 specimens were collected from 1894 cases during the 7 years. Of the 1894 cases, specimens from 1804 (95%) were subjected to virus isolation. A total of 2272 specimens were subjected to virus isolation, thus ensuring that the majority of specimens from the majority of cases were tested (88.4% of specimens from 95% of cases). An analysis of the proportion of the different types of specimens and the virus yield obtained is shown in Table 2 . More than 2000 specimens from outbreak and nonoutbreak periods were tested from 1998 to 2005. Enteroviruses were grown from 21.6% of those tested. Throat swabs comprised 72.3% of the total number of specimens tested and 25.4% of these yielded enteroviruses. Detailed information about the enterovirus serotypes isolated during this surveillance programme is also provided [see additional file 2] . Although on the whole, the virus isolation success rate was much lower than anticipated from the results for the first 18 months, it remained the case that throat swabs were more useful than the rectal swabs which yielded non-polio enteroviruses in only 5.8% of the samples tested. It should be noted however, that the first 18 months of the study coincided with an inter HEV71 epidemic period, with mostly CVA16 and non-HEV71 species A human enteroviruses causing HFMD. We have compared the virus isolation yields during HEV71 outbreak (2000 and 2003) years with the yields during an HFMD outbreak caused by non-HEV71 enteroviruses (2002) and we found that we successfully isolated virus from 40% of specimens collected in 2002 but only 20% of viruses during the HEV71 outbreak years, suggesting that HEV71 is more difficult to isolate than CVA16 and other species A enteroviruses. The virus isolation rate in the first 18 months (44%) is therefore comparable to that obtained later, when HEV71 was not circulating. There were 491 children from whom a non-polio enterovirus was isolated. Of these, 8 were excluded from the analysis because of missing information on their age at presentation. The children ranged in age from 18 days to 155 months, with a mean of 32.2 months and a median of 27.5 months. There were 3 dominant serotypes of enteroviruses isolated from these 483 children and we asked the question if different serotypes of enteroviruses caused infection in children of different ages. Table 3 shows the mean ages of the children who had CVA16, HEV71 and CVA10 infection. Comparison of means for each pair using an unpaired t test at an alpha of 0.05, showed that there was no significant difference in the mean ages of the children in the different groups (CVA10 versus CVA16: P = 0.0872; CVA10 versus HEV71: P = 0.1800; CVA16 versus HEV71: P = 0.6992). Following the 1997 outbreak of EV71 associated HFMD in Sarawak, Malaysia, the Health Department installed a sentinel surveillance programme with the expectation that we would be able to study epidemiological trends and begin to predict when to expect outbreaks with sufficient accuracy in order to implement public health interventions to reduce the burden of the disease. Although the surveillance programme is still ongoing in Sarawak, we have sought to glean some preliminary information from the data generated over the first 7 years of the programme. two HEV71 outbreaks. A recent report on a similar surveillance programme in Yamagata Prefecture in Japan (1998) (1999) (2000) (2001) (2002) (2003) suggests that in Yamagata there is frequent importation of HEV71 from surrounding countries seeding the clusters of cases seen annually in this community [10] . The HEV71 strains in this study were isolated from small clusters of cases that tended to be seen in the summer months while in our situation we observed outbreaks of HEV71 every 3 years with cases being seen much earlier in the year, well before the northern summer. In 2003 both Sarawak and Yamagata experienced a large outbreak and in both situations, a genogroup shift from C to B was noted. It is interesting that in Sarawak, of the genogroup C viruses, only genogroup C1 strains have been observed, while genogroup B viruses appear to be changing from outbreak to outbreak, suggesting that it is likely that genogroup B viruses are evolving within Borneo and that the outbreaks we have experienced are being seeded from within rather than from imported viruses. Since the outbreaks in Sarawak typically begin early in the year, it is also possible that genogroup B strains generated in Sarawak may seed HEV71 outbreaks in the region, which typically occur later than the Sarawak outbreaks. This temporal sequence of regional outbreaks is also true of those occurring in Singapore and in Peninsula Malaysia. The data we have obtained through 7 years of our sentinel surveillance programme for HFMD in Sarawak have provided useful clues to understanding the epidemiology of HEV71 in the state. It is clear that the appearance of HEV71 associated HFMD in sentinel clinics signals the start of an outbreak, but the rise in the number of cases is so rapid that this approach is not a suitable early warning system. In 2003 there were only 5 weeks between the time the first HEV71 cases were seen and the peak of the outbreak. Clearly this could be explained by rapid and effective response by the public health teams, but we have no way to know. Alternatively, the 3-year cycle of HEV71 outbreaks we have observed could, if verified in the coming years, provide public health officials with the relevant information to plan and to implement their intervention programmes to reduce the disease burden in the years when an HEV71 outbreak is expected. Although this is not expected to prevent the outbreaks entirely, effective public health measures put into place early enough can limit the spread, reduce mortality and reduce the burden on the community and the health system. It is important to note that epidemiological curves showing HFMD alone, without distinguishing the infecting agent for each case, can stretch broadly over many months, with non-HEV71 enteroviruses continuing to be isolated after cessation of HEV71 activity. This was especially evident in 2000, when HEV71 associated fatal cases were reported in neighbouring Singapore in September and October 2000 [29] , and the media attention surrounding these events generated a high index of suspicion in Sarawak as well. No HEV71 was isolated in Sarawak after August that year, but numerous CVA16 continued to be isolated until the end of 2000. Thus even though sociological factors affect the shape of the HFMD epidemiological curves in Sarawak, epidemiological curves specifically showing genogroup B strains of HEV71 were consistently sharp and well defined in 2000 and 2003. The mean and median age of children with HFMD was 36 months and 30 months respectively, but the mean ages did not differ between the groups infected with the different serotypes. It is thus intriguing that HEV71 has caused much larger and sharper outbreaks than either CVA16 or CVA10. This suggests that HEV71 has the capacity to spread rapidly through the susceptible population and then become quiescent in the community. In the third year after any HEV71 outbreak, the whole cohort of children under 3 years of age has not been exposed to HEV71 and all of these children are then susceptible, providing the conditions for another sweeping transmission of HEV71 through the community. The annual birth cohort in Sarawak is 48 to 49 thousand and thus in 3 years there are up to 150,000 susceptible children in the state. According to the trends we have reported, we expect that the next outbreak of HEV71 in Sarawak will be in 2006. At the time of writing we have already begun to pick up HEV71 cases in our sentinel programme and from past experience, an outbreak in Sarawak is often followed by outbreaks in other countries in the region. We have therefore decided to put our data into the public domain in order that other public health practitioners in the Asia Pacific region may benefit from this experience and prepare for a spread of HEV71 in the region once again in the months to come. The main conclusions arising out of this preliminary report are described below: a. HEV71 outbreaks have occurred every 3 years in Sarawak starting in 1997. All the 3 outbreaks (1997, 2000 and 2003) have been caused by genogroup B viruses and furthermore, each of the 3 outbreaks has been associated with genogroup B viruses that are genetically distinct from each other. b. HEV71 of subgenogroup C1 has been isolated throughout the 7 years of the surveillance programme and are closely related to each other and to genogroup C1 viruses isolated elsewhere. Sarawak has so far not experienced large HFMD outbreaks caused by HEV71 of genogroup C1. Indeed HEV71 of subgenogroup C1 behave much like other species A enteroviruses, occurring sporadically throughout the surveillance period. c. In Sarawak, occurrence of HEV71 genogroup B infections is tightly clustered, with cases rising and falling very rapidly.
56
Model-Based Design of Growth-Attenuated Viruses
Live-virus vaccines activate both humoral and cell-mediated immunity, require only a single boosting, and generally provide longer immune protection than killed or subunit vaccines. However, growth of live-virus vaccines must be attenuated to minimize their potential pathogenic effects, and mechanisms of attenuation by conventional serial-transfer viral adaptation are not well-understood. New methods of attenuation based on rational engineering of viral genomes may offer a potentially greater control if one can link defined genetic modifications to changes in virus growth. To begin to establish such links between genotype and growth phenotype, we developed a computer model for the intracellular growth of vesicular stomatitis virus (VSV), a well-studied, nonsegmented, negative-stranded RNA virus. Our model incorporated established regulatory mechanisms of VSV while integrating key wild-type infection steps: hijacking of host resources, transcription, translation, and replication, followed by assembly and release of progeny VSV particles. Generalization of the wild-type model to allow for genome rearrangements matched the experimentally observed attenuation ranking for recombinant VSV strains that altered the genome position of their nucleocapsid gene. Finally, our simulations captured previously reported experimental results showing how altering the positions of other VSV genes has the potential to attenuate the VSV growth while overexpressing the immunogenic VSV surface glycoprotein. Such models will facilitate the engineering of new live-virus vaccines by linking genomic manipulations to controlled changes in virus gene-expression and growth.
Infections caused by viruses persistently threaten human health. For example, 40 million, 350 million, and 170 million people in the world are carrying human immunodeficiency virus type 1 (HIV-1), hepatitis B virus (HBV), and hepatitis C virus (HCV), respectively [1] [2] [3] . Annually 5% to 15% of the global population is infected with influenza, resulting in 250,000 to 500,000 deaths [4] . Protection against viral infections may be provided by inoculations with live-virus, killed-virus, or subunit vaccines. Live-virus vaccines offer advantages because they activate both humoral and cellmediated immunity, require only a single boosting, and generally provide longer immune protection than other forms of vaccines. However, they must be adequately attenuated in their growth to minimize the possibility of vaccine-induced pathogenic effects while retaining their immunogenicity. Attenuation of live viruses has traditionally been achieved by serially passaging viruses in tissue or cell culture and adapting them to grow well on new cell types or at elevated or reduced temperatures [5] , a process that tends to reduce their replicative ability and virulence in humans or animals [6] . Such attenuation has historically been a highly empirical process, where its mechanisms are often neither known nor elucidated. During the last decade the emergence of reverse genetics techniques has created unprecedented opportunities to better control viral attenuation [7] [8] [9] . Reverse genetics enables the production of RNA viruses from cloned cDNA, so specific mutations can be relatively easily introduced into viruses. The challenge to engineering viruses for attenuation then shifts from creating the variants to predicting how specific genetic changes define or correlate with measurable effects on virus growth. Such a challenge can be addressed through the development of quantitative and mechanistic models that map genome-level changes to the dynamics of virus growth under different environmental conditions. Models of intracellular virus growth aim to predict how rapidly a virus-infected cell will produce virus progeny. Inputs to such models include rates of constituent processes such as entry of the virus into the cell, transcription of viral mRNAs, translation of viral proteins, replication of viral genomes, assembly of intermediates, and finally, production and release of viral progeny. Decades of detailed biochemical, biophysical, and genetic studies have, for diverse viruses, contributed toward a level of mechanistic understanding of viral functions and interactions. Various intracellular models of virus growth have been developed for phage Qb [10] , phage T7 [11, 12] , HIV-1 [13] , and influenza A virus [14] . How can a detailed model for the intracellular growth of a virus be used to explore the behavior of mutant viruses that encode alternative designs? As a starting point, one can create alternative designs that reorder or rearrange the wildtype genes or regulatory elements. Such genomic changes can alter the timing and level of expression of different viral genes, and thereby impact growth because the production of viral progeny depends on the dynamic expression of viral genes. Preliminary models of such alternative genome designs can use the ''language'' of the wild-type virus. They retain the parameters that characterize wild-type molecular interactions, wild-type average rates of viral polymerase elongation, and wild-type composition of progeny viruses, but they apply them in a manner that reflects the reordering of wild-type components in the engineered genome. For example, the timing of expression for the genes of phage T7 during infection maps closely to their sequential order on the T7 genome [15, 16] . By relocating an essential early gene, encoding the T7 RNA polymerase to downstream positions, one delays initiation of transcription by the highly efficient T7 RNA polymerase and thereby attenuates phage growth [17] . Preliminary models for the growth of phage carrying the altered genomes, based on wild-type parameters, capture the overall observed trends in attenuation. Here we expand this approach to a mammalian virus of biomedical and agricultural relevance: vesicular stomatitis virus (VSV). VSV is a prototype negative-sense single-stranded RNA (Mononegavirales, (-)ssRNA) virus and a member of the family Rhabdoviridae [18, 19] , which includes rabies virus. Each VSV particle has a single copy of an 11-kb genome carrying five genes that encode nucleocapsid (N), phospho (P), matrix (M), envelope (G), and polymerase (L) proteins. VSV is economically important because it can cause symptoms like those of foot-and-mouth disease in livestock [19] . It offers several advantages as a vaccine vector including low seropositivity in humans, a capacity to accommodate foreign genes up to 40% of its own genome size, and an established reverse genetics [18, 20] . Recombinant forms of VSV carrying foreign virus genes that encode immunogenic proteins have been proposed as potential vaccines against HIV, influenza, and respiratory syncytial virus [21] [22] [23] [24] [25] . Less pathogenic but more immunogenic VSV-based vaccines against infection by VSV or other viruses are being sought. Here we develop an in silico model of a VSV infection cycle, incorporating known regulatory interactions and mechanisms and relevant quantitative data from the literature of the past 40 years. These interactions and the corresponding equation formulations are described in detail in the model development section of Materials and Methods. Using the model, we first quantitatively analyze how the intracellular growth of wild-type VSV directs host biosynthetic resources toward VSV gene expression, synthesis of progeny genomes, and pathway switching from the synthesis of VSV intermediates to the production of VSV progeny. We then reveal that the model captures experimental results showing progressive attenuation of virus growth associated with moving the N gene downstream from its wild-type position. Finally, we use the model to predict how altering the positions of other VSV genes and promoters may attenuate the growth of VSV while increasing its potential capacity to activate an adaptive immune response. Using our model with the established parameter set (Tables 1 and 2) , we first analyzed quantitatively and systematically how the intracellular growth of VSV is regulated. The improved understanding of the virus infection by this model-based analysis may guide us to identify the key regulatory components to manipulate for developing virus mutants as possible vaccine or vector candidates. Attenuation mechanism leads to unequally distributed synthesis of viral mRNAs and proteins. The partial transcription termination mechanism (or attenuation) is common in (-)ssRNA viruses. This mechanism is important to satisfy the different needs of each viral protein during its infection cycle. Five attenuation factors for each intergenic region of the VSV genome (Table 1 and Equation 8 ) were obtained from the literature [18, 19, 26, 27] and incorporated into our model. Owing to the step-wise release of polymerases from each gene junction, our simulations estimated the gradual decrease of VSV mRNA synthesis in the order of N . P . M . G . L ( Figure 1A ). Compared with the most abundant N mRNA, L mRNA is very scanty in infected baby hamster kidney (BHK) cells (40 ; 140-fold less). The relative production level of each protein matched the relative availability of the corresponding mRNAs ( Figure 1B ) even though different proteins degrade at different rates ( Table 1 ). Because of the different level of incorporation of each protein into a single virion particle, as defined by the protein stoichiometry (Table 2 , [28, 29] ), the relative levels of free viral proteins in the cytoplasm develop differently from their mRNA levels ( Figure 1C ). P protein is most abundant owing to its low content in the virion, and L and N proteins are least abundant. The persistently low level of N protein is related to its immediate complexation with nascent genomes and antigenomes to make nucleocapsid particles during the replication step. Owing to the cyclic switching between transcription and replication by the encapsidation process, the N protein level is predicted to oscillate as shown in Figure 1C . The engineering of viral genomes provides ways not only to explore viral regulatory mechanisms at a genomic level, but also to produce recombinant viruses that may serve as vaccines, gene delivery vectors, and oncolytic (tumor-killing) agents. However, the complexity of interactions among viral and cellular components involved in the life cycle of a virus can make it challenging to anticipate how altering viral components will influence the overall behavior of the virus. Lim, Lang, Lam, and Yin have developed a computer model that begins to mechanistically account for key virus-cell interactions in its predictions of viral intracellular development. Lim et al.'s model was able to capture experimentally observed effects of gene rearrangements on the levels and timing of viral protein expression and virus progeny production, aspects that are important for the design of live-virus vaccines. Refinement and extension of their approach to current and other virus systems has the potential to advance the application of viruses as therapeutic agents. After 9 h post-infection, our simulation predicts a significant decrease of free M proteins. This arises from the dominance of the virion assembly process, which depletes M proteins, compared with transcription and replication. In infected murine delayed brain tumor (DBT) cells, similar distributions of viral mRNAs and proteins were obtained (unpublished data). Higher demands for genomes are satisfied by the stronger promoter of the anti-genome template relative to that of the genome template. Anti-genome templates are only utilized to amplify genomes, while genome templates are used to amplify both anti-genomes and mRNAs, and they are also incorporated into virion progeny particles. The higher demands for genome by these multiple tasks are satisfied by the stronger promoter of the anti-genome template compared with that of the genome template [30] [31] [32] . More polymerases bind to the stronger promoter of the anti-genome, ultimately enhancing the production of genomes over anti-genomes ( Figure 2 ). In our model the parameter S prom measures the strength of the anti-genomic promoter relative to that of the genomic promoter ( Table 1 ). The production ratio of genomes to anti-genomes was estimated to be dynamically changed [33] , varying from one to 30 (wild-type VSV case (S prom ¼ 5.4), Figure S1 ). Such oscillatory changes in the production ratio shown in Figure S1 follow from the on-off use of the genomes for transcription or replication. They also arise owing to the staggered shifting of dominant templates between genomes and anti-genomes during replication. A large value of S prom favors use of anti-genome templates to replicate genomes. However, as genome templates accumulate in large excess relative to anti-genome templates, they successfully compete for replication resources. Synthesis of anti-genomes then dominates until they accumulate and serve again as the dominant templates. The virion production rate in BHK cells is at maximum 5-10 h post-infection. In infected DBT cells, similar simulation results were obtained except that the synthesis of genomesized viral RNAs continued for longer time (active until 15 h post-infection, Figure 2 ). Optimal utilization of genomic nucleocapsids. Genomic nucleocapsids can either be used as templates for RNA synthesis or they may be incorporated into progeny virions. Their fate depends on levels of polymerase and M protein, which respectively favor RNA synthesis or virion production pathways, as well as on the extent to which association of the nucleocapsid with M protein will dominate over association with polymerase, described with the parameter S cond in our model (Table 1) . Because both RNA synthesis and virion production are essential processes of the infection, extreme values of S cond that favor one process over the other will be [49] S pol ¼ 170 nt Strength of anti-genomic promoter relative to that of genomic promoter a S prom ¼ 5.43 Rate constant ratio of the associations of M protein and polymerase with the genomic nucleocapsid a S cond ¼ 6.2 3 10 À5 Fraction of M protein bound to the plasma membrane [51] cond M ¼ 0.1 Attenuation factors [18, 19, 26, 27] / N / / P / / M / / G / / L ¼ 0.0 / 0.25 / 0.25 / 0.25 / 0.95 Degradation rate constants N protein [58] k dp,N ¼ 3.5 3 10 À5 sec À1 P protein [59] k dp,P ¼ 1.4 3 10 À6 sec À1 M protein [58] k dp,M ¼ 1.5 3 10 À4 sec À1 G protein b k dp,G ¼ 5.7 3 10 À5 sec À1 L protein [59] (if concentration P protein ! 10 3 concentration L protein ) k dp,L ¼ 1.2 3 10 À5 sec À1 (if concentration P protein , 10 3 concentration L protein ) k dp,L ¼ 4.3 3 10 À5 sec À1 mRNA [60] k d,mRNA ¼ 1.9 3 10 À4 s À1 nucleocapsid c k d,nc ¼ 1.9 3 10 À5 s À1 Host-dependent parameters Number of ribosomes in host cell [61] nrib ¼ 5 3 10 6 Elongation rate of ribosome [62] k e,rib ¼ 6 aa/s Spacing between neighboring ribosomes a (BHK/DBT) Srib ¼ 238. 5 Owing to their increased stability by encapsidation, nucleocapsids are less degradable than naked mRNAs. It was assumed that the nucleocapsids of VSV are degraded 10-fold slower than its mRNAs, and our simulation results were insensitive to this parameter. DOI: 10.1371/journal.pcbi/0020116.t001 detrimental for growth. For excessively large S cond , newly synthesized genomic nucleocapsids would tend to be prematurely incorporated into virion particles before they could serve as templates for transcription and replication. On the other hand, for extremely small S cond , genomic nucleocapsids would be utilized primarily to produce viral RNA without being packaged into viral progeny. Hence, an intermediate parameter value is expected to be optimal for viral growth. We estimated a possible range for the wild-type value of S cond by fitting our simulation results to previous experimental observations by others (2.5 3 10 À5 to 1.0 3 10 À4 , Figure S4 ). Our simulations further indicate that this range of S cond is near-optimal and optimal for VSV growth on BHK and DBT cells, respectively (unpublished data). Diversion and inhibition of host translation machinery create a time window of opportunity for translation of viral proteins. During the infection cycle, virus actively and passively competes with the host for limited translation resources by inhibiting host transcription and by amplifying viral mRNAs, respectively. Viral leader-mRNA and M protein play key roles in this inhibition [19, 28, [34] [35] [36] . As viral components accumulate in the cytoplasm from the initiation of infection, an ever-increasing fraction of host ribosomes are available for viral mRNAs ( Figure 3A , the fraction of ribosomes associated with viral mRNAs is defined by 1-rib_host in Equation 26 ). However, the inhibition of host macromolecular synthesis causes a failure to supply accessory factors needed for initiation and elongation steps of translation, resulting in a reduction in the fraction of active ribosomes over time ( Figure 3A , as described by f dec in Equation 25 ). These two mechanisms create a time window when active ribosomes are maximally available for viral translation in infected cells ( Figure 3A , refer to the term (1-rib_host)*f dec in Equation 27 ). The abundance of viral mRNAs and the limitation imposed by ribosomal spacing determine the fraction of the active ribosomes involved in translating viral mRNAs (occupied ribosomes in Figure 3B , refer to Equations 11 and 27) . In our model, if the occupied active ribosomes are less than the available ones (in this case the number of free active ribosomes . 0), viral translation is fully supported without any limitation of host machinery. In the early infection stages up to 7 h and 13 h post-infection for BHK and DBT cells, respectively, the host machinery is in excess ( Figure 3B ). However, at later times viral translation becomes limited by the host resources (in this case the number of free active ribosomes ¼ 0). This limitation may cause a transition from replication-dominant to assembly-dominant infection stages because the replication requires the continuous protein synthesis. As shown in Figure 3B , a small fraction of ribosomes as active forms (less than 5% out of 5 3 10 6 ribosomes, Table 1 ) are utilized for viral translation. Experiments and simulations of VSV gene-order mutants. For vaccine applications, one seeks to minimize viral pathogenicity and maximize its immunogenicity. Based on observed correlations between in vitro and in vivo results, we assume here that the pathogenicity and the immunogenicity of a virus are directly linked to the levels of progeny production [20, 37, 38] and G protein expression in infected cells [39, 40] , respectively. In the previous section we have showed that various VSV regulatory mechanisms are involved in maintaining balances, during infection, among viral synthesis processes, which indirectly indicates the importance of such balances for viral growth. Perturbations of such balances by genetic or genomic manipulations could provide ways to obtain viral phenotypes favorable to vaccine applications. We first test the predictive ability of our model by comparing simulated protein expression and growth of several gene-rearranged VSV strains with experimental results. Later we employ the model to predict how various genomic manipulations could attenuate virus growth and increase G protein expression. Protein expression rates of gene-rearranged viruses. The stepwise decline in the transcription of genes more distant from the 39-end region promoter highlights how gene order affects gene expression in VSV. Advances in reverse genetics have made it possible to create gene-rearranged virus strains where the transcriptional attenuation mechanism then creates altered levels of gene expression [7, 9, 18, 20] . In one study [18] the three internal genes, P, M, and G, were permuted, and the resulting six possible VSV strains were characterized. Relative rates of viral protein expression in BHK cells were experimentally measured based on their incorporation of [ 35 S]-labeled methionine for a one-hour window at 4 h post-infection [18] . We extended our model to simulate this experiment for mutants representing each geneorder permutation and compared the model prediction with the published results [18] , as shown in Figure 4 . All rates are expressed relative to the synthesis rate of N protein, whose corresponding gene was in position 1, closest to the 39 end of the genome in all strain cases. Expression of gene L, in position 5, was minimal in both simulations and experiments, and the expression of all other genes was above 40%, a feature of the experimental data that the simulation also captured. Most of the points fall close to the parity line, indicating agreement between the simulation and experiment. Noteworthy are two subsets of points. First, the four circled points are exceptions to the general rule that gene order determines the level of gene expression. These were genes in the second position of the genomes that were expressed essentially at the same rate as gene N, in the first position [18] . This result highlights that the expression rate of protein is affected not only by its gene order, or corresponding rate of mRNA production, but also by its length and degradation rate. For a fixed average rate of translational elongation, longer gene products will tend to be produced more slowly. Further, the net rate of protein production will reflect the rates of both protein synthesis and protein degradation. The model accounts for these contributions, and for the circled genes such accounting appears to capture unexpected high translation levels of genes in the second position, which were measured by Ball et al. [18] . The second sets of points, shown in two boxes, indicate mismatches ( Figure 4 ) that, in the most challenging scenario, could reflect unknown strain-specific mechanisms that are not present in our general gene-permutation model. However, one should also note that the experiment is based on labeling and quantifying proteins about 4 h post-infection. This relatively early time point allows one to minimize potentially confounding influences of virion particle assembly and production on cytoplasmic levels of viral proteins, but it also represents a point before the majority of viral proteins have been made ( Figure 1B) . Growth of gene-rearranged viruses. We also employed our model to predict the growth of VSV strains having the N gene first position on the genome grows best, followed by N2, N3, and N4. This result is consistent with the previously suggested hypothesis that relocation of the N gene to 39-distal positions on the genome would be an efficient way to attenuate VSV for vaccine use [20] . The reduction in growth that follows from moving the N gene likely reflects, at least in part, an imbalance between replication and transcription. Insufficient production of N protein would reduce the extent of encapsidation of nascent anti-genome and thereby allow transcription to dominate over genome replication [41] . While the simulation matches the growth ranking, it did not quantitatively match the experimental data. The predicted variation in virion production (N1 ! N4: 4.7-fold decrease) is smaller than the experimentally observed variation (N1 ! N4: 16.7-fold decrease). A potential source of this quantitative difference was our neglect of mass action effects of N proteins on the encapsidation process in our model; encapsidation was simulated as an instantaneous process when free N proteins were available. As shown above, our model could capture the major effects of gene rearrangement on viral growth and protein expression. Effects of relative promoter strength on viral growth. The genome and anti-genome of VSV are synthesized in unequal amounts, determined by the differing strengths of their promoters [19, 31, 33, 39] . The stronger promoter of the antigenome allows viral polymerases to produce more genomes to meet their demands as components of virion particles and as templates for transcription and replication. To explore how VSV growth is influenced by differences in the relative strength of the genomic and anti-genomic promoters, we predicted the yield of virus on BHK and DBT cells over a broad range of S prom , as shown in Figure 6A . Small S prom virus cannot grow well because most polymerases would be associated with genomes and tend to synthesize primarily anti-genomes. For example, for S prom equal to 0.1, infected BHK and DBT cells make 5-fold and 26-fold fewer progeny than wild-type VSV infected cells, respectively. However, large S prom virus also cannot grow well because most polymerases would preferentially bind to newly synthesized anti-genomes, producing few of the anti-genomes that are needed as templates to amplify genomes. Our simulations predicted that values of 30 and 50 for S prom would be optimal for VSV growth in BHK and DBT cells, respectively ( Figure 6A ). The estimated wild-type value of S prom of 5.4 gives VSV yields higher than 80% of their maximum yields for both cell types ( Figure 6A) . We speculate that a rational way to attenuate the pathogenicity of the live wild-type virus would be to swap its two promoters, giving an S prom of (5.4) À1 . For this promoter swap we predict virion production would be decreased by 3.3fold and 14.5-fold for infected BHK and DBT cells, respectively, relative to wild-type. However, the extent of growth attenuation by the promoter swapping can be higher than the model prediction because the swapping may also perturb viral transcription and virion assembly and budding processes modulated by the signals encoded in the 39and 59 termini of the genome [42, 43] that are not yet sufficiently defined to be included in the simulation. Rational vaccine attenuation by double genomic manipulations. Several variant VSV strains, including N1 through N4, have been made by Ball and Wertz [18, 20] . With an aim to generate a potentially broader diversity of growth phenotypes, we created and tested in silico VSV mutants by combining N gene relocations with a range of S prom . This is a computationally simple task, but experimentally nontrivial. For example, the VSV N4 with S prom ¼ 0.1 produced a simulated 38-fold fewer virus progeny than wild-type in BHK Figure 6B ). It is interesting to note that this 38-fold degree of growth attenuation is greater than the product of the constituent attenuations (4.7 3 5.4 ¼ 25) that one calculates by assuming that the effects were uncoupled. Such nonmultiplicative effects of double genomic manipulations on growth would be challenging to predict in the absence of a quantitative model. Modulation of VSV immunogenicity by gene shuffling. To elicit a systematic immune response, live viral vaccines must present or display neutralizing epitopes, typically through the expression of viral surface proteins. Higher levels of antigen expression have been found to correlate with more rapid and potent induction of anti-viral antibodies [39, 40] . As Flanagan et al. suggested, the gene encoding G may be moved to other positions in the VSV genome to modulate the expression of the VSV surface glycoprotein G [39] . We generated in silico five gene-shuffled VSV strains, having gene orders for the three internal genes, MPG, MGP, PGM, GMP, and GPM, and simulated levels of G protein in BHK cells infected with those strains. Our simulation results were consistent with the idea that the location of G gene affects the production of G protein. The GPM strain gave the highest concentration of G protein in the cytoplasm (almost 2-fold higher than that of wild-type, Figure 7 ). However, in many cases effects of such gene rearrangements can be difficult to anticipate because of the complexity of the involved interactions among viral components. For example, the PGM strain showed only a level of G protein expression similar to those of PMG (wt) and MPG strains even though it has the G gene at an earlier position than the other two strains (Figure 7) . For vaccine use we might aim to maximize the immunogenicity of VSV or a VSV-based vector through the expression increase of VSV G gene or inserted foreign gene while minimizing their potential pathogenicity by growth attenuation. Given such design goals, specifically for a VSV vaccine, we might prefer strain GPM, which showed the highest expression of G protein (Figure 7 ) and the lowest production of virions [18] . Toward such favorable features, Flanagan et al. previously constructed three VSV strains having the following gene orders: 39-G-N-P-M-L-59 (G1N2), 39P-M-G-N-L-59 (G3N4), and 39-G-P-M-N-L-59 (G1N4) [39] . These genome constructions were based on their intuitive idea that translocations of G gene and N gene to earlier and later positions, respectively, compared with wild-type, could not only increase the expression of G protein, but also attenuate virus growth [39] . This idea was supported by their experimental results [39] . Seeking a more detailed correlation between locations of the two genes and the viral phenotypes relevant to vaccine application, we simulated in silico the growth of all mutants that retain the gene order P -M -L of the wild-type, but allow G and N to move, criteria that define 20 possible geneorder permutations. The viral growth and the level of G protein in infected BHK cells mainly depend on the locations of N gene and G gene, respectively ( Figure 8A and 8B) , which is consistent with the experimental results of Flanagan et al. [39] . Further, if gene G is fixed, then moving gene N closer to the 39 promoter is predicted to increase protein G expression ( Figure 8B ). Consistent with this prediction, Flanagan et al. also observed a higher G protein expression for the G1N2 strain than for the G1N4 strain [39] . Enhanced replicative ability of VSV strain by locating its N gene at an earlier position in its genome can contribute to increasing the level of G protein. If either gene N or gene G is located at the fifth position, then both levels of virus growth and G protein expression are very low (Figure 8) , because with such genome organizations the stoichiometric amounts of N and G proteins required for replication and assembly (Table 2 , [28, 29] ) cannot be reached. Our simulations with the BHK cell parameters (Table 1 ) overall captured the experimentally established relative growth of the VSV strains in BHK-21 cells, but the growth of the G3N4 strain was significantly overestimated compared with the experimental results [39] ( Table 3 ). The relevant mechanism for such a large discrepancy between the simulation and the experimental results remains to be elucidated. The changes of protein expression levels by gene shuffling can be a rational means to modify the viral features for vaccine use. Robust synthesis of antigen by a highly attenuated strain appears to be an effective vaccine strategy as Flanagan et al. previously suggested. In addition to controlled attenuation of virus growth, a potent vaccine should ideally elicit a strong humoral or cell-mediated immune response. In the era of highly advanced genetic technologies, we have witnessed a turning point for the development of live viral vaccines. Conventional empirical vaccine development processes are now being replaced by more rational reversegenetics-based ones. With this trend, much attention will be focused on mechanism-based design of less pathogenic and more immunogenic virus stains. Mathematical models for intracellular virus growth can support this design process by providing a tool to systematically analyze the viral infection regulatory network, identify critical regulatory mechanisms or components for redirecting viral phenotypes, and reverse engineer desirable phenotypes. One-step infection of cell monolayers. Cells were harvested, resuspended in growth medium, and plated into six-well plates at a concentration of 5 3 10 5 cells per 2 ml per well. Plated cells were returned to the incubator and allowed to grow overnight. The next day, two representative cell monolayers were harvested and counted to give an approximate number of cells per well. Each monolayer was then incubated with 200 ll of virus inoculum (MOI 3) for 1 h to allow virus adsorption. The plates were rocked gently every 20 min to evenly distribute virions on the monolayers during the adsorption step. After the adsorption period, the monolayers were rinsed twice with 1 ml of HBSS and then placed under 2 ml of infection medium for incubation. Medium samples of 200 ll including virion particles were taken from each well at 2, 3, 4, 6, 8, 10, and 20 h postinoculation. Samples were kept frozen at À90 8C until their analysis by the plaque assay. Plaque assay. BHK cells were plated into six-well plates and cultured to 90% confluence. Culture medium was removed from each well and replaced with 200 ll of serially diluted viral samples. The inoculated monolayers were returned to the incubator for 1 h to allow virus adsorption. The plates were rocked gently every 20 min. At the end of the adsorption period, the inoculum was removed from each monolayer sample and then replaced with 2 ml of agar overlay. The agar overlay consisted of 0.6% weight/volume (w/v) agar (Agar Nobel, Difco, BD Diagnostic Systems, http://www.bd.com). 5-Bromo-29-deoxyuridine (B5002, Sigma) was added, at 100 lg/ml, to the agar overlay of N3-and N4-infected samples to enhance plaque formation. Following agar addition, the plates were allowed to cool at room temperature for 30 min, returned to the incubator and incubated for 24 h, and then each sample was fixed with 2 ml of fixative for 3 h at room temperature. The fixative consisted of 4% (w/v) paraformaldehyde (VWR) and 5% (w/v) sucrose (Sigma) in 10 mM phosphate buffered saline (PBS, Sigma) of pH 7.4. The agar overlay was then removed, and each sample was rinsed twice with 2 ml of PBS. Gentian violet diluted in methanol (0.01% (w/v), Sigma) was used, at 1 ml each, to stain the samples. Model development. Using algebraic and differential forms of equations, our mathematical model aims to account for established molecular processing steps in the development of VSV. Most model parameters were extracted from the literature. However, five parameters were obtained by fitting our simulation results to experimental data that were from the literature and our own experiments. Key model parameters are given in Table 1 , and detailed descriptions of the model and parameter estimation process are provided below, and in Protocol S1 and Figures S2 and S3 , respectively. Virus binding and penetration. As shown in Figure 9A , VSV initiates an infection by binding to a receptor such as phosphatidyl serine, a lipid component in the plasma membrane [19, 44] . After the binding step, the VSV particle is endocytosed via a clathrin-coated pit, and then penetrates intracellular vesicles such as endosomes by membrane fusion [19, [45] [46] [47] . The penetration leads to the release of the encapsidated negative-sense viral genome and virus proteins into the cytoplasm of the host cell. By assuming the binding step is ratedetermining [48] , we lump these early steps from the binding to the penetration into a first-order expression: where V b and V ex are the concentrations of bound and extracellular virus particles, respectively, t is time, and k b is the apparent rate constant for virus binding. After binding, we assume the bound virus is immediately endocytosed and fused, and its genome and protein components are instantaneously released into the cytoplasm at the expense of the fused virus particle. The protein stoichiometry of a single VSV particle and the lengths of each viral gene and protein are summarized in Table 2 , [28, 29] . Population distribution of polymerases and nucleocapsids. Following the release of the encapsidated genome and proteins into the cytoplasm, VSV transcription is initiated. The viral transcription was assumed to be independent of host-cell functions such as replication [33] . Instead, the viral complex of L and P proteins, with a stoichiometry of 1-3.6, was taken to function as polymerase in transcription and replication [49] . In the absence of P protein, L protein cannot bind to the genome or anti-genome [50] . After binding to the 39 promoter regions of the genomic and anti-genomic templates, the viral polymerase starts to synthesize its own RNA transcription and replication products by elongating along the templates. During transcription a fraction of elongating polymerases terminate transcription by dissociating from the templates as they encounter regulatory signals at intergenic regions [19, 26] . In addition to the regulated polymerase dissociation, time-dependent concentration changes of the polymerases and the viral templates in the cytoplasm influence the distribution of polymerases on the viral templates during transcription and replication. Hence, the distribution of polymerases continuously varies over the viral templates, ultimately determining the relative synthesis levels of mRNAs and genome-size RNAs. We simulate the transcription and replication processes by considering the spatial-temporal distribution of template-associated polymerases. We first partition the viral genome and anti-genome templates into 40 segments, excluding their 39 and 59 end regions, which are the leader (Le) and trailer regions (Tr) for the genome, and the complementary trailer (Trc) and complementary leader regions (Lec) for the anti-genome, respectively ( Figure 9B ). For the genome template that is used for transcription as well as for replication, we specially grouped the segments into five genes ( Figure 9B ). We chose 40 as a minimum number for total segments which allows each gene to be split into a specific integer number of segments, proportional to the length of the gene. By considering the mechanisms for the interactions between polymerase and the intergenic regulatory sequences of the templates, as described below in the Transcription section, we simulated the polymerase flux into each segment over the time elapsed from the initiation of transcription on each template. Then we correlated the level of polymerase occupying each geneencoding section of the template with the synthesis rate of each corresponding viral mRNA. In a similar way, the distribution of polymerases on the replication templates was correlated with the synthesis rate of viral genome-sized RNA. Such explicit treatment of polymerase spatial distributions on the viral genome and antigenome templates was central to modeling the growth of wild-type and gene-rearranged virus strains. This treatment systematically accounts for polymerization-associated time delays and the polymerase fluxes into each template segment. Before estimating the polymerase flux, we need to figure out how the polymerase complex and M protein compete with each other for binding to the genomic nucleocapsids as well as how the polymerases bound to nucleocapsids are subsequently distributed to one of three possible tasks: transcription, replication of genome, or replication of anti-genome. In our model we assume that the genomic templates (negative-sense nucleocapsids) whose promoters (leader regions) are free of polymerases are available for association with free polymerase or M protein. We further assume that the associations of the free genomic templates by M proteins or polymerases take place instantaneously: where (-)nc, (-)nc M,new , and (-)nc pol,new are the concentrations of total genomic nucleocapsids and subsets of genomic nucleocapsids whose promoters are newly occupied by M protein and polymerase, respectively. S pol is the spacing between neighboring polymerases on the genomic or anti-genomic template, pol l is the concentration of polymerases bound to the promoter region (Le) of the genomic template, and l l is the length of the promoter region. Specifically, the second term in the left-hand side of the equation denotes the concentration of the genomic templates whose promoters are currently occupied by polymerases. In our model the concentration of the genomic nucleocapsids whose promoters are bound to polymerases and the concentration of the polymerases bound to the promoters of the genomic nucleocapsids are interchangeable with each other by the factors (l l /S pol ) and (S pol /l l ), respectively. The binding of M protein or polymerase initiates reactions leading to virion assembly or RNA synthesis, respectively. Because the initiation of RNA synthesis by the polymerase requires a finite time, a space between adjacent polymerases on the template (S pol ) would be maintained during infection, assuming a fixed elongation rate. With these considerations, one may expect that at any time the concentration of nucleocapsids available for the new binding of the free proteins is inversely proportional to the concentration of polymerases currently bound to the leader region of the genomic nucleocapsids (pol l ) and the polymerase spacing (S pol ) as shown in the second term of Equation 2. The ratio of (-)nc M,new to (-)nc pol,new is determined by the ratio of the association rates of M protein and polymerase with the genomic nucleocapsid, which is further a function of the rate constants and relative amounts of the corresponding free components in the cytosol: where r asso,M and r asso,pol are the rates of the associations of M protein and polymerase with the genomic nucleocapsid, respectively, and k M and k pol are the rate constants for each association reaction, respectively. S cond denotes the ratio of the two rate constants (¼k M / k pol ). Unlike L protein, 10% of synthesized M proteins are associated with the plasma membrane [51] . In Equation 3, cond M is the fraction of M proteins associated with the plasma membrane, trans is the fraction of L proteins satisfying the polymerase stoichiometry with P protein, pol total is the total concentration of polymerases associated at the time with nucleocapsids, and M and L are the total concentrations of M and L proteins not assembled into viral progeny. If the concentration of P protein (P) is larger than 3.6-fold concentration of L protein, then trans is equal to 1. Otherwise, trans is equal to P/(3.6L). In our model, M and L proteins compete for free genomic nucleocapsids, and the condensed nucleocapsids, owing to their association with M proteins, cannot be utilized for transcription or replication [19] . From Equations 2 and 3, the newly occupied nucleocapsids by polymerases ((-)nc pol,new ) can be calculated: In the same way, given S pol , the concentration of positive-sense antigenomic nucleocapsids available for binding to polymerases would be ((þ)nc -pol trc (S pol /l trc )), where (þ)nc is the total concentration of antigenomic nucleocapsids, pol trc is the concentration of the polymerases bound to the promoter region (Trc) of the anti-genomes, and l trc is the length of the promoter region. Because the anti-genome has a stronger promoter than the genome [19, 31] , which is quantified by S prom in our model, S prom -fold, more polymerases bind to the promoter of the anti-genome than to that of the genome. Under the limitation of free polymerase complex, the concentration of the polymerases newly binding to the promoters of the genomes or the anti-genomes (pol term new ) could be described as follows: where pol trc new is the concentration of the polymerases newly binding to the complementary trailer region (promoter) of the anti-genome. The polymerases newly binding or already bound to the promoters of the genomes and anti-genomes start viral RNA synthesis as transcription or replication process. Transcription. The viral polymerase on the leader region of the genome starts either transcription or replication. If there are sufficient N proteins, transcription is inhibited by the encapsidation of nascent positive-sense RNAs by N proteins; then replication dominates transcription [41, 52, 53] . In contrast, if there are insufficient free N proteins, then transcription dominates replication. In the model we correlate the extent of transcription dominance with the availability of N proteins by introducing a factor, encap. This factor is defined as the ratio of the number of free N proteins to the number required to encapsidate all available nascent genome-sized viral RNAs. Only nocap (¼ 1 À encap) of the polymerases bound to the genomic promoters can start the transcription: dpol N;1 dt ¼ k e;pol ðð1 À / N Þ nocap pol l l l À n sec;N l mRNA;N pol N;1 Þ ð 7Þ where pol N,1 is the concentration of the polymerases located at the first segment of the N gene ( Figure 9C) , k e,pol is the elongation rate of polymerase, / N is the attenuation factor for N gene, n sec,N is the total number of the segments of N gene, and l mRNA,N is the length of N mRNA (Table 2 , [28, 29] ). The genome segments are continuously charged with incoming polymerases and discharged with outgoing polymerases with a rate of k e,pol (Equations 7-9). If the polymerase input to the leader region of the genome is decreased owing to a lack of free polymerases, then the polymerase concentrations downstream of the leader region will be subsequently reduced ( Figure 9C ). There are conserved intergenic sequences involved in letting a fraction of viral polymerases release from the genome template at intergenic sections during transcription, which is so-called partial transcription termination or attenuation ( Figure 9C ) [18, 19] . Because the transcription is initiated from the 39 end promoter, the attenuation mechanism causes genes more proximal to the 39 end to be more highly expressed, which ultimately leads to an unequal concentration distribution of viral mRNAs. The extents of partial transcription termination are quantified by the attenuation factors, / i , in our model ( Figure 9C ). These are 0/0.25/0.25/0.25/0.05 for leader-N/N-P/ P-M/M-G/G-L intergenic regions, respectively [18, 19, 26, 27] . / i fraction of polymerases are released at intergenic region i. With Equations 7-9, we simulate the polymerase flux into each gene segment, which is proportional to the elongation rate of polymerase, but inversely proportional to the extent of attenuation: dpol i;j dt ¼ k e;pol ðð1 À / i Þ n sec;iÀ1 l mRNA;iÀ1 pol iÀ1;nsec;iÀ1 À n sec;i l mRNA;i pol i; j Þ j ¼ 1; i ¼ P; M; G; L ð8Þ dpol i;j dt ¼ k e;pol n sec;i l mRNA;i ðpol i; jÀ1 À pol i; j Þ j 6 ¼ 1; where pol i,j is the concentration of the polymerases located at the jth segment of gene i, and iÀ1 indicates the prior gene of gene i. The amount of newly synthesized mRNAs for each gene is determined by the concentration of polymerases occupying each gene section on the genome template and the decay rates of the mRNAs: where mRNA i is concentration of mRNAs for gene i, k d,mRNA is the decay rate constant of mRNA that is the same for all five viral mRNAs [54] , and pol t,i is the total concentration of the polymerases occupying on the ith gene. Our formulation for transcription assumes that the synthesis of viral mRNAs is rate-controlled by the transcription initiation as well as the elongation of polymerase. Transcription initiation rate is parameterized by the spacing between neighboring polymerases in our model. At a given polymerase elongation rate, the larger polymerase spacing indicates the lower rate of transcription initiation. Transcription initiation modulates the input of polymerases to the leader region of the genome. Translation. We consider that both translation initiation and polypeptide chain elongation contribute to the rate of viral protein synthesis. The translation initiation rate is parameterized by the ribosomal spacing. In our model we first calculated the number of ribosomes involved in viral translation by considering the maximum concentration of the ribosomes bound to viral mRNAs at a fixed ribosomal spacing: where rib and n rib,avail are the concentrations of the ribosomes actually involved in viral translation and the ribosomes available for viral translation, respectively, and S rib is the spacing between neighboring ribosomes. The ribosomes involved in viral translation (rib) are allocated to the five types of viral mRNAs according to their length and abundance, assuming that each viral mRNA has the same efficiency of translation initiation [55] : where rib i is the concentration of the ribosomes assigned to mRNA i. The synthesis rate of each viral protein depends on the elongation rate of the ribosome, linear density of ribosomes on its corresponding mRNA, and its first-order decay rate: dp i dt ¼ k e;rib l p;i rib i À k dp;i p i i ¼ P; M; G; L ð13Þ where p i is the concentration of protein i, k e,rib is the elongation rate of ribosome, l p,i is the length of protein i, and k dp,i is the decay rate constant of protein i. We also accounted for the consumption of free N proteins during the encapsidation of genome-length nascent RNAs and assumed that the degradation of nucleocapsids yielded intact N proteins: dp i dt ¼ k e;rib l p;i rib i À k dp;i p i À n i ðencap Á k e;pol pol tr l tr þ pol lec l lec where n N is the stoichiometry of N protein in a single nucleocapsid or virion progeny (Table 2 , [28, 29] ), pol tr and pol lec are the concentrations of the polymerases located on the trailer and complementary leader regions of the genomes and the anti-genomes, respectively, l lec (¼ l l ) is the length of the complementary leader region, and k d,nc is the decay rate constant of nucleocapsid. As progeny virions are assembled, the concentration of each protein is reduced by the amount corresponding to its stoichiometry in a single virion particle. Replication. We assumed that N protein regulates the switch of the role of polymerase between transcription and replication by encapsidating the newly synthesized RNAs [41, 52] . The polymerase that starts the replication at the leader region of the genome requires further supply of N proteins to skip the attenuation signals at each gene junction and thereby to complete each round of replication. Depending on the availability of N proteins, nocap(¼1-encap) fraction of polymerases terminate the replication at each gene junction in our model: dpol r;n;N;1 dt ¼ k e;pol ðencap pol l l l À n sec;N l mRNA;N pol r;n;N;1 Þ ð 15Þ where pol r,n,N,1 is the concentration of the replicating polymerases on the first segment of the N gene section in the negative-sense genomic nucleocapsid. dpol r;n;i;1 dt ¼ k e;pol ðencap n sec;iÀ1 l mRNA;iÀ1 pol r;n;iÀ1;nsec;iÀ1 À n sec;i l mRNA;i pol r;n;i;1 Þ i ¼ P; M; G; L ð16Þ where pol r,n,i,1 and pol r,n,i-1,nsec,i-1 are the concentrations of the replicating polymerases on the first segment of gene i, and on the last segment of gene iÀ1, respectively. The level of polymerases that scan through the whole genome (pol tr ) determines the amount of newly synthesized anti-genomic nucleocapsids, (þ)nc: dpol tr dt ¼ k e;pol n sec;i l mRNA;i pol r;n;i; j À pol tr l tr i ¼ L; j ¼ n sec;L ð17Þ dðþÞnc dt ¼ k e;pol encap pol tr l tr À k d;nc ðþÞnc ð18Þ where l tr (¼ l trc ) is the length of the trailer region of the genome. We also considered the first-order kinetics for the decay of anti-genomic nucleocapsid. The synthesis and decay of genomic nucleocapsids are described in the same way as for those of the anti-genomic nucleocapsids except that the polymerases on the anti-genomic templates are not released at intergenic regions: dpol r;p;j dt ¼ k e;pol ðencap pol trc l trc À n sec l À l trc À l lec pol r;p; j Þ j ¼ 1 ð19Þ where pol r,p,j is the concentration of the replicating polymerases on the jth segment of the positive-sense anti-genomic nucleocapsids, l is the total length of the genome, and n sec is the total number of segments of the genome ( Figure 9B ). dpol r;p; j dt ¼ k e;pol n sec l À l trc À l lec ðpol r;p; j À 1 À pol r;p; j Þ j ¼ 2; . . . n sec ð20Þ dpol lec dt ¼ k e;pol n sec l À l trc À l lec pol r;p;nsec À pol lec l lec ð21Þ dðÀÞnc dt ¼ k e;pole ncap Á pol lec l lec À k d;nc ðÀÞnc ð22Þ In our model, non-encapsidated nascent genome and anti-genome fragments are released from polymerases and immediately degraded. As polymerases leave the promoter regions by moving toward the downstream sequences, the concentration of polymerases on the promoters will decrease. The dynamic changes of the polymerase concentrations on the promoters of the genomic and the antigenomic templates are finally described, respectively: pol l;nþ1 ¼ pol l;n þ pol new l;n À pol lÀleave;n ð23Þ where pol l-leave is the concentration of the polymerases leaving the genomic promoters, O ,n and O ,nþ1 are the concentrations of a component (O) at time n and time nþ1 (in our numerical integration, time nþ1 À time n ¼ Dt), respectively. pol trc;nþ1 ¼ pol trc;n þ pol new trc;n À pol trcÀleave;n ð24Þ where pol trc-leave is the concentration of the polymerases leaving the anti-genomic promoters. Assembly and budding. We assume that the condensation of negative-sense nucleocapsid by M protein initiates the virion assembly and the condensed nucleocapsids are not degraded in the same manner as virion progeny. Whenever the requirement for the stoichiometric amounts of proteins is satisfied, progeny virions are instantaneously assembled and released to the extracellular space. The time required for the condensation of the negative-sense nucleocapsid, the assembly, and the budding of progeny virion was assumed to be negligible relative to the preceding steps. Host cell. In our model, the host cell provides unlimited building blocks such as nucleoside triphosphates and amino acids for the growth of virus. However, as viral components accumulate during the course of infection, some key host components for translation such as initiation and elongation factors may be depleted [28, 63, 56] . Two main viral products, leader-mRNA and M protein, contribute to the deficiency by inhibiting the synthesis of host macromolecules at the transcription level [19, 28, 35, 36] . Because leader-mRNA starts to accumulate soon after the initiation of infection, and a small amount of the component is enough to trigger the inhibition [28, 35] , the pool of host factors is continuously reduced from the onset of infection. We quantify this reduction with a single decay rate constant specific to the type of host cell: where f dec is the level of host translation factors at time t, relative to that of the initial state of the cell before infection (at t ¼ 0), and k d,host is the decay rate constant. The inhibition by the leader mRNA causes a first-order decay of the host factors, resulting in a shortage of the ribosomes equipped with the accessory factors for viral translation in the late infection stage in our model. Unlike viral transcription and replication, viral translation is directly affected by the decay of host factors since it depends entirely on host machinery. In the early infection, host mRNAs outnumber viral mRNAs and thereby successfully compete for the host translation machinery. However, the newly synthesized M proteins inhibit the host transcription initiation and the export of host mRNAs from the nucleus to the cytoplasm [57] , thereby causing a gradual shift in translation from host mRNAs to viral mRNAs. For our model we assumed that the potency of the inhibition by the M protein was independent of the type of cell and its differentiation state [36] , and we developed an empirical formula using available experimental data from the literature [36] to account for the competition between host and viral mRNAs for ribosomes. Lyles at al. cotransfected the host cells with VSV M mRNA and chloramphenicol acetyl transferase (CAT) plasmid DNA, and then they quantified the expression of CAT based on its activity, as a function of the expression of VSV M protein [36] . In their experiment the gene expression of CAT was more reduced at higher M protein expression levels. We assume that the decrease of the expression of CAT (or its activity decrease) is proportional to the decrease of the occupancy of host mRNAs by the translation machinery. Using their experimental data, the occupancy of host mRNAs by the translation machinery is correlated with the number of newly synthesized M proteins in the cytoplasm: where rib_host is the fraction of the translation machinery associated with host mRNAs, and M cell is the total number of newly synthesized M proteins per cell. Considering the decay of host factors and the competition between host and viral mRNAs, we could derive a formula to quantify the number of the fully functional ribosomes that are available for the viral protein synthesis over time post-infection (nrib avail ): where nrib denotes the total concentration of ribosomes whether or not they incorporate all the required accessory factors for their translation function. Although the ribosomes distribute into membrane-bound and cytoplasmic forms, each class supporting the syntheses of the viral G protein and the other four viral proteins (N, P, M, and L), respectively, we treated the ribosomes in our model as one population. Initial condition for simulation. The initial condition for our simulation is set by a fixed number of infectious extracellular virus particles per cell (V ex (0)). At time zero (t ¼ 0), the number of bound virus particles and the level of all viral components within cells are zero. In our model, binding of extracellular virus particles to cells reduces their level (Equation 1), and an encapsidated genome and stoichiometric amounts of viral proteins (Table 2 , [28, 29] ) are then immediately released from each bound virus particle to the cytoplasm. Specifically, we assume that all N proteins from a bound virus particle are released as a form of encapsidated genome complex. Downstream processes, beginning with transcription, are then initiated. In our simulation, viral infection starts with rib_host ¼ 0.99 (with rib_host ¼ 0.9, ; 0.9995 simulations showed the same results). Other key model parameters for simulation are summarized in Table 1 . In addition, a nomenclature list is shown in Table 4 . Figure S1 . Ratio of Genome to Anti-Genome During Simulated VSV Infection Depends on Relative Promoter Strengths The relative strength of the anti-genomic promoter relative to the genomic promoter is given by S prom . PI stands for time post-infection. Figure S4 . Productivity Ranking of Six Gene-Shuffled Viruses Provides Parameter Constraint For each value of S cond and its corresponding host parameter values (k d,host and S rib ), the virus productivity of each gene-shuffled virus in BHK cells was determined and normalized by the highest productivity. Simulated rankings of six strains matched experimentally observed rankings [9] over a narrow range of S cond , indicated by the bar. Initial number of infectious extracellular virus particles per cell was three. Found at DOI: 10.1371/journal.pcbi/0020116.sg004 (872 KB TIF). Protocol S1. Parameter Estimation
57
Delivery Systems for the Direct Application of siRNAs to Induce RNA Interference (RNAi) In Vivo
RNA interference (RNAi) is a powerful method for specific gene silencing which may also lead to promising novel therapeutic strategies. It is mediated through small interfering RNAs (siRNAs) which sequence-specifically trigger the cleavage and subsequent degradation of their target mRNA. One critical factor is the ability to deliver intact siRNAs into target cells/organs in vivo. This review highlights the mechanism of RNAi and the guidelines for the design of optimal siRNAs. It gives an overview of studies based on the systemic or local application of naked siRNAs or the use of various nonviral siRNA delivery systems. One promising avenue is the the complexation of siRNAs with the polyethylenimine (PEI), which efficiently stabilizes siRNAs and, upon systemic administration, leads to the delivery of the intact siRNAs into different organs. The antitumorigenic effects of PEI/siRNA-mediated in vivo gene-targeting of tumor-relevant proteins like in mouse tumor xenograft models are described.
Altered expression levels of certain genes play a pivotal role in several pathological conditions. For example, in many cancers the upregulation of certain growth factors or growth factor receptors, or the deregulation of intracellular signal transduction pathways, represents key elements in the process of malignant transformation and progression of normal cells towards tumor cells leading to uncontrolled proliferation and decreased apoptosis. Since these processes may result in the direct, autocrine stimulation of the tumor cell itself as well as the paracrine stimulation of other cells, including the stimulation of tumor-angiogenesis, many novel therapeutic strategies focus on the reversal of this effect, that is, the inhibition of these proteins or the downregulation of their expression. Likewise, several other diseases have been firmly linked to the (over-)expression of endogenous wildtype or mutated genes. Taken together, in addition to strategies based on the inhibition of target proteins, for example, by low molecular weight inhibitors or inhibitory antibodies, this opens an avenue to gene-targeting approaches aiming at decreased expression of the respective gene. The first method to be introduced for the specific inhibition of gene expression was the use of antisense oligonucleotides in the late 1970s [1, 2] . Upon their introduction into a cell, antisense ODNs are able to hybridize to their target RNA leading to the degradation of the RNA-DNA hybrid double strands through RNAase H, to the inhibition of the translation of the target mRNA due to a steric or conformational obstacle for protein translation and/or to the inhibition of correct splicing. In the early 1980s, the discovery of ribozymes, that is, catalytically active RNAs which are able to sequence-specifically cleave a target mRNA, further expanded gene-targeting strategies [3] [4] [5] . Subsequently, both methods were extensively studied and further developed with regard to the optimization of targeting efficacies and antisense-ODN/ribozyme delivery strategies in vitro and in vivo. Most recently, another naturally occurring biological strategy for gene silencing has been discovered and termed RNA interference (RNAi). Since RNAi represents a particularly powerful method for specific gene silencing and is able to provide the relatively easy ablation of the expression of any given target gene, it is now commonly used as a tool in biological and biomedical research. This includes the RNAimediated targeting in vitro and in vivo for functional studies of various genes whose expression is known to be upregulated as well as the development of novel therapeutic approaches based on gene targeting. double-stranded RNA molecules as described first in C elegans by Fire et al [6] who then introduced the name RNA interference. These findings also explained earlier observations in petunias which turned white rather than purple upon the introduction of the "purple gene" in form of dsRNA [7] , and on gene silencing by antisense oligonucleotides as well as by sense oligonucleotides in C elegans [8] . Subsequent studies demonstrated that RNAi, while described under different names (posttranscriptional gene silencing (PTGS), cosuppression, quelling), is present in most eukaryotic organisms with the response to dsRNA, however, being more complicated in higher organisms. RNAi relies on a multistep intracellular pathway which can be roughly divided into two phases, that is, the initiation phase and the effector phase. In the initiation phase, double-stranded RNA molecules from endogenous or exogenous origin present in the cell are processed through the cleavage activity of a ribonuclease III-type protein [9] [10] [11] [12] into short 21-23 nucleotide fragments termed siRNAs. These effector siRNAs, which contain a symmetric 2 nt overhang at the 3 -end as well as a 5 -phosphate and a 3hydroxy group, are then in the effector phase incorporated into a nuclease-containing multiprotein complex called RISC (RNA-induced silencing complex) [13] . Several structural and biochemical studies have shed light on the processing of double-stranded RNA and the formation of the RISC complex (see, eg, [14] for a recent review). Through unwinding of the siRNA duplex by an RNA helicase activity [15] , this complex becomes activated with the single-stranded siRNA guiding the RISC complex to its complementary target RNA. Upon the binding of the siRNA through hybridization to its target mRNA, the RISC complex catalyses the endonucleolytical cleavage of the mRNA strand within the target site, which, due to the generation of unprotected RNA ends, results in the rapid degradation of the mRNA molecule. With the RISC complex being recovered for further binding and cleavage cycles, the whole process translates into a net reduction of the specific mRNA levels and hence into the decreased expression of the corresponding gene. For an overview of the RNAi pathway, see Figure 1 . While from this mechanism it becomes obvious that siRNA molecules complementary to the target mRNA and thus being able to serve as a guide sequence for the RISC complex play a pivotal role in this process, they need not be derived from long double-stranded precursor molecules. Rather, omitting the initiation phase, they can be delivered directly into the target cell ( Figure 1 , upper right arrow). Several studies have led to the development of guidelines for the generation of siRNAs which are optimal in terms of efficacy and specificity [12, 16] . This includes the initial definition of the preferable length (19-25 bp) combined with a low G/C content in the range between 36% and 52% and the requirement of symmetric 2 nt overhangs at the 3 -end [16] [17] [18] . Later studies on synthetic siRNA molecules, however, revealed an up to 100-fold higher targeting efficacy in the case of even longer duplexes (25-30 nucleotides) which act as a substrate for Dicer and which therefore allow the direct incorporation of the newly produced siRNAs into the RISC complex [19] . As to be expected, intramolecular foldback structures which can result from internal repeats or palindrome sequences decrease the numbers of functional siRNA molecules with silencing capability [20] . Additional silencing-enhancing criteria include an A in position 3 and a G at position 13 of the sense strand, the absence of a C or G at position 19 and, most importantly, a U in position 10 of the sense strand. Since nucleotides 10-11 represent the site of the RISC-mediated cleavage of the target mRNA, this indicates that RISC is comparable to most other endonucleases in preferentially cleaving 3 of U rather than any other nucleotide [20, 21] . Furthermore, it was shown more generally that the thermodynamic flexibility of the positions 15-19 of the sense strand correlates with the silencing efficacy and that the presence of at least one A/U base pair in this region improves siRNA-mediated silencing efficacy due to a decreased internal stability of its 3 -end [20] . Still, different siRNA sequences may display differing efficacies, which suggest additional still unknown criteria for optimal siRNA selection and emphasize the influence of target mRNA accessibility. In fact, several studies also correlate the siRNA efficacy with the mRNA secondary structure [18, [22] [23] [24] [25] [26] [27] . In conclusion, apart from the selection criteria defined above, the individual screening of different siRNAs for highly efficient and specific duplexes, or the pooling of multiple siRNAs, is the most effective approach to increase siRNAmediated targeting efficacy. For the design of effective siRNAs, several algorithms on publicly accessible web sites are available (see [28] for review). To reduce the risk of nonspecific ("off-target") effects of the siRNAs, a homology search of the targeting sequence against a gene database is necessary and already incorporated in some of these web sites. Nevertheless, it has also been shown that siRNAs may cross-react with targets of limited sequence similarity when regions of partial sequence identity between the target mRNA and the siRNA exist. In fact, in some cases regions comprising of only 11-15 contiguous nucleotides of sequence identity were sufficient to induce gene silencing [29] . The prediction of these off-target activities is difficult so far. An additional mechanism that may lead to nonspecific effects in vivo relies on the interferon system [30] [31] [32] [33] which is induced when double-stranded RNA molecules enter a cell activating a multi-component signalling complex. This effect is particularly true for long dsRNA molecules and essentially prevents them from being used as inducers of RNA interference in mammalian systems. The development of synthetic siRNAs [10, 12, 33, 34] largely circumvents this problem since they seem to be too small. However, some synthetic siR-NAs do induce components of the interferon system which seems to be dependent on their sequence [31, 32, 35] as well as, in the case of in vitro transcribed siRNAs, on the 5 initiating triphosphate [36] . Thus, strategies to avoid as far as possible the unwanted interferon response upon application of siRNAs in vivo will include a design of siRNAs without known interferon-stimulating sequences, the use of the lowest possible siRNA dose to still achieve the desired effect and optimized siRNA delivery methods. Based on the known mechanisms of antisense technology, ribozyme-targeting or RNAi, small oligonucleotides or plasmid-based expression vectors can be used to specifically downregulate the expression of a given gene of interest or of pathological relevance in vitro. In principle, this also applies to the in vivo situation leading to novel, potentially relevant therapeutic approaches. For the delivery of therapeutic nucleic acids, viral vectors have been used which have the advantage of high transfection efficacy due to the inherent ability of viruses to transport genetic material into cells. On the other hand, however, viral systems show a limited loading capacity regarding that the genetic material are rather difficult to produce in a larger scale and, most importantly, pose severe safety risks due to their oncogenic potential and their inflammatory and immunogenic effects which prevent them from repeated administration [37] [38] [39] [40] . In the light of these problems, concerns, and limitations, nonviral systems have emerged as a promising alternative for gene delivery. Main requirements are the protection of their nucleic acid "load" as well as their efficient uptake into the target cells with subsequent release of the DNA or RNA molecules and, if necessary, their transfer into the nucleus. Several strategies can be distinguished, mainly lipofection and polyfection relying on cationic lipids or polymers, respectively (see, eg, [41] [42] [43] ). The efficient protection against enzymatic or nonenzymatic degradation is particularly important for RNA molecules including siRNAs. In fact, while the therapeutic potential of siRNAs for the treatment of various diseases is in principle very promising, limitations of transfer vectors may turn out to be rate-limiting in the development of RNAi-based therapeutic strategies. One approach to solve this problem is the use of DNA expression plasmids which encode palindromic hairpin loops with the desired sequence. Upon transcription and folding of the RNA, the doublestranded short hairpin RNAs (shRNAs) are recognized by Dicer and cleaved into the desired siRNAs. Additionally, an in vitro method has been described recently which is based on the expression of shRNAs in E coli and their delivery via bacterial invasion [44] . While all these different DNAbased systems offer the advantage of siRNA expression with a longer duration and a probably higher level of gene silencing, they still rely on (viral or nonviral) delivery of DNA molecules and again raise safety issues in vivo. Hence, the direct delivery of siRNAs molecules, derived from in vitro transcription or chemically synthesized, offers advantages over DNA-based strategies and may be preferable for in vivo therapeutic use. In the last years, a large body of studies has been published which describe different strategies for the systemic or local application of siRNAs in vivo. Tables 1-3 give an overview. The probably largest number of papers focuses the use of unmodified siRNAs (Table 1 ) whose administration is often performed IV by hydrodynamic transfection (high pressure tail vein injection). While this method is widely used and in some cases led to efficient target gene inhibition in the liver and, to a lesser extent, in lung, spleen, pancreas, and kidney, it may suffer from certain technical and practical limitations at least in a therapeutical setting since it relies on the rapid IV injection of a comparably large volume (>= 1 ml/mouse/injection, in theory equivalent to a ∼ 3 l IV bolus injection in man). Alternative strategies for the application of naked siRNAs include various delivery routes which, however, often provide an only local administration or rely on an administration at least close to the target tissue or target organ, thus restricting the number of target organs which may not be relevant for certain diseases. It should also be noted that several studies described here and below use rather large amounts of siRNAs and that upon intravenous injection of siRNAs the liver is the primary site of siRNA uptake. As an alternative approach for the application of siR-NAs in vivo, their delivery by liposomes/cationic lipids has been described. For liposome-based siRNA formulations, a wide variety of modes of application allowing local or systemic delivery has been used (Table 2) . Finally, several other strategies for local or systemic siRNA administration have been explored, including chemical modifications of siRNA molecules, electropulsation, polyamine, or other basic complexes, atelocollagen, virosomes, and certain protein preparations (Table 3 ). An alternative approach relies on the complexation of unmodified siRNA molecules with a cationic polymer, polyethylenimine (PEI). Polyethylenimines (PEIs) are synthetic polymers available in branched or linear forms (Figure 2 , upper panels) and in a broad range of molecular weights from < 1000 Da to > 1000 kd. Commercial PEI preparations, although labelled with a defined molecular weight, consist of PEI molecules with a broad molecular weight distribution [45] [46] [47] . PEIs possess a high cationic charge density due to a protonable amino group in every third position [48, 49] . Since no quarternary amino groups are present, the cationic charges are generated by protonation of the amino groups and hence are dependent on the pH in the environment (eg, 20% at pH 7.4, see [50] for review). Due to its ability to condense and compact the DNA into complexes, which form small colloidal particles allowing efficient cellular uptake through endocytosis, PEI has been introduced as a potent DNA transfection reagent in a variety of cell lines and in animals for DNA delivery (for review, see [51, 52] and references therein). In fact, in several studies PEI has been shown to be able to deliver large DNA molecules such as 2.3 Mb yeast artificial chromosomes (YACs) [53] as well as plasmids or small oligonucleotides [48, [54] [55] [56] into mammalian cells in vitro and in vivo. The N/P ratio, which indicates the ratio of the nitrogen atoms of PEI to DNA phosphates in the complex and thus describes the amount of PEI used for complex formation independent of its molecular weight, influences the efficiency of DNA delivery. A positive net charge of the complexes, resulting from high N/P ratios, inhibits due to electrostatic repulsion their aggregation and improves their solubility in aqueous solutions as well as their interaction with the negatively charged extracellular matrix components and thus their cellular uptake [57] . Additionally, the strong buffer capacity, described by the "proton sponge hypothesis" which postulates enhanced transgene delivery by cationic polymer-DNA complexes (polyplexes) containing H + buffering polyamines due to enhanced endosomal Cl − accumulation and osmotic swelling/lysis [48] , seems to be responsible for the fact that PEI-based delivery does not require endosome disruptive agents for lysosomal escape. This tight condensation of the DNA molecules as well as the buffering capacity of PEI in certain cellular compartments like endosomes and lysosomes also protects DNA from degradation [48, 49, 58, 59] . PEIs have been successfully used for nonviral gene delivery in vitro and in vivo. While initial publications showed increased transfection efficacies when using high molecular weight PEIs [45] , more recent studies demonstrated the advantages of certain low molecular weight PEIs [47, 60, 61] . The higher transfection efficacy of low molecular weight PEIs may be due to a more efficient uptake of the resulting PEI/DNA complexes, a better intracellular release of the DNA and/or lower in vitro cytotoxicity as compared to high molecular weight PEI [60] [61] [62] [63] . In fact, a decrease in the molecular weight of the PEI leads to an increase in complex size which may be favourable at least for in vitro use [64, 65] . On the other hand, other PEIs with very low molecular weight (< 2 kd) display little or no transfection efficacy even at very high N/P ratios which may be attributed to the fact that a decrease in the molecular weight of PEI has been shown to translate into an increasingly lower ability to form small complexes [63] . Therefore, low molecular weight PEIs require higher N/P ratios for optimal transfection efficacies as compared to higher molecular weight PEIs since higher N/P ratios lead to an increase in compaction with reduced complex sizes and a reduced tendency of the complexes to aggregate due to hydrophobic interactions [61, 63, 64] . Nevertheless, while several parameters have been extensively studied, some precise determinants for transfection efficacy remain to be elucidated (see [50, 66] for review). Also, the mechanism of the cytotoxic effects of PEI complexes is only poorly understood. It may rely on the formation of large aggregates in the range of up to 2 μm which, when formed on the cell surface, impairs membrane functions finally leading to cell necrosis [60] . Clearly, there is a trend towards low molecular weight PEIs as rather nontoxic delivery reagents in vitro and in vivo, which combine high biocompatibility and reduced side-effects thus also allowing to employ larger PEI/DNA complex amounts without significant cytotoxicity. More recently, the use of polyethylenimines has been extended towards the complexation and delivery of RNA molecules, especially small RNA molecules like 37 nt all-RNA ribozymes [67] [68] [69] and siRNAs [70] (Figure 2 ). While chemically unmodified RNA molecules are very instable and prone to rapid degradation, the PEI complexation has been shown to lead to an almost complete protection against enzymatic or nonenzymatic degradation. In fact, PEI-complexed siR-NAs, which are [ 32 P]-labeled for better detection, remain intact in vitro for several hours even in the presence of RNase A or fetal calf serum at 37 • C, while non-complexed siRNAs are rapidly degraded (Figure 3(a) ). This indicates that siRNA molecules are efficiently condensed and thus fully covered and protected by PEI. Indeed, the analysis of PEI/siRNA complexes by atomic force microscopy showed the absence of free siRNAs or siRNA molecule ends and thus confirms these findings regarding an efficient complexation (Grzelinski et al, submitted). However, while the complex stability seems to be sufficient for siRNA protection with all PEIs tested (Werth et al, in press; Aigner et al, unpublished data), several of these complexes do not show any targeting efficacy at all. In fact, only when using certain polyethylenimines, PEI/siRNA complexes are efficiently delivered into target cells in vitro, where siRNAs are released and display bioactivity (Figures 1 and 2) . In general and as seen before for PEI/DNA complexes (see above), the transfection efficacy is dependent on the PEI used, also indicating that the siRNA targeting efficiency mainly depends on the endocytotic uptake of the complex and/or its intracellular decomposition rather than on the in vitro complex stability. Good results were obtained with commercially available JetPEI [70] while the in vivo JetPEI from the same supplier showed only poor siRNA delivery efficacies [71] . Likewise, a novel low molecular weight PEI based on the fractionation of a commercially available polyethylenimine demonstrates high siRNA protection and delivery efficacies in vitro (Werth et al, in press). Under certain conditions, the PEI/RNA (siRNA or ribozyme) complexes retain their physical stability and biological activity also after lyophilization ( [72] and Werth et al, in press). Although the PEI transfection is only transient, data from our lab show that PEI/siRNA effects are stable for at least 7 days (Urban-Klein and Aigner, unpublished results). Finally, another study has explored the use of siRNA nanoplexes comprising of PEI that is PEGylated with an RGD peptide ligand attached at the distal end of the PEI. Again, siRNA nanoplexes protect siRNAs against serum degradation and show in vitro activity [73] . The ultimate goal is the application of siRNAs in vivo which has been explored in some studies in different mouse models. Ge et al showed that PEI-complexed siRNAs targeting conserved regions of influenza virus genes are able to prevent and treat influenza virus infection in mice. Upon IV injection, PEI promoted the delivery of siRNAs into the lungs where, either given before or after virus infection, siRNA reduced influenza virus production in the lungs [74] . Most biological effects of the systemic application of PEIcomplexed siRNAs, however, have been determined in different mouse tumor models and by targeting different proteins which have been shown previously to be tumor-relevant. This includes the epidermal growth factor receptor HER-2 (c-erbB-2/neu), the growth factor pleiotrophin (PTN), and vascular endothelial growth factor (VEGF) and its receptor (VEGF R2), and the fibroblast growth factor-binding protein FGF-BP. The in vivo administration of PEI complexed, but not of naked siRNAs, through IP or subcutaneous injection resulted in the detection of intact siRNAs even hours after injection (Figure 3(b) ). Radiolabeled siRNA molecules were found in several organs including subcutaneous tumors, muscle liver, kidney and, to a smaller extent, lung and brain. It is important to note that the siRNAs were actually internalized by the tissues as indicated by the fact that blood was negative for siRNAs ( Figure 3(b) ). Overexpression of the HER-2 receptor has been observed in a wide variety of human cancers and cancer cell lines. Since HER-2 displays strong cell growth-stimulating and antiapoptotic effects especially through heterodimer formation with other members of the EGFR family, its overexpression has been established as a negative prognostic factor and linked to a more aggressive malignant behaviour of tumors (eg, [75] ). Consequently, HER-2 qualifies as an attractive target molecule for antitumoral treatment strategies including anti-HER-2 antibodies, low molecular weight inhibitors, or HER-2-specific gene-targeting approaches. In fact, the relevance of HER-2 (over-)expression in tumor growth has been established in several in vitro HER-2 targeting studies including the use of ribozymes [76, 78, 79] or siRNAs [80, 81] . proposed mechanism of PEI-mediated siRNA transfer. Due to electrostatic interactions, PEI is able to complex negatively charged siRNAs leading to a compaction and the formation of small colloidal particles which are endocytosed. The "proton sponge effect" exhibited by PEI complexes leads to osmotic swelling and ultimately to the disruption of the endosomes. siRNAs are protected from degradation due to their tight condensation in the complex and the buffering capacity of PEI. Upon their release from the PEI-based complex, intact siRNAs are incorporated into the RISC complex and induce RNAi (see Figure 1 ). It was demonstrated that HER-2 reduction in vitro leads, among others, to the inhibition of cell proliferation and increased apoptosis. The systemic treatment of athymic nude mice bearing subcutaneous SKOV-3 ovarian carcinoma tumor xenografts through IP injection of PEI-complexed HER-2-specific siRNAs led to marked antitumoral effects as seen by a significant reduction tumor growth (Figure 4 ) [70] . PEIcomplexed nonspecific siRNAs or HER-2-specific, naked siR-NAs had no effects. This was paralleled by the detection of intact HER-2-specific siRNAs in the tumors of the specific treatment group already 30 min after administration and for at least 4 h, and by the downregulation of HER-2 on mRNA and protein levels [70] . Another receptor, VEGF R2, was targeted in a study employing self-assembling nanoparticles based on siRNAs complexed PEI which is PEGylated with an RGD peptide ligand attached at the distal end of PEG. While the PEGylation allows steric stabilization and reduces nonspecific interactions of the complexes, the RGD motif provided tumor selectivity due to their ability to target integrins expressed on activated endothelial cells in the tumor vasculature. Upon IV administration into mice bearing subcutaneous N2A neuroblastoma tumor xenografts, a selective tumor uptake and a VEGF R2 downregulation were observed, resulting in decreased tumor growth and tumor angiogenesis [73] . The receptor ligand, VEGF, is a mitogenic and angiogenic growth factor stimulating tumor growth and angiogenesis in several tumors including prostate carcinoma. Thus, it may represent attractive target molecule for RNAi-based genetargeting strategies also bearing in mind the double antitumoral effect due to reduction of tumor cell proliferation as well as tumor angiogenesis. The subcutaneous or intraperitoneal injection of VEGF-specific siRNAs complexed with a novel PEI obtained through fractionation of a commercially available PEI (Werth et al, in press) resulted in the reduction of tumor growth due to decreased VEGF expression levels (Höbel and Aigner, unpublished results). The same was true for PEI/siRNA-mediated targeting of FGF-BP (Dai and Aigner, unpublished results), which has been established Figure 3 : Protection and in vivo delivery of siRNAs upon PEI complexation. In [70] (a) in vitro protection of siRNAs against nucleolytic degradation. [ 32 P] end-labeled siRNAs, complexed (upper panel) or not complexed (lower panel) with PEI, were subjected to treatment with 1 % fetal calf serum at 37 • C. At the time points indicated, the samples were analysed by agarose gel electrophoresis, blotting, and autoradiography. The bands represent full-length siRNA molecules indicating that PEI complexation leads to the efficient protection of siRNAs while noncomplexed siRNAs are rapidly degraded. (b,c) In vivo delivery of intact siRNAs upon PEI complexation. [ 32 P]-labeled siRNAs, complexed (+) or not complexed (−) with PEI, were injected IP into mice bearing subcutaneous SKOV-3 ovarian carcinoma cell tumor xenografts, and after 30 min (b) or 4 h (b) total RNA from various organ and tissue homogenates was prepared and subjected to agarose gel electrophoresis prior to blotting and autoradiography. The bands represent intact [ 32 P]-labeled siRNA molecules which for several hours are mainly found in tumor and muscle as well as in liver and, time-dependently, in kidney. Only little siRNA amounts are detected in the lung and traces in the brain. previously as "rate-limiting" for tumor growth and angiogenesis in several tumors ( [82, 83] , see [84] for review). Finally, PEI/siRNA-mediated targeting of pleiotrophin (PTN) exerted strong antitumoral effects. PTN is a secreted growth factor which shows mitogenic, chemotactic, angiogenic and transforming activity [85] [86] [87] [88] [89] [90] [91] [92] [93] and which is markedly upregulated in several human tumors including cancer of the breast, testis, prostate, pancreas, and lung as well as in melanomas, meningiomas, neuroblastomas, and glioblastomas. The in vivo treatment of nude mice through systemic subcutaneous or IP application of PEI-complexed PTN siRNAs led to the delivery of intact siRNAs into subcutaneous tumor xenografts and a significant inhibition of tumor growth. Likewise, in a clinically more relevant orthotopic mouse glioblastoma model with U87 cells growing intracranially, the injection of PEI-complexed PTN siRNAs into the CNS exerted antitumoral effects. This establishes, also in a complex and relevant orthotopic tumor model, the potential of PEI/siRNA-mediated PTN gene targeting as a novel therapeutic option in GBM, and further extends the modes of delivery of PEI/siRNA complexes intrathecal strategies as employed in the therapy of glioblastomas with antisense oligonucleotides. Only a few years after their discovery, siRNAs are catching up with ribozymes and antisense oligonucleotides as efficient tools for gene targeting in vitro and, more recently, also in vivo. This includes the exploration of their potential as therapeutics which will lead to the development of siRNA-based therapeutic strategies. Their ultimate success, however, will Figure 4 : Systemic treatment of mice with PEI-complexed HER-2-specific siRNAs leads to reduced growth of subcutaneous SKOV-3 tumor xenografts due to decreased HER-2 expression. In [70] athymic nude mice bearing subcutaneous tumor xenografts were injected IP with 0.6 nmoles HER-2-specific naked (open circles) or PEI-complexed (closed circles) siRNAs 2-3 times per week and tumor sizes were evaluated daily from the product of the perpendicular diameters of the tumors. Mean +/-standard error of the mean (SEM) is depicted and Student's unpaired t test was used for comparisons between data sets ( * * P < .03, * * * P < .01). Differences in tumor growth reach significance at day 5 indicating the antitumoral effects of the PEI-complexed HER-2-specific siRNAs. strongly depend on the development of powerful and feasible siRNA delivery strategies which need to address several issues including the stability/stabilization of siRNA molecules while preserving their efficacy and maintaining their genesilencing activity, an efficient delivery into the target organ(s) as well as a sufficiently long siRNA half life in the organism and particularly in the target organ. Thus, siRNA delivery strategies must provide siRNA protection and transfection efficacy, the absence of toxic and nonspecific effects, they must be efficacious also when using small amounts of siRNAs and must be applicable in various treatment regimens and in various diseases even when this requires to overcome biological barriers after their administration to reach their target tissue or target organ. The research done on DNA-based gene delivery, ribozyme-targeting, and antisense technology will facilitate this process since it already provides a basis of established technologies. This is also true for the complexation of siRNAs with polyethylenimine, which may represent a promising avenue for siRNA applications in vivo. This may eventually lead to novel therapeutic strategies. The work of A. Aigner is supported by the Deutsche Forschungsgemeinschaft (AI 24/5-1) and by the Deutsche Krebshilfe. The author would like to apologize to the authors whose primary works have not been cited due to length considerations.
58
Identification of new participants in the rainbow trout (Oncorhynchus mykiss) oocyte maturation and ovulation processes using cDNA microarrays
BACKGROUND: The hormonal control of oocyte maturation and ovulation as well as the molecular mechanisms of nuclear maturation have been thoroughly studied in fish. In contrast, the other molecular events occurring in the ovary during post-vitellogenesis have received far less attention. METHODS: Nylon microarrays displaying 9152 rainbow trout cDNAs were hybridized using RNA samples originating from ovarian tissue collected during late vitellogenesis, post-vitellogenesis and oocyte maturation. Differentially expressed genes were identified using a statistical analysis. A supervised clustering analysis was performed using only differentially expressed genes in order to identify gene clusters exhibiting similar expression profiles. In addition, specific genes were selected and their preovulatory ovarian expression was analyzed using real-time PCR. RESULTS: From the statistical analysis, 310 differentially expressed genes were identified. Among those genes, 90 were up-regulated at the time of oocyte maturation while 220 exhibited an opposite pattern. After clustering analysis, 90 clones belonging to 3 gene clusters exhibiting the most remarkable expression patterns were kept for further analysis. Using real-time PCR analysis, we observed a strong up-regulation of ion and water transport genes such as aquaporin 4 (aqp4) and pendrin (slc26). In addition, a dramatic up-regulation of vasotocin (avt) gene was observed. Furthermore, angiotensin-converting-enzyme 2 (ace2), coagulation factor V (cf5), adam 22, and the chemokine cxcl14 genes exhibited a sharp up-regulation at the time of oocyte maturation. Finally, ovarian aromatase (cyp19a1) exhibited a dramatic down-regulation over the post-vitellogenic period while a down-regulation of Cytidine monophosphate-N-acetylneuraminic acid hydroxylase (cmah) was observed at the time of oocyte maturation. CONCLUSION: We showed the over or under expression of more that 300 genes, most of them being previously unstudied or unknown in the fish preovulatory ovary. Our data confirmed the down-regulation of estrogen synthesis genes during the preovulatory period. In addition, the strong up-regulation of aqp4 and slc26 genes prior to ovulation suggests their participation in the oocyte hydration process occurring at that time. Furthermore, among the most up-regulated clones, several genes such as cxcl14, ace2, adam22, cf5 have pro-inflammatory, vasodilatory, proteolytics and coagulatory functions. The identity and expression patterns of those genes support the theory comparing ovulation to an inflammatory-like reaction.
In fish, as in other lower vertebrates, the post-vitellogenic period is very important for the completion of the oogenetic process. During this step, the follicle-enclosed postvitellogenic oocyte undergoes several key events such as the final acquisition of the ability to resume meiosis in response to the maturation-inducing steroid (MIS), the resumption of the meiotic process and, finally, its release from the surrounding follicular layers. In addition, the whole follicle (oocyte and surrounding follicular cells) undergoes a progressive differentiation ultimately leading to the release of a metaphase 2 oocyte. The key hormonal and molecular events involved in the control of meiosis resumption have been thoroughly studied and many studies have been dedicated to the action of gonadotropins, the regulation of steroidogenenic events and the action of the MIS (see [1] [2] [3] [4] [5] [6] for review). However, the associated follicular or extra-follicular events involved in concomitant processes such as oocyte-follicular cells cross talk and ovulationmechanisms have received far less attention. Nevertheless, several researchgroups have studied the periovulatory ovarian physiology using classical biochemical or histological tools and, later, molecular approaches. Thus, several studies have dealt with ovarian proteases in their participation in the ovulatory process [7] [8] [9] . Differential display PCR and suppressive subtractive hybridization (SSH) approaches have also been developed in order to identify new differentially regulated genes in the fish periovulatory ovary [10] [11] [12] [13] . In addition, numerous candidate gene studies have also been performed in the fish periovulatory ovary. Apart from genes related to hormonal controls, these studies were mostly dedicated to some specific gene families such as TGF beta family [14, 15] or connexins [16, 17] . Finally, fewer studies have simultaneously analyzed the expression profiles of several genes belonging to different families [18, 19] . However, in contrast to other biological processes, such as immune response [20] , the post-vitellogenic period has never benefited from genome-wide transcriptomic studies that could provide a global view of the molecular events occurring in the post-vitellogenic ovary undergoing oocyte maturation. In this context, the present study aimed at performing a transcriptomic analysis of the postvitellogenic rainbow (Oncorhynchus mykiss) trout ovary. In order to do so, 9152-gene rainbow trout cDNA microarrays were hybridized using RNA samples originating from rainbow trout ovarian tissue collected during late vitellogenesis, post-vitellogenesis and oocyte maturation. A statistical analysis was performed in order to identify all the genes exhibiting a differential expression over this period. In addition, a supervised clustering analysis was performed using only the differentially expressed genes in order to identify groups (or clusters) of genes exhibiting similar expression profiles. Furthermore, as a first step in a long-term transcriptomic analysis of the rainbow trout post-vitellogenic ovary, we deliberately chose to focus on 3 gene clusters exhibiting the most remarkable expression patterns. Finally, specific genes were selected in each cluster based on the novelty of their putative identity and/or function. For each gene, a real-time PCR analysis of their ovarian expression profiles was performed using additional ovarian RNA samples. Investigations were conducted according to the guiding principles for the use and care of laboratory animals and in compliance with French and European regulations on animal welfare. Two year old female rainbow trout (Oncorhynchus mykiss) were obtained during their first reproductive season from our experimental fish farm (Sizun, France) and held under natural photoperiod in a re-circulated water system in INRA experimental facilities (Rennes, France). The water temperature was kept constant at 12°C. Ovaries were sampled from individual females during late vitellogenesis (N = 6), post-vitellogenesis (N = 6) and during oocyte maturation (N = 6). Oocyte developmental stage was assessed under binocular microscope according to previously described criteria [21, 22] . Late vitellogenic samples were collected at the end of the vitellogenic process, approximately 3-4 weeks before expected ovulation. At this stage, germinal vesicle is not visible and no polarized cytoplasm area can be observed. Post-vitellogenic samples were collected 2-3 weeks later but before any noticeable morphological changes in yolk structure due to the process of meiosis resumption. At this stage, oocytes can display a subperipheral or peripheral germinal vesicle. When germinal vesicle is not visible, a dark mass of polarized cytoplasm can be observed. Oocyte maturation samples were collected after meiosis resumption. Those samples were thus collected after yolk clarification and around the time of germinal vesicle breakdown (GVBD). For tissue collection, trout were deeply anesthetized in 2-phenoxyethanol, killed by a blow on the head and bled by gill arch section. Ovaries were then dissected out of the body cavity under sterile conditions. Ovarian aliquots were frozen in liquid nitrogen and stored at -80°C until RNA extraction. Low oligonucleotide signals (lower than three times the background level) were excluded from the analysis. After this filtering step, signal processing was performed using the vector oligonucleotide data to correct each spot signal by the actual amount of DNA present in each spot. After correction, signal was normalized by dividing each gene expression value by the median value of the array. A statistical analysis was performed in order to identify differentially expressed genes between late vitellogenic, post-vitellogenic and maturing groups using SAM software [28] . Three 2-by-2 statistical analyses were performed in order to compare each group with the two other ones. In addition, a comparison was performed between samples taken prior to meiosis resumption (from late and post-vitellogenic females, N = 7) and during oocyte maturation (N = 6). For each comparison, the lowest false discovery rate (FDR) was used to identify differentially abundant genes. All genes identified in at least one of the above comparisons were kept for clustering analysis in order to characterize the expression profiles of statistically relevant genes. For supervised clustering analysis [29] , data was log transformed, median-centered and an average linkage clustering was performed using CLUSTER software [29] . Clusters were visualized using TREEVIEW software [29] . Rainbow trout sequences originating from INRA Agenae [24] and USDA [30] EST sequencing programs were used to generate publicly available contigs [31] . The 8th version (Om.8, released January 2006) was used for BlastX [32] comparison against the Swiss-Prot database (January 2006) [33] . The score of each alignment was retrieved after performing a BlastX comparison. In addition, for each EST spotted onto the membrane, the accession number of the corresponding rainbow trout cluster (Unigene Trout, January 2006), if any, was retrieved from the UniGene database [34] . Real-time PCR was performed using all RNA (N = 18) samples including those used for microarray analysis. Several over and under expressed clones belonging to three selected remarkable clusters, were selected according to their putative identity and/or function for analysis. Reverse transcription and real time PCR were performed as previously described [19] . Briefly, 3 μg of total RNA were reverse transcribed using 200 units of Moloney murine Leukemia virus (MMLV) reverse transcriptase (Promega, Madison, WI) and 0.5 μg random hexamers (Promega) per μg of total RNA according to manufacturer's instruction. RNA and dNTPs were denatured for 6 min at 70°C, then chilled on ice for 5 min before the reverse transcription master mix was added. Reverse transcription was performed at 37°C for 1 hour and 15 min followed by a 15 min incubation step at 70°C. Control reactions were run without MMLV reverse transcriptase and used as negative controls in the real-time PCR study. Real-time PCR experiments were conducted using an I-Cycler IQ (Biorad, Hercules, CA). Reverse transcription products were diluted to 1/25, and 5 μl were used for each real-time PCR reaction. Triplicates were run for each RT product. Real-time PCR was performed using a real-time PCR kit provided with a SYBR Green fluorophore (Euro-gentec, Belgium) according to the manufacturer's instructions and using 600 nM of each primer. After a 2 min incubation step at 50°C and a 10 min incubation step at 95°C, the amplification was performed using the following cycle: 95°C, 20 sec; 60°C, 1 min, 40 times. For all primer pairs, the relative abundance of target cDNA within sample set was calculated from a serially diluted ovarian cDNA pool using the I-Cycler IQ software. This dilution curve was used to ensure that PCR efficiency was within an 80-100% range and that amplification was linear within sample set. After amplification, a fusion curve was obtained using the following protocol: 10 sec holding followed by a 0.5°C increase, repeated 80 times and starting at 55°C. The level of 18S RNA in each sample was measured and used for target genes abundance normalization within sample set. In addition to the genes identified from the transcriptomic analysis, a widely used standard gene, elongation factor 1 alpha (ef1α), was monitored using the same sample set to validate the normalization procedure. GenBank accession number and primer sequences are shown in table 1. Statistical analyses were performed using Statistica 7.0 software (Statsoft, Tulsa, OK). Differences between ovarian developments stages were analyzed using non parametric U tests. After signal processing, 8263 clones out of 9152 were kept for further analysis. From the statistical analysis, 310 clones were found to exhibit a differential abundance between at least 2 of the studied ovarian stages (late vitellogenesis, post-vitellogenesis and oocyte maturation). For all SAM analyses performed, the false discovery rate (FDR) was always lower than 0.7%. Among the 310 identified clones, 90 were up-regulated during oocyte maturation while 220 exhibited an opposite pattern. A clustering analysis was performed using only expression data of the 310 identified clones in order to characterize the expression profiles of those genes. The clustering analysis clearly separated the over from the under expressed genes ( Figure 1 ). The number of each clone in the clustering analysis ( Figure 1 ) was kept in subsequent tables 1, 2, 3, 4 and in the text. Within down-regulated genes, a cluster of 32 genes (cluster 1, Figure 1 ) was characterized by high expression levels during late vitellogenesis, low levels during oocyte maturation and intermediate or variable levels during post-vitellogenesis ( Figure 1 ). Within up-regulated clones, a cluster of 44 genes (cluster 2, Figure 1 ) was characterized by a strong over expression at the time of meiosis resumption while a cluster of 14 genes (cluster 3, Figure 1 ) exhibited a very low expression during late and post-vitellogenesis and an up-regulation before meiosis resumption ( Figure 1 ). The rainbow trout (Oncorhynchus mykiss) genome has not been sequenced and the number of characterized rainbow trout proteins and mRNAs is limited. The identity of studied transcripts was therefore based on the most significant hit obtained after performing a BlastX search against the SwissProt database. For the clones belonging to cluster 1-3, the results of this blast search is presented in tables 2, 3, 4. For each clone, the identity of the best hit in SwissProt and the score value of the BlastX comparison are given. However, this similarity search was performed using all EST sequences available in public databases and not using fully characterized cDNAs displaying the full coding sequence of the transcript. For some of the clones spotted on the trout array, the corresponding mRNA was previously characterized and made available in public databases. The identity of those clones is therefore unambiguous. In contrast, for some other clones, the best hit in SwissProt only gives significant, but incomplete, information. This is especially true for protein family members for which only a phylogenetic analysis will allow a more relevant identification of the gene. However, the name of the best hit was used in the text for clarity reasons. This large cluster of 32 clones (# 189-220) was characterized by a clear under expression at the time of oocyte maturation. Among those 32 clones, 29 belonged to a UniGene cluster and 30 had a significant hit in Swiss-Prot ( (Table 3 ). In addition, 39 clones exhibited a significant hit in SwissProt while 5 clones had no significant sequence similarities with known genes (Table 3) . Within this cluster, several genes exhibited inflammation or ovulation-related functions. Thus some of the clones exhibited sequence similarities with human chemokine cxcl14 (clone # 250), clawed frog adam22 (clone # 258) and coagulation factor V (cf5) (clone # 235). In addition, one clone (# 245) exhibited strong sequence similarity with human angiotensin-converting enzyme 2 precursor (ace2). Two clones (# 238 and 239) exhibited strong sequence similarity with salmon (Oncorhynchus keta) vasotocin-neurophysin (avt) and isotocin-neurophysin respectively. Finally, cluster 3 also contained clones exhibiting sequence similarity with, human Forkhead box protein O3A and human pendrin, also know as solute carrier family 26 member 4 (slc26) (clone # 236). Within cluster 2, cxcl14, adam22, slc26, avt, ace2 and cf5 genes were kept for real-time PCR analysis. This small cluster of 14 clones (# 296-309) was characterized by an over expression occurring earlier than for the genes belonging to cluster 3. Among those 14 clones, 12 belonged to a UniGene cluster and 11 had a significant hit in SwissProt (Table 4 ). Two clones (# 305 and 306) were most similar to rat and human aquaporin 4 (aqp4) respectively. These 2 clones belonged to the same UniGene cluster (Omy.23866). In addition, one clone (# 296) was most similar to mouse serine protease 23 (sp23). Within cluster 3, aqp4 and sp23 genes were kept for real-time PCR analysis. For all the genes selected for the real-time PCR analysis, a similar up or down regulation was observed between microarray and real-time PCR experiments. We observed a dramatic under expression of aromatase (cyp19a1, clones # 196 and 198) in the ovary during the preovulatory period ( Figure 2 ). The mRNA abundance of cyp19a gene during oocyte maturation was more than 200 times lower than during late vitellogenesis. In addition, successive decreases of cyp19a gene expression levels were observed during post-vitellogenesis and during oocyte maturation ( Figure 2 ). The mRNA abundance of vitamin K-dependent protein S precursor gene (clones # 199 and 200) was lower during oocyte maturation than during late or post-vitellogenesis. In contrast, no significant differences were observed between late and post-vitellogenesis ( Figure 2) . A similar expression profile was observed for Cytidine monophosphate-N-acetylneuraminic acid hydroxylase (cmah) gene ( Figure 2 ). We observed a strong over expression of aquaporin 4 (aqp4) gene during post-vitellogenesis and at the time of oocyte maturation (Figure 3) . The mRNA abundance of aqp4 gene exhibited a 6-fold increase during post-vitellogenesis and a further 12-fold increase during oocyte maturation. In addition, the mRNA abundance of pendrin (slc26) gene exhibited a 1500-fold increase during oocyte maturation while no significant differences were observed between late and post-vitellogenesis. Similarly, vasotocin (avt) mRNA abundance exhibited a 500-fold increase at the time of oocyte maturation ( Figure 3 ). Angiotensinconverting enzyme 2 (ace2) gene expression levels exhibited a 215-fold increase between late vitellogenesis and oocyte maturation (Figure 3) . A similar profile was observed for the chemokine cxcl14 gene. The mRNA abundance of this gene exhibited a 35-fold increase between late vitellogenesis and oocyte maturation ( Figure 3 ). The mRNA abundance of coagulation factor V (cf5) gene exhibited a 177-fold increase between late or postvitellogenesis and oocyte maturation while adam22 mRNA abundance exhibited a 6-fold increase between late or post-vitellogenesis and oocyte maturation ( Figure 3) . Finally, the mRNA abundance serine protease 23 (sp23) gene monitored during oocyte maturation was higher than in the late vitellogenic ovary. However, this difference was not significantly different (p = 0.078). The mRNA abundance of elongation factor 1 alpha (ef1α), a translation regulatory protein commonly used as a stable reference, did not exhibit any significant difference over the preovulatory period ( Figure 3 ). The hybridization of radiolabeled cDNAs with cDNAs deposited on nylon membranes has been used for several decades. However, the use of nylon cDNA microarrays is not very common in comparison to glass slide microarray technology. Nevertheless, this technology has successfully been used for several years [27, 35] . In the present study we used similar cDNA manufacturing and hybridization protocols. While most of the 9152 clones used to generate the microarray putatively correspond to distinct genes, a small proportion of genes are represented by 2 distinct clones (e.g clones belonging to the same UniGene cluster). In our data, it is noteworthy that those clones are usually found in the same gene clusters (e.g clones #196 and 198, #199 and 200, #305 and 306). Since the position of clones in the clustering analysis is based on the correlation between their profiles, this indicates that they display very similar expression profiles. In addition, for all genes selected for real-time PCR analysis, the over or under expression observed was always consistent with microarray data. Furthermore, the expression of ef1α, a widely used reference gene, was stable over the preovulatory period. Together, these observations suggest that our overall microarray analysis is extremely robust and reliable. In the present study, we identified 310 genes exhibiting a differential expression during the preovulatory period. Among them, 220 were down-regulated during oocyte maturation while 90 exhibited an opposite pattern. However, because we decided, as a first step, to focus our anal-ysis on the genes exhibiting the most differential regulation in the periovulatory period, we only present the identity of the 90 genes belonging to 3 specific clusters exhibiting the most remarkable patterns. Among those 90 transcripts we have chosen to discuss the most informative or novel genes based on their identities and/or putative involvement in the rainbow trout preovulatory ovarian functions. Among the 32 clones belonging to cluster 1, two clones correspond to rainbow trout ovarian aromatase (cyp19a1). The real-time PCR study confirmed that cyp19a1 was dramatically under expressed during the preovulatory period. This observation is in total agreement with existing data on aromatase expression during this period [19, 36] . In addition, a clone putatively encoding for a NADPH-cytochrome P450 reductase (EC 1.6.2.4) was also located in cluster 1. The aromatase enzyme complex is formed from 2 principal protein components. CYP19a1 contains the catalytic domain that binds C19 steroid substrates in the proximity of the heme prosthetic group critical in the activation of molecular oxygen and subsequent substrate hydroxylation. The other essential component is the redox partner flavoprotein, NADPH cytochrome P450 reductase. Interestingly, present data show that both transcripts exhibited an under expression during the rainbow trout preovulatory period, although it should be confirmed that the identified clone is coding for the oxydoreductase protein involved in the aromatase complex. Other cytochrome P450 genes Two other cytochrome P450 genes, exhibiting similar expression profiles were found in the same cluster. One clone (# 194) was most similar to rat cytochrome P450 2J3 while the other one (# 202) putatively corresponded to rainbow trout cytochrome P450 1A3 (cyp1a3). Cytochrome P450 1A proteins are ubiquitous proteins that have been associated with the detoxification of several organic compounds such as PCB (polychlorinated biphenyl), PAH (polyaromatic hydrocarbons), and dioxin [37] . In fish, these compounds are able to induce cyp1a gene expression in a variety of tissues. In the rainbow trout immature ovary, a constitutive expression of CYP1A protein was previously reported [38] . Together, previous and present observations suggest that a CYP1A-related detoxification activity in the rainbow trout ovary. From the under expression of cyp1a3 gene observed in the ovary immediately prior to ovulation we could speculate that a decrease of the detoxification activity of the ovary is required before the beginning of the ovulation process. In addition, it was previously shown in rat C6 glioma cells that epoxygenases could inhibit prostaglandin E2 production [39] . Interestingly, C6 cells express epoxygenase mRNAs, CYP1A1, CYP2B1 and CYP2J3, which convert arachidonic acid to epoxyeicosatrienoic acids; those epoxyeicosatrienoic acid being able to inhibit the activity of cyclooxygenase [39] . The role of prostaglandins in the ovulatory process has been thoroughly studied (see [40] for review). Thus, in rainbow trout, prostaglandin F2α was able to induce in vitro ovulation [21, 41] . Therefore, the observed down-regulation of cyp1a1 and cyp2j3 genes in the ovary prior to ovulation is therefore totally consistent with available data on the participation of prostaglandins in the ovulatory process. In the present transcriptomic analysis, two aquaporin 4 (aqp4) clones were found in cluster 3. Real-time PCR data confirmed that rainbow trout aqp4 gene exhibited a strong over expression in the preovulatory ovary. In mammals, AQP4 is also known as mercurial insensitive water channel (MIWC). It was previously shown that water permeability was strongly increased in African clawed frog oocytes expressing MIWC [42] . In marine fish, a strong oocyte hydration occurs during oocyte maturation [43, 44] . In addition, it was recently shown that this oocyte hydration involves an aquaporin1-like protein in seabream [45] . In freshwater species, data on oocyte hydration is more controversial. However, a limited but Ovarian expression profiles of aromatase (cyp19a1), vitamin K dependent protein S (proteinS) and cytidine monophosphate-N-acetylneuraminic acid hydroxylase (cmah) genes during rainbow trout late oogenesis (mean ± SEM) Figure 2 Ovarian expression profiles of aromatase (cyp19a1), vitamin K dependent protein S (proteinS) and cytidine monophosphate-Nacetylneuraminic acid hydroxylase (cmah) genes during rainbow trout late oogenesis (mean ± SEM). Ovaries were sampled from separate females during late vitellogenesis (LV, N = 6), post-vitellogenesis (PV, N = 6) and oocyte maturation (MAT, N = 6). The mRNA abundance of each gene was determined by real-time PCR and normalized to the abundance of 18S. Abundance was arbitrarily set to 1 for LV stage and data are expressed as a percentage of the transcript abundance at this stage. Bars sharing the same letter(s) are not significantly different (p < 0.05). Relative mRNA abundance Ovarian expression profiles of angiotensin-converting enzyme 2 (ace2), coagulation factor V (cf5), CXC chemokine L14 (cxcl14), aquaporin 4 (aqp4), pendrin (slc26), vasotocin (avt), serine protease 23 (sp23), ADAM22 (adam22), and elongation fac-tor 1 alpha (ef1α) genes during rainbow trout late oogenesis (mean ± SEM) Figure 3 Ovarian expression profiles of angiotensin-converting enzyme 2 (ace2), coagulation factor V (cf5), CXC chemokine L14 (cxcl14), aquaporin 4 (aqp4), pendrin (slc26), vasotocin (avt), serine protease 23 (sp23), ADAM22 (adam22), and elongation factor 1 alpha (ef1α) genes during rainbow trout late oogenesis (mean ± SEM). Ovaries were sampled from separate females during late vitellogenesis (LV, N = 6), post-vitellogenesis (PV, N = 6) and oocyte maturation (MAT, N = 6). The mRNA abundance of each gene was determined by real-time PCR and normalized to the abundance of 18S. Abundance was arbitrarily set to 1 for LV stage and data are expressed as a percentage of the transcript abundance at this stage. Bars sharing the same letter(s) are not significantly different (p < 0.05). significant hydration was also observed in several freshwater species including rainbow trout [46] . Our data suggest that, similarly to marine species, the oocyte hydration occurring during oocyte maturation could also be aquaporin-mediated in freshwater species such as rainbow trout. In addition to aqp4 gene, we also observed a dramatic over expression of slc26 gene at the time of meiosis resumption. Solute carrier family 26 member 4 (slc26) is also known as sodium-independent chloride/ iodide transporters or pendrin. The over expression of slc26 gene at the time of oocyte maturation is dramatic, as demonstrated by real-time PCR. Together, the strong upregulation of aqp4 and slc26 genes at the time of meiosis resumption stresses the importance of water and ion transports in the rainbow trout preovulatory ovarian functions. In marine species, the major oocyte hydration occurring before ovulation is probably important for adjusting egg buoyancy. In contrast, in freshwater species laying demersal eggs such as rainbow trout, it has been hypothesized that the limited (25%) oocyte hydration occuring before ovulation could be necessary for the completion of the ovulation process [46] . Thus, the increase of oocyte volume could facilitate the rupture of the follicular walls and subsequently, the release of the oocyte from its follicular layers. The neurophysial hormones arginine vasotocin (AVT) and isotocin (IT) are the fish counterparts of argininevasopressin and oxytocin respectively. Vasotocin precursor and isotocin precursor cDNAs were previously cloned in several fish species including chum salmon [47, 48] . In fish, AVT is involved in several physiological processes including water conservation and excretion of electrolytes [49] . However, existing data in fish correspond to the local effect, in various tissues, of circulating AVT [49] . Surprisingly, we observed that AVT precursor (avt) mRNA is expressed in the rainbow trout preovulatory ovary. To the best of our knowledge, there is no evidence of non-neural expression of avt mRNA in fish. In addition, it is noteworthy that we also observed a similar over expression of isotocin mRNA precursor in the ovary at the time of oocyte maturation. Further investigations are needed to elucidate the role of AVT and IT in the trout preovulatory ovarian functions. Ovulation is a complex process resulting in the release of the oocyte from surrounding follicular layers. Since the early eighties, the similarities between ovulatory and inflammatory processes have been thoroughly discussed [50] [51] [52] and it is now well accepted that mammalian ovulation is an inflammatory-like reaction. In fish, despite numerous studies on the hormonal control of spawning, the ovulatory process has been far less documented. In mammals, ovulation is accompanied by broad-spectrum proteolysis and the implication of several classes of proteases is well documented (see [53] for review). In salmonid fish, several proteases have been identified in the periovulatory ovary [54] . In mammals, there is evidence that mature ovarian follicles contain proteolytic enzymes, including serine proteases. Indeed, serine proteases have been implicated in both ovulatory and inflammatory reactions (see [50] for review). In the present study, serine protease 23 (sp23) gene appears progressively up-regulated during the preovulatory period. To our knowledge, sp23 gene expression was never reported in the periovulatory ovary of any vertebrate species. However, we could speculate that this protease participates in the rainbow trout ovulatory process. Interestingly, our data showed that adam22 metalloprotease-disintegrin gene was sharply up-regulated at the time of oocyte maturation. The metalloprotease-disintegrin protein family (also known as ADAMs: A Disintegrin And Metalloproteinases) is thought to function in cell-cell interactions and in the proteolysis of luminal or extracellular protein domains. In mammals, several ADAMs family members are involved in the ovulatory process. In brook trout (Salvelinus fontinalis), metalloprotease activity increases in the ovary prior to ovulation [8, 9] . Together, these observations also suggest that adam22 also participates in the rainbow trout ovulatory process. Mammalian CXC chemokines, named after a conserved pattern of conserved cysteine residues, have been initially identified as potent mediators of neutrophil chemotaxis [55, 56] and are also involved in chemotaxis of monocytes and lymphocytes. They have also been implicated in angiogenesis and, later, in a large variety of functions [57, 58] . In mammals, 16 CXC have been described. In Fish, however, several CXC have been identified but only CXCL12 and CXCL14 exhibit unambiguous orthologues [59] . In the present study, we showed that cxcl14 gene expression strongly increases during the preovulatory period. In catfish, RT-PCR data showed that cxcl14 gene was expressed in a wide variety of tissues, including the ovary [60] . In carp, quantitative PCR data showed that cxcl14 was predominantly expressed in the brain [61] . Despite its good conservation throughout vertebrate evolution [59] , the number of studies addressing the in vivo role(s) of CXCL14 is limited. As a consequence, a lot of information is still unavailable in fish. In a murine model used to study Crohn's disease, cxcl14 expression is induced during inflammation [62] . Together, these observations suggest that cxcl14 gene expression induction contributes to the inflammatory-like events occurring in the rainbow trout at the time of ovulation. To date the participation of this gene in preovulatory ovarian functions was unsuspected. In mammals, coagulation factor V participates in the coagulation process. In zebrafish, a coagulation factor V (cf5) cDNA was previously characterized [63] . According to these authors, several lines of evidences including biochemical and phylogenetic analyses suggest that the modern coagulation pathways found in mammals could also be functional in fish. Furthermore, it was previously shown that cultured rabbit macrophages were able to generate factor V procoagulant activity [64] . In the present study, we observed a dramatic increase of cf5 gene expression in the ovary during oocyte maturation. However, no significant difference was observed between late and postvitellogenesis. From these observations we could speculate that, immediately prior to ovulation, the trout ovary secretes coagulation factors in order to prevent bleeding from ruptured ovarian follicles at the time of ovulation. Interestingly, the transcriptomic analysis showed that a transcript exhibiting sequence similarity with clotting factor C (Clone # 251, Table 3 ) was also over expressed immediately prior to ovulation. Angiotensin-converting enzyme (ACE) cleaves Angiotensin I (Ang I) to form Angiotensin II (Ang II). Angiotensin-converting enzyme 2 (ACE2) is a recently described ACE homolog [65] . Both ACE and ACE2 are zinc-dependent peptidases of the M2-metalloprotease family. Within the renin-angiotensin system (RAS), ACE2 competes with ACE because it is capable of hydrolyzing Ang I into the nonapeptide Ang(1-9) [65] . In humans, ace2 gene expression was predominantly detected by Northern blot analysis in kidney, heart and testis [65, 66] . In addition, a moderate expression was also observed in several other tissues including the ovary [66] . Using semiquantitative RT PCR, a wide distribution was observed in rat tissues [67] . In mammals, previous observations suggested that the renin-angiotensin system was functional in the ovary. In cattle, a greater expression of Ang II was observed in large follicles. In addition, several lines of evidence supported the idea of Ang II in blocking the inhibitory effect of theca cells on meiosis resumption of bovine oocytes [68] . In brook trout (Salvelinus fontinalis) salmon Ang I and human Ang II were both able to increase the level of in vitro spontaneous ovulation [69] . In the present transcriptomic study, we observed a dramatic increase of ace2 gene expression during the preovulatory period. This observation was confirmed by real-time PCR data. Together, these observations suggest that the dramatic upregulation of ace2 gene immediately prior to ovulation is important for the ovulatory process. In mammals, little is know about the role of ACE2 in the ovary. However, it is known in mammals that ACE2 can function as an Ang II degrading enzyme, forming the vasodilatator peptide Ang(1-7) [70, 71] . Interestingly, a local vasodilatation is a key characteristics of the inflammatory response that is also observed during the mammalian ovulatory process (see [50] for review). Therefore, it can be hypothesized that the observed increase of ace2 gene expression in the trout preovulatory participates in the vascular dynamics changes that are putatively occurring during the ovulatory process. Genes involved in the synthesis of egg components Cytidine monophosphate-N-acetylneuraminic acid hydroxylase (CMAH) is the key enzyme for the synthesis of N-glycolylneuraminic acid. In salmonid eggs, cortical alveoli contain polysialoglycoproteins (PSGP). In rainbow trout, it was previously shown that those PSGP contain N-glycolylneuraminic acid residues [72] . In the present study we observed a significant decrease of cmah gene expression at the time of oocyte maturation. While the presence of cmah gene expression in the ovary is totally consistent with the presence of N-glycolylneuraminic acid in rainbow trout cortical alveoli content, it seems however difficult to speculate on the dynamics of PSGP accumulation in the oocytes. Our observations further confirmed that a progressive shut down of estrogen synthesis genes expression occurs in the ovary prior to meiosis resumption. In addition to already well studied genes such as aromatase, the present work shows that other genes exhibit a similar down-regulation, thus suggesting their participation in the preovulatory decrease of circulating estrogen levels. In addition, we observed a strong up-regulation of ion/ water transport genes in the preovulatory ovary. The identity of those genes is consistent with the recent identification of aquaporin mediated mechanisms in the fish oocyte hydration process and further supports the recent description of a limited but significant oocyte hydration occurring in the rainbow trout preovulatory ovary. Finally, in addition to oocyte hydration-related genes, we also observed a strong over expression of several genes such proinflammatory factors, coagulation/clotting factors, vasodilatation factors and proteases in the ovary immediately prior to ovulation. Together, these observations suggest that, similarly to the theory developed in mammals, fish ovulation could also be compared ton an inflammatory-like reaction. In addition, the identification of those genes will allow specific studies leading to a better understanding of the ovulatory process in fish. In the future, a global analysis of differentially regulated genes, based on their ontologies, is needed to satisfyingly describe preovulatory ovarian mechanisms. In addition, a cellular localization of gene expression will contribute in the understanding of their respective roles in the preovulaory ovarian physiology. Nevertheless, the present study clearly demonstrates that distinct (i.e. steroidogenic, proteolytic, proinflammatory) but concomitant events occur in the preovulatory ovary. Together, all those events concur to achieve the same goal which is the release, at the time of ovulation, of a fully competent oocyte, ready to be fertilized.
59
An object simulation model for modeling hypothetical disease epidemics – EpiFlex
BACKGROUND: EpiFlex is a flexible, easy to use computer model for a single computer, intended to be operated by one user who need not be an expert. Its purpose is to study in-silico the epidemic behavior of a wide variety of diseases, both known and theoretical, by simulating their spread at the level of individuals contracting and infecting others. To understand the system fully, this paper must be read together in conjunction with study of the software and its results. EpiFlex is evaluated using results from modeling influenza A epidemics and comparing them with a variety of field data sources and other types of modeling. EpiFlex is an object-oriented Monte Carlo system, allocating entities to correspond to individuals, disease vectors, diseases, and the locations that hosts may inhabit. EpiFlex defines eight different contact types available for a disease. Contacts occur inside locations within the model. Populations are composed of demographic groups, each of which has a cycle of movement between locations. Within locations, superspreading is defined by skewing of contact distributions. RESULTS: EpiFlex indicates three phenomena of interest for public health: (1) R(0 )is variable, and the smaller the population, the larger the infected fraction within that population will be; (2) significant compression/synchronization between cities by a factor of roughly 2 occurs between the early incubation phase of a multi-city epidemic and the major manifestation phase; (3) if better true morbidity data were available, more asymptomatic hosts would be seen to spread disease than we currently believe is the case for influenza. These results suggest that field research to study such phenomena, while expensive, should be worthwhile. CONCLUSION: Since EpiFlex shows all stages of disease progression, detailed insight into the progress of epidemics is possible. EpiFlex shows the characteristic multimodality and apparently random variation characteristic of real world data, but does so as an emergent property of a carefully constructed model of disease dynamics and is not simply a stochastic system. EpiFlex can provide a better understanding of infectious diseases and strategies for response.
The most commonly used measure in public health, R 0 , is estimated from historical data and derived from SIS/SIR type models (and descendents) for forward projection [14, 15] R 0 is the basic reproductive ratio for how many individuals each infected person is going to infect [16] R 0 is often used on its own in public health as an indicator of epidemic probability; if R 0 < 1 then an epidemic is not generally considered possible, for R 0 > 1, the larger the value, the more likely an epidemic is to occur. R 0 is a composite value describing the behavior of an infectious agent. Hence, R 0 can be decomposed classically, for example, as: p d c, where p is probability of infection occurring for a contact, d is duration of infectiousness, and c is number of contacts [17] . However, R 0 in the classical decomposition above, while it is one of the best tools we have, does not account for age segregation of response, existing immunity in population, network topology of infectious contacts and other factors. These observations were significant in the motivation for developing EpiFlex. The EpiFlex model was designed to create a system that could incorporate as much realism as possible in an epidemic model so as to enable emerging disease events to be simulated. There are limitations, described below in a separate section, but the model is quite effective as it stands. In most cases, the limitations of EpiFlex are shared by other modeling systems. There are a variety of methods used for mathematical modeling of diseases. The most common of these are the SIR (susceptible, infected, recovered) of Kermack and McKendrick [15] , SIS (susceptible, infected, susceptible), SEIR (susceptible, exposed, infected, recovered), and SIRP (susceptible, infected, recovered, partially immune) as developed by Hyman et al. [18] and further developed by Hyman and LaForce [19] . The SIRP model was used as the starting point for development of the object model of Epi-Flex. In SIRP, the SIR model is extended to include partial immunity (denoted by P) and the progressive decline of partial immunity to allow influenza to be modeled more accurately. (See Appendix.) There is a need for experimentation in more realistic discrete modeling, since the lattice type of discrete modeling is understood to skew in favor of propagation, as discussed by Rhodes and Anderson [20] and Haraguchi and Sasaki [21] . Others such as Eames and Keeling [22] and Edmunds et al. [12] have explored the use of networks to model interactions between infectable entities, and Ferguson et al. [23] and others have called for more balance in realism for epidemiology models. Since EpiFlex was completed, Lloyd-Smith et al. [17] have shown the importance of superspreading in disease transmission for the SARS epidemic. EpiFlex is designed to take these issues into account. There are known weaknesses in SIS-descended models, some of which are discussed by Hyman and LaForce [14] . They suggested that a model dealing with demographics and their subgroups would be useful and described a start Theoretical Biology and Medical Modelling 2006, 3:32 http://www.tbiomed.com/content/3/1/32 toward conceiving such a model, creating a matrix of SIRP flows for each demographic group within a "city" and modeling contacts between these groups. Thus, the possibility of building an entirely discrete model using the object-oriented approach, essentially setting the granularity of the Hyman-LaForce concept at the level of the individual, together with Monte Carlo method, was attractive. The object method of design seemed to be a good fit, since object-oriented programming was invented for discrete simulations [24] . An Object-Oriented (OO) design defines as its primitive elements "black box" subunits that have defined ways of interacting with each other [25] . The OO language concept originally was conceived for the Simula languages [24] for the purpose of verifiable simulation. Enforcement of explicit connections between objects is fundamental to OO design, whereas procedural languages such as FORTRAN and COBOL do not because data areas can be freely accessed by the whole program. OO languages wrap data in methods for accessing the data. If each "black box" (i.e. object) has a set of specified behaviors, without the possibility of invisible, unnoticed interactions between them, then the simulation can potentially be validated by logical proof in addition to testing. (It would take an entire course to introduce OO languages and concepts, and there is not space to do so here. Interested readers are suggested to start with an implementation of Smalltalk. There are excellent free versions downloadable. Smalltalk also has an enthusiastic and quite friendly user community. See: http:// www.smalltalk.org/main/.) The design of EpiFlex is described more completely in the appendix. Design proceeded by establishing the definition of a disease organism as the cornerstone, then defining practical structures and objects for simulating the movement of a disease through populations. The disease object was assigned a set of definitions drawn from literature that would allow a wide spectrum of disease-producing organisms to be specified. The aim was to minimize the number of configuration parameters that require understanding of mathematical models. The hosts that are infected became the second primary object. A host lives and works in some area, where hosts are members of some demographic group, which together determine what of n types of contacts they might have to spread an infectious disease. The hosts move about the area in which they live between locations at which they interact. In EpiFlex, an area contains some configured number of locations, and locations are containers for temporary groups of hosts. Since people travel between metro areas, the model supports linkages between areas to move people randomly drawn from a configurable set of demographic groups. The remainder of this section presents the disease model adopted, an overview of each component, an overview of program flow, and a description of the core methods. This is followed by discussion of results from the EpiFlex software system. This model has up to four stages during the infection cycle: the Incubation, Prodromal, Manifestation, and Chronic stages; to this is added a fatality phase. I have named this 'extended-SIRP'. Fig. 1 shows a diagram of this model. The model of Fig. 1 allows us to track the different phases of the disease process separately, and to define variable infectiousness, symptoms, fatality, recovery and transition to chronic disease at each stage as appropriate. This allows us to model the progress of a disease in an individual more realistically. For diseases that have no identifiable occurrence of a particular stage, this stage can be set to length zero to bypass it entirely. The 8 contact types designed into EpiFlex are drawn from literature in an attempt to model spread of infection more accurately. These contact types are: blood contact by needle stick, blood to mucosal contact, sexual intercourse, skin contact, close airborne, casual airborne, surface to hand to mucosa, and food contact. The probability of infection for a contact type is input by the user as estimated from literature or based on hypothetical organism characteristics. Durations of disease stages are chosen uniformly at random from a user-specified interval [R low , R high ]. Random numbers, denoted by ξ, on [0, 1] are used to seed the determination of the infected disease stage periods (denoted I Incubation , I Prodromal , I Manifestation , I Chronic ). R low and R high are taken from medical literature and describe a range of days for each stage of an illness. These calculations are simply: (ξ × (R high -R low )) + R low = D, where D is days for a particular stage. (This may be extended in the future to include ability to define a graph to determine the flatness of distribution and the normative peak. This will make a significant difference in modeling of diseases such as rabies, which can, under unusual circumstances, have very long incubations.) One of the following three equations describing immunity decay is chosen; L is the current level of partial immunity, P is the level of partial immunity specified as existing Random values on [0, 1] are then used to decide whether an infection occurs during the partial immunity phase P shown in the chart above. This decision uses the output of the immunity level algorithm, L, which is a number on [0, 1], as is the random value ξ: EpiFlex uses a dynamic network to model the interactions between hosts at a particular location based on the skew provided and the demographic segments movement cycles. The networks of contacts generated in this version Extended-SIRP disease model of Epiflex Figure 1 Extended-SIRP disease model of Epiflex. S: susceptible I: Infected R: recovered P: partially immune F: fatality Extended SIRP breaks the infected stage I into 4: I Incubation , I Prodroma l, I Manifestation , I Chronic , and adds a fatality terminating stage. of EpiFlex are not made visible externally; they can only be observed in their effects. (See: Limitations of EpiFlex modeling.) Their algorithms were carefully designed and tested at small scales, observing each element. A location describes a place, the activities that occur there, and the demographic groups that may be drawn there automatically. A location can have a certain number of cells, which are used to specify N identically behaving locations concurrently. This acts as a location repetition count within an area when the location is defined. The user sets an average number of hosts inhabiting each cell, and a maximum. There is also a cell exchange fraction specifiable to model hosts moving from cell to cell. The algorithm for allocating hosts in cells is semi-random. It randomly puts hosts into cells in the location. If a cell hits the average, then it does another random draw of a cell. If all locations are at maximum, then it overloads cells. Interactions are within the cell. So a host must be exchanged to another cell in order to be infective. See the appendix for 'Location component', and also with an open model look at how hospitals were defined. Households are modeled at this time using a cell configuration. EpiFlex is implemented with a Monte Carlo algorithm such that each host in a location is assigned a certain number of interactions according to the Cauchy distribution parameter setting for that location. This distribution describes a curve with the y axis specifying the fraction of the maximum interactions for the location and x axis specifying the fractional ordinal within the list of hosts in the location. The distribution can be made nearly flat, or severely skewed with only a few actors providing nearly all contacts, as desired by the user of EpiFlex. Note that the structure of the network formed also depends on what locations are defined, what demographic groups are defined for the population, and how demographic groups are moved between locations. Each location has a maximum number of interactions specified per person, which is used as the base input. Initially, a Gaussian equation was used, but it was discarded in favor of a Cauchy function since this better fits the needs of the skew function and computes faster. The algorithm iterates for each infectious host, and selects other hosts to expose to the infected party in the location, by a Monte Carlo function. This results in a dynamically allocated network of interactions within each location. The exposure cycle also makes use of Monte Carlo inputs. Each location has a list of contact types that can take place at a particular location, and a maximum frequency of interactions. This interaction frequency determines how many times contacts that can spread a disease will be made, and the contact specification defines the fractional efficacy of infection by any specific route. Modeling the effect of different types of contacts has been discussed in the literature, e.g. Song et al. [26] . EpiFlex attempts to make a more generalized version. For each host infection source, target hosts are drawn at random from the location queue. A contact connection is established with the target as long as the contact allocation of that target has not been used up already. Contact connections made to each target are kept track of within the location to prevent over-allocation of contacts to any target. Thus, for each randomly established connection, a value is set on both ends for the maximum number of connections that can be supported. Once the maximum for either end of the link is reached, the algorithm will search for a different connection. The location algorithm is described below in more detail. The user specifies the maximum number of connections for a location; the σ output from a Cauchy distribution function determines how many connections an individual will have. This allows variations in the degree of skewness for superspreading in a population to be modeled, which has been shown to be of critical importance by Lloyd-Smith et al. [17] . If p = position in queue, q = number of hosts in queue for location: X = p/q, where X denotes the proportional fraction of queue for position. If K is a constant chosen for the location to express skew distribution, the Cauchy distribution function is: If κ is the number of contacts for a particular host and κ max is maximum number of contacts for any given host in the location: When hosts move from one location to another within the model, they tend to maintain a rough order of ordinal position. Consequently, when there is a high σ for a location, the high connection host in one location tends to be a high connection host in another. This reflects real-world situations, (though not perfectly) and corresponds better than persistently maintaining high connection individuals from location to location, since host behavior changes from place to place. The Cauchy distribution function is fairly fast in execution. The function can be used to approximate the often radical variations seen in epidemiology studies; as an extreme example, one active super-spreader individual might infect large numbers, when one or even zero is typical [17] . This type of scale-free network interaction has been explored by Chowell and Chavez [27] . The Cauchy function allows networks to be generated dynamically within each type of location in a very flexible manner, such as corresponding to super-spreader dynamics [17] . In addition to the specification of skew within a location, the network of contacts is also defined by (a) what locations are present and (b) the movement cycles defined for each demographic group within the model. Processing time increases with population. This slowing is an expected characteristic of an object modeling system and is the price paid for the discrete detail of the EpiFlex model. The primary source of this increase in processing time is the sum of series of possible infectious events that are modeled for each iteration. It therefore scales as a series sum not as a log, based on the contagiousness of the disease and the number of potential hosts in a location with an infected host. This is minimized by only processing infectious host contacts. The increase stems from the characteristics of networks in which each node has n connections to other nodes. When iteration is done for a location containing infectable hosts, it is the number of infected hosts that creates an element of the series. The infected hosts are put into a list, and each one interacts randomly with other hosts (including other infected ones) in the location. Thus, considered as a network with m nodes, each of the m nodes is a host. A temporary connection to another host is made to n other nodes where n <m, and n<k. The value of k is determined by a randomized input that returns the number of contacts of this infected host in this location. Consequently, the series consists of all the temporary connections made for contact modeling for each cycle. In the interest of completeness, the limitations of the Epi-Flex model are described here. The plan is to address these elements for implementation in future versions. Only one infectious disease can be occurring at a time. Thus, competitive inhibition [28] and synergistic effects will not be seen. Hosts do not reproduce Hosts do not reproduce within a model. Removal and addition rates are defined for the population as a whole, and the basis is US Census data. To meet the specifications for removal and addition within the model, hosts are removed from randomly chosen locations, and similarly added to randomly chosen locations. Demographic group is also randomly assigned. For long-term modeling, and modeling of alternative short-lived hosts, a reproduction cycle is desirable. However, EpiFlex is a practical way of modeling periods of a few years. There is no explicit definition of an age distribution for the host population, which can be quite significant [29] . To a degree, age is taken into account through the demographic segmentation of populations. A demographic can be defined with a fraction or multiple of baseline susceptibility. However, hosts do not age, nor do they move from one demographic to another as they age. No provision is made to define a previous exposure profile for hosts [30, 31] . In real populations previous exposures can have significant effects on the spread of a disease and dramatic effects on mortality where infection does occur [32] . Proper implementation of previous exposure profiles is intertwined with age definition. There is no implementation of mutation rate for diseases. Mutation rates vary considerably by type, particularly for viruses [33] . Decay of immunity is modeled, and immunity decay can act as a fair surrogate for antigenic change. Pass-through events must be defined as part of surface contacts For efficiency, EpiFlex eliminates pass-through infection events from being modeled: for example, an infected A shakes the hand of a non-infected B, who then shakes the hand of another non-infected C, but B washes hands and does not become infected while C does. Therefore, the model definition must account for this through "Surface to hand to mucosa" contacts, where a person can also be a surface. The model does not at this time record the contact network that is dynamically created except in the log file at this time. Those that are logged are only potential infectious contacts. To get at that data requires looking at the log file and writing an extract program. Making the network visible is an item for the future. Currently, EpiFlex has no way of accounting for seasonal damping. Similarity of results is due to the settings of the rate at which immunity declines. Addition of a seasonal damping function would be expected to cause EpiFlex results to synchronize with a yearly cycle. Seasonal damping would result in loss of interesting epidemic behavior with an overriding function that would virtually be guaranteed to drown out other behaviors. Public health response to epidemics is not optimally modeled A public health response definition component is present in EpiFlex. Testing of this component, and more thorough review of literature, indicate that the method used is not optimal. Current public health responses are centered on contact tracing, ring vaccination and quarantine [34] , with mass vaccination as a backup when it is available. Closures of schools, daycare and travel restrictions are also used. These methods are not modeled in EpiFlex' response component. Their importance has recently been underscored by Lloyd-Smith et al. [17] . As a consequence, results from the current system that defines across-theboard cuts in the probability of infection should be considered in this light. It is not clear whether any other object system can model all current techniques properly. Diseases have ranges of times for each stage that can be drawn from literature. Probability of a specific disease stage time period for an infected host being chosen within the range is equal. This is reasonably adequate for most diseases where times are measured in a few days, however, some, such as rabies have a quite unequal distribution, and their very long tail makes a difference in modeling. The discussion is presented in three parts: (1) a brief set of examples of native EpiFlex displays to develop a better feel for the system; (2) comparisons of EpiFlex results with real world data; (3) a set of examples of observations made using EpiFlex. The purpose of these examples is to serve as a guide to others who may want to experiment and analyze results. Different views of the epidemic data for simulated influenza in two different populations are shown in Figure 2 . Figures 2a and 2b show graphs of the second and third epidemics in the population. These graphs show the kinds of commonly-seen deviations from a smooth curve that occur in real world data [35] . In the EpiFlex model, this is attributed to less synchronization of immunity combined with the formation of small world networks among demographic groups as they move from location to location. Figures 2c and 2d are alternate views of a simple influenza epidemic occurring within a naïve population (Figure 3 ). Comparison of EpiFlex with WHO/NREVSS surveillance Comparing EpiFlex with surveillance data, we see that WHO/NREVSS surveillance data [36] have a qualitatively similar graph form to EpiFlex for influenza, as shown in Figures 4 and 5. The width of the primary curves per season for EpiFlex is 3.5 to 4.5 months while that of the NREVSS data is approximately 5 to 7 months, which can be explained by the NREVSS data being collected nationally from surveillance centers, whereas the EpiFlex data shown are for a single area. EpiFlex runs executed with multiple cities connected by transport, such as the 3.5 million population 35 city model, have a combined graph for all cities showing self-similarity to the graphs for individual areas, becoming wider, matching the NREVSS data graph formation. The NREVSS data consist of diagnostics of samples sent in by physicians. Comparisons of absolute numbers in terms of quantity are therefore not applicable. A percentage of population comparison is done below. In Table 1 , EpiFlex indicates that roughly 48% of the population has been infected before herd immunity stops the epidemic, though this depends on population size. Total morbidity is obtainable from EpiFlex by adding maximum immune level to deaths, although deaths contribute such a small amount to influenza morbidity that for practical purposes the immune level is used as a proxy for morbidity. Moreover, true morbidity itself is relatively prone to inaccuracy, whereas better measures of immune fractions for influenza are available. The California state average for 2000-2003 is 25.4% infected in a range from 12.7 to 44.6 depending on county [37] . Thus, EpiFlex is above the high end of the state of California estimated morbidity range. Dowdle [32] gives serological influenza data categorized by age. For influenza A/Swine/15/30 H1, seroprevalence ranges from roughly 25% to over 95%. For influenza A/ Hong Kong/68 H3, the range is from 5% to 99%. EpiFlex figures fall within this latter pair of ranges, and EpiFlex immune fraction is more properly comparable. What is most notable in Figure 7 is the relationship between the rough sine wave form of the classically derived SIRP [19] once the initial startup period is over for influenza, a repeating wave develops that is similar in overall shape and variability to real world data such as those for Milwaukee, at a roughly similar scale. These two graphs refer to populations that differ in size by about 1 order of magnitude (i.e. Milwaukee is 9 times the size of the model run shown). We can also see a similar number of peaks. Owing to the need to compare these two graphs natively, these two figures are not optimum. However, they show essential features. Total morbidity rate linked to population size The smaller a population over the range 1,000 to ~3.5 million, the higher the total morbidity rate, given identical organisms ( Figure 6 ). It is intuitively expected that population size will affect morbidity since, for any given network of contacts connecting individuals in populations, the chance of the epidemic spreading during the window prior to the development of immunity in parts of the population increases as the population size decreases. This This is the simulation that was imported for comparison Figure 3 This is the simulation that was imported for comparison. Vertical scale 1000 per demarcation. Horizontal scale one year per demarcation. Upper white line is total population with standard removal rate. effect is most striking when very small populations in the order of 1,000 are examined. Literature data regarding this in real world populations are sparse. However, there are indications from historical accounts of small populations in the new world that a link between population size and morbidity is observable in real world populations [38] [39] [40] [41] [42] [43] . The most recent such account is from Heyerdahl in the Pacific in the mid 20 th century [44] . In the graphs of Figure 6 , the immune fraction at completion is used as a proxy for total morbidity on a log scale of population. The longer an epidemic takes to progress within an enclosed population, the greater the number of potentially infectious contacts that hit a dead end because the host is already immune. Since very small populations will mostly function within the window when there is no host immunity, the infection will spread to a larger fraction. This effect has public health implications because, clearly, the structure of the network is highly significant in determining the likelihood that an infected host will contact naïve hosts. Essentially what this EpiFlex result indi-cates is that during the period prior to the development of an immune subpopulation, a disease has a functionally higher R 0 . (i.e. R 0 is variable through the course of an epidemic.) In the Figure 6 graphs, EpiFlex is also suggesting that there are more asymptomatic infected spreaders of influenza in our populations than surveillance data estimate. This is also suggested by the discussion above regarding comparison with seroprevalence. A minor experimental result is that for a repeating illness such as influenza, when a continuously active initiating disease vector tries to infect 3 people per day, it will develop higher peaks after the initial event than a vector that tries to infect 30 people a day (where both are randomly distributed through the population.) This makes intuitive sense, because there is lower probability that a subpopulation of susceptible hosts will become large when there are more attempts to infect them. Similarly, in a system of cities interconnected by transport linkages, later peaks tend to be smaller and more variable than earlier ones. This is due to two things. First, a degree of lowgrade infection linking back through the system provides a higher total level of infection events in the whole system than the formally defined initiating disease vector. Second, as time passes, the mix of immune versus susceptible becomes unsynchronized for the population as a whole, since hosts that escaped infection during one epidemic may be infected during the next, and some whose immunity has declined may also become infected. Thus, it is expected that we would see the development of a complex non-repeating waveform with some similarity. This type of waveform is what EpiFlex shows with longer simulations in large populations, as illustrated in Figure 8 . A variety of results is obtained when one or just a few index cases are provided to seed a single city's susceptible population when those index cases are not repeated. This is expected, owing to random interactions that break the chain of contacts in some percentage of cases. This effect would be expected to increase with a higher skew on super-spreaders. The significance of this for modeled epidemics, particularly in the light of recent work [17] , is that in some cases (the proportion would be expected to vary in rough accordance with R 0 ) the infection dies out owing to random chance. Thus, a Monte-Carlo model such as EpiFlex, when used in multiple trials, clearly reveals the potential range of variation in epidemics given apparently identical conditions. EpiFlex shows wave propagation of epidemics through its transport network that are similar to real world epidemic studies such as those of Viboud et al. [45] . Viboud et al. state a mean duration of 5.2 weeks to spread across the United States, with a range from 2.7 to 8.4 weeks. The Epi-Flex results shown using a simplified city configuration of 35 major airline hub cities, with a total 3.5 million population among all 35 cities, shows a propagation wave of 1.8 weeks. While this is shorter than real world data, several factors account for the difference. First, in the current EpiFlex "vanilla" configuration, the transport network is flat in terms of the numbers of persons moved from city to city. Second, each city contains the same population of only 100,000 due to practical limitations. For the propagation histogram of Figure 9a , 1000 manifesting cases or more was used as the data point. Please note, however, this flatness of transport and population is purely a matter of the configuration of the specific model used. The Epi-Flex system allows separate specification of all parameters for each city, and any kind of transport level between any two cities that is desired. Figure 9a shows a histogram of cities in which 1,000 manifesting cases are first occurring. Figure 9b shows a histogram of cities in which the first occurrences of at least 10 incubating cases of influenza appear. Note that in Figure Upper graph shows graphed points for population versus total morbidity as estimated from immunity Figure 6 Upper graph shows graphed points for population versus total morbidity as estimated from immunity. Lower two graphs show sample graphs for 3500 (lower left) and 35000 (lower right). The green line in the lower two graphs shows immunity. One vertical demarcation on the x axis is one month in the lower two graphs. EpiFlex is useful for doing in-silico experiments with epidemic behavior and easy to configure. It can run on an ordinary computer without special configuration. Data can be imported from it into other tools and worked with there. It is effective for showing factors that are invisible or difficult to access under normal conditions, such as visibility of incubation and prodromal stages, true morbidity and estimates of immunity. EpiFlex has capabilities that were not discussed owing to time and space constraints. This system is effective in duplicating the kind of multimodality and apparent stochastic variation [13] that are seen in real populations. EpiFlex results can provoke thought about the nature of epidemics and infectious disease spread in interesting ways by providing an experimental test environment that is not as abstracted from its subject as most mathematical models are. Items in bold below are entities that are objects in the Epi-Flex system. They are the "nouns", each of which became a class when the program was written in C++. Items in italics are actions, the "verbs" that represent interactions between objects. Hosts go through the infectious stages, are members of Demographic groups and move between locations Locations contain hosts (For example a hospital, home or workplace.) Areas contain locations (A city, typically) and contain special locations that can move hosts to locations in other areas. Disease Vectors introduce disease into host populations at some location in some area, in either a limited way or on a regular cycle. Epidemic responses modify the spread of an infection. They represent actions taken to respond to the epidemic that will damp the spread. SIRP is a simple flow chart, which is then elaborated into a multi-city system. Three graphics are reproduced in Figure 10 concluding with output of the model, because the SIRP paper is in a specialty publication and the work was important to the conception of EpiFlex. As seen in the "SIRP diagram" of Figure 10 , a host begins as susceptible (S) in the upper left. They can become infected (I), and then they either recover (R) or are removed (removal not shown). After recovery (R), they have an immunity level that starts at or near complete as in SIR. This immunity declines, in the case of influenza, over a period of roughly 2 years, with a steep drop-off beginning after the first year, and hence is termed partial immunity(P). (It is not actually the case that the immune system's response to influenza antigens degrades so rapidly; rather, the virus mutates rapidly, and host immunity declines. The net result is similar.) This decline in immunity models the real world antigenic mutation of influ-enza viruses. In a more perfect model, the influenza virus would change and multiple viruses would be modeled. At some point the immunity declines so that it is effectively zero in the model (which corresponds to large antigenic change) and the person is returned to the fully susceptible(S) population. A person can become reinfected (diagonal arrow), though their probability of reinfection is lower, during this period of declining immunity. In the "Multi-city SIRP" flows of Figure 10 center, m [x,y] denote movement between the cities in a four city example. Moving further to the right of Figure 10 and looking at the "Actual versus model graph", one can see how a sinusoid curve, with some general correlation with period, results from the conventional model. There are four files used or generated by EpiFlex. All three generated files have the name of the configuration from A 5.5 year, 9 city simulation linked by transport corridors. 100,000 people per city Comparison of distribution of days when 1,000 manifesting cases or more first appear in a city (upper graph -9a) versus distri-bution of days when 10 or more incubating cases first appear in a city (lower graph -9b) Figure 9 Comparison of distribution of days when 1,000 manifesting cases or more first appear in a city (upper graph -9a) versus distribution of days when 10 or more incubating cases first appear in a city (lower graph -9b). Model files end in .EPDM. These files are an XML format. This is the file edited by Epiflex using the various panels available from the menu. An EPDM file can also be edited directly in WordPad or a good XML editor by more sophisticated users. Reading through them should be selfexplanatory to most users who familiarize themselves with the software. Run record files end in .RPX. These files are in text format, comma, and semicolon delimited. This format is intended to be as easy to import as possible. RPX files can be imported into Excel, read by SAS, SPSS, etcetera. At the top of this file is a descriptor of the fields. The most common problem is to import using spaces as a delimiter and have spaces in area names define new columns. Log files end in .LOG. Everything that happens in a model is logged. These files can get fairly large. Potentially, they could be parsed by a secondary piece of software to match them up with RPX files. However, that is an exercise for the user at this time. There is a viewer for log files under the View menu. Snapshot files end in .SNAP. Whenever a run completes, a SNAP file is written that is read only. This file is a duplicate of the model that was used to run -as that model existed in memory at completion of the run. This is done as a "lab notebook" aid because it was recognized early on that remembering exactly what was contained in a model was well nigh impossible and a lot of work to document. Note that if your model does not complete its run, the RPX file will still be there, but the SNAP file will not be. In that case it is up to the experimenter to make a record of the state of the model. The SNAP file can be copied outside of Epiflex, manually changed from read-only and the extension renamed to .EPDM. The resulting model can then be used and run just like any other. Software engineering details such as internal class definitions and structure of components will not be presented since the UI does a better job of educating, and is much more compact. Disease component Looking at screen shots of the current UI in Figure 11 , one can see what the functional parametric elements of a disease specification are. This particular example is selected for influenza. The parameters set reproduce the known characteristics of the disease. The disease stages of Figure 11 correspond to the model diagram shown in Figure 1 . One can see that this definition is superior in several ways to what is available with SIRP modeling as shown in the previous section. (See Figure 10 ). In addition to multiple stages, one can define the level of infectiousness for each type of contact that an infected host might have. On the lower right, the partial immunity stage is specified. Here, one is able to define a wide variety of immune responses. Asymptotic decline seems to fit available data and theory on immune system (from left) SIRP Diagram, Multi-city SIRP, Actual versus model Figure 10 (from left) SIRP Diagram, Multi-city SIRP, Actual versus model. function the best for many diseases. Use of the slider allows one to "eyeball" the asymptotic curve to approximate what seems reasonable. In this case, immunity is set to begin dropping after approximately half the 730 day period has passed. It makes sense to do this because data in this area are fairly sparse, and one may want to perform multiple runs with slightly different immune drop characteristics. Keep in mind that the width of the graph is for the number of days entered on this panel. A host is implicitly one of the number in the population of an area. In the EpiFlex system, there is one object defining each disease. A host infected by a disease receives a pointer to this archetypal disease object, and stores their disease stages internally as the disease progresses. Most of the functionality of hosts consists of an implementation of disease stage transition. Figure 12 is a screen shot of the UI for a set of demographic definitions selected from US Census data for the nation as a whole. In the resulting model, one can over-ride the default fractions later when using them in populations for specific areas. No fractions are represented that could not be derived from recent census data. Each demographic group can be set up to specify variations in overall susceptibility to disease, or modify it by specific type of contact if that is desired. In the example, no fractional modifiers have been created for this demographic group. There is no necessity to stick to the demographic groups defined here. One can add new ones as desired. A location describes a place, the activities that occur there, and the demographic groups that may be drawn there automatically. A location can have a certain number of cells, which are used to specify N identically behaving locations concurrently. This acts as a location repetition count within an area when the location is defined. A technical point with homes as cells is that the model cannot yet be set to create a number of homes based on population size with a distribution of household sizes. Figure 11 Disease definition panel. Consequently, one must, at present, define the number of home cells appropriate to hold the population for each area. If a distribution function were available to specify household sizes and households were definable by population size, it would then be practical to provide canned profiles of household sizes based on US census data. The types of contacts that are available among hosts in the location are defined, together with the average number of such contacts per host during one cycle of the model; see Figure 13 . To the right of the contact definitions is a skew function that enables one to specify the degree to which contacts vary. Within any host subpopulation, there are high contact individuals and lower contact individuals. The example shown sets the difference rather high to describe the way that service workers interact more strongly than customers. When running the simulation, during determination of which hosts become infected in this cycle, the position of the host in the queue is used together with the contact frequency to decide how many contacts an individual will get in a particular cycle. Think of this adjustable skew graph as the histogram of contact counts. For each demographic group, a cycle of movement between locations in an area is defined as illustrated in Figure 14 . In the model used as an example, days are divided into three cycles. So a sequence of 21 movements defines a 7 day week. Once a cycle completes, it restarts at the top. This is how hosts in the model are moved from one location to the next. An area is made up of a list of locations that it contains, as shown in Figure 15 . An area specifies a particular population level, and the population fractions for each demographic group making up the area are allocated at runtime. Demographic population, removals and additions can be overridden at the area level. An area also has linkages to other areas, through special locations that are eligible as links. In the example of Fig- Group demographic population definition panel Figure 12 Group demographic population definition panel. ure 16, the Anchorage airport is going to send 2% of its population to Chicago, Illinois, each cycle. In each cycle, 70% of the population of the Anchorage airport location is going to be shipped off somewhere. A 'Vector' is defined in EpiFlex as n infection events to attempt at a specific location. A vector can repeat each cycle, or have a limited input. A vector can represent an arthropod or snail, or it can represent people arriving on aircraft from Hong Kong. The vector shown in Figure 17 is a startup vector for influenza used for these modeling runs. This initiating disease vector will operate for the first 3 cycles, then stop, attempting to produce 9 cases at a location in Chicago. An initiation disease vector can have delay parameters, run continuously, or run once. It can force the infection of a specific number of people or infect them at a particular rate. When a disease occurs, the medical system will respond to it in some way. Some illnesses, such as influenza A, will result in very little intervention. Others, such as SARS or smallpox, would result in a great deal of intervention very rapidly. Consequently, a reasonable modeling system will allow response to an epidemic occurrence to be modeled. In the example shown in Figure 18 , we see the response specification for an emergent illness. A response has an alert trigger, which in this case is the appearance of fatalities. This results in a period of heightened awareness of the disease and its symptoms, which is set in this example to 100 days. The probability of noticing the triggering event is set, to unity in this example, which may be overly optimistic. This specification says that after three fatalities, an alert would be triggered. Since this alert trigger is for fatalities, not symptoms or instruments, the detection fraction for each disease stage is not applicable. Once the alert has been triggered, the detection method for the illness switches primarily to symptoms. We can then specify what fraction of each occurrence will be diagnosed and at what stage. If we detect an event, then we can mitigate it. In this implementation, one has to specify the degree of infectiousness remaining after intervention is in process. Some methods, for some diseases, can stop disease spread completely. This is not true for all diseases, so it must be specifiable. Figure 13 Location definition panel. Note that this component is the one with which the author is least comfortable. None of the examples for this paper have used responses as a consequence. The primary reason is that the way response is modeled is over-simplified and does not conform well to the response that happens in the real world; see: Limitations of EpiFlex modeling in the introduction. EpiFlex is now composed of 83 major classes, of which 45 are core internal model functionality, the rest being infrastructure and UI. A functional walkthrough of how the system operates is presented below. A model starts running by verifying the model data. It looks for any references to things that were deleted and definitions that are impossible. It writes an error log file with this information that can be audited after the run. If errors are found, you will be informed of their severity and given the option to cancel. Many of these errors are non-fatal, and may be modified with a warning, but they may change the results of a run. The EpiFlex system iterates through all areas in a model and allocates hosts, putting them in their initial locations, per the movement definitions for the demographic group. Each group steps through its location movement list to determine whether the area to which it is attached has the locations the group needs. If it does not, EpiFlex will increment the pointer when it comes to this location, but it will leave the hosts in their previous real location until a good one is found. This was done for ease of practical use, allowing the same demographic to be used when a location may not exist in an area. Step 1 -disease stage pass. For each infected host, it will update the stage of the illness for that host. This is also where the system checks to see if disease response conditions have been met yet. Step 2 -vector pass. EpiFlex will iterate through each vector and apply the disease to the locations in each area in the vector according to the rules defined. Step 3 -infection pass. Iterate through each location in each area, and figure out the contacts between infected and uninfected hosts. For each contact between an Group movement cycle definition panel Figure 14 Group movement cycle definition panel. infected host and a host not infected with the disease, a probabilistic determination will decide whether or not this illness is communicated. Step 4 -iterate areas looking for locations that have group draw factors. Randomly pull hosts from the groups specified, and move them to the group draw location. This feature is used to simulate various things; generally an airport will have this type of specification. Step 5 -iterate all areas looking at area links. Area links push population from one location to another area's location as specified. Step 6 -Areas will be iterated, and each location in each area will be iterated using the random exchange factor to move hosts assigned to cells of a location between cells in that location. Step 7 -Normal addition and removal of population groups is applied to model. This allows the user to model normal birth and death rate plus immigration and emigration across the outer boundary of the model. Step 8 -iterate all areas, and iterate through each location in each area. For each location, it will iterate each host attached to the location. Each host has a location pointer for its group that indicates where it is in its movement cycle. It increments the pointer, and if the location name is different from the one it is in, it will find a location of that name in the area to put itself into. If the location has N cells, it will put itself into the cells in a semi-randomized fashion according to the parameters defined. Note that currently the hosts do not remember which cell they are put into from one iteration to another. So they will not return to the same cell next time they come back around in their movement cycle. This is a limitation at present, which would require greatly increased space allocations to change. Step 9 -An audit is done to verify that all hosts occur only once in a location. Step 10 -Areas are iterated, locations are iterated, to total results. Results are then either written to the log file and/ or displayed on the main window in a graph form depending on how parameters were specified. Figure 15 Area definition panel. Area linkages panel Figure 16 Area linkages panel. Initiating disease vector definition panel Figure 17 Initiating disease vector definition panel.
60
Multiplexed Genetic Analysis Using an Expanded Genetic Alphabet
BACKGROUND: All states require some kind of testing for newborns, but the policies are far from standardized. In some states, newborn screening may include genetic tests for a wide range of targets, but the costs and complexities of the newer genetic tests inhibit expansion of newborn screening. We describe the development and technical evaluation of a multiplex platform that may foster increased newborn genetic screening. METHODS: MultiCode(®) PLx involves three major steps: PCR, target-specific extension, and liquid chip decoding. Each step is performed in the same reaction vessel, and the test is completed in ~3 h. For site-specific labeling and room-temperature decoding, we use an additional base pair constructed from isoguanosine and isocytidine. We used the method to test for mutations within the cystic fibrosis transmembrane conductance regulator (CFTR) gene. The developed test was performed manually and by automated liquid handling. Initially, 225 samples with a range of genotypes were tested retrospectively with the method. A prospective study used samples from >400 newborns. RESULTS: In the retrospective study, 99.1% of samples were correctly genotyped with no incorrect calls made. In the perspective study, 95% of the samples were correctly genotyped for all targets, and there were no incorrect calls. CONCLUSIONS: The unique genetic multiplexing platform was successfully able to test for 31 targets within the CFTR gene and provides accurate genotype assignments in a clinical setting.
Few multiplexed technologies with sufficient specificity to identify small changes within the human genome are available for clinical use. Line probe or linear array technology uses nitrocellulose paper strips as the support matrix (1 ) . A benefit of the line probe is that expensive instrumentation is not required. A technology termed "oligonucleotide ligation assay" uses target-specific ligation to generate mutationspecific products that are analyzed by gel electrophoresis (2 ) . Other genetic analysis systems are coupled to microparticle flow cytometry (3) (4) (5) (6) (7) (8) . Limitations of these methods include complicated procedures that require extensive training (washings, centrifugations, and transfers), long completion times, expensive instrumentation, and technician-dependent result analysis. In addition, these technologies are not easily amenable to automation. Many hospitals and clinics currently send molecular diagnostics tests to large reference laboratories. This is in part attributable to the limitations described. Because there are benefits in having local laboratories perform diagnostic tests (9 -12 ) , there appears to be a need for molecular testing methods that circumvent the limitations posed by current testing methods. We have constructed a three-step platform technology called MultiCode ® PLx that eliminates the most technically challenging steps involved in multiplexed genetic analysis and compared it with the line probe assay in a clinical setting. MultiCode PLx uses one additional nucleobase pair constructed from the complementary nucleobases 2Ј-deoxy-isoguanosine (diG) 5 and 5-Me-iso-cytosine (iC). These nucleobases specifically recognize each other based on a different pattern of hydrogen bondings. We chose this pair because no other such pair is available commercially and their chemistries have been well explored. For example, iG and iC have been successfully used for both molecular recognition (13) (14) (15) and for site-specific incorporation (16 -18 ) . In addition, because iG:iC recognition is "orthogonal" to the naturally occurring nucleobase pairs, a strand of DNA that contains several iC/iG components can be constructed so that it will not hybridize to natural DNA. In complex reactions in which high concentrations of natural DNAs of known or unknown sequence exist, this orthogonal attribute allows specific molecular recognition to take place without interference. These additional nucleobases are used in each step of the PLx process: PCR, extension labeling, and liquid decoding. PCR primers are first designed to be target-specific and contain single iCs. The amplicons act as labeling templates for the target-specific extension (TSE) step. During that step, labels attached to diG triphosphate (diGTP) are incorporated site specifically into coded target-specific extenders (Fig. 1 ). The tags are short sequences (typically 10 nucleotides in length) assembled by use of a mixture of natural and nonnatural bases. The tags are designed to hybridize only to their perfect complements encoded on color-addressed microspheres. In the final step, decoding of the extension reactions is accomplished at room temperature by capture of the coded extenders on the addressed microspheres and reading of the reporter signal on a Luminex 100 instrument. All steps are carried out in the same reaction vessel without transfers or washings. To demonstrate the approach, we developed a screening test for carriers of cystic fibrosis (CF) that tests for all 25 mutations and the reflex targets in the widely used panel (19 ) and also tests for 394delTT and 3199del6. 394delTT is the second most prevalent mutation in populations with Nordic heritage and may be important in Wisconsin, where testing was performed. 3199del6 was added because several lines of evidence indicate that 3199del6 is a disease-causing mutation and may replace I148T (20 ) . The Wisconsin State Laboratory of Hygiene, in conjunction with the University of Wisconsin, was chosen as the major testing site because of its long history of screening newborns for CF (21 ) and study of screening methods (22) (23) (24) (25) . In March 2002, DNA testing of the top 4% of the daily immunoreactive trypsinogen results expanded to include all 25 mutations recommended by the American College of Medical Genetics, the American College of Obstetricians and Gynecologists, and the NIH. In this report, we first present MultiCode PLx testing data on a variety of genomic DNAs with various genotypes prepared from several sample types for the presence or absence of mutations in the cystic fibrosis transmembrane conductance regulator (CFTR) gene. Finally, data are presented on a prospective study comparing testing results obtained for freshly prepared blood spots by the PLx assay and the commercially available CF Gold Linear Array TM system (Roche Molecular). ) . DNA was extracted from whole blood and dried blood spots on Guthrie cards. The dried blood spots were stored at room temperature for 2-3 days after receipt until routine testing was completed. For long-term storage, dried blood spots were stored at 4°C. DNA extraction from Guthrie cards was performed with the Generation Capture Card Kit (Gentra Systems) according to the manufacturer's guidelines for DNA purification and elution. DNA was extracted from whole blood by use of the Puregene DNA Isolation Kit (Gentra Systems) or the Whatman FTA system (Whatman Inc., Clifton, NJ). All genomic DNAs from these preparations were diluted 1:10 in water, and 1L was added to each PCR reaction (1-50 ng of total DNA). As a cross-reference and in the prospective study, genotypes were also determined by CF Gold linear array or DNA sequencing if discrepancies were detected. PCR and TSE oligonucleotides were manufactured inhouse with a 48-column DNA Synthesizer (Northwest Engineering; now the 3900 from Applied Biosystems Inc.), using standard ␤-cyanoethyl phosphoramidite chemistry (for sequence information, see Tables 1, 2, and 3 in the Data Supplement accompanying the online version of this article at http://www.clinchem.org/content/vol50/ issue11/). Synthesis of sequences that contain iC and iG can also be purchased from IDT, Eurogentec, or EraGen. A demonstration assay encompassing this specific system can be purchased from EraGen Biosciences. Amino-labeled EraCode oligonucleotides were manufactured with an ABI 394 DNA Synthesizer (Applied Biosystems) using 3Ј-PT-Amino-Modifier C6 CPG (Glen Research). Target-specific extender primers were manufactured with the spacer phosphoramidite C3 (Glen Research) placed between the EraCode complement and the target-specific region. The isoG and isoC phosphora- , respectively] were coupled and deprotected under the conditions used for the natural base phosphoramidites. Postsynthesis work-up consisted of ammonium hydroxide deprotection followed by ethanol precipitation (PCR and TSE oligonucleotides) or 2-propanol precipitation (EraCode oligonucleotides). EraCode bead coupling For each conjugation reaction, a mixture (400 L) containing ϳ5 ϫ 10 6 carboxylated microspheres (Luminex Corporation) was centrifuged at 8000g for 1 min. The supernatant was removed, and the beads were resuspended in 50 L of 0.1 mol/L MES, pH 4.5. After resuspension, 3.3 L of 300 mol/L 5Ј-amine-modified EraCode DNA oligonucleotide and 5.0 L of freshly prepared 10 g/L EDC (Pierce Chemical) were added. The reaction was mixed thoroughly and incubated at room temperature in the dark for 30 min. A second 5.0 L of 10 g/L EDC was added, and the mixture was incubated for 30 min. Era-Code microspheres were washed by the addition of 1.0 mL of 10 mmol/L Tris (pH 8.0) containing 0.2 mL/L Tween 20 and pelleted by centrifugation at 8000g for 1 min; the supernatant was then removed. EraCode microspheres were washed a second time with 1 g/L sodium dodecyl sulfate in 10 mmol/L Tris (pH 8.0) and resuspended in 100 L of bead buffer (10 mmol/L MOPS, pH 7.5; 1 mmol/L EDTA, 0.5 g/L sodium dodecyl sulfate, 200 mmol/L sodium chloride, and 0.1 g/L sonicated herring sperm DNA). EraCoded microspheres were combined and diluted in bead buffer to generate a 50-code mixture at a concentration of 800 microspheres/L for each class of microsphere. Capture at room temperature of a tag to its specific complementary EraCode was tested by use of a model extension system where one target with a 5Ј-iC was constructed and 50 tags were interchanged onto a targetspecific extender sequence. EraCodes were constructed by use of 9 -10 bases comprising 2-3 iCs. In an effort to minimize interference from naturally occurring sequences, EraCodes were designed to have no more than four naturally occurring bases in a row. In an effort to minimize code-to-code cross-hybridization, no two codes were allowed to contain the same series of more than four nucleobases in a row. Fifty EraCodes were validated by use of 50 extenders, all containing the identical 3Ј target-specific sequence. The extenders were individually labeled using the target, standard concentrations of all four naturally occurring deoxynucleotide triphosphates, TiTaq, and diGTP-biotin at 1.5 mol/L. Each EraCode sequence was separately coupled to a specific color-coded fluorescent microsphere type (Luminex Corp). Each microsphere type contains a unique color or signature that can be determined by the Luminex 100 . After coupling, each EraCode bead conjugate was pooled to create an equalized bead master mixture (bead array: 50 different EraCode-labeled beads). The bead array was distributed into 50 separate reaction vessels. To each of the vessels, a specific extension reaction from above was added at room temperature along with streptavidin-phycoerythrin and was read on the Luminex 100 . In a 10-L extension reaction, an excess of a template oligonucleotide (5Ј-iC-CGTATGGGCTCACCTCGCTGTG-3Ј) at 150 nmol/L was combined with a control extension sequence at 100 nmol/L in EraGen MC TSE solution 35 (prod. no. PN1236; EraGen Biosciences, Inc.) with 100 mol/L each deoxynucleotide triphosphate, 1.5 mol/L diGTP-biotin, and 1ϫ Titanium Taq (Clontech). Each control extension oligonucleotide was composed of a 5Ј sequence complementary to each of the EraCode sequences, a three-carbon spacer, and a 3Ј sequence complementary to the template oligonucleotide (5Ј-CACAGC-GAGGTGAGCCCA-3Ј). Reactions were subjected to the following thermal cycling protocol: 30 s at 95°C and 1 min at 65°C. Each 10-L extension reaction was then hybridized to the EraCode bead mixture according to the described hybridization protocol. In addition, to mimic MultiCode conditions, unextended TSEs to the other EraCodes were also added. Hybridization of extension products in each 10-L reaction was performed by mixing 40 L of EraGen MC Hyb Solution A (prod. no. PN1237; EraGen Biosciences) and 0.25 L of EraCoded bead mixture. To each hybridization reaction, 0.5 L of 2 g/L streptavidin-R-phycoerythrin conjugate (Prozyme) plus 39.5 L of sheath fluid (Luminex Corp.) were added, and 65 L of the final reaction was injected into the Luminex 100 . pcr PCR primers were designed using Primer 3 (http:// www.broad.mit.edu/cgi-bin/primer/primer3_www.cgi) or Hyther (http://ozone2.chem.wayne.edu). Primer pairs were designed to flank the polymorphism(s) of interest and have a melting temperature of 53-57°C (see Table 2 in the online Data Supplement). A single iC was added to the 5Ј end of the PCR primers. PCR reactions were performed in a final volume of 8.0 L; each reaction contained EraGen MC PCR solution 35 (prod. no. PN1235 or PN1305 for the 579T system; Clinical Chemistry 50, No. 11, 2004 EraGen), 0.4 L of 20ϫ EraGen CFTR PCR primer mixture, 1ϫ Titanium Taq, and 1 L of diluted genomic target DNA. Reactions were cycled according to the following profile: 1 cycle of 30 s at 95°C, and 40 cycles of 1 s at 95°C, 10 s at 55°C, and 30 s at 72°C. Each of the primer mixtures contained forward and reverse primers for a subset of the mutations at concentrations of 100 -200 nM. tse TSE was used to determine the exact mutations present. TSE requires TiTaq, diGTP-biotin, and extenders that contain 3Ј target-specific bases and 5Ј codes specific for the EraCode beads described above. During TSE, TiTaq incorporates labels on extenders only when specific targets having 5Ј iCs are present. The nonnatural base pair formed by iG and iC has been successfully used for both molecular recognition (13) (14) (15) and for site-specific enzymatic incorporation (16 -18 ) . To date, we have covalently coupled a series of different reporter groups to diGTP, using succinimidyl ester chemistry and confirmed enzymatic incorporation into nucleic acids by use of several commercially available polymerases (data beyond the scope of this report). In many cases, single amplicons spanning regions that incorporated multiple mutations were used as templates for multiple extenders. After TSE, EraCode-labeled beads and streptavidin-phycoerythrin were added, and the reactions were analyzed on the Luminex 100 . Data are reported in median fluorescence intensity (MFI). For each polymorphism, pairs of TSE oligonucleotide primers were designed by use of Primer 3 and Hyther (see Table 3 in the online Data Supplement). Pairs were designed to have balanced melting temperatures ranging from 63 to 67°C. Each TSE oligonucleotide was composed of three components: a 5Ј sequence tag complementary to one of the EraCodes; a three-carbon spacer; and a 3Ј target-specific sequence fully complementary to the allele of interest. The polymorphic sequence for each pair of TSE oligonucleotides was the 3Ј base. Extension reactions were performed by bringing the 8-L PCR reactions to 10 L with addition of 2.0 L of EraGen MC TSE solution 35 (prod. no. PN1236; EraGen Biosciences) including primers for the 579T system (prod. no. PN1310), and 0.4 L of 20ϫ EraGen CFTR TSE primer mixture; final concentrations of TSEs were 25-100 nmol/L (see Table 3 in the online Data Supplement). Reactions were cycled according to the following profile: 1 cycle of 30 s at 95°C; 5 cycles of 1 s at 95°C and 2 min at 65°C. For this report, an assignment is defined as a genotype for a given target within the CFTR gene made for any given sample. Each sample tested will have 30 possible assignments because 30 targets for each sample are analyzed. To rapidly analyze the data obtained, we developed software to present the raw data from the instrument. The software directly imports the Luminex 100 data and organizes them by target mutation. Fractions of wild-type to wild-typeplus-mutant and raw MFI signals can be visualized graphically (Fig. 2) . The program sets default fraction windows at 1.0 -0.8 for wild-type samples, 0.6 -0.4 for heterozygous samples, and 0.0 -0.2 for homozygous mutant samples. Default windows can be changed by the end user. The area between the windows is designated as the no-call window. The program also sets no-amplification windows where wild-type and mutant MFIs are Ͻ200. Template set-up within the Luminex LabMap software was required, and importation to analysis software was achieved by dragging the Luminex file folder into the Analysis Software. Data files were parsed, and the resulting raw MFI values were organized by target and sample. After data acquisition from all 225 clinical samples tested, default cutoff windows for each target were empirically determined and set in a blinded fashion. Once determinations were made, reports were generated for offline analysis. The analysis software was written in Java and is compatible with Windows, Mac OS X, and Linux operating systems. Data importation is achievable with Luminex LabMap 100 (software versions 1.7, 2.1, and 2.2) and Bio-Rad BioPlex 3.0. The software displays scatter plots and fraction plots of each given target for all samples. automation Automation was performed on a Tecan Genesis Model 200 robotic workstation (TRP) starting from either the PCR amplicon formation or from isolated genomic DNA and continuing through data analysis. Samples used were a subset of the 225 used in the initial manual testing. Instructions for the automation program were written with Gemini software (Tecan). Initialization and launch of the bead-reading protocol on the Luminex 100 was performed automatically by use of a custom image-recognition and command-scripting program written in Java. After the read, the results were remotely analyzed by MultiCode Analysis software. Genomic DNA was purified manually before automation as described above. The TRP was fitted with the low-volume eight-channel dispense pipette system, TeStack, to supply disposable tips, a MJ Research thermocycler equipped with remote Alpha docks with Power Bonnet, and a robotic manipulator arm. Liquid reagents and disposable tips were also placed on the deck before runs were started. Reactions and additions for the automated process were the same as those described for the manual method unless noted. Reactions were performed in low-profile hard-skirted polypropylene well microplates, which allowed the robotic arm to grip and move the plates as needed. Each reaction was overlaid with 15 L of light mineral oil (Sigma) to prevent evaporation during thermal cycling. The robotic manipulator was used to transfer the plates to and from the thermocycler block and the Luminex 100 mounted off deck. Initialization and launch of the bead-reading protocol on the Luminex 100 were performed automatically by use of a custom image-recognition and command-scripting pro-gram written in Java. After bead-reading of each reaction, the results were analyzed by the MultiCode Analysis software as described above. A schematic diagram of the MultiCode system is shown in Fig. 3 . The MultiCode process has three steps. For the first step, regions of the target of interest are amplified by PCR using primers that contain a single 5Ј-iC. In the second step, the amplicons from the first step are used as templates for the TSEs, and iC directs specific diGTP-biotin incorporation. The TSEs are coded on their 5Ј ends with sequences that specifically recognized short complementary sequences (EraCodes) attached to an addressed solid matrix. The molecular recognition of any given code with its EraCode is specific enough so that increased temperatures, incubation steps, or washings are not required (typical requirements for hybridization capture using only naturally occurring base chemistries). During this second step, labels are placed site specifically on the TSEs. This process requires the presence of iC-containing amplicons that are complementary to the TSE. When the correct target is available, the extension reaction continues to the 5Ј end of the amplicon, where the diGTP-biotin is used to insert the biotin into the TSE. Specificity and signal consistency arise when coded TSEs are covalently attached to a single biotin reporter in the presence of the correct target. In the final step, the TSEs are captured on the solid matrix and detected. The presence of the target is then determined by where the reporters are located spatially on the liquid matrix (created by the differentially colored microspheres). Because each EraCode should be specific only for its complementary tag, only the bead that contains the specific EraCode should report a high MFI. Results show that for the correct complementary code, specific signal was between 2130 and 4357 MFI, with a mean (SD) MFI of 3040 (532; see Table 4 in the online Supplemental Data). Signals generated on microspheres that contained Era-Codes that were not completely complementary varied from 0 to 320 MFI with a mean (SD) MFI of 92 (28) . This set gave signal-to-noise ratios that ranged from 32:1 and 10:1 for the worst cases (i.e., where the specific signal was compared with the nonspecific signal from the microsphere presenting the highest noise). To our understanding, there have been no reports of a genetic-based molecular recognition system that was successfully implemented at room temperature without washing. mutation analysis using the manual cf multicode system PCR amplicons were generated for 17 regions within the CFTR gene, and 63 tagged extenders were used to analyze 27 mutations and 4 reflex targets. To improve the robustness of the overall results, the test was divided into four reactions with one additional test targeting the 5T7T9T reflex target. Two additional targets were included late in the development phase to include 394delTT and 3199del6. The test was transferred to the Wisconsin State Laboratory of Hygiene, where newborn CFTR genetic screening has been ongoing since 1994 and multimutation testing used since 2002. For the CFTR-specific PLx assay, positive signal intensities varied between 2000 and 5000 MFI and background signal varied from 50 to 1500 MFI. The variation in signal depended strictly on the target sequence. We observed that determination windows varied from 0.99 to 0.7 for wild-type samples, from 0.4 to 0.6 for heterozygous samples, and from 0.0 to 0.2 for homozygous mutant samples. The calling results are summarized in Table 1 [also see Table 4 in the online Data Supplement for additional details]. The 225 samples included 65 wild-type samples; 93 heterozygous samples representing 22 types; 36 compound heterozygous samples for 9 types; and 31 homozygous mutant samples for 6 types. Of the 225 samples tested manually, 203 (90.2%) samples reported all correct assignments. The failure rate of 10% was most likely a Step 2), tagged target-specific extenders are site-specifically labeled with diGTP-label when the correct amplicon is present. For this study, the label was biotin, and streptavidin-phycoerythrin was used for signaling purposes. ( Step 3), the tags are captured on EraCode-modified Luminex microspheres. The captured reactions are analyzed with the Luminex 100 system. product of DNA quality and not assay sensitivity. As stated in the Materials and Methods, the DNA samples were not fresh, and signal intensities indicated gross failure consistent with PCR failure. There were no discrepancies for genotype classification. For these 225 samples, there were a total of 6750 potential assignments (225 ϫ 27 mutations ϩ 3 reflex tests; 5T7T9T was not included). Correct assignments were made 99.3% (6704 of 6750) of the time. For the separate 5T7T9T-specific reflex test, 194 samples were tested and included at least 10 separate samples from each of the six possibilities (see Table 4 and Fig. 1 in the online Data Supplement). Test results showed that there were zero incorrect determinations and one no amplification failure. mutation analysis using the automated cf multicode system The final automated processes for the CFTR gene were conducted on a total of 86 samples. The results are reported from two separate processes: (1) PCR through data acquisition (20 samples) and (2) TSE through data acquisition (66 samples; Table 1 ; also see Table 5 in the online Data Supplement for further sample and testing information). Because workstation space is limited on the Tecan Genesis Model 200, tip usage is high when disposable tips are implemented, and amplicon carryover may be a concern. Process 1 may be applied for higher throughput scenarios where the PCR step could be implemented on yet another workstation or manually. For process 1, 20 (100%) samples reported calls correct. For process 2, 64 (97%) samples reported all correct assignments. There were no incorrect determinations. After assay validation, which included the testing of various mutations and sample types, the PLx assay for CFTR gene analysis was directly compared with the CF Gold linear array method. Using freshly obtained blood cards from newborns that had immunoreactive trypsinogen scores in the upper 4%, we used both tests to analyze CFTR for the recommended core panel of mutations and reflex targets. The testing was performed on 419 samples, and 100% concordance was recorded (no incorrect assignments; Table 1 , column B). Of the samples tested, there were 21 samples (5%) in which the complete lists of calls could not be obtained with the PLx system. The failure rate of 5% was documented as errors resulting from pipetting errors. These samples were retested, and correct assignments were made. Of the possible 12 570 potential assignments, 12 521 (99.6%) were correct. Run time after sample preparation for the CF Gold test was ϳ7 h compared with Ͻ3 h for the PLx system. Total hands-on time for the CF Gold test was ϳ3.5 h compared with 0.5 h for the PLx system. The data output for the CF Gold test was visual inspection of the nitrocellulose strips, whereas the PLx system has a computerized calling system as described. We have described a novel high-throughput platform that incorporates nucleic acid chemistries that simplify genetic analysis. PLx uses one additional base pair not found in nature in combination with microsphere flow cytometry technology to eliminate the most cumbersome steps commonly found in established multiplexed systems. The EraCode molecular recognition system described is particularly suited for the Luminex 100 instrumentation because the solid support can be dispensed into the complex reaction and decoding can be done at room temperature without washing, creating a true "Liquid Chip" type of system. Our study indicates that the chemistry is transferable to a clinical setting. In addition, although not required, the process can be performed with a robotic workstation for higher throughput. We demonstrated the principles of the chemistry by use of an essential target and selected mutations found within the CFTR gene recommended by leading medical organizations. Data from the Wisconsin State Laboratory of Hygiene confirmed that the chemistry is accurate. For example, there were no incorrect determinations, and the percentage of correct determinations was high. The data presented here were tabulated by use of a broad sample set that included every mutation possible within the panel targeted with the exception of the 3199del6. For 24 samples, the entire manual process takes ϳ3 h to obtain final results. An upgraded prototype version of the CFTR test where all mutations are analyzed in a single well takes ϳ2.5 h to complete analysis of 96 patient samples (validation data not yet available). To demonstrate the system in a fully automated mode, we elected to use the Tecan genesis liquid-handling system. The entire automated process also takes ϳ3 h from prepared sample to final read for an entire 96-well plate or 24 samples. We presented data from two possibilities (when automation started at the PCR step and when it started at the TSE step) because there may be a need to perform PCR offline. We believe that the expanded genetic alphabet enabled us to eliminate many of the steps that are required in technologies based on just the two base pairs provided by nature. The steps eliminated include solid support washings, transfers, aspirations, and high-temperature hybridizations. Specifically, we believe that there are many reasons that the extra base pair is beneficial. For example, we used short code sequences in the capture step to decode the complex extension mixtures. This may be possible with code sets constructed from only A, G, C, or T, but optimization would surely be more difficult because of cross-hybridization with sequence complements or close complements found in nature. The additional informational content that iC:iG builds into DNA also allows for greater diversity when compared with their all-natural counterparts. Specifically, if the four naturally occurring bases are used to make a library of 10mers, there are 4 10 , or 10 6 possibilities. In contrast, a six-letter system generates 6 10 , or 10 8 unique 10mers. As demon-Clinical Chemistry 50, No. 11, 2004 strated here, specific assembly can occur at conditions such as room temperature, which is in effect more amenable to and simplifies molecular diagnostics. We also demonstrated for the first time that biotin labels attached to diGTP could be site specifically incorporated into extension primers. More generally, we demonstrated that labeled diGTP can be incorporated enzymatically into DNA. This could allow for broad use of such compounds in molecular biology. Whenever a reporter group is required at a specific site within an amplicon, transcript, or extension product, the use of an additional base pair may be helpful. Recently we reported that site-specific enzymatic incorporation of a quencher attached to diGTP can be used to create a new and highly sensitive real-time PCR system (26 ) . Polymerase incorporation of labels only at sites where expanded genetic alphabet bases are positioned will also have other uses, such as 3Ј end labeling. Polymerase incorporation of EraCodes within PCR amplicons has now also been reported (27 ) . The EraCode system that uses additional nonnatural base pairing may have advantages over other DNA-based molecular recognition systems (28 -30 ) . These short DNA codes are simple to produce, specifically hybridize at room temperature, do not require wash steps, can be amplified with standard polymerases, and do not cross-hybridize with natural DNA. Taken together, we believe that the expansion of the genetic alphabet is a true paradigm shift to the ways molecular diagnostics will be developed in the future.
61
Antisense-induced ribosomal frameshifting
Programmed ribosomal frameshifting provides a mechanism to decode information located in two overlapping reading frames by diverting a proportion of translating ribosomes into a second open reading frame (ORF). The result is the production of two proteins: the product of standard translation from ORF1 and an ORF1–ORF2 fusion protein. Such programmed frameshifting is commonly utilized as a gene expression mechanism in viruses that infect eukaryotic cells and in a subset of cellular genes. RNA secondary structures, consisting of pseudoknots or stem–loops, located downstream of the shift site often act as cis-stimulators of frameshifting. Here, we demonstrate for the first time that antisense oligonucleotides can functionally mimic these RNA structures to induce +1 ribosomal frameshifting when annealed downstream of the frameshift site, UCC UGA. Antisense-induced shifting of the ribosome into the +1 reading frame is highly efficient in both rabbit reticulocyte lysate translation reactions and in cultured mammalian cells. The efficiency of antisense-induced frameshifting at this site is responsive to the sequence context 5′ of the shift site and to polyamine levels.
The standard triplet readout of the genetic code can be reprogrammed by signals in the mRNA to induce ribosomal frameshifting [reviewed in (1) (2) (3) ]. Generally, the resulting trans-frame protein product is functional and may in some cases be expressed in equal amounts to the product of standard translation. This elaboration of the genetic code (4, 5) demonstrates versatility in decoding. Requirements for eukaryotic ribosomal frameshifting include a shift-prone sequence at the decoding site and often a downstream secondary structure in mRNA. The majority of À1 programmed frameshift sites consist of a heptanucleotide sequence X XXY YYZ [where X can be A, G, C or U; Y can be A or U; and Z can be any nucleotide (6) ]. In this configuration, the P-and A-site tRNAs can re-pair with at least 2 out of 3 nt when shifted 1 nt towards the 5 0 end of the mRNA. Similarly, for +1 frameshift sites, the identity of the codons in the P-and A-sites of the ribosome is critical for efficient frameshifting. One factor affecting +1 frameshift efficiency is the initial stability of the P-site tRNA-mRNA interaction in the 0 frame (7) . High-efficiency frameshifting occurs when the P-site tRNA does not form standard codon-anticodon interactions (8) . In some studies, a correlation between +1 frameshift efficiency and the final stability of the P-site tRNA-mRNA interaction in the +1 frame has been shown previously (9, 10) . However, in other systems there appears to be little correlation (11) . In addition, competition between decoding of the 0 frame and +1 frame codons in the A-site may affect frameshifting efficiency (7) . Slow to decode 0 frame codons such as stop codons or those decoded by low abundance tRNAs favor frameshifting, as do +1 frame codons with high levels of corresponding cognate tRNAs (12) (13) (14) (15) (16) . High levels of frameshifting are often achieved by the stimulatory action of a cis-acting element located downstream of the shift site. A wide variety of structures, most commonly H-type pseudoknots (17) , have been identified which stimulate À1 frameshifting in eukaryotes [for reviews see (18, 19) ]. Mutagenic and structural data for several of the frameshift stimulators have demonstrated that each pseudoknot has key structural features required for frameshift stimulation (20) (21) (22) (23) (24) (25) (26) (27) (28) . However, unifying structural feature essential for frameshifting has not yet been identified. This observation combined with recent reports that simple antisense oligonucleotides can functionally mimic cis-acting 3 0 stimulators of À1 frameshifting (29, 30) demonstrates that many different structures can stimulate frameshifting. Although it should be noted that not all structures of equal thermodynamic stability can stimulate frameshifting (Discussion). RNA pseudoknots have also been shown to stimulate programmed +1 frameshifting in many eukaryotic antizyme genes (31, 32) . Antizyme is a negative regulator of cellular polyamine levels through its ability to target ornithine decarboxylase (the rate-limiting enzyme in polyamine biosynthesis) for degradation (33) (34) (35) , inhibits polyamine import (36, 37) and stimulates export (38) . Antizyme expression is induced by high-intracellular polyamine levels, and decreased with lowered levels. The polyamine sensor is a programmed +1 frameshift event that is required for antizyme synthesis. *To whom correspondence should be addressed. Tel: +1 801 585 1927; Fax: +1 801 585 3910; Email: mhoward@genetics.utah.edu Ó 2006 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. At low polyamine levels, termination at the end of open reading frame 1 (ORF1) is efficient, whereas at high levels of polyamines, a substantial proportion of ribosomes shift to the +1 reading frame and then resume standard decoding to synthesize the full-length and active antizyme protein. Frameshifting at the mammalian antizyme mRNA shift site, UCC UGA, is stimulated by two cis-acting signals (39, 40) . One of these, the 5 0 element, encompasses $50 bases upstream of shift site and is important for the polyamine effect (39) (40) (41) . The other cis-acting element is a pseudoknot located 3 0 of the shift site. The mammalian antizyme pseudoknot and a structurally distinct counterpart in a subset of invertebrate antizyme mRNAs (31) are the only pseudoknots known to act as stimulators for +1 frameshifting in eukaryotes. Although it is unknown if pseudoknots stimulate À1 frameshifting and +1 frameshifting by different mechanisms, one notable difference is found in positioning of the downstream structure relative to the shift site. Naturally occurring pseudoknots or stem-loop stimulators of À1 frameshifting typically begin $6-9 nt downstream of the A-site codon of the shift site (18) , whereas +1 frameshift pseudoknots are located closer with only a 2-3 nt separation from the A-site codon (31) . Mutagenic studies have revealed that altering the size of the spacer affects frameshifting and, in general, reduces efficiency (27, 31, (42) (43) (44) . Here we have tested the ability of antisense oligonucleotides, annealed downstream of the shift-prone site, UCC UGA, to induce shifting of the ribosome to the +1 reading frame. The directionality of frameshifting (either into the +1 or À1 reading frame) is shown to be dependent upon the position of the duplex region relative to the shift site, and the efficiency of frameshifting is responsive to polyamine levels and enhanced by the inclusion of stimulatory sequences found upstream of the human antizyme +1 programmed frameshift site. Complementary oligonucleotides, to construct the sequences described in this paper, were synthesized at the University of Utah DNA/Peptide Core Facility such that when annealed they would have appropriate ends to ligate into the SalI/ BamHI sites of the dual luciferase vector, p2luc (45) . Dual luciferase constructs were prepared and their sequence was verified as described previously (46) . Insert sequences with shift site in boldface is given as follows: P2lucAZ1wt: TCGACGGTCTCCCTCCACTGCTGTAG-TAACCCGGGTCCGGGGCCTCGGTGGTGCTCCTGATG-CCCCTCACCCACCCCTGAAGATCCCAGGTGGGCGAG-GGAATAGTCAGAGGGATCACAACGGATC; P2lucAZ10sp: TCGACGGTCTCCCTCCACTGCTGTAG-TAACCCGGGTCCGGGGCCTCGGTGGTGCTCCTGAC-CCTCACCCACCCCTGAAGATCCCAGGTGGGCGAGG-GAATAGTCAGAGGGATCACAACGGATC; P2lucAZ1hp: TCGACGGTCTCCCTCCACTGCTGTAG-TAACCCGGGTCCGGGGCCTCGGTGGTGCTCCTGATG- P2lucAZ1PKdel: TCGACGGTCTCCCTCCACTGCTG-TAGTAACCCGGGTCCGGGGCCTCGGTGGTGCTCCT-GATGCCCCTGGATC; P2lucAZ1PKm1: TCGACGGTCTCCCTCCACTGCTGT-AGTAACCCGGGTCCGGGGCCTCGGTGGTGCTCCTG-ATGCCCCTCACCCACCGGGATCACAAGGATC; P2lucAZ1sl: TCGACGGTCTCCCTCCACTGCTGTAGT-AACCCGGGTCCGGGGCCTCGGTGGTGCTCCTGATG-CCCCTCACCCACCCGGATC; P2lucAZ1FS: TCGACGTGCTCCTGATGCCCCTG-GATC; P2lucAZ1FSUGG: TCGACGTGCTCCTGGTGCCCCTG-GATC. The dual luciferase constructs (0.1 mg) described above were added directly to TNT coupled reticulocyte lysate reactions (Promega) with 35 S-labeled methionine in a volume of 10 ml. Reactions were incubated at 30 C for 1 h. Radiolabeled proteins were separated by SDS-PAGE and the gels were fixed with 7.5% acetic acid and methanol for 20 min. After drying under vacuum, the gels were visualized using a Storm 860 PhosphorImager (Molecular Dynamics) and radioactive bands quantified using ImageQuant software. Percent frameshifting was calculated as the percentage of full-length (frameshift) product relative to the termination product and the full-length product combined. The value of each product was corrected for the number of methionine codons present in the coding sequence. The reported values are the average and standard deviations obtained from at least three independent measurements. Tables showing percent frameshifting and standard deviations can be found in Supplementary Data. Plasmid p2lucAZ1PKdel was co-transfected into CV-1 cells with varying concentrations of AZ1B 2 0 -O-Methyl antisense oligonucleotides under the following conditions. CV-1 cells (1.5 · 10 4 ) in 50 ml of DMEM + 5% fetal bovine serum were added to wells (1/2 area 96-well tissue culture treated plates) containing 25 ng of DNA, varying amounts of AZ1B antisense oligonucleotides and 0.4 ml Lipofectamine 2000 (Invitrogen) in 25 ml of Optimem. Cells were incubated at 37 C (5% CO 2 ) for 20 h. Media were then removed from the cells and the transfected cells were lysed in 12.5 ml lysis buffer and luciferase activity determined by measuring light emission following injection of 25 ml of luminescence reagent (Promega). Percent frameshifting was calculated by comparing firefly/Renilla luciferase ratios of experimental constructs with those of control constructs: (firefly experimental RLUs/Renilla experimental RLUs)/(firefly control RLUs/Renilla control RLUs) · 100. The ability of cis-acting RNA structures or trans-acting 2 0 -O-Methyl antisense oligonucleotides to induce ribosomal frameshifting was determined by in vitro transcription and translation of a dual luciferase reporter vector, p2Luc. p2Luc contains the Renilla and firefly luciferase genes on either side of a multiple cloning site, and can be transcribed using the T7 promoter located upstream of the Renilla luciferase gene (45) . Sequences containing shift-prone sites were cloned between the two reporter genes such that the downstream firefly luciferase gene is in the +1 reading frame. The resulting constructs were then transcribed and translated in vitro with or without complementary cis-Acting stimulators of frameshifting at the antizyme shift site Initially, three dual luciferase reporter vectors were generated containing the human antizyme 1 frameshift cassette (p2luc-AZ1wt) with the 5 0 and 3 0 stimulators of frameshifting, with the pseudoknot deleted (p2luc-AZ1PKdel), or replaced with a stem-loop (p2luc-AZ1hp) (Figure 1 ). Each constructs was then subjected to coupled transcription and translation reactions in the presence of increasing amounts of spermidine, and the 35 S-labeled products separated by SDS-PAGE. Table 2 ). Maximal levels of frameshifting were found to occur when 2-4 mM of antisense oligonucleotide was added to the transcription/translation reactions (Supplementary Table 3 ). In the presence of 0.4 mM exogenous spermidine, highly efficient shifting of ribosomes into the +1 reading frame (higher than that observed in the wild-type antizyme frameshift cassette) was observed with the addition of AZ1A (26.1%), AZ1B (51.8%) and AZ1C (31.8%) (Supplementary Table 2 ). The most efficient frameshifting is observed with the antisense oligonucleotide AZ1B which anneals such that spacing between the shift site and the beginning of the duplex region is the same as that observed between the shift site and the beginning of stem 1 of the natural antizyme 3 0 pseudoknot structure (i.e. each has a 3 nt spacer). To verify that the antisense oligonucleotide was activating ribosomal frameshifting and not transcription slippage, RNA was transcribed from p2luc-AZ1PKdel in the absence of oligonucleotide and added to reticulocyte lysate translations in the presence of increasing amounts of 2 0 -O-Methyl AZ1B oligonucleotide. Frameshifting levels were increased to the same level as that observed in coupled transcription and translation reactions demonstrating that the oligonucleotide acts to induce frameshifting during translation (Supplementary Figure) . Surprisingly, the addition of AZ1A (0 spacer) also induced high-level frameshifting into the À1 reading frame in a manner which was modestly inhibited by the addition of spermidine (19% in the absence and 10% in the presence of 0.4 mM exogenous spermidine) ( Figure 3A and Supplementary Table 2 ). No À1 frameshift product was observed when the wild-type antizyme cassette was examined in the absence of antisense oligonucleotide addition (Figure 2 ; AZwt). As the AZ1A antisense oligonucleotide was designed to anneal directly adjacent to the UGA codon of the shift site, it was of interest to determine whether the wild-type antizyme pseudoknot could induce À1 frameshifting when located in the equivalent position. To address this, a new construct p2luc-AZ1-0sp ( Figure 1A ) was made by deleting the 3 nt spacer between the pseudoknot and the shift site of p2luc-AZ1wt. In this case, the wild-type pseudoknot is directly 3 0 adjacent to the shift site. The products of in vitro transcription and translation were separated by SDS-PAGE. No À1 frameshift product was observed and levels of the +1 frameshift product were significantly reduced to $3% ( Figure 3D and Supplementary Table 2) . AZ1A, AZ1B and AZ1C were designed to complement RNA sequences encoded by the originating vector. To determine if duplexes formed between the antisense oligonucleotide and 3 0 adjacent antizyme sequences would result in more efficient frameshift stimulation, reporter vectors were designed to contain a portion of the antizyme 3 0 stimulator. Construct p2luc-AZ1PKm1 contains sequences from the 5 0 half of the axis formed by the stacking of stem 1 and stem 2 of the pseudoknot ( Figure 1A) . Two complementary 2 0 -O-Methyl antisense oligonucleotides were designed. First, PKm1 has perfect complimentarity to the region starting 3 nt and ending 21 nt downstream of the UGA shift site codon. Second, PKm2 is the same except that a mispaired C and bulged A were located at positions 9 and 16, respectively. These two alterations were included to more closely mimic the natural pseudoknot which also contains a mispaired C and bulged A at equivalent positions along the extended stem formed by the stacking of pseudoknot stems 1 and 2 ( Figure 1 ; compare p2luc-AZ1wt with the duplex formed between p2luc-PKm1 and antisense oligonucleotide PKm2). PKm1 and PKm2 induced 30 and 22% frameshifting, respectively, when added to coupled transcription and translation reactions of p2luc-AZ1PKm1 in the presence spermidine ( Figure 4A and B, and Supplementary Table 4 ). Neither PKm1 nor Pkm2 induced frameshifting to the same levels seen with AZ1B, suggesting that the sequence content of the duplex region can affect the efficiency of frameshift stimulation and that native antizyme sequences are not required. A second construct, p2luc-AZ1sl, was designed to contain only the 5 0 half of stem 1 of the antizyme pseudoknot downstream from the UCC UGA shift site ( Figure 1A ). 2 0 -O-Methyl antisense oligonucleotides were designed to anneal between 3 and 15 nt (SL1) or 3 and 22 nt (SL2) downstream from the UGA codon of the shift site. Frameshift efficiency induced by these two antisense oligonucleotides, 8 and 22% respectively, was somewhat lower than that observed with PKm1 and PKm2 ( Figure 4C and D and Supplementary Table 4 ). In these cases frameshift efficiency was higher for the longer antisense oligonucleotide (SL2), suggesting that frameshift efficiency most probably correlates with stability of the duplex. As was seen with AZ1A, AZ1B and AZ1C, frameshifting efficiency stimulated by antisense oligonucleotides PKm1, PKm2, SL1 and SL2 was also strongly correlated with the concentration of exogenously added spermidine (Supplementary Table 5) . The importance of the antizyme 5 0 sequence context to antisense oligonucleotide induced ribosome frameshifting was examined by testing the frameshift site, UCC UGA, without the 5 0 and 3 0 stimulatory antizyme sequences. To this end, the 5 0 antizyme stimulatory sequences were deleted from p2luc-AZ1PKdel to make p2luc-AZ1FS. Each of the antisense oligonucleotides AZ1A, AZ1B or AZ1C was added to coupled transcription and translation reactions with p2luc-AZ1FS in the presence or absence of spermidine. Frameshift efficiency was measured at 11, 8 and 4%, in the presence of spermidine and 3, 0.4 and 0.2% in its absence for AZ1A, AZ1B and AZ1C, respectively ( Figure 5A and B) . To determine whether the stop codon of the shift site is essential for frameshifting, the UGA codon of p2luc-AZ1FS was altered to UGG such that the shift site was UCC UGG (p2luc-AZ1-UGG). Frameshift efficiency was significant, but reduced, compared to the shift site UCC UGA, and shows little stimulation by the addition of spermidine; AZ1A, AZ1B and AZ1C induced 3, 1 and 1.4% frameshifting in the presence of spermidine, and 1.9, 0.8 and 1.7% frameshifting in its absence, respectively ( Figure 5C and D) . The ability of antisense oligonucleotides to induce frameshifting in cultured mammalian cells was examined by co-transfection of CV-1 cells with p2lucAZ1PKdel and increasing amounts of 2 0 -O-Methyl antisense oligonucleotides AZ1B as described in Materials and Methods. In the absence of antisense oligonucleotide frameshifting levels were determined to be 1.1%, whereas a graded increase in frameshift levels was observed upon the addition of AZ1B ( Figure 6 ). Maximal frameshifting levels were 13% in the presence of 2 mM AZ1B in the transfection media. Several models attempting to explain pseudoknot stimulation of programmed À1 frameshifting have been proposed [for reviews see (18, 19) ]. Most models invoke a pausing mechanism whereby the ribosome is paused over the shift site such that time is allowed for the tRNAs to reposition in the new reading frame. This explanation is clearly too simplistic as stem-loops and pseudoknots of similar thermodynamic stability that cause ribosome pausing are not necessarily effective frameshift stimulators (47) (48) (49) . In addition, variations of the IBV pseudoknot have demonstrated a lack of correlation between the extent of pausing and the efficiency of frameshifting (47) . A recent publication by Brierley and co-workers (50) presents structural data demonstrating that the IBV frameshift stimulating pseudoknot blocks the mRNA entrance tunnel and leads to a structural deformation of the P-site tRNA. The resulting movement of the tRNA displaces the anticodon loop towards the 3 0 end of the mRNA. A model is presented in which this movement results in disruption of the codon-anticodon interactions, thus allowing for tRNA slippage relative to the mRNA. Similar tRNA movements were not observed with non-frameshift stimulating stem-loop structures. This model provides a feasible mechanistic explanation for the ability of some downstream structures to induce frameshifting. The ability of antisense oligonucleotides to induce highlevel À1 frameshifting (29, 30) demonstrates that elaborate tertiary structures are not required, and that a duplex formed by complementary antisense oligonucleotides (with a variety of chemistries, including RNA, 2 0 -O-Methyl, morpholino) is sufficient to induce high-level frameshifting. Here we demonstrate for the first time that trans-acting antisense oligonucleotides may stimulate ribosome shifting to the +1 reading frame at surprisingly high levels, levels which are greater than those achieved by natural 3 0 cis-acting mRNA pseudoknot structures in programmed +1 frameshifting. Structural studies indicating that the mRNA begins to enter the ribosome 7-9 nt downstream from the A-site codon is of direct relevance to this study (50, 51) . Our results indicate that maximal frameshifting is induced when the antisense-mRNA duplex begins 3 nt downstream of the UGA of the shift site, in agreement with the distance found between the UGA of the shift site and the beginning of stem 1 of the pseudoknot stimulator found in antizyme genes. Given this distance, the implication is that the stimulatory secondary structure would be encountered by the ribosome when the UCC codon enters the A-site of the ribosome. Perhaps as suggested by the structural studies of the IBV-1 frameshift inducing pseudoknot, the codon-anticodon interactions between the UCC codon and Ser-tRNA Ser are disrupted during translocation to the P-site. Given the importance of the UGA codon during frameshifting at the UCC UGA shift site, subsequent events following translocation of the UCC codon to the P-site and UGA to the A-site must influence frameshifting efficiency. This latter event most probably involves competition between termination and +1 frame decoding when the UGA codon is in the A-site. Various discussions have been presented for the importance of A-site and P-site events during ribosomal frameshifting (7, 52) and clearly, further investigations of this topic are warranted. The observation presented here that the antisense oligonucleotide, AZ1A, which anneals directly adjacent to the UGA stop codon can induce ribosome frameshifts to either the +1 or À1 reading frame is surprising. In light of the above discussion of spacing for naturally occurring cis-acting frameshift stimulators, it is possible that frameshifting may occur at codons upstream of the known UCC UGA shift site. However, visual examination of upstream codons does not reveal an obvious À1 or +1 frameshift site. The ability of spermidine to stimulate antisense oligonucleotide induced ribosome frameshifting to the +1 reading frame at the UCC UGA shift site in the absence of the natural 3 0 stimulator demonstrates that this cis-acting element is not required for polyamine responsiveness. Similarly, spermidine stimulation was observed in the absence of the 5 0 element but virtually eliminated by altering the UGA codon of the shift site to UGG. These observations are in agreement with previous studies examining the importance of cis-acting elements for polyamine induced frameshifting during expression of antizyme genes (39) (40) (41) . Finally, the ability to direct ribosomes to the +1 reading frame in living cells ( Figure 6 ) suggests a potential therapeutic application for antisense oligonucleotides. Directed frameshifting to the +1 reading frame near a disease causing À1 frameshift mutation would cause some ribosomes to resume decoding in the wild-type ORF, thus restoring partial production of full-length protein from mutant alleles. The importance of the stop codon for efficient frameshifting suggests that the stop codon following the frameshift mutation presents a promising target for antisense induced phenotypic suppression, and that modulation of intracellular polyamine levels, although not essential, may increase the effectiveness of this approach. Further experiments are required to determine the therapeutic potential of this approach in vivo including the generality and efficiency of frameshift induction at non-programmed frameshift sites.
62
The gene of an archaeal α-l-fucosidase is expressed by translational frameshifting
The standard rules of genetic translational decoding are altered in specific genes by different events that are globally termed recoding. In Archaea recoding has been unequivocally determined so far only for termination codon readthrough events. We study here the mechanism of expression of a gene encoding for a α-l-fucosidase from the archaeon Sulfolobus solfataricus (fucA1), which is split in two open reading frames separated by a −1 frameshifting. The expression in Escherichia coli of the wild-type split gene led to the production by frameshifting of full-length polypeptides with an efficiency of 5%. Mutations in the regulatory site where the shift takes place demonstrate that the expression in vivo occurs in a programmed way. Further, we identify a full-length product of fucA1 in S.solfataricus extracts, which translate this gene in vitro by following programmed −1 frameshifting. This is the first experimental demonstration that this kind of recoding is present in Archaea.
Translation is optimally accurate and the correspondence between the nucleotide and the protein sequences are often considered as an immutable dogma. However, the genetic code is not quite universal: in certain organelles and in a small number of organisms the meaning of different codons has been reassigned and all the mRNAs are decoded accordingly. More surprisingly, the standard rules of genetic decoding are altered in specific genes by different events that are globally termed recoding (1) . In all cases, translational recoding occurs in competition with normal decoding, with a proportion of the ribosomes not obeying to the 'universal' rules. Translational recoding has been identified in both prokaryotes and eukaryotes. It has crucial roles in the regulation of gene expression and includes stop codon readthrough, ribosome hopping and ±1 programmed frameshifting [for reviews see (2) (3) (4) ]. In stop codon readthrough a stop codon is decoded by a tRNA carrying an unusual amino acid rather than a translational release factor. Specific stimulatory elements downstream to the stop codon regulate this process (5) . Hopping, in which the ribosome stops translation in a particular site of the mRNA and re-start few nucleotides downstream, is a rare event and it has been studied in detail only in the bacteriophage T4 (6) . In programmed frameshifting, ribosomes are induced to shift to an alternative, overlapping reading frame 1 nt 3 0 -wards (+1 frameshifting) or 5 0 -wards (À1 frameshifting) of the mRNA. This process is regulated and its frequency varies in different genes. The ±1 programmed frameshifting has been studied extensively in viruses, retrotransposons and insertion elements for which many cases are documented (7) (8) (9) . Instead, this phenomenon is by far less common in cellular genes. A single case of programmed +1 frameshifting is known in prokaryotes (10, 11) while in eukaryotes, including humans, several genes regulated by this recoding event have been described previously [(4) and references therein]. Compared to +1 frameshifting, À1 frameshifting is less widespread with only two examples in prokaryotes (12) (13) (14) and few others in eukaryotes (15) (16) (17) . The programmed À1 frameshifting is triggered by several elements in the mRNA. The slippery sequence, showing the X-XXY-YYZ motif, in which X can be any base, Y is usually A or U, and Z is any base but G, has the function of favouring the tRNA misalignment and it is the site where the shift takes place (3, 18) . Frameshifting could be further stimulated by other elements flanking the slippery sequence: a codon for a low-abundance tRNA, a stop codon, a Shine-Dalgarno sequence and an mRNA secondary structure. It has been *To whom correspondence should be addressed. Tel: +39 081 6132271; Fax: +39 081 6132277; Email: m.moracci@ibp.cnr.it Ó 2006 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. reported that these elements, alone or in combination, enhance frameshifting by pausing the translating ribosome on the slippery sequence (4, 18) . Noticeably, known cases of recoding in Archaea [recently reviewed in (19) ] are limited to termination codon readthrough events that regulate the incorporation of the 21st and 22nd amino acids selenocysteine and pyrrolysine, respectively (20) (21) (22) (23) . No archaeal genes regulated by translational programmed frameshifting and ribosome hopping have been identified experimentally so far; therefore, if compared with the others domains of life, the study of translational recoding in Archaea is still at its dawn. We showed that the a-L-fucosidase gene from the crenarchaeon Sulfolobus solfataricus is putatively expressed by programmed À1 frameshifting (24) . This gene, named fucA1, is organized in the open reading frames (ORFs) SSO11867 and SSO3060 of 81 and 426 amino acids, respectively, which are separated by a À1 frameshifting in a 40 base overlap ( Figure 1A ). We have reported previously that the region of overlap between the two ORFs had the characteristic features of the genes expressed by programmed À1 frameshifting including a slippery heptanucleotide A-AAA-AAT (codons are shown in the zero frame) flanked by a putative stem-loop and the rare codons CAC ( Figure 1A ) resembling the prokaryotic stem-loops/hairpins and the Shine-Dalgarnolike sites (24) . We showed that the frameshifting, obtained by mutating by site-directed mutagenesis the fucA1 gene exactly in the position predicted from the slippery site, produced a full-length gene, named fucA1 A , encoding for a polypeptide of 495 amino acids ( Figure 1B ). This mutant gene expressed in Escherichia coli a fully functional a-L-fucosidase, named Ssa-fuc, which was thermophilic, thermostable and had an unusual nonameric structure (24, 25) . More recently, we determined the reaction mechanism and the function of the residues of the active site of the mutant enzyme (26, 27) . The functionality of the product of the mutant gene fucA1 A does not provide direct experimental evidence that programmed À1 frameshifting occurs in vivo and in S.solfataricus. To address these issues, we report here the study of the expression of the wild-type split gene fucA1 and of its mutants in the slippery sequence. We demonstrate here that fucA1 is expressed by programmed À1 frameshifting in both E.coli and S.solfataricus. This is the first experimental demonstration that this kind of recoding is present in the Archaea domain of life. The relevance of programmed À1 frameshifting in Archaea is also discussed. Figure 1 . The a-fucosidase gene. (A) Region of overlap in the wild-type split fucA1 gene. The N-terminal SSO11867 ORF is in the zero frame, the C-terminal SSO3060 ORF, for which only a fragment is shown, is in the À1 frame. The slippery heptameric sequence is underlined; the rare codons are boxed and the arrows indicate the stems of the putative mRNA secondary structure. The amino acids involved in the programmed À1 frameshifting and the first codon translated after this event in the À1 frame are shadowed. (B) Fragment of the full-length mutant fucA1 A gene. The small arrows indicate the mutated nucleotides. Analysis of the a-fucosidase expression S.solfataricus cells were grown, and cell extracts obtained, as described previously (24, 28) . The expression in the E.coli strain BL21(RB791) of the wild-type gene fucA1 and of the mutant genes fucA1 A [previously named FrameFuc in (24) ], fucA1 B , fucA1 sm and fucA1 tm as fusions of glutathione S-transferase (GST) and the purification of the recombinant proteins were performed as reported previously (23) . The nomenclature used in this paper for the different a-fucosidase genes is listed in Table 1 . For the western blot studies, equal amounts of E.coli cultures expressing the wild-type and mutant fucA1 genes, normalized for the OD 600 , were resuspended in SDS-PAGE loading buffer containing 0.03 M Tris-HCl buffer, pH 6.8, 3% SDS (w/v), 6.7% glycerol (w/v), 6.7% 2-mercaptoethanol (w/v) and 0.002% blue bromophenol (w/v). The samples were incubated at 100 C for 5 min (unless otherwise indicated) and were directly loaded on to the gel. Western blot analyses were performed by blotting SDS-PAGEs of the concentrations indicated on Hybond-P polyvinylidenfluorid filters (Amersham Biosciences, Uppsala, Sweden); polyclonal anti-Ssa-fuc antibodies from rabbit (PRIMM, Milan, Italy) and anti-GST antibodies (Amersham Biosciences) were diluted 1:5000 and 1:40 000, respectively. The filters were washed and incubated with the ImmunoPure anti-rabbit IgG antibody conjugated with the horseradish peroxidase (HRP) from Pierce Biotechnology (Rockford, IL, USA). Filters were developed with the ECL-plus Western Blotting Detection system (Amersham Biosciences) by following the manufacturer's indications. The molecular weight markers used in the western blot analyses were the ECL streptavidin-HRP conjugate (Amersham Biosciences). The protein concentration of the samples was measured with the method of Bradford (29) and the amounts of sample loaded on to the SDS-PAGEs are those indicated. The quantification of the bands identified by western blot was performed by using the program Quantity One 4.4.0 in a ChemiDoc EQ System (Bio-Rad, Hercules, CA, USA) with the volume analysis tool. The frameshifting efficiency was calculated as the ratio of the intensity of the bands of the frameshifted product/frameshifted product + termination product. The mutants in the slippery sequence of the wild-type gene fucA1 were prepared by site-directed mutagenesis from the vector pGEX-11867/3060, described previously (24, 27) . The synthetic oligonucleotides used (PRIMM) were the following: FucA1sm-rev, 5 0 -TTTAGGTGATATTGGTGTT-CTGGTCTATCT-3 0 ; FucA1sm-fwd, 5 0 -GAACACCAATAT-CACCTAAAGAATTCGGCCCA-3 0 ; FucA1tm-rev, 5 0 -AGG-TGATATTGGTGTTCTGGTCTATCTGGC-3 0 ; FucA1tm-fwd, 5 0 -CCAGAACACCAATATCACCTCAAGAACTCGGCCCA-GT-3 0 , where the mismatched nucleotides in the mutagenic primers are underlined. Direct sequencing identified the plasmids containing the desired mutations and the mutant genes, named fucA1 sm and fucA1 tm , were completely re-sequenced. The mutant Ssa-fuc B was prepared by site-directed mutagenesis from the vector pGEX-11867/3060, by using the same site-directed mutagenesis kit described above. The synthetic oligonucleotides used were FucA1sm-rev (described above) and the following mutagenic oligonucleotide: Fuc-B, 5 0 -GAACACCAATATCACCTAAAGAAGTTCGGCCC-AGT-3 0 , where the mismatched nucleotides are underlined. Direct sequencing identified the plasmid containing the desired mutations and the mutant gene, named fucA1 B , was completely re-sequenced. The enzymatic characterization of Ssa-fuc B was performed as described previously (24, 27) . Samples of the proteins expressed in E.coli from the wildtype gene fucA1 and the mutants fucA1 A and fucA1 sm , purified as described, were fractionated on an SDS-PAGE. Protein bands were excised from the gel, washed in 50 mM ammonium bicarbonate, pH 8.0, in 50% acetonitrile, reduced with 10 mM DTT at 56 C for 45 min and alkylated with 55 mM iodoacetamide for 30 min at room temperature in the dark. The gel pieces were washed several times with the buffer, resuspended in 50 mM ammonium bicarbonate and incubated with 100 ng of trypsin for 2 h at 4 C and overnight at 37 C. The supernatant containing peptides was analysed by MALDIMS on an Applied Biosystem Voyager DE-PRO mass spectrometer using a-cyano-4-hydroxycynnamic acid as matrix. Mass calibration was performed by using the standard mixture provided by manufacturer. Liquid chromatography online tandem mass spectrometry (LCMSMS) analyses were performed on a Q-TOF hybrid mass spectrometer (Micromass, Waters, Milford, MA, USA) coupled with a CapLC capillary chromatographic system (Waters). Peptide ions were selected in the collision cell and fragmented. Analysis of the daughter ion spectra led to the reconstruction of peptide sequences. Genomic DNA from S.solfataricus P2 strain was prepared as described previously (24) . A DNA fragment of 1538 nt containing the complete fucA1 gene, was prepared by PCR, by using the following synthetic oligonucleotides (Genenco, Florence, Italy): FucA1-fwd, 5 0 -CTGGAGGCGCGCTAA-TACGACTCACTATAGGTCAGTTAAATGTCACAAAA-TTCT-3 0 ; FucA1-rev, 5 0 -GACTTGGCGCGCCTATCTAT-AATCTAGGATAACCCTTAT-3 0 , in which the sequence corresponding to the genome of S.solfataricus is underlined. In the FucA1-fwd primer, the sequence of the promoter of the T7 RNA polymerase is in boldface and the sequence of the BssHII site is shown in italics. The PCR amplification was performed as described previously (24) and the amplification products were cloned in the BssHII site of the plasmid pBluescript II KS+. The fucA1 gene was completely re-sequenced to check if undesired mutations were introduced by PCR and the recombinant vector obtained, named pBlu-FucA1, was used for translation in vitro experiments. The plasmids expressing the mutant genes fucA1 A , fucA1 sm and fucA1 tm for experiments of translation in vitro were prepared by substituting the KpnI-NcoI wild-type fragment, containing the slippery site, with those isolated from the mutants. To check that the resulting plasmids had the correct sequence, the mutant genes were completely re-sequenced. The mRNAs encoding wild-type fucA1 and its various mutants were obtained by in vitro run-off transcription. About 2 mg of each plasmid was linearized with BssHII and incubated with 50 U of T7 RNA polymerase for 1 h 30 min at 37 C. The transcription mixtures were then treated with 10 U of DNAseI (RNAse free) for 30 min. The transcribed RNAs were recovered by extracting the samples twice with phenol (pH 4.7) and once with phenol/chloroform 1:1 followed by precipitation with ethanol. The mRNAs were resuspended in DEPC-treated H 2 O at the approximate concentration of 0.6 pmol/ml. In vitro translation assays were performed essentially as described by Condò et al. (28) . The samples (25 ml final volume) contained 5 ml of S.solfataricus cell extract, 10 mM KCl, 20 mM Tris-HCl, pH 7.0, 20 mM Mg acetate, 3 mM ATP, 1 mM GTP, 5 mg of bulk S.solfataricus tRNA, 2 ml of [ 35 S]methionine (1200 Ci/mmol at 10 mCi/ml) and $10 pmol of each mRNA. The mixtures were incubated at 70 C for 45 min. After this time, the synthesized proteins were resolved by electrophoresis 12.5% acrylamide-SDS gels and revealed by autoradiography of the dried gels on an Instant Imager apparatus. Cells of S.solfataricus, strain P2, were grown in minimal salts culture media supplemented with yeast extract (0.1%), casamino acids (0.1%), plus glucose (0.1%) (YGM) or sucrose (0.1%) (YSM). The extraction of total RNA was performed as reported previously (24) . Total RNA was extensively digested with DNAse (Ambion, Austin, TX, USA) and the absence of DNA was assessed by the lack of PCR amplification with each sets of primers described below. The RT-PCR experiments were performed as reported previously (24) by using the primers described previously that allowed the amplification of a region of 833 nt (positions 1-833, in which the A of the first ATG codon is numbered as one) overlapping the ORFs SSO11867 and SSO3060 (24) . For real-time PCR experiments total cDNA was obtained using the kit Quantitect RT (Qiagen GmbH, Hilden, Germany) from 500 ng of the same preparation of RNA described above. cDNA was then amplified in a Bio-Rad LightCycler using the DyNAmo HS Syber Green qPCR Kit (Finnzymes Oy, Espoo, Finland). Synthetic oligonucleotides (PRIMM) used for the amplification of a region at the 3 0 of the ORF SSO3060 were as follows: 5 0 -Real: 5 0 -TAAATGGC-GAAGCGATTTTC-3 0 ; 3 0 -Real: 5 0 -ATATGCCTTTGTCGC-GGATA-3 0 for the gene fucA1. 5 0 -GAATGGGGGTGATA-CTGTCG-3 0 and 5 0 -TTTACAGCCGGGACTACAGG-3 0 for the 16S rRNA gene. For each amplification of the fucA1 gene was used $2500fold more cDNA than that used for the amplification of the 16S rRNA. Controls with no template cDNA were always included. PCR conditions were 15 min at 95 C for initial denaturation, followed by 40 cycles of 10 s at 95 C, 25 s at 56 C and 35 s at 72 C, and a final step of 10 min at 72 C. Product purity was controlled by melting point analysis of setpoints with 0.5 C temperature increase from 72 to 95 C. PCR products were analysed on 2% agarose gels and visualized by ethidium bromide staining. The expression values of fucA1 gene were normalized to the values determined for the 16S rRNA gene. Absolute expression levels were calculated as fucA1/16S ratio in YSM and YGM cells, respectively. Relative mRNA expression levels (YSM/YGM ratio) were calculated as ( fucA1/ 16S ratio in YGM cells)/(fucA1/16S ratio in YSM). Each cDNA was used in triplicate for each amplification. The wild-type fucA1 gene, expressed in E.coli as a GST-fused protein, produced trace amounts of a-fucosidase activity (2.3 · 10 À2 units mg À1 after removal of GST), suggesting that a programmed À1 frameshifting may occur in E.coli (24) . The enzyme was then purified by using the GST purification system and analysed by SDS-PAGE revealing a major protein band ( Figure 2A ). The sample and control bands were excised from the gel, digested in situ with trypsin and directly analysed by matrix-assisted laser desorption/ ionization mass spectrometry (MALDIMS). As shown in Figure 2B and C, both spectra revealed the occurrence of an identical mass signal at m/z 1244.6 corresponding to a peptide (Peptide A) encompassing the overlapping region of the two ORFs. This result was confirmed by liquid chromatography online tandem mass spectrometry (LCMSMS) analysis of the peptide mixtures. The fragmentation spectra of the two signals showed the common sequence Asn-Phe-Gly-Pro-Val-Thr-Asp-Phe-Gly-Tyr-Lys in which the amino acid from the ORF SSO11867 is underlined. These results unequivocally demonstrate that the protein containing the Peptide A is produced in E.coli by a frameshifting event that occurred exactly within the slippery heptamer predicted from the analysis of the DNA sequence in the region of overlap between the ORFs SSO11867 and SSO3060 ( Figure 1A) . Remarkably, the MALDIMS analysis of the products of the wild-type fucA1 gene revealed the presence of a second Peptide B at m/z 1258.6 that is absent in the spectra of the Ssa-fuc control protein ( Figure 2B and C). The sequence of Peptide B obtained by LCMSMS ( Figure 2D ) was Lys-Phe-Gly-Pro-Val-Thr-Asp-Phe-Gly-Tyr-Lys. This sequence differs only by one amino acid from Peptide A demonstrating that the interrupted gene fucA1 expresses in E.coli two fulllength proteins originated by different À1 frameshifting events. Polypeptide A results from a shift in a site A and it is identical to Ssa-fuc prepared by site-directed mutagenesis (24) , suggesting that the expression occurred with the simultaneous P-and A-site slippage. Instead, polypeptide B, named Ssa-fuc B , is generated by frameshifting in a second site B as the result of a single P-site slippage ( Figure 2E) . To measure the global efficiency of frameshifting in the two sites of the wild-type gene fucA1 we analysed the total extracts of E.coli by western blot using anti-GST antibodies ( Figure 2F ). Two bands with marked different electrophoretic mobility were observed: the polypeptide of 78.7 ± 1.1 kDa migrated like GST-Ssa-fuc fusion and was identified as originated from frameshifting in either site A or B of fucA1. The protein of 38.1 ± 1.2 kDa, which is not expressed by the mutant gene fucA1 A (not shown), had an electrophoretic mobility compatible with GST fused to the polypeptide encoded by the ORF SSO11867 solely (27 and 9.6 kDa, respectively). This polypeptide originated from the translational termination of the ribosome at the OCH codon of the fucA1 N-terminal ORF ( Figure 1A) . The calculated ratio of frameshifting to the termination products was 5%. To test if the full-length a-fucosidase produced by the À1 frameshifting event in site B (Ssa-fuc B ), resulting from the single P-site slippage has different properties from Ssa-fuc, whose sequence arises from the simultaneous P-and A-site slippage, we prepared the enzyme by site-directed mutagenesis. The slippery sequence in fucA1 A-AAA-AAT was mutated in A-AAG-AAG-T where mutations are underlined. The new mutant gene was named fucA1 B . The first G, producing the conservative mutation AAA!AAG, was made to disrupt the slippery sequence and hence reducing the shifting efficiency. The second G was inserted to produce the frameshifting that results in the amino acid sequence of Peptide B. Therefore, the sequence of the two full-length mutant genes fucA1 A and fucA1 B differs only in the region of the slippery sequence: A-AAG-AAT-TTC-GGC and A-AAG-AAG-TTC-GGC, respectively (the mutations are underlined, the nucleotides in boldface were originally in the À1 frame) ( Table 1) . The recombinant Ssa-fuc B was purified up to $95% (Materials and Methods). Gel filtration chromatography demonstrated that in native conditions Ssa-fuc B had the same nonameric structure of Ssa-fuc with an identical molecular weight of 508 kDa (data not shown). In addition, Ssa-fuc B had the same high substrate selectivity of Ssa-fuc. The two enzymes have high affinity for 4-nitrophenyl-a-L-fucoside (4NP-Fuc) substrate at 65 C; the K M is identical within the experimental error (0.0287 ± 0.005 mM) while the k cat of Ssa-fuc B (137 ± 5.7 s À1 ) is $48% of that of Ssa-fuc (287 ± 11 s À1 ). In addition, 4-nitrophenyl-a-L-arabinoside, -rhamnoside, 4-nitrophenyl-a-D-glucoside, -xyloside, -galactoside and -mannoside were not substrates of Ssa-fuc B as shown previously for Ssa-fuc (24) . This suggests that the different amino acid sequence did not significantly affect the active site. Both enzymes showed an identical profile of specific activity versus temperature with an optimal temperature higher than 95 C (data not shown). The heat stability and the pH dependence of Ssa-fuc and Ssa-fuc B are reported in Figure 3 . At 80 C, the optimal growth temperature of S.solfataricus, the half-life of Ssa-fuc B is 45 min, almost 4-fold lower than that of Ssa-fuc ( Figure 3A ). The two enzymes showed different behaviour at pH <6.0 at which Ssa-fuc B is only barely active and stable ( Figure 3B) ; however, the two enzymes showed similar values of specific activity at pHs above 6.0, which is close to the intracellular pH of S.solfataricus (30) . Characterization of the slippery sequence of fucA1 in E.coli The experimental data reported above indicate that the predicted slippery heptanucleotide in the region of overlap between the ORFs SSO11867 and SSO3060 of the wildtype gene fucA1 could regulate in cis the frameshifting events observed in E.coli. To test this hypothesis, we mutated the sequence A-AAA-AAT into A-AAG-AAT and C-AAG-AAC (mutations are underlined) obtaining the fucA1 single mutant ( fucA1 sm ) and triple mutant ( fucA1 tm ) genes, respectively. It is worth noting that the mutations disrupt the slippery sequence, but they maintain the À1 frameshift between the two ORFs (Table 1) . Surprisingly, the expression of fucA1 sm in E.coli produced a full-length polypeptide that, after purification by affinity chromatography and removal of the GST protein, showed the same electrophoretic migration of Ssa-fuc and Ssa-fuc B ( Figure 4A ). This protein was then characterized by mass spectrometry analyses following in situ tryptic digestion. Interestingly, the MALDI spectra revealed the presence of a single peptide encompassing the overlapping region between the two ORFs with a mass value of 1259.7 Da (peptide C; Figure 4B ). The sequence of peptide C, determined from the fragmentation spectra obtained by LCMSMS analysis, was Glu-Phe-Gly-Pro-Val-Thr-Asp-Phe-Gly-Tyr-Lys ( Figure 4C ). Remarkably, apart from the Glu residue, this sequence is identical to that of peptide B produced from fucA1, indicating that in the mutant gene fucA1 sm only one of the two frameshifting events observed in the wild-type fucA1 gene had occurred. The presence of a Glu instead of Lys was not unexpected. The mutation A-AAA-AAT!A-AAG-AAT in fucA1 sm was conservative in the zero frame of the ORF SSO11867 (AAA!AAG, both encoding Lys), but it produced the mutation AAA!GAA (Lys!Glu) in the À1 frame of the ORF SSO3060. It is worth noting that the frameshifting efficiency of the gene fucA1 sm , calculated by western blot as described above, was 2-folds higher (10%) if compared to fucA1 (5%) ( Figure 4D ). This indicates that the mutation cancelled the frameshifting site A and, in the same time, enhanced the frameshifting efficiency of site B. In contrast, the triple mutant fucA1 tm produced in E.coli only the low molecular weight band resulting from translational termination ( Figure 4D ). No full-length protein could be detected in western blots probed with either anti-GST ( Figure 4D ) or anti-Ssa-fuc antibodies ( Figure 4E ). These data show that the disruption of the heptameric slippery sequence completely abolished the frameshifting in E.coli confirming that this sequence has a direct role in controlling the frameshifting in vivo. To test whether fucA1 is expressed in S.solfataricus we analysed the extracts of cells grown on yeast extract, sucrose and casaminoacids medium (YSM). Accurate assays showed that S.solfataricus extracts contained 3.4 · 10 À4 units mg À1 of a-fucosidase activity. These very low amounts hampered the purification of the enzyme. The extracts of S.solfataricus cells grown on YSM revealed by western blot a band of a molecular mass >97 kDa and no signals were detected with the pre-immune serum confirming the specificity of the anti-Ssa-fuc antibodies ( Figure 5A) . The different molecular mass may result from post-translational modifications occurred in the archaeon or from the incomplete denaturation of a protein complex. In particular, the latter event is not unusual among enzymes from hyperthermophilic archaea (31, 32) . To test which hypotheses were appropriate, cellular extracts of S.solfataricus were analysed by western blot extending the incubation at 100 C to 2 h. Interestingly, this treatment shifted the high-molecular mass band to 67.6 ± 1.2 kDa ( Figure 5B and C) , which still differs from that of the recombinant Ssa-fuc, 58.9 ± 1.2 kDa, leaving the question on the origin of this difference unsolved. To try to shed some light we immunoprecipitated extracts of S.solfataricus with anti-Ssa-fuc antibodies and we analysed the major protein band by MALDIMS. Unfortunately, we could not observe any peptide compatible with the fucosidase because the heavy IgG chain co-migrated with the band of the expected molecular weight (data not shown). To test if the scarce amounts of the a-fucosidase in S.solfataricus extracts was the result of reduced expression at transcriptional level, we performed a northern blot analysis of total RNA extracted from cells grown either on YSM or YGM media. We could not observe any signal by using probes matching the 3 0 of the ORF SSO3060 (data not shown). These results suggest that fucA1 produced a rare transcript; therefore, we analysed the level of mRNA by RT-PCR and by real-time PCR. A band corresponding to the region of overlap between the ORFs SSO11867 and SSO3060 was observed in the RNA extracted from cells grown on YSM 2 mg) , the product of the gene fucA1 sm (2 mg), and Ssa-fuc (4 mg). The bands with faster electrophoretic mobility result from the proteolytic cleavage of the full-length protein (25) . (B) Partial MALDIMS spectrum of the tryptic digest from mutant fucA1 sm expressed in E.coli. The mass signal corresponding to peptide C encompassing the overlapping region is indicated. (C) LCMSMS analysis of peptide C. The amino acid sequence inferred from fragmentation spectra is reported. (D) Western blot of E.coli cellular extracts expressing fucA1 A , the wild-type fucA1, fucA1 sm and fucA1 tm genes (Materials and Methods). The blot was probed with anti-GST antibodies. (E) Western blot of partially purified protein samples expressed in E.coli fused to GST from wild-type and mutant fucA1 genes. Cellular extracts were loaded on GST-Sepharose matrix. After washing, equal amounts of slurries (30 ml of 300 ml) were denaturated and loaded on a 8% SDS-PAGE. Extracts of E.coli cells expressing the parental plasmid pGEX-2TK were used as the negative control (pGEX). The blot was probed with anti-Ssa-fuc antibodies. and YGM media, demonstrating that under these conditions the two ORFs were co-transcribed ( Figure 6A) . The experiments of real-time PCR shown in Figure 6B demonstrated that rRNA16S was amplified after $17 cycles while the amplification of fucA1 mRNA was observed after 38 cycles, despite the fact that we used $2500-fold more cDNA for the amplification of fucA1. This indicates that the gene fucA1 is transcribed at very low level. No significant differences in the fucA1 mRNA level were observed in cells grown in YSM or YGM media. This is further confirmed by the analysis by western blot of the extracts of the same cells of S.solfataricus used to prepare the total RNAs, which revealed equal amounts of a-fucosidase in the two extracts ( Figure 6C) . Therefore, the low a-fucosidase activity observed under the conditions tested is the result of the poor transcription of the fucA1 gene. To determine whether, and with what efficiency, the À1 frameshifting could be performed by S.solfataricus ribosomes, mRNAs obtained by in vitro transcription of the cloned wild-type fucA1 gene and the mutants thereof were used to program an in vitro translation system prepared as described by Condò et al. (28) . To this aim, a promoter of T7 polymerase was inserted ahead of the gene of interest to obtain RNA transcripts endowed with the short 5 0untranslated region of 9 nt observed for the natural fucA1 mRNA (24) . Autoradiography of an SDS-PAGE of the translation products (Figure 7 ) revealed that the wild-type fucA1 transcript produced a tiny but clear band whose molecular weight corresponded to that of the full-length Ssa-fuc obtained by site-directed mutagenesis (24); the latter was translated quite efficiently in the cell-free system in spite of being encoded by a quasi-leaderless mRNA. Judging from the relative intensity of the signals given by the translation products of the wild-type fucA1 and the full-length mutant fucA1 A , the efficiency of the À1 frameshifting in the homologous system was $10%. No signals corresponding to the polypeptides expected from the separated ORFs SSO11867 and SSO3060 (9.6 and 46.5 kDa, respectively) were observed. However, it should be noted that the product of SSO11867, even if synthesized, is too small to be detected in the gel system employed for this experiment. The larger product of ORF SSO3060, on the other hand, is certainly absent. These data unequivocally demonstrate that the ribosomes of S.solfataricus can decode the split fucA1 gene by programmed À1 frameshifting with considerable efficiency producing a full-length polypeptide from the two ORFs SSO11867 and SSO3060. Remarkably, under the same conditions at which fucA1 drives the expression of the full-length protein, we could not observe any product from the fucA1 sm and fucA1 tm constructs. These data demonstrate that the integrity of the heptanucleotide is essential for the expression of the fucA1 gene in S.solfataricus, thus further confirming that the gene is decoded by programmed À1 frameshifting in this organism. In addition, the lack of expression of fucA1 sm by translation in vitro in S.solfataricus contrasts with the efficient expression of this mutant in E.coli, indicating that the two organisms recognize different sequences regulating the translational frameshifting. The identification of genes whose expression is regulated by recoding events is often serendipitous. In the framework of our studies on glycosidases from hyperthermophiles, we identified in the genome of the archaeon S.solfataricus a split gene encoding a putative a-fucosidase, which could be expressed through programmed À1 frameshifting (24) . We tackled this issue by studying the expression of fucA1 in S.solfataricus and in E.coli to overcome the problems connected to the scarcity of expression of the a-fucosidase gene and to the manipulation of hyperthermophiles. As already reported by others, in fact, it is a common strategy to study recoding events from different organisms in E.coli (23, 33) . The expression in E.coli of the wild-type split gene fucA1 led to the production by frameshifting of two full-length polypeptides with an efficiency of 5%. This is a value higher than that observed in other genes expressed by translational frameshifting in a heterologous system such as the proteins gpG and gpGT (0.3-3.5%) (33) . The gene fucA1 is expressed in S.solfataricus at very low level under the conditions tested. In particular, the transcriptional analysis of the gene revealed that it is expressed at very low level in both YSM and YGM media. Similarly, no differences in the two media could be found by western blot probed with anti-Ssa-fuc antibodies, indicating that the low expression of the enzyme in S.solfataricus is the result of scarce transcription rather than suppressed translation. Western blots allowed us to identify a specific band $8.7 kDa heavier than that of the recombinant Ssa-fuc and experiments of translation in vitro showed that the wild-type gene expresses a full-length polypeptide exhibiting the same molecular mass of the recombinant protein. This demonstrates that the translational machinery of S.solfataricus is fully competent to perform programmed frameshifting. It seems likely that the observed discrepancy in molecular mass might arise from post-translational modifications that cannot be produced by the translation in vitro. Further experiments are required to characterize the a-L-fucosidase identified in S.solfataricus. MALDIMS and LCMSMS analyses of the products in E.coli of the wild-type split gene fucA1 demonstrated that two independent frameshifting events occurred in vivo in the proposed slippery site. In particular, the sequences obtained by LCMSMS demonstrate that peptide A results from a simultaneous backward slippage of both the P-and the A-site tRNAs ( Figure 8A) . Instead, the sequence of peptide B is the result of the re-positioning on the À1 frame of only the P-site tRNA; in fact, the next incorporated amino acid is specified by the codon in the new frame ( Figure 8B) . Therefore, the expression by À1 frameshifting of the wild-type gene fucA1 in E.coli follows the models proposed for ribosomal frameshifting (34) . We confirmed the significance of the slippery heptanucleotide in promoting the programmed frameshifting in vivo by mutating the putative regulatory sequence. The triple mutant fucA1 tm gave no full-length products; presumably, the mutations in both the P-and in the A-site of the slippery sequence dramatically reduced the efficiency of the À1 frameshifting as observed previously in metazoans (35) . This result confirms that the intact slippery sequence in the wild-type gene fucA1 is absolutely necessary for its expression in E.coli. In contrast, surprisingly, the single mutant fucA1 sm showed an even increased frequency of frameshifting (10%) if compared to the wild-type and produced only one polypeptide by shifting specifically in site B. We explained this result observing that the mutation in the P-site of the slippery sequence A-AAA-AAT!A-AAG-AAT created a novel slippery sequence A-AAG identical to that controlling the expression by programmed À1 frameshifting of a transposase gene in E.coli (36) . Therefore, apparently, the single mutation inactivated the simultaneous P-and A-site tRNA re-positioning and, in the same time, fostered the shifting efficiency of the tRNA in the P-site. It is worth noting that, instead, in S.solfataricus, only the simultaneous slippage is effective ( Figure 8B ) and even the single mutation in the slippery sequence of fucA1 sm completely annulled the expression of the gene. This indicates that this sequence is essential in the archaeon and that programmed frameshifting in S.solfataricus and E.coli exploits different mechanisms. Furthermore, since the only difference between the enzymes produced by the frameshifting sites A and B, Ssa-fuc and Ssa-fuc B , respectively, is the stability at 80 C, which is the S.solfataricus physiological temperature, the functionality of Ssa-fuc B in the archaeon appears questionable. The reason why fucA1 is regulated by programmed À1 frameshifting is not known. However, the physiological significance of programmed frameshifting has been assigned to a minority of the cellular genes while for most of them it is still uncertain [see (4) and reference therein; (16) ]. This mechanism of recoding is exploited to set the ratio of two polypeptides such as the t and g subunits of the DNA polymerase III holoenzyme in E.coli (12) . Alternatively, programmed frameshifting balances the expression of a protein, as the bacterial translational release factor 2 and the eukaryotic ornithine decarboxylase antizyme [see (4) and (18) and references therein]. In the case of fucA1, the polypeptide encoded by the smaller ORF SSO11867 could never be detected by western blots analyses. In addition, the modelling of Ssa-fuc on the high-resolution crystal structure of the a-L-fucosidase from Thermotoga maritima (25, 37) showed that the fucA1 N-terminal polypeptide is not an independent domain. Moreover, we have shown recently that SSO11867 includes essential catalytic residues (27) , excluding the possibility that a functional a-fucosidase can be obtained from the ORF SSO3060 alone. Therefore, several lines of evidence allow us to exclude that programmed À1 frameshifting is used to set the ratio of two polypeptides of the a-fucosidase from S.solfataricus. More probably, this translational mechanism might be required to control the expression level of fucA1. Noticeably, this is the only fucosidase gene expressed by programmed À1 frameshifting. Among carbohydrate active enzymes, the only example of expression through this recoding mechanism is that reported for a gene encoding for a a(1,2)-fucosyltransferase from Helicobacter pylori that is interrupted by a À1 frameshifting (38) . In this case, the expression by programmed frameshifting would lead to a functional enzyme synthesizing components of the surface lipopolysaccharides to evade the human immune defensive system. It is hard to parallel this model to fucA1. Nevertheless, the monosaccharide fucose is involved in a variety of biological functions (39) . Therefore, the a-L-fucosidase might play a role in the metabolism of fucosylated oligosaccharides; experiments are currently in progress to knockout the wild-type fucA1 gene and to insert constitutive functional mutants of this gene in S.solfataricus. FucA1 is the only archaeal a-L-fucosidase gene identified so far; hence, it is probably the result of a horizontal gene transfer event in S.solfataricus. However, since there are no a-fucosidases genes regulated by programmed frameshifting in Bacteria and Eukarya, it is tempting to speculate that this sophisticated mechanism of translational regulation preexisted in S.solfataricus and it was applied to the fucosidase gene for physiological reasons. The identification of other genes interrupted by À1 frameshifts in S.solfataricus would open the possibility that they are regulated by programmed À1 frameshifting. Recently, the computational analysis of prokaryotic genomes revealed that seven Archaea harbour interrupted coding sequences, but S.solfataricus is not included in this study (40) . A computational analysis on several archaeal genomes revealed that 34 interrupted genes are present in the genome of S.solfataricus, 11 of these genes are composed by two ORFs separated by À1 frameshifting and could be expressed by recoding (B. Cobucci-Ponzano, M. Rossi and M. Moracci, manuscript in preparation). We have experimentally shown here, for the first time, that programmed À1 frameshifting is present in the Archaea domain. This finding is the missing piece in the puzzle of the phylogenetic distribution of programmed frameshifting demonstrating that this mechanism is universally conserved.
63
Role of RNA helicases in HIV-1 replication
Viruses are replication competent genomes which are relatively gene-poor. Even the largest viruses (i.e. Herpesviruses) encode only slightly >200 open reading frames (ORFs). However, because viruses replicate obligatorily inside cells, and considering that evolution may be driven by a principle of economy of scale, it is reasonable to surmise that many viruses have evolved the ability to co-opt cell-encoded proteins to provide needed surrogate functions. An in silico survey of viral sequence databases reveals that most positive-strand and double-stranded RNA viruses have ORFs for RNA helicases. On the other hand, the genomes of retroviruses are devoid of virally-encoded helicase. Here, we review in brief the notion that the human immunodeficiency virus (HIV-1) has adopted the ability to use one or more cellular RNA helicases for its replicative life cycle.
Helicases are enzymes that separate in an energy-dependent manner stretches of duplexed DNA and/or RNA into singlestranded components. Currently, based on characteristic motifs and the sequence comparisons, three superfamilies (SF1 through 3) and two smaller families (F4, F5) of helicases have been identified (1) . Superfamily 1 and 2 contain helicases which share seven or more recognized signature amino acid motifs while SF3 and F4 and F5 helicases are characterized generally by three conserved motifs (2) ; the F4 and F5 proteins are largely bacterial and bacteriophage proteins. Currently, it should be cautioned that many 'helicases' are not bona fide helicases, but may only function as RNA translocases, perhaps to fulfill functions in the remodeling of ribonucleoprotein complexes (RNP). DEAD-box and the related DEAH, DExH and DExD (3) helicases are the most numerous members of SF2 and are ubiquitously present in eukaryotic genomes. These helicases share eight conserved motifs and are commonly refered to as the DExH/D family of helicases. Humans, Arabidopsis and Saccharomyces have 38, 55 and 25 such entities, respectively (4) . Differing from DNA helicases and DExH proteins, DEAD helicases are poor in unwinding long nucleic acid duplexes and are best suited for separating short RNA hybrids. DEAD-box proteins bind with high-affinity RNAprotein complexes while exhibiting little RNA sequence preference. This suggests that the specificity determinants for DEAD helicases may be through the recognition of protein factors. In this regard, a better understanding of the roles for DEAD proteins depends on the clear characterization of their respective interacting proteins. Although the precise substrate for most helicases awaits definition, DEAD helicases are generally thought to participate pleiotropically in many aspects of RNA metabolism including transcription, mRNA splicing, mRNA export, translation, RNA stability and mitochondrial gene expression (5) (6) (7) (8) . Some examples of helicases and their attributed functions include the following. UAP56, Brr2, Prp16, Prp22 and Prp43 play roles in RNA-splicing (4, 9) , while Dbp5 (10, 11) and DDX3 (12) chaperone RNAs from the nucleus into the cytoplasm. eIF4a and Ded1 serve for translation of mRNAs while Rh1B, Ski2, Dob1, Dhh1 helicases contribute to mRNA stability (4) . Other DEAD helicases act in ribosome biogenesis through regulation of small nucleolar RNAs and ribosomal RNAs (rRNAs) interactions (13, 14) . Finally, Neurospora and Trypanosoma DEAD proteins contribute to mitochondrial gene expression (15, 16) ; a Cryptococcus DEAD helicase is required for cryptococcosis pathogenesis (17) , and the dipteran Chironomus tentans uses a hrp84 DEAD helicase to regulate mRNA transport from the nucleus into the cytoplasm onto polyribosomes (18) . Given that helicases significantly contribute to normal cellular metabolism, are they similarly essential to viruses? The operational answer appears to be a qualified 'yes'. Indeed, when DEAD/DEAH-box helicase motif (InterPro IPR001410) was used to search the EMBL-EBI database, 1561 matches to individual viral sequence entries were found (http://www.ebi.ac.uk/interpro/DisplayIproEntry?ac¼ IPR001410), suggesting that many viruses have evolved to encode directly helicase or helicase-like proteins. The strongest biological evidence which supports the importance of a helicase in the virus life cycle comes from those viruses with an RNA genome. Hence, all positive-strand RNA viruses encode one or more helicase/helicase-like open reading frame (ORF) which, aside from the RNA-dependant RNA polymerase, is the most highly conserved viral sequence. Although less ubiquitous, helicases are also found in other types of viruses (see some examples listed in Table 1 ). Direct mutagenesis studies have confirmed that a helicase function is biologically required for the replication of many viruses including vaccinia virus (19) , poliovirus (20) , alphaviruses (21) , brome mosaic virus (22) , nidoviruses (23, 24) and flaviviruses (25) (26) (27) . In 1981, the first cases of acquired immunodeficiency syndrome (AIDS) were described in American homosexual men. Thereafter, within three short years, French and American scientists confirmed that the human immunodeficiency virus (HIV) is the causative agent for AIDS. In the ensuing 20 years, >20 million individuals have died from AIDS; and currently, in 2006, 50 million people worldwide are infected by HIV-1 with 3 million incremental AIDS deaths and 4-5 million new infections occurring annually. The magnitude of this burden casts urgency to medical research on HIV/AIDS. HIV-1 is a retrovirus of the lentivirus genus with an RNA genome of 9 kilobases which encodes nine polypeptides. The major HIV-1 structural proteins are encoded by three genes, gag (group-specific antigen), pol (polymerase) and env (envelope), while the accessory proteins, Vif, Vpu, Vpr and Nef, and the regulatory proteins, Tat and Rev, are the primary translation products of multiply-spliced mRNA. HIV-1 infects CD4+ human T-cells and macrophages and integrates as a provirus into the host cell's DNA. Gene expression of HIV-1 is governed transcriptionally by a viral protein, Tat (28, 29) , via its binding to a nascent viral TAR RNA (30) , and post-transcriptionally by a second viral protein Rev (31,32) through its association with the viral RRE RNA. Both Tat and Rev interact with several host cell proteins in their transcriptional and post-transcriptional functions (33) . HIV-1 does not encode for any RNA helicase; however, findings suggest that host cell RNA helicases may be involved in the reverse transcription of HIV-1 RNA, in HIV-1 mRNA transcription and in the nucleus-to-cytoplasm transport of viral mRNA. A recent unexpected finding revealed the possibility that an RNA helicase may potentially contribute roles in HIV-1 particle assembly and reverse transcription (34) . Using proteomic analyses, Roy et al. (34) reported that the DEAH protein RNA helicase A (RHA) was found associated with HIV-1 Gag and packaged into HIV-1 virions in an RNAdependent manner. When RHA was knocked down in cells, HIV-1 particles which were produced from these cells were significantly less infectious. This appears to be compatible with two possible explanations. First, it is conceivable that RHA participates in the formation of infectious virus particles either by shaping Gag-RNA interaction during viral particle assembly or by budding. Failure of RHA to properly restructure viral RNP could explain the observed reduced infectivity. Second, Roy et al. (34) reported evidence that HIV-1 particles that do not contain RHA showed reduced virionendogenous reverse transcriptase activity. In this respect, it may be that RHA assists HIV-1 reverse transcriptase to more efficiently copy RNA by unwinding RNA secondary structure or by promoting the interaction of viral RNA with the nucleocapsid protein in order to assemble a better reverse transcription complex. Separate from reverse transcription, the unwinding of highly structured RNAs might also be reasoned to be important for transcription (35) . However, direct evidence for an RNA helicase role has been somewhat elusive. There are several examples which seemingly support an activity for RNA helicase in transcription. First, in vaccinia virus, it has been postulated that the NPH-II helicase assists transcription by strand-separating duplexed RNA structures to prevent R-loop formation behind the elongating RNA polymerase (36) . Second, RHA has been invoked to provide a factorrecruitment role, bridging at the promoter the CREB-binding protein (37) and RNA polymerase II (37) . Third, the p68 DEAD-box helicase was shown recently to be a novel transcriptional co-activator for p53's transcriptional function (38) . Interestingly, in the latter two instances, neither the ATPase nor the helicase activity of RHA and p68 is apparently required for their attributed transcriptional roles. For HIV-1, two recent studies provide clues that RNA helicases may also serve co-factor function for transcription from the viral long terminal repeat (LTR). Fujii et al. (39) observed that RHA conserves in its N-terminus two double-stranded RNA-binding (dsRBD) domains characterized previously for the TAR RNA-binding protein, TRBP (40, 41) . These investigators found in both reporter and virus replication assays that RHA activated, in a TAR RNA-binding dependant manner, HIV-1 LTR-directed transcription (39 roles in transcription or indirectly influence the milieu of polymerase II initiation/elongation at the LTR. Downstream from transcription, the fate of HIV-1 encoded RNA is regulated at the step of export of unspliced/partially spliced moieties from the nucleus into the cytoplasm. Unspliced and partially spliced viral RNAs code for genomic RNAs that are packaged into progeny virions and structural proteins. Hence, the egress of these RNAs from the nucleus into the cytoplasm is critical to the life cycle of the virus. Exit of HIV RNAs from the nucleus is a significant issue because unspliced/partially spliced cellular mRNAs are routinely retained in and not permitted export from the nucleus (43) (44) (45) (46) . A large body of work has suggested an elegant solution to this conundrum. Thus, it was established that the HIV-1 encoded Rev protein binds a highly secondary structured element (Rev responsive element; RRE) present in all unspliced and partially spliced HIV transcripts (47) (48) (49) (50) (51) (52) (53) (54) (55) (56) (57) ; and this binding specifically distinguishes, for purposes of nuclear export, viral transcripts from cellular RNAs. New evidence now suggests that RNA helicases are also co-factors for Rev-directed export of HIV-1 mRNAs (58) . In its role of transporting unspliced and incompletely spliced viral RNAs from the nucleus, Rev directly interacts with nuclear export receptor CRM1 (59, 60) , and CRM1 is required for Rev-mediated export of HIV RNAs (59, 61, 62) . A recent report provides data that an RNA helicase, DDX3, is an additional player in the Rev-CRM1-RRE complex (12) . Thus, it was shown that DDX3 over-expression enhanced Revdependent, but not other export, pathway; and that DDX3 is a nucleo-cytoplasmic shuttling protein which binds CRM1 and Rev. Moreover, DDX3's necessity for Rev/RRE/CRM1 function was demonstrated by knock-down of cell endogenous DDX3. Finally, because DDX3 locates to nuclear pore complexes (NPC), Yedavalli et al. (12) further proposed that this human helicase, like the analogous yeast Dbp5p (11), may function with Rev/CRM1 to remodel and 'thread' large unspliced HIV-1 RNAs through the nuclear pore, facilitating their final release to the cytoplasmic side of the NPC. The above DDX3 results are consistent with two additional papers which described similar findings for a related RNA helicase, DDX1. Thus, Pomerantz and co-workers (63) showed that DDX1 binds directly to the N-terminus of Rev and to the RRE-RNA motif and participates in the export of unspliced HIV-1 RNA from the nucleus to the cytoplasm. Additionally, they illustrated that reduced expression of DDX1 in astrocytes explains the previously observed tissue restricted function of HIV-1 Rev (64). Fully spliced viral mRNAs encoding for viral Tat, Rev and Nef, proteins have been shown previously to exit the nucleus using the cellular mRNA export pathway. Export of these mRNA may require the RNA helicase Dbp5 (65, 66) . As yet, the involvement of Dbp5 in export of spliced HIV-1 viral RNA has not been fully clarified. The story of HIV-1 and RNA helicases is, however, likely to be more complex than and unlikely to conclude simply with RHA, RH116, DDX1 and DDX3 (Figure 1 ). HIV-1 RNAs are extensively regulated through splicing. Splicing is a multiple-step process requiring the recognition of splice sites by spliceosomes. It is generally believed that remodeling of RNA-RNA and RNA-protein interactions within the spliceosome is catalyzed by a family of DEAD/DExH box RNA helicases. To date, seven mammalian proteins that are RNA helicases have been implicated in mRNA splicing (67, 68) . Whether there is specific preference by subclasses of RNA helicases for viral mRNA splicing remains to be clarified. Moreover, how cellular RNA helicases might contribute to the translation of viral mRNAs also require further investigation. Recently both Van't Wout et al. (69) and Krishnan and Zeichner (70) have provided evidence that the expression of several cellular RNA helicases including DDX24, DDX21, DDX18, DDX11 and DDX9 is modulated during HIV-1 infection; however, the precise cellular role and significance of these helicases for HIV-1 pathogenesis have not be characterized. Interestingly, Krishnan and Zeichner reported microarray data which examined the transition of HIV-1 infection from latency to productive replication, and found that several cellular RNA helicases were upregulated (71) . For future understanding of functions, it will be important to design experiments which can segregate helicases which serve direct, although perhaps overlapping and redundant, roles on HIV-1 from those that might participate indirectly in the viral life cycle. Nevertheless, the convergence of evidence would support that several discrete cellular RNA helicases contribute importantly to the efficient execution of several steps in the HIV-1 replicative cycle. Given that the HIV-1/AIDS disease burden has reached pandemic global proportions, new antiviral strategies that target molecularly delineated mechanisms used by this virus are urgently needed (72) . Is there a possibility that host cell helicases can be therapeutic targets for anti-HIV-1 chemotherapy? Implicit within this question is the concept that one could attack a host cell protein in order to treat an infecting pathogen. Although targeting a cellular protein involved in a viral pathway risks obvious cytotoxicity, this approach avoids the inherent problem posed by rapid HIV-1 mutation to all currently utilized chemotherapeutics targeted to virus-encoded proteins. We note that inhibition of cell-encoded enzymes in medical therapy is not an unprecedented strategy. Suppression of angiotensin-converting enzyme (ACE) is widely used to treat hypertension, congestive heart failure, myocardial infarction, endothelial dysfunction and renal disease (73) . Elsewhere, aromatase inhibitors have been used to treat hormone-dependent breast cancer (74) , and inhibitors of cellular secretory proteases are contemplated for Alzheimer's disease (75) . We recently inhibited the cellular polyprotein convertase, furin, at minimal toxicity to the cell in order to block HIV-1 replication (76). Thus, a priori exclusion of cellular helicase as an antiviral target is not warranted. Guarded optimism that small molecule helicase inhibitors can be developed against viruses arises from encouraging progress in non-retroviral systems. Unlike HIV-1, human herpesviruses physically encode helicases. The herpes simplex virus UL5 and UL9 genes are helicases in superfamily 1 and 2, respectively (77) . HSV UL5 together with UL8 and UL52 form a heterotrimeric helicase-primase complex responsible for unwinding duplex viral DNA at replication forks. Two recent studies provide proof-of-concept that the HSV helicase-primase can be targeted at low host cell toxicity by two new classes of drugs, amino-thiazolyphenylmolecules (78) and thiazole amide derivatives (79) . In addition, other studies suggest that the NS3 protein, a RNA helicase encoded by Hepatitis C virus and related West Nile virus and Japanese Encephalitis virus can be targeted to inhibit viral replication (80) (81) (82) . This conceptual break through in drug development is important because it indicates that target discrimination between different helicases by small molecule inhibitors is possible. Of relevance to HIV-1, a synthetic immunomodulator Murabutide was shown recently to suppress HIV-1 replication in macrophages and T cells. Murabutide was shown to inhibit the activity of RNA helicase RH116, blocking its positive transcriptional activity for HIV-1 gene expression (42) . If one looks beyond the signature motifs conserved amongst helicases, then it becomes clear that the different proteins are widely divergent in their coding sequences. In principle, this suggests that individual helicases can be abrogated with specificity in a knowledge-directed manner. In theory, a helicase can be attacked by (i) inhibition of NTPase activity through direct competition for NTP binding, (ii) inhibition of substrate binding through direct competition at active site, (iii) allosteric mechanisms to affect NTPbinding/NTP hydrolysis and/or polynucleotide binding, and (iv) inhibition of unwinding activity by steric hinderance of helicase translocation along the polynucleotide substrate (83, 84) . Because the NTP-binding and substrate-binding pockets may be sufficiently similar between various helicases, specificity of inhibition through these sites will likely be extremely difficult, although perhaps not impossible. On the other hand, the tremendous variations in sequence and sizes of helicases, in their oligomerization states, in their discrete domains responsible for protein-protein interactions and/or for targeting to specific nucleic acids (85) , and in their differential localizations within cells (86) offer interventional possibilities outside of the NTP-or polynucleotidebinding sites. We are in the preliminary stages of screening ringexpanded nucleoside analogs found previously to be successful NTPase/helicase inhibitors of West Nile virus, Hepatitis C virus and Japanese encephalitis virus (81, 82) . We have observed that a few of these candidate inhibitors have substantial anti-HIV-1 activity at doses that do not incur cytotoxicity to cells treated in tissue culture for 1 week. Further studies are needed before concluding that these compounds exert specific inhibition of DDX3, one of the other cellular helicases, or some other target altogether. There is another area where a cellular helicase activity and HIV-1 are likely to intersect. An emerging research focus is the role of small interfering RNAs (87) and microRNAs (miRNA) as innate cell defenses against viruses including HIV-1 (87) (88) (89) (90) . In human cells, the precursor for miRNA (pre-miRNA) is processed by DICER (Figure 2A) which is a ribonuclease with a bona fide RNA helicase domain (91, 92) . A surprising recent finding revealed that the human TRBP, which has been shown to be a potent binder of the HIV-1 TAR RNA RNA (40, 41) , is an indispensable dsRNA-binding partner of DICER which allows the latter to associate with pre-miRNA (91, 92) . Without TRBP, DICER's miRNA processing activity is lost. Thus an intriguing scenario can potentially unfold. Accordingly, whereas the ribonuclease-helicase protein DICER requires TRBP to process duplex-structured miRNAs in order that the cell can use such matured miRNAs for antiviral defense, it could be speculated that HIV-1 has evolved to restrict this defense by the ability to transcribe viral TAR RNA to squelch TRBP away from DICER (Y. Bennasser, M. L. Yeung and K. T. Jeang, manuscript submitted) ( Figure 2B ). If this thinking is correct, then HIV-1 has developed mechanisms not only to co-opt the active functions of a virus-propitious cellular helicase (i.e. DDX3) but also to inactivate the role of a second virus-pernicious helicase (i.e. DICER) for purposes of selfish gain. In a separate perspective, virus infection can trigger through double-stranded viral RNAs an innate antiviral immune response. Thus viral dsRNAs can be recognized by cellular proteins [pattern-recognition receptors (PRRs)] which initiate antiviral responses by inducing the production of a variety of cytokines including type I interferons (IFN-a and IFN-b) and initiating additional inflammatory and adaptive immune responses. Recently, DExD/H RNA helicases such as RIG-1 (retinoic acid inducible gene-1) (93) and Mda5 (melanoma differentiation-associated gene 5) (94) have been identified as suppressors of viral replication by binding to virus associated dsRNA and activating type I interferon-dependent antiviral immunity. Over-expression of RIG-1 and Mda5 was found to enhance dsRNA induced type I interferon antiviral response. Currently, it remains speculative whether helicases like RIG-1 and Mda5 may recognize HIV-1 dsRNA and trigger an innate immune response. Intriguingly, several reports exist in the literature that HIV-1 infection does induce activation of type 1 interferons (95, 96) . In conclusion, by studying helicase proteins one can gain insights into normal cellular metabolic processes, abnormal inherited human diseases (e.g. Bloom syndrome, Werner syndrome, Cockayne's syndrome and xeroderma pigmentosum; all diseases with mutations in cellular helicases), and remarkably also the biology of viruses.
64
MIMOX: a web tool for phage display based epitope mapping
BACKGROUND: Phage display is widely used in basic research such as the exploration of protein-protein interaction sites and networks, and applied research such as the development of new drugs, vaccines, and diagnostics. It has also become a promising method for epitope mapping. Research on new algorithms that assist and automate phage display based epitope mapping has attracted many groups. Most of the existing tools have not been implemented as an online service until now however, making it less convenient for the community to access, utilize, and evaluate them. RESULTS: We present MIMOX, a free web tool that helps to map the native epitope of an antibody based on one or more user supplied mimotopes and the antigen structure. MIMOX was coded in Perl using modules from the Bioperl project. It has two sections. In the first section, MIMOX provides a simple interface for ClustalW to align a set of mimotopes. It also provides a simple statistical method to derive the consensus sequence and embeds JalView as a Java applet to view and manage the alignment. In the second section, MIMOX can map a single mimotope or a consensus sequence of a set of mimotopes, on to the corresponding antigen structure and search for all of the clusters of residues that could represent the native epitope. NACCESS is used to evaluate the surface accessibility of the candidate clusters; and Jmol is embedded to view them interactively in their 3D context. Initial case studies show that MIMOX can reproduce mappings from existing tools such as FINDMAP and 3DEX, as well as providing novel, rational results. CONCLUSION: A web-based tool called MIMOX has been developed for phage display based epitope mapping. As a publicly available online service in this area, it is convenient for the community to access, utilize, and evaluate, complementing other existing programs. MIMOX is freely available at .
Since the pioneering work of Smith and co-workers [1] [2] [3] , phage display technology has been widely used in both basic research such as the exploration of protein-protein interaction sites and networks [2] [3] [4] [5] , and applied research such as the development of new drugs, diagnostics, and vaccines [6] [7] [8] . Phage display has also become a promising epitope mapping method, which has been applied in many fields such as allergology [9] and oncology [10] . The phage display based epitope mapping is usually accomplished through comparing the sequence of mimotopes (antibody-selected phage displayed peptides) to the anti-gen. In some cases, the mimotope sequence is identical or very similar to a sequence in the antigen [2] , there by indicating the location of the native epitope. These cases are rare however, and usually the mimotope sequence has little, if any, similarity with the antigen sequence. Compared with traditional epitope mapping methods such as solving the crystal structure of the antigen-antibody complex or scanning overlapping peptides of the antigen, phage display based epitope mapping is generally much cheaper and less arduous. Though epitope mapping based on phage display can be done manually [11] , it is quite tedious and time-consuming to compare a set of mimotopes to the antigen without computational support. The low sequence similarity between the mimotope and the antigen often makes the mapping even harder. To solve these problems, several groups have researched algorithms and programs that assist and automate phage display based epitope mapping [12] [13] [14] [15] [16] [17] . According to their dependency on antigen structure, the existing programs for phage display based epitope mapping can be classified into three categories. Program in the first category such as FINDMAP, only work with sequence data from the mimotopes and antigen [13] . The second category needs both the sequence data and the antigen structure. SiteLight [12] , 3DEX [14] , and Mapitope [16, 17] belong to this category. A very recently published work: MIMOP [15] makes the third category, which integrates the two different approaches and can work with or without the antigen structure. Though implemented differently, all the existing programs have succeeded in given cases. However, most of the existing tools have not been implemented as a freely available online service until now, making it less convenient for the community to access, utilize, and evaluate them. In the present study, we describe a web-based tool for phage display based epitope mapping named MIMOX. It was coded with Perl as a CGI program and can be used to align a set of mimotopes and derive a consensus sequence. The consensus sequence, or a single mimotope sequence, can then be mapped on to the antigen structure, and potential epitopes determined by spatial clustering of the mapped residues. The results mapped on to the antigen's 3D structure can then be viewed interactively. To validate this web-based tool, we compared the results from MIMOX with the results from other computational tools and experimentally identified native epitopes in several case studies. Overall architecture of MIMOX MIMOX was coded with Perl using modules from the Bioperl [18] project. The whole online service provided by MIMOX is accomplished through a set of CGI scripts. The MIMOX service can be divided into two main sections. In the first section, MIMOX provides a simple interface for ClustalW [19] to align a set of mimotope sequences; this is implemented as the script mimosa.pl. The alignment can then be used to derive the consensus sequence through a simple statistical method; this is implemented as the script mimocs.pl. The alignment can also be viewed and managed through an embedded Java applet version of JalView [20] ; this is implemented as the script jalviews.pl. In the second section, MIMOX tries to map the user supplied sequence on to the given antigen structure. This is implemented as the script mimox.pl. The program NAC-CESS [21] is also wrapped into mimox.pl and used to calculate the surface accessibility of the mapping results. All mapping results are ranked based on their solvent accessible surface. Each mapping result has detailed information of the accessibility of each candidate residue, which is parsed through the script parsa.pl and displayed as a table in a new window. Each mapping result can also be viewed interactively on the antigen structure. This is implemented as the script jmol.pl, which wraps a Java applet version of Jmol [22] . The overall architecture of MIMOX is shown schematically in Figure 1 . As described above, MIMOX wraps ClustalW to align a set of mimotope sequences and then allows the alignment to be viewed, edited, and analyzed through an embedded version of JalView. Based on the review by Smith et al [3] , we also implemented a simple statistical method in the script mimocs.pl to derive a consensus sequence from the alignment. Firstly, the script counts the appearance of each kind of amino acid at each position in the alignment and calculates the percentage frequency of each one. The frequency of a given amino acid X at the position i of the alignment (f xi ) is defined as where Xi means the times that the given amino acid X appears at the position i of the alignment and N is the number of sequences in the alignment. All frequencies are compared to a threshold value, which is 25% by default. If a frequency is more than the threshold, the corresponding residue is considered as a motif residue at that position. If the sum frequency of similar residue at the same position is above the threshold, the similar residues are also regarded as motif residues. In MIMOX, there are five similar residue groups (L, I, V; T, S; E, D; Q, N; K, R; F, W); other residues are considered unique. This classification scheme is the same as that used by Mapitope [16, 17] . If no motif residue is found at a given position, then X is used to stand for any amino acid residue. Motif residues at all positions of the alignment are then displayed in a table. Overall architecture of MIMOX Figure 1 Overall architecture of MIMOX. MIMOX has two sections. The first section has 3 perl scripts. The script mimosa.pl aligns a set of mimotope sequences powered by ClustalW. The script jalviews.pl wraps JalView to view and manage the alignment. The script mimocs.pl derives a consensus sequence from the alignment. The second section also has 3 perl scripts. The script mimox.pl maps the user supplied sequence on to the given antigen structure and utilizes NACCESS to calculate the accessibility. The script parsa.pl displays the detailed accessibility information of each mapping result. The script jmol.pl wraps Jmol to view the mapping result interactively on the antigen structure. Thus a consensus sequence is suggested by the program. The script also creates a 3D bar figure based on the statistical analysis above, where the X axis represents the 20 amino acid types and gap type, Y axis is the frequency and Z axis stands for the position of the aligned sequences (Shown in Figure 2 ). Since the mimotopes and the native epitope on the antigen bind to the same antibody, it is assumed that the mimotopes and the native epitope have similar physicochemical properties and similar spatial organization. This assumption is the basis of the MIMOX algorithm. The mapping process of MIMOX is based on the input sequence (such as the consensus sequence) and the uploaded antigen structure. A fragment of the sequence can also be used as input. Firstly, for each position in the input sequence, MIMOX searches the uploaded PDB structure for matching residues and places them into an array of candidate residues for that position. Two matching modes are available at present. One is strict mode, which means the type of mimotope residue must match the antigen residue exactly. The other is called conservative mode, which means similar residues are also included in the candidate residue array. There are 5 groups of similar residues (L, I, V; T, S; E, D; Q, N; K, R; F, W) in MIMOX, which has been described in previous section. Web interface of MIMOX section 1 Figure 2 Web interface of MIMOX section 1. Mimotopes selected out with trastuzumab [10] are input and aligned with wrapped Clus-talW. The frequency of a given amino acid at each position of the alignment is calculated and displayed in a table. A 3D bar figure is also created, where the X axis represents the 20 amino acid types and gap, Y axis is the frequency and Z axis stands for the position of the aligned sequences. A consensus sequence is then suggested, which can be used in further mapping with MIMOX. The alignment can also be managed with the embedded JalView [20] . The array of candidate residues for each position is then added to an array of arrays. MIMOX finds all the residue neighbour pairs between consecutive candidate residue arrays in the array of arrays. Whether two residues are neighbours is determined by the distance between the two residues and the distance threshold value. If the distance between two residues is below the threshold, the two residues are taken as a neighbour pair. MIMOX provides three methods to calculate neighbour residue pairs. One method is to take the distance between the Cα atoms of the two amino acids as the distance between the two residues. Using Cα atoms may better reflect the backbone positions. The second method is to use the distance between the Cβ atoms, which may better reflect the side chain position (Cα atom is still used when it is a glycine because it does not have a Cβ atom). The third method described below, is based on the distances between all the heavy atoms of the two amino acids. All of the distances mentioned above are Euclidean distances, calculated as: where D 21 means the distance between atom 2 and atom 1 and x 2 , y 2 , z 2 , x 1 , y 1 , z 1 are coordinates of atom 2 and atom 1. When the methods based on Cα or Cβ atom position are used, the default distance threshold is 7.0 angstroms, as it approximates the upper limit for noncovalent interactions in macromolecular structures [23] . When the third method is used, the distance threshold is calculated as where DT is the distance threshold, DF is the Distance Factor (given by user), and vdwAtom is the Van der Waals radius of the atom. The default DF value is 1.11. If two residues have a pair of heavy atoms which are nearer than the distance threshold calculated from their Van der Waals radius, the two residues are taken as a neighbour pair. MIMOX then recursively links the neighbour pairs until all possible ways of forming the input sequence are made. Each result is then ranked according to the sum of the absolute residue accessibility of each residue calculated from the NACCESS result file. In the end, the results are displayed in a table with hyperlinks to call the script parsa.pl which can parse and display the accessibility data in detail, and the script jmol.pl to view the result interactively mapped onto the antigen structure. Web interface of MIMOX MIMOX has successfully been implemented as an online service, which has a simple web interface both for input and output. As described previously, MIMOX can be divided into two sections; we show here the input and output of the two sections in Figure 2 and Figure 3 respectively. To test MIMOX, we have applied it to several cases taken from other similar research and literature. We compared the results from MIMOX with the results from other computational tools and the native epitope itself if the epitope is known in the CED database [24] . It should be pointed out that cases using monoclonal antibodies are most appropriate for testing [15] . However, in order to compare with previously published tools, some less appropriate cases (using polyclonal antibodies) taken from the corresponding literature are also used. More case studies [see Additional file 1] can also be found on the test dataset page of MIMOX[25]. The first case is taken from FINDMAP [13] . In 1999, Jesaitis and co-workers used an anti-actin polyclonal antibody to select a phage displayed random peptide library; VPHPTWMR was one of the consensus sequences they derived from the selected mimotopes. They manually mapped VPHPTWMR to the known structure of actin [PDB: 1ATN] and suggested that it might correspond to residues: V129, P130, H101, P102, T358, W356, M355, R372 [11] . In 2003, Mumey et al used FINDMAP to align VPHPTWMR to the actin sequence without utilizing information on the antigen structure. The result from FIND-MAP shows VPHPTWMR can be mapped to residues as V129, P130, H101, P102, T103, W356, M355, R372 [13] . FINDMAP mapped the input sequence to a slightly different set of residues (using T103 instead of T358). When running MIMOX with all parameters as defaults, we got no result. However, after the distance threshold is changed to 12 Å (the maximum distance allowed in MIMOX), we find that the two mappings above are returned as candidate cluster 5 and candidate cluster 17. As the side chain of some amino acids (such as arginine) can span a distance as great as 12 Å, MIMOX takes this value as the maximum allowable distance. This distance restriction is also used by Mapitope [16, 17] . In this case, the need for the higher distance threshold is due to R372 which lies some distance from the other mapped residues. MIMOX also suggested other possibilities such as cluster 1(V96, P102, H101, P130, T358, W356, M355, R372) which has a bigger solvent accessible surface, and cluster 26 (V96, P98, H101, P102, T103, W356, M355, R372), which clearly has 3 sequential segments, i.e. VPHPT, WM, and R. The second case is taken from work by Enshell-Seijffers [16] . They used monoclonal antibody 17b, which is against HIV gp120 envelope glycoprotein, to select a phage displayed random peptide library and got a set of Figure 4 . As the latter two mapping results suggested by MIMOX are more exposed, they might be able to bind to the antibody more easily. The last case is taken from MIMOP [15] . BO2C11 is a human monoclonal antibody against human coagulation factor VIII. Villard et al selected two phage displayed random peptide libraries with BO2C11 and got a set of 27 mimotopes [26] . Very recently, Moreau et al have applied their newly developed tool MIMOP to analyze these mimotopes. Combining the two methods MimAlign and Mim-Cons in MIMOP, the BO2C11 epitope is predicted be composed of a segment YFTNMF (2195-2200) and residues T2202, K2207, R2215, R2220, Q2222. The structure of human coagulation factor VIII in complex with Comparison of three mapping results. MIMOX was used to map LLTTNKD to HIV gp120 with three different methods. Taking together, our initial case studies show that MIMOX can fully or partially repeat results from manual mapping, other existing tools, and also provide novel suggestions. MIMOX is designed to be a tool which is more interactive than automatic. We acknowledge that tuning the probe sequences and parameters are often required to get good results. This interactive process gives hints to users step by step and greatly decreases the load of the server and prevents the loss of some reasonable results. MIMOX lists all the matched results with no prediction threshold. This allows users to find the reasonable results by themselves based on their background knowledge on a given antibody, a given antigen and a given phage display experiment. Nevertheless, according to the test dataset page of MIMOX, the true epitope (or its segments) often falls in the top 5 (if the there are only a few result entries) or top 10% (if the there are many result entries) of the results. Where the real epitope is unknown, we would suggest running MIMOX with a range of parameters and consensus sequence derived fragments to find overlapping or otherwise promising (high surface accessibility) candidate. As we have mentioned previously, several groups have researched algorithms and programs that may assist and automate phage display based epitope mapping. Based on the dependency on antigen structure, the existing programs can be classified into three categories. FINDMAP belongs to the first category, which is independent of any structural information. FINDMAP has been implemented as a C++ program. It aligns a probe (e.g. a consensus sequence derived from a set of mimotopes) to the sequence of native antigen, allowing any permutation of the probe sequence. It uses a two-part scoring system to evaluate the quality of alignments and a branch-andbound algorithm to find an alignment with maximum score [13] . The programs in the second category include SiteLight, 3DEX, Mapitope, and MIMOX. SiteLight was implemented in C++ and it has been tested on Red Hat Linux. First, the program divides native protein surface into overlapping patches based on geodesic distances between residues; then aligns each mimotope in the library with each patch and scores and sorts them; finally, high scoring matches are selected iteratively until 25% of the native protein is covered [12] . Another program 3DEX was implemented in Visual Basic and could only run on Windows. It divides a sequence into a set of overlapping subsequences with a user-defined length (3-maximum length of mimotope). Then, it searches for matching residues at each position of the above subsequences against the sequence or PDB structure of native protein and links the neighbours iteratively until the first subsequence is complete. This is repeated for the following subsequences to complete the mimotope and return the result [14] . Mapitope was also implemented in C++ and its algorithm was first described by Enshell-Seijffers in 2003. Briefly, Mapitope deconvolutes a set of mimotope sequences into a set of overlapping amino acids pairs (AAP). Then a set of major statistically significant pairs (SSP) are identified based on the AAP. Later, the SSP are mapped and clustered in the antigen structure. Finally, the most elaborate and diverse clusters on the antigen surface are identified and regarded as the predicted epitope candidates [16, 17] . [15] . It seems that MIMOP can work with or without the antigen structure from the published case studies. However, the sequence of the only case that is independent of antigen structure is just a continuous subsequence of the antigen sequence. Thus, more studies are still needed to prove that MimAlign can work without antigen structure information. All the existing programs described above have succeeded in given cases. However, a systematic evaluation on these tools is absent. Moreover, as shown in the Table 1 , most of the existing tools have not been implemented as a publicly accessible online service until now, making it less convenient for the community to access, utilize, and evaluate them. Like all software, bugs will have crept into MIMOX during the programming. We expect users will send their feedback to help us maintain and improve MIMOX in the future. A new version of MIMOX with more user definable options and supporting multiple-chain antigens will be implemented in the future, allowing epitopes formed by residues from different polypeptide chains to also be predicted. A systematic evaluation and comparison study with all available tools including MIMOX that assist phage display based epitope mapping is also under our consideration. MIMOX, a web application for phage display based epitope mapping has been coded with Perl. It is helpful for molecular biologists to identify the native epitope of an antibody based on the antigen structure and a set of mimotope sequences they get through phage display technology. As a publicly accessible web tool in this area, MIMOX is very convenient for the community to access, utilize, and evaluate, complementing other existing programs.
65
Endogenous Cell Repair of Chronic Demyelination
In multiple sclerosis lesions, remyelination typically fails with repeated or chronic demyelinating episodes and results in neurologic disability. Acute demyelination models in rodents typically exhibit robust spontaneous remyelination that prevents appropriate evaluation of strategies for improving conditions of insufficient remyelination. In the current study, we used a mouse model of chronic demyelination induced by continuous ingestion of 0.2% cuprizone for 12 weeks. This chronic process depleted the oligodendrocyte progenitor population and impaired oligodendrocyte regeneration. Remyelination remained limited after removal of cuprizone from the diet. Fibroblast growth factor 2 (FGF2) expression was persistently increased in the corpus callosum of chronically demyelinated mice as compared with nonlesioned mice. We used FGF2(−/−)mice to determine whether removal of endogenous FGF2 promoted remyelination of chronically demyelinated areas. Wild-type and FGF2(−/−)mice exhibited similar demyelination during chronic cuprizone treatment. Importantly, in contrast to wild-type mice, the FGF2(−/−)mice spontaneously remyelinated completely during the recovery period after chronic demyelination. Increased remyelination in FGF2(−/−)mice correlated with enhanced oligodendroglial regeneration. FGF2 genotype did not alter the density of oligodendrocyte progenitor cells or proliferating cells after chronic demyelination. These findings indicate that attenuating FGF2 created a sufficiently permissive lesion environment for endogenous cells to effectively remyelinate viable axons even after chronic demyelination.
In central nervous system (CNS) demyelinating diseases such as multiple sclerosis (MS), myelin damage impairs impulse conduction along denuded axons. Limited remyelination occurs in MS lesions (1Y3) but is typically insufficient to prevent long-term neurologic disability. After a demyelinating event, improved remyelination could maximize functional recovery of viable axons and prevent associated axonal damage and degeneration. In rodent models, extensive spontaneous remyelination and functional recovery is possible after an episode of acute transient demyelination (4, 5) . Proliferation of immature cells and differentiation into myelinating oligodendrocytes is required for this extensive remyelination (6, 7) . Similarly, immature oligodendrocyte lineage cells persist in the adult human CNS and may proliferate to increase in number within and near MS lesions (8, 9) . However, these immature cells often fail to differentiate sufficiently to remyelinate throughout the extent of MS lesions. Therefore, the repair capacity of endogenous cells may be limited by nonpermissive signals in chronic MS lesions. Growth factors, cytokines, and cell adhesion molecules may regulate oligodendrocyte progenitor (OP) differentiation into remyelinating oligodendrocytes in the environment of a demyelinated lesion. Among these potential signaling molecules, we have examined the in vivo role of endogenous fibroblast growth factor 2 (FGF2), which is upregulated during postnatal development and in acute demyelination (10Y 12). During remyelination after acute demyelination, FGF2 null mice exhibit enhanced oligodendrocyte repopulation of demyelinated lesions (12) . This improved oligodendrocyte regeneration was further studied using in vivo retroviral lineage analysis in wild-type versus FGF2 null mice, which demonstrated that the predominant effect of FGF2 in vivo during remyelination is inhibition of OP differentiation (11) . The current study examines whether this FGF2 effect on OP differentiation and oligodendrocyte regeneration has a significant impact on remyelination. Acute demyelination models in rodents typically exhibit spontaneous remyelination that may be so robust as to prevent appropriate evaluation of strategies for improving remyelination. Therefore, the current study challenges the capacity of endogenous cells to regenerate oligodendrocytes and remyelinate chronically demyelinated lesions. We induced active demyelination in mice by adding 0.2% cuprizone to the diet so that return to normal chow would allow analysis of spontaneous remyelination without ongoing disease pathogenesis. Cuprizone reproducibly demyelinates the corpus callosum of mice (13) . Spontaneous remyelination is greatly reduced after chronic cuprizone demyelination in contrast to acute demyelination (14 Y 16) . Importantly, after chronic cuprizone demyelination, axons remain viable and can be remyelinated by transplanted OP cells (16) . In this chronic demyelination model, the current study demonstrates dramatically improved remyelination by endogenous cells in mice with genetic deletion of FGF2 as compared with wild-type mice. Thus, FGF2 expression in chronic lesions may limit remyelination and attenuation of FGF2 may generate a sufficiently permissive environment for spontaneous remyelination by endogenous cells. Mice were bred and maintained in the Uniformed Services University of the Health Sciences (USUHS) animal housing facility and all procedures were performed in accordance with guidelines of the National Institutes of Health, the Society for Neuroscience, and the USUHS Institutional Animal Care and Use Committee. FGF2 knockout mice on the 129 Sv-Ev:Black Swiss genetic background were obtained from breeding heterozygous pairs (provided by Dr. Doetschman, University of Cincinnati). This FGF2 knockout was generated by a targeted deletion replacing a 0.5-kb portion of the FGF2 gene, including 121 bp of the promoter and the entire first exon with an Hprt minigene (17) . Cuprizone ingestion results in a reproducible pattern of extensive corpus callosum demyelination (12, 13) . Cuprizone treatment was started at 8 weeks of age and only male mice were used. Cuprizone (0.2% (w/w), finely powdered oxalic bis(cyclohexylidenehydrazide) (Sigma-Aldrich, St. Louis, MO), was thoroughly mixed into chow (diet TD.01453; Harlan Teklad, Madison, WI), which was available ad libitum. Mice were maintained on the cuprizone diet until perfused for analysis or returned to normal chow after 6 weeks or 12 weeks of cuprizone ingestion. Mice were perfused with 4% paraformaldehyde and then brains were dissected before overnight postfixation in 4% paraformaldehyde (18) . Brain tissue was cryoprotected overnight at 4-C in 30% sucrose and frozen in OCT compound for immunostaining and in situ hybridization. In situ hybridization and preparation of digoxigeninlabeled riboprobes were performed as previously detailed (18, 19) . Antisense riboprobes were used to detect mRNA transcripts for proteolipid protein (PLP; gift from Dr. Lynn Hudson; National Institutes of Health [20] ) and PDGF>R (gift from Dr. Bill Richardson; University College London [21] ). The digoxigenin-labeled riboprobes were hybridized to 15-Km-thick coronal brain sections. Digoxigenin was detected with an alkaline phosphatase-conjugated sheep antidigoxigenin antibody (Boehringer Mannheim, Indianapolis, IN) followed by reaction with NBT/BCIP substrate (DAKO, Carpinteria, CA). To identify OPs in situ, 15-Km coronal brain sections were immunostained for NG2 and PDGF>R (12, 19) . Primary antibodies used were rabbit polyclonal antiNG2 antibody (gift from Dr. William Stallcup, La Jolla, CA) and rat monoclonal antiPDGF>R antibody (APA5; Pharmingen, San Diego, CA). Donkey antirabbit IgG F(ab') 2 conjugated with Cy3 (Jackson Immunoresearch, West Grove, PA) was used to detect NG2, whereas the PDGF>R was detected with biotinylated donkey antirat IgG F(ab') 2 (Jackson Immunoresearch) followed by coumarin tyramide amplification (New England Nuclear, Boston, MA). Myelin was immunostained with monoclonal antibody 8-18C5, which recognizes myelin oligodendrocyte glycoprotein (MOG). Hybridoma cells were provided by Dr. Minetta Gardinier, University of Iowa (22) . MOG immunolabeling was detected with donkey antimouse IgG F(ab') 2 conjugated with Cy3 (Jackson Immunoresearch). Cell proliferation was estimated with immunostaining for Ki-67 antigen, which is expressed in the nuclei of actively dividing cells but absent at G 0 (23). Ki-67 was recognized with a rat monoclonal antibody to mouse Ki-67 antigen (DAKO) followed by detection with the ABC elite kit using 3,3-diaminobenzidine (DAB) as a substrate (Vector Labs, Burlingame, CA). Images of in situ hybridization and immunostaining results were captured with a Spot 2 CCD digital camera using Spot Advanced image acquisition software (Diagnostic Instruments, Sterling Heights, MI) on an Olympus IX-70 microscope. Images were prepared as panels using Adobe Photoshop (San Jose, CA). For comparing cell densities, all quantification was performed by an investigator blinded to the treatment condition. Cells expressing PLP mRNA were quantified using unbiased stereologic morphometric analysis (12) (Stereologer System from Systems Planning and Analysis, Inc., Alexandria, VA). Analysis was restricted to the corpus callosum region from the midline and extending laterally to below the cingulum in 15-Km-thick coronal sections. Using the Stereologer System, the specimen thickness contributes to the sampled volume so that measurements reflect cells/mm 3 . The unbiased stereologic method could not be used appropriately for conditions with relatively few cells of interest in any chosen category. Therefore, quantification of cells in the corpus callosum expressing PDGF>R or Ki-67 required counting all labeled cells and using the Spot 2 CCD camera and software to measure the area sampled, resulting in density units of cells/mm 2 (12) . Quantification of corpus callosum myelination was estimated from MOG immunofluorescence detected with a Spot 2 CCD camera. Using Metamorph software, pixel intensity values were normalized between sections by thresholding to exclude values below the level of immunoreactivity in the dorsal fornix, which was selected as an adjacent white matter tract that is not demyelinated by cuprizone. The percent area of the corpus callosum (midline bilaterally to a point under the cingulum apex) with MOG immunoreactivity above the threshold level was then used as an estimate of the myelinated area. Each category analyzed included three or more tissue sections per mouse and three or more mice per condition, except where larger sample sizes are noted in text and/or figure legends. One-way analysis of variance (ANOVA) with post hoc Tukey's multiple comparison test was used to determine significant differences among stages of disease progression or treatment. Unpaired Student t-test was used to compare between FGF2 genotypes in nonlesioned mice. Significance of an FGF2 genotype effect across multiple treat-ment conditions was calculated using a two-way ANOVA. No statistical comparisons were made between mice with different genetic backgrounds (i.e. C57Bl/6 mice and FGF2 mice). Chronic Cuprizone Demyelination Provided a Relevant Model of Insufficient Remyelination to Characterize the Repair Capacity of Endogenous Cells C57Bl/6 mice were used to establish parameters for analyzing, and later manipulating, the capacity of endogenous FIGURE 1. Spontaneous remyelination of the corpus callosum was compromised after chronic demyelination of C57Bl/6 mice. (A) Corpus callosum myelination was estimated by immunofluorescence for myelin oligodendrocyte glycoprotein (MOG). Pixel intensity values were normalized between tissue sections by thresholding to exclude values below the immunoreactivity in the dorsal fornix, which was not demyelinated by cuprizone (see DF in [C]). The percent area of the corpus callosum with MOG immunoreactivity above the threshold level was then used as an estimate of myelinated area. Bar colors indicate treatment conditions: no cuprizone (white), acute cuprizone (up to 6 weeks, gray), or chronic cuprizone (9 weeks and above, black). At least three sections were quantified per mouse and at least four mice per condition. After acute cuprizone, myelination returned to near nonlesioned values (p > 0.05; no cuprizone, n = 4; 6 weeks cuprizone 6 weeks off, n = 5). Recovery to nonlesioned levels did not occur after chronic cuprizone (p G 0.001; 12 weeks cuprizone 6 weeks off; n = 6). Values for 12-week cuprizone three off were not significantly different than either 12-week cuprizone or 12-week cuprizone 6 off. Significant differences, p G 0.05 or less, are noted by an asterisk (*) for comparisons with no cuprizone. oligodendrocyte lineage cells to repopulate and remyelinate chronically demyelinated lesions. We used the cuprizone model to take advantage of the ability to stop active demyelination simply by returning the mice to a normal diet. As characterized in C57Bl/6 mice, continuous cuprizone ingestion results in persistent demyelination with a variable degree of partial remyelination, indicating a regenerative potential of endogenous cells even while mice are still fed cuprizone (24) . However, after 12 weeks of continuous 0.2% cuprizone ingestion in C57Bl/6 mice, spontaneous remyelination remained limited through the 6-week recovery period (Fig. 1 ). This poor remyelination after chronic demyelination contrasted with the almost complete remyelination seen within the same recovery period after acute demyelination (i.e. 6 weeks on cuprizone followed by 6 weeks off cuprizone; Fig. 1 ). To better evaluate the underlying causes of limited remyelination after chronic demyelination, we characterized the oligodendrocyte lineage cell responses relative to disease progression with a focus on the repair potential. Oligodendrocyte and OP populations were characterized during multiple stages of disease progression: nonlesioned (no cuprizone), acute demyelination (3Y5 weeks of continuous cuprizone), chronic demyelination (12 weeks of continuous cuprizone), and recovery after chronic demyelination (12 weeks of continuous cuprizone followed by normal chow for 3 or 6 weeks). In situ hybridization was used to identify oligodendrocytes expressing PLP mRNA transcripts (Fig. 2) and OP cells expressing FIGURE 2 . Oligodendroglial repopulation of the corpus callosum was compromised after chronic demyelination of C57Bl/6 mice. (A) Quantification of the density of oligodendrocytes in the corpus callosum of C57Bl/6 mice. In situ hybridization for PLP mRNA, which identifies premyelinating and myelinating oligodendrocytes, was quantified using unbiased stereology with at least three sections per mouse and at least four mice per condition. Bar colors indicate no cuprizone (white), acute cuprizone (gray), or chronic cuprizone (black). The oligodendrocyte density after 12 weeks of cuprizone followed by 6 weeks for recovery was still significantly below nonlesioned values (p > 0.05; n = 5 for both conditions). Significant differences, p G 0.05 or less, are noted by an asterisk (*) for comparisons with no cuprizone and with a carrot (^) for comparisons of recovery stages with 12 week cuprizone. platelet-derived growth factor > receptor (PDGF>R) mRNA transcripts (Fig. 3) . Quantitative analysis of the oligodendrocyte population revealed ongoing partial regeneration during and after chronic demyelination (Fig. 2) . The oligodendrocyte density was lowest after the initial 3 weeks of cuprizone in the acute phase. After 12 weeks of continuous cuprizone, the oligodendrocyte density continued to be severely reduced relative to values from nontreated mice. Ongoing partial regeneration during chronic demyelination was indicated by the fact that the oligodendrocyte density after 12 weeks was higher than after the initial 3 weeks. After ending the chronic cuprizone treatment, this oligodendrocyte regeneration improved during the recovery period. Interestingly, as shown in Figure 1 , remyelination did not progress similarly during this 6-week recovery period. The PLP mRNA in situ hybridization used to identify oligodendrocytes should detect both premyelinating and myelinating oligodendrocytes because PLP transcription precedes myelin formation (25) . These results indicate that after chronic demyelination, a proportion of premyelinating oligodendrocytes that are generated during the recovery period may fail to differentiate into myelinating oligodendrocytes. The OP population dynamics were dramatically different in response to acute versus chronic demyelination (Fig. 3) . In response to the initial episode of acute demyelination, the OP population was amplified almost fourfold (Fig. 3) . However, by the end of the chronic demyelination period, the OP density was greatly reduced and remained low during the recovery period. In the chronic lesions, similar results were observed with OP identification using PDGF>R (Fig. 3A ) and using NG2 (12 weeks cuprizone = 155 cells/mm 2 , n = 5; 12 weeks cuprizone plus 3 weeks recovery Figure 1 ) of corpus callosum myelination estimated by immunostaining for myelin oligodendrocyte glycoprotein (MOG) in cuprizone-treated FGF2 j/j mice (A) and FGF2 +/+ mice (B). The corpus callosum showed persistent demyelination with 0.2% cuprizone ingestion throughout 9 weeks and 12 weeks. After 12 weeks of cuprizone treatment, on return to normal chow, remyelination in FGF2 j/j mice significantly improved (p G 0.001; 12 weeks cuprizone, n = 4; 12 weeks cuprizone 6 weeks off, n = 4) and recovered to nonlesioned levels (p > 0.05; no cuprizone, n = 5; 12 weeks cuprizone 6 weeks off). In contrast, values in FGF2 +/+ mice did not increase after chronic demyelination (p > 0.05; 12 weeks cuprizone, n = 4 vs. 12 weeks cuprizone 6 weeks off, n = 3) and remained significantly below nonlesioned levels (p G 0.001; no cuprizone, n = 5 vs. 12 weeks cup). = 169 cells/mm 2 , n = 6; 12 weeks cuprizone plus 6 weeks recovery = 115 cells/mm 2 , n = 5). At 12 weeks of cuprizone treatment, very few cells in the corpus callosum were identified as actively undergoing mitosis using DAPI counterstaining to reveal mitotic chromatin figures, but each actively dividing cell was also immunostained for NG2 (inset, Fig. 3A ). Immunostaining for Ki-67 nuclear antigen was used to more broadly identify actively cycling cells within the corpus callosum (Fig. 3B) . Ki-67-labeled cycling cells were present at less than half the density of OP cells. This cycling population, although still relatively small, was significantly increased within the corpus callosum of mice fed cuprizone for 12 weeks as compared with age-matched controls or with mice that had recovered for 6 weeks after chronic demyelination. Together, these findings demonstrated compromised but continued OP cycling and regeneration of oligodendrocytes during chronic demyelination and the subsequent recovery period (Figs. 2, 3 ). FGF2 expression is increased during acute cuprizone demyelination and can inhibit OP differentiation and oligodendrocyte regeneration during subsequent remyelination (11, 12) . Using in situ hybridization for FGF2 mRNA transcripts, we have shown that elevated FGF2 expression persists within chronic cuprizone lesions, as compared with nonlesioned corpus callosum (Fig. 4) . FGF2 mRNA transcript abundance related to changes in FGF2 protein detection in developing white matter and after acute demyelination (10, 19) . We predicted that in chronic lesions, FGF2 inhibition of differentiation could contribute to the limited capacity to generate remyelinating oligodendrocytes (Figs. 1, 2) , especially in the context of a reduce OP pool (Fig. 3A) . Although the endogenous populations did not efficiently remyelinate after chronic cuprizone demyelination in C57Bl/6 mice (Fig. 1) , transplantation of OP cells has shown that the axons remain viable for remyelination (16) . Therefore, we used this chronic demyelination model in FGF2 FIGURE 6 . Oligodendroglial repopulation of the corpus callosum after chronic demyelination was dramatically improved in FGF2 j/j mice as compared with FGF2 +/+ mice. (A) Quantification of the density of oligodendrocytes in the corpus callosum of FGF2 mice. In situ hybridization for PLP mRNA, which identified premyelinating and myelinating oligodendrocytes, was quantified using unbiased stereology with at least three sections per mouse and at least four mice per condition. The oligodendrocyte density after 12 weeks of continuous cuprizone was significantly below nonlesioned values (noted by asterisks [*]; p G 0.01 for FGF2 +/+ ; p G 0.05 for FGF2 j/j mice; n = 4 for each condition). After 12 weeks of cuprizone and a subsequent 3-and 6-week period for recovery on normal chow, oligodendrocyte densities in FGF2 j/j mice were no longer significantly different as compared with nonlesion values (p > 0.05 at 3 weeks and at 6 weeks; n = 4 for each condition). Comparison of the FGF2 +/+ values with the matching conditions in FGF2 j/j mice demonstrated a significant effect of genotype (p = 0.0196). knockout mice to test whether repair from the endogenous OP population might be improved by removing endogenous FGF2 from the lesion environment. The extent of demyelination after 12 weeks of cuprizone ingestion was similar in C57Bl/6 mice and mice of both FGF2 wild-type (FGF2 +/+ ) and null (FGF2 j/j ) genotypes (Figs. 1, 5) . The percentage of the corpus callosum area with myelin immunolabeling was 45.8 T 14.1% in C57Bl/6 mice, 50.05 T 7.8% in FGF2 +/+ mice, and 46.9 T 4.3% in FGF2 j/j mice. After the chronic demyelination, only in FGF2 j/j mice did myelin immunostaining in the corpus callosum significantly increase during the recovery period (Fig. 5) . Surprisingly, the myelin immunostaining recovered to approximately nonlesion levels by 6 weeks after removal of cuprizone from the diet of FGF2 j/j mice. In contrast, in C57Bl/6 mice (Fig. 1) or FGF2 +/+ mice (Fig. 5) , the myelinated proportion of corpus callosum area did not significantly increase during the recovery period. This striking improvement in remyelination corresponded with a dramatic increase in oligodendrocyte repopulation of chronic lesions in FGF2 j/j mice, which was not found in FGF2 +/+ mice (Fig. 6) . Before cuprizone treatment, nonlesioned FGF2 +/+ and FGF2 j/j mice had a similar density of oligodendrocytes in the corpus callosum at 26 weeks of age, which corresponded with the longest cuprizone treatment and recovery protocol. In addition, the cuprizone-induced loss of oligodendrocytes in FGF2 +/+ and FGF2 j/j mice was similar at the end of the chronic cuprizone treatment. In conjunction with the quantitative myelin immunostaining analysis, these findings indicate that FGF2 +/+ and FGF2 j/j mice experienced a similar chronic disease severity. Importantly, during the 6-week recovery period after chronic demyelination, the oligodendrocyte density returned to nonlesioned values in the FGF2 j/j mice, but not in FGF2 +/+ mice. Analysis of an additional intermediate recovery period in FGF2 j/j mice confirmed this increased oligodendrocyte regeneration and demonstrated that this repopulation could occur within 3 weeks after removal of cuprizone from the diet. We previously showed that the absence of FGF2 did not significantly alter OP amplification in response to acute cuprizone demyelination (12) . A robust OP proliferative response occurred after 5 weeks of cuprizone in FGF2 null and wild-type mice that was similar to the previously mentioned data for C57Bl/6 mice (Fig. 3) . However, the OP population was depleted during chronic demyelination in contrast to the amplified OP response to acute demyelination in C57Bl/6 ( Fig. 3) . Therefore, we examined the effect of FGF2 genotype on the density of OP cells and cycling cells in the corpus callosum after chronic demyelination. FGF2 genotype did not appear to alter the OP population dynamics of chronically demyelinated mice or age-matched nonlesioned mice (Fig. 7) . The density of OP cells, identified by in situ hybridization for PDGF>R mRNA, was similar in FGF2 +/+ mice and FGF2 j/j mice (Fig. 7A) . Somewhat unexpectedly, mice of both FGF2 null and wild-type genotypes, which are on a 129 Sv-Ev:Black Swiss genetic background, exhibited an increase in OP cell density after 12 weeks of cuprizone that was not observed in C57Bl/6 mice (Fig. 3A ). This strain difference in OP accumulation after chronic demyelination corresponded with fewer oligodendrocytes observed during the recovery period in FGF2 +/+ mice (Fig. 6 ) as compared with C57Bl/6 mice (Fig. 2) . Therefore, the FGF2 background appears to be even less favorable than the C57Bl/6 background for OP differentiation and oligodendrocyte regeneration, yet the detrimental effects of the chronic lesion environment can still be overcome in the FGF2 j/j mice. Ki-67 immunostaining (Fig. 7C ) indicated that the density of cycling cells in the corpus callosum was less than 20% of the density of OP cells, indicating a relatively low level of ongoing proliferation. Cells with nuclear Ki-67 immunoreactivity were often observed as doublets, as is appropriate for confirming detection of a cycling population. The density of cells immunolabeled for Ki-67 nuclear antigen increased in chronically lesioned corpus callosum with a return to nonlesioned levels during the recovery phase. The density of endogenous cells that were proliferating in the corpus callosum of FGF2 j/j mice was similar to the values for FGF2 +/+ mice (Fig. 7B ) and for C57Bl/6 mice (Fig. 3) . The similarity of FGF2 j/j mice and FGF2 +/+ mice in our analysis of OP cells and cycling cells indicates that the dramatic differences observed in oligodendrocytes and remyelination may result from a permissive effect of FGF2 removal on OP differentiation into remyelinating oligodendrocytes. In MS lesions, an initial episode of transient demyelination may be followed by spontaneous remyelination. However, remyelination typically fails with recurring or chronic myelin damage. The current studies show the advantages of the cuprizone model of chronic demyelination for focusing on this compromised repair response. Importantly, our analysis in FGF2 j/j mice demonstrates that it is possible to overcome chronic lesion effects on endogenous oligodendrocyte lineage cells to increase remyelination. More specifically, taken together with our previous work (11, 12) , our findings in FGF2 j/j mice indicate that removing FGF2 inhibition of OP differentiation promotes spontaneous remyelination from a depleted pool of endogenous progenitors that persists after chronic demyelination. The current findings in C57Bl/6 mice are consistent with previous reports of cuprizone treatment causing extensive demyelination of the corpus callosum in this mouse strain (13, 16, 24, 26) . To take further advantage of this model for examining mechanisms to promote remyelination from endogenous cells, we have further characterized the endogenous oligodendrocyte lineage population responses during the recovery period. We used C57Bl/6 mice for this part of our study to facilitate comparison with other studies, because differences in mouse strain can influence the response to cuprizone as well as overall cellular responses to CNS injury (13, 27) . After a transient episode of demyelination from 6 weeks of cuprizone ingestion, spontaneous remyelination is effective throughout the corpus callosum. In contrast, after a prolonged period of demyelination, from 12 weeks of continuous cuprizone, the remyelination remains limited and does not progress during a recovery interval that was examined out to 6 weeks. This limited remyelination corresponds with depletion of the OP pool ( [16] current study). However, the number of oligodendrocytes increases somewhat during the recovery period without a corresponding increase in remyelination of the corpus callosum. Taken together, these findings indicate that the depleted OP pool has a limited capacity to generate oligodendrocytes, and available OP cells may fail to fully differentiate into remyelinating oligodendrocytes in the environment of a chronic lesion. In lesions of some MS cases, immature oligodendrocyte progenitors and premyelinating oligodendrocytes may be present in relatively normal densities without efficiently remyelinating (9, 28) . Therefore, further studies will be important for better understanding regulation of the OP pool size and the transition of OP cells into premyelinating and then myelinating oligodendrocytes. Robust OP proliferation in response to acute demyelination appears to be required for spontaneous remyelination (29) . During acute demyelination, the OP pool is dramatically amplified so that efficient differentiation into mature oligodendrocytes may not be required for extensive remyelination. PDGF-AA ligand, acting through PDGF>R activation, is an important mitogen for this OP amplification in response to acute cuprizone demyelination (11, 30) . In PDGF>R heterozygous mice, amplification of the OP pool in response to acute cuprizone demyelination is compromised and the corresponding generation of oligodendrocytes is reduced (11) . Crossing these PDGF>R heterozygous mice to FGF2 j/j mice actually increased oligodendroglial repopulation of lesions after acute cuprizone demyelination (11) . Therefore, in both acute and chronic lesion environments, FGF2 removal may promote the generation of oligodendrocytes from OP cells during remyelination. Endogenous FGF2 has been predicted to contribute to proliferation of neonatal and adult OPs, especially when present in combination with PDGF-AA, based on in vitro studies (31Y33). In our current analysis, the densities of OP cells and Ki-67-immunolabeled cells were similar between FGF2 wild-type and null genotypes. These results are consistent with our previous BrdU incorporation studies indicating that endogenous FGF2 is not a predominant mitogen for OPs in the acute cuprizone model (11) . However, subtle changes in the proliferation rate among an asymmetrically dividing OP pool could be relatively difficult to detect in a small population of cells over a prolonged disease and recovery course. FGF2 is also predicted to stimulate neural stem cells in the subventricular zone (SVZ) to contribute to repopulation of corpus callosum lesions (34) . During the prolonged demyelination period of the current study, the contribution of cells derived from the SVZ should have been evident in the dynamics of the OP pool in the corpus callosum. The lack of a detrimental effect of FGF2 absence on the OP pool may indicate that FGF2 effects on the SVZ cells may not be a significant contribution to the remyelination of the overlying corpus callosum. However, more direct analysis of the response of cells in the SVZ is required to make this determination. In addition, endogenous elevation of FGF2 in lesions may elicit specific effects that are not replicated with in vivo methods of elevating exogenous FGF2 levels (34Y37). FGF2 is among the most effective neuroprotective growth factors in diverse models of CNS injury, including stroke, trauma, excitotoxicity, and axotomy (38, 39) . In contrast to this expected protective effect of FGF2 on neurons, removal of endogenous FGF2 actually improved regenerative parameters in lesions that involved remyelination. Improved regeneration of oligodendrocytes was observed in FGF2 j/j mice after acute cuprizone demyelination of the corpus callosum and after murine hepatitis virus strain A59 demyelination of spinal cord ( [11, 12] current study). In an example from the peripheral nervous system, FGF2 knockout mice were used to prevent the normal upregulation of FGF2 and FGF receptor (FGFR) activation associated with Schwann cells and macrophages at a site of sciatic nerve crush injury (40) . Compared with wild-type mice, the FGF2 knockout mice exhibited improved myelination and increased axon diameter during regeneration from sciatic nerve crush (40) . These results indicate that a detrimental effect of endogenous FGF2 on remyelination may outweigh FGF2 protection from acute axon damage in these lesion models. Indeed, the improved remyelination in nerves of the FGF2 knockout mice may provide significant protection of axons from damage subsequent to the initial crush injury. Correlative support for a similar role of FGF2 in a chronic autoimmune model of MS can be found in a study of neural stem cell transplantation improved into mice with experimental allergic encephalomye-litis (EAE) (41) . The transplanted mice exhibited improved remyelination, with a major contribution from endogenous cells, and had less axonal loss in lesion areas. Neural stem cell transplantation was also associated with a significant decrease in FGF2 expression and reduced astroglial scar formation in the EAE lesions. FGF2 levels may modulate scar formation in demyelinated lesions because expression of FGF2 and FGFR1 corresponds with scarring, but not nonscarring, astroglial responses after other forms of CNS injury (42) . It is not yet clear which FGFR type, or types, mediates the effect of endogenous FGF2 in the context of demyelination and remyelination. Differential effects of FGF2 at different stages of the oligodendrocyte lineage may occur through differential expression and activation of FGFR isoforms (43) . Oligodendrocyte lineage cell expression of high-affinity FGFRs and coreceptors varies with developmental stage and FGF2 exposure (44) . Furthermore, expression of multiple FGFR types is significantly increased in response to demyelination (19) . FGF2 may also have differential effects based on changing interactions with other signaling pathways that regulate oligodendrocyte lineage differentiation and myelination such as Notch1 (45, 46) and neuregulin (47, 48) . Interpreting a direct effect of FGF2 is complicated because neurons and glial cells express multiple FGFR types within and near demyelinated lesions (19) . Therefore, future studies will be required to identify the specific signaling components mediating endogenous FGF2 signaling in the in vivo context of demyelination and remyelination. In MS lesions, the pathology is heterogeneous and the effect on the oligodendrocyte lineage population varies dramatically (49) . Oligodendrocyte density is severely compromised in demyelinated lesions in some patients with MS. Further studies are imperative to optimize regenerative responses from immature oligodendrocyte lineage cell populations that have been observed in some MS lesions (9, 28) . Targeting inhibitory signals that are upregulated in lesions should be a viable strategy for therapeutics to promote differentiation as needed near lesions. FGF2 expression has been reported in reactive astrocytes of acute and chronic MS plaques (50) . In addition, in cultures from adult human brain white matter, FGF2 inhibited the differentiation of preoligodendrocytes into oligodendrocytes (51) . Strategies to promote remyelination by attenuating FGF2 inhibition of OP differentiation may take advantage of reagents to modulate FGF2 signaling, which are currently being developed for angiogenesis and cancer treatments. In our in vivo analysis, absence of FGF2 was not detrimental in the normal adult or throughout the disease process. Therefore, inhibition of FGF2 signaling may be acceptable across demyelinating disease stages, which would improve treatment feasibility in diseases such as MS that have an unstable disease course. Treatments to promote remyelination should be a valuable complement to strategies that abrogate ongoing causes of demyelination such as immunomodulatory therapies.
66
Automated identification of multiple micro-organisms from resequencing DNA microarrays
There is an increasing recognition that detailed nucleic acid sequence information will be useful and even required in the diagnosis, treatment and surveillance of many significant pathogens. Because generating detailed information about pathogens leads to significantly larger amounts of data, it is necessary to develop automated analysis methods to reduce analysis time and to standardize identification criteria. This is especially important for multiple pathogen assays designed to reduce assay time and costs. In this paper, we present a successful algorithm for detecting pathogens and reporting the maximum level of detail possible using multi-pathogen resequencing microarrays. The algorithm filters the sequence of base calls from the microarray and finds entries in genetic databases that most closely match. Taxonomic databases are then used to relate these entries to each other so that the microorganism can be identified. Although developed using a resequencing microarray, the approach is applicable to any assay method that produces base call sequence information. The success and continued development of this approach means that a non-expert can now perform unassisted analysis of the results obtained from partial sequence data.
For both surveillance and diagnostic applications, fine-scale pathogen identification and near-neighbor discrimination is important; therefore, an assay that monitors at this very specific level is desirable for many types of samples such as clinical and environmental (1) (2) (3) . To successfully use any method based on DNA or RNA detection, these assays must be coupled with large databases of nucleic acid sequence information for assay design to ensure that the desired information is provided and for the interpretation of raw data. Several well-established techniques use PCR to amplify individual target pieces of sequenced genomes to provide detection of organisms (4) . These methods can roughly be divided into approaches that target individual short sequence lengths or probes (<40 bp) and methods that examine longer probes. The advantage of using short probes is that when the uniqueness of the probe has been assured and unique primers are also selected, this method gives good specificity. This approach is capable of providing fine-scale identification of several genetically close organisms by selecting a sufficient number of probes. However, this can rapidly lead to a very large number of total probes being required to detect all organisms of interest. In addition these selected probes, which in the initial selection process were determined to be unique, are often later found to be less specific as more organisms are sequenced or are less specific under conditions that differ from the original conditions. This is particularly a problem for organisms belonging to a family with a high mutation rate and also for pathogens that have relatively few neighboring pathogens sequenced. In addition, PCR approaches focused on short unique probes are not capable of detecting the presence of new significant mutations nor can they easily resolve base sequence details. Approaches that use longer individual probes avoid many of these issues at the cost of being less specific. This issue means most of these approaches are not suitable for providing the information desired, providing impetus to this work. High-density resequencing microarrays produce variable length segments, 10 2 -10 5 bp, of direct sequence information. This target sequence falls in the longer target regime of PCR approaches but rather than being hybridized to a longer lessspecific probe on the microarray, many shorter specific probes are placed on the microarray to allow more detailed determinations from the entire PCR amplicon. This also means that the specificity of the primers used can be relaxed. They have been successfully used to detect single nucleotide polymorphism (SNP) and genetic variants from viral, bacterial and eukaryotic genomes (5) (6) (7) (8) (9) (10) (11) (12) . Their use for SNP detection has clearly established their ability to provide reliable quality sequence information. In most cases, the microarrays were designed to study a limited number of genetically similar target pathogens and for many cases, the detection methods relied only on recognizing hybridization patterns for identification (6, 9, 10, 13, 14) . Taking advantage of the sequential base resolution capability of resequencing microarrays that is required for SNP detection, resequencing has recently been successfully adapted recently using a different approach for organism identification of multiple bacterial and viral pathogens while allowing for fine detailed discrimination of closely related organisms and tracking mutations within the targeted pathogen (15) (16) . The new methodology differed from earlier work by using the resolved bases as the query of a similarity search of DNA databases to identify the most likely species and variants that match the base calls from the hybridization observed. The system was capable of testing for 26 pathogens simultaneously and could detect the presence of multiple pathogens. A software program, resequencing pathogen identifier (REPI), was used to simplify data analysis by performing similarity searches of a genetic database using basic local alignment search tool (BLAST) (17) . The REPI program used BLAST default settings and would only return sequences that might represent the hybridization if the expect value, a quantity calculated by the BLAST program that indicates the likelihood that the sequence match found would have occurred by random chance in the database, was <10 À9 . This screened out all cases that had insufficient signal; however, the final determination of what pathogen(s) was detected and to what degree discrimination was possible required manual examination of the returned results. This method successfully allowed fine discrimination of various adenoviruses and strain identifications of Flu A and B samples in agreement with conventional sampling results (15, 16) . Two important advantages of this approach were that the information was always recovered at the most detailed level possible and that it was capable of still recognizing organisms with recent mutations. This approach also maintained specificity well, as it was not dependent on the uniqueness of a few individual short probes. Although this analysis method has utility, there are several shortcomings: it is time consuming, not optimized to maximize sensitivity, has complicated results, is suitable only for an expert, and contains redundant or duplicate information. The process was time consuming because only the initial screening was handled automatically while the remaining steps required manual interpretation before the detection analysis was complete. Because a simple criterion (expect value cutoff of 10 À9 ) and non-optimized BLAST parameters were used to consider a pathogen detected, the REPI algorithm provided a list of candidate organisms but did not make a final simple conclusion or relate the results of one prototype sequence to another. Instead a manual process was used to make the final determination, but because the REPI program provided all similar results and the use of public nucleic acid databases containing redundant entries, a large amount of data was presented to a user that was not useful. In addition, with a manual process it was not possible to establish that the algorithm developed was generally applicable for any organism where nucleic acid base resolved sequence information has been provided. In this paper, we describe a new software expert system, Computer-Implemented Biological Sequence Identifier system 2.0 (CIBSI 2.0), that successfully uses resolved base sequence information from custom designed Affymetrix resequencing microarrays to provide a simple list of organisms that are detected. This algorithm addresses the most important shortcoming of previous methods by incorporating new features to completely automate pathogen identification. We have demonstrated the effectiveness of this algorithm for identification via several examples. The single program is capable of making correct decisions for all 26 pathogens contained on the Respiratory Pathogen Microarray v.1 (RPM v.1), whether detected alone or in combinations, with improved sensitivity. Although the program is currently applied to resequencing microarrays, the methodologies developed remain generally applicable. Only the first portion of the algorithm handles issues specific to microarrays while the remainder deals with sequences that are suitable for use as a query by the BLAST algorithm. In developing the general identification algorithm, we have identified and resolved issues specific to resequencing microarrays that complicate their use. Because the entire decision process for what is detected has been automated, it is straightforward to test whether the rules used to make identifications are rigorous and applicable to any pathogen. With this efficient program, resequencing based assays can provide a competitive method to test simultaneously for many possible pathogens, providing output that can be interpreted by a nonexpert. The details of the RPM v.1 design and the experimental methods have been discussed in previous work (15, 16, 18) (Lin et al., submitted for publication). Briefly, the RPM v.1 chip design includes 57 tiled regions allowing resequencing of 29.7 kb of sequences from 27 respiratory pathogens and biothreat agents. These were selected based upon clinical relevance for the population of immediate interest (United States military recruit in training) (19) (20) (21) . Partial sequences from the genes containing diagnostic regions were tiled for the detection of these pathogens. The experimental microarray data used in the present analysis were obtained using a variety of purified nucleic acid templates and clinical samples culture (throat swabs and nasal washes) using random and multiplexed RT-PCR amplification schemes (for more detail description of amplification methods see Supplementary Data). Resequencing microarrays provide base call resolution by comparing the intensities between a set of four 25mer probes that differ from each other at the same position (13th base). An amplicon or target sequence is represented by numerous overlapping probe sets. GCOSÔ software v1.3 (Affymetrix Inc., Santa Clara, CA) was used to align and scan hybridized microarrays to determine the intensity of each probe in every probe set. Base calls were made based on the intensity data of each probe set using GDAS v3.0.2.8 software (Affymetrix Inc.) which used an implementation of the ABACUS algorithm (5) . The sequences were represented in FASTA format for later analysis steps. In this paper, target pathogens are the organisms the assay was specifically designed to detect. The sets of probes that represent reference sequence selected from target pathogen genomes are referred to as a Prototype Sequence or 'ProSeq' for brevity. The set of resolved bases that result from hybridization of genomic material to a ProSeq is referred to as the hybridized sequence or 'HybSeq'. The HybSeq is split into possible subsequences or 'SubSeqs'. The CIBSI 2.0 program implemented in Perl described in this study handled a hierarchy of three tasks ( Figure 1 ): (I) Pro-Seq identification; (II) ProSeq grouping; and (III) pathogen determination. The most developed and important portion of the algorithm deals with ProSeq identification Task(I) and is handled in three important subtasks: initial filtering of individual HybSeqs into SubSeqs suitable for sequence similarity comparisons ( Figure 2 ), database querying of individual SubSeqs ( Figure 3 ) and taxonomic comparison of BLAST returns for each SubSeq (Figure 4 ). The NCBI BLAST and taxonomy databases were used for the queries and images were obtained on February 7, 2006. For the ProSeq grouping Task(II), ProSeqs were compared to determine if they supported the same identified organism. In the pathogen determination Task(III), detected organisms were compared to the list of target pathogens the assay was designed for in order to determine if any were positively detected or were possibly related close genetic near neighbors. The level of discrimination that a particular sample supported was automatically determined. An initial filtering algorithm, REPI, was developed previously (16) and the general concepts with revisions were incorporated into the current (automated detection) algorithm used in the CIBSI 2.0 program. Filtering and subsequence selection were used to remove potential biasing caused by reference sequence choice and by other sources (i.e. primers). When PCR amplification was used, microarrays were hybridized in the presence of only primers to determine locations where they resulted in hybridization. Any portions of the ProSeqs that hybridized with the primers were masked as N calls so that the HybSeq did not contain biased information. Normally the primers are designed to be outside the Pro-Seq region to minimize the interference caused by primers, and so minimize the bases to be masked. There is still the chance that some bases require masking because with the large number of primers used in the multiplex, short stretches of a ProSeq not corresponding to primer locations may still hybridize with the primers. Such regions could be removed from the reference sequences and so not appear on the microarray. However, determining such locations are a difficult and time-consuming task that for most cases is not worth the effort. The first subtask of ProSeq identification Task(I) is noted in Figure 1 and shown schematically in detail in Figure 2 . This subtask uses a procedure to examine a Hyb-Seq to find the longest possible subsequence of base calls (SubSeq) that can be submitted as a query to BLAST. It produces a group of SubSeq that contain all portions of a HybSeq that have a chance of producing a limited list of returns from a BLAST query. When a HybSeq has two regions separated by a long stretch of continuous N calls, the relational positioning of the two regions cannot be trusted and so must be sent as separate queries. In addition for shorter subsequences, the number of base calls that must be made is dependent on the length. It was also recognized that for very long sequences a longer WORD size in BLAST may be used. A detailed description of the criteria and process used for each step is contained in Supplementary Data. Upon completion, the algorithm returned to the Task(I) loop and performed the BLAST subtask. The database query subtask performed a batch similarity search of a database using SubSeq as the queries. The BLAST program used was the NCBI Blastall -p blastn version 2.12 with a defined set of parameters. The masking of low complex regions was performed for the seeding phase to speed up the query; however, low complexity repeats were included in the actual scoring. The entire nucleotide database from NCBI acquired on February 7, 2006 was used as the reference database. (Note that earlier images of the database were used during development but all experiments were rerun with the algorithm as described with the Figure 3 . Detailed schematic representation of the second subtask of ProSeq identification Task(I), organism identification for an individual SubSeq. Each SubSeq sent to BLAST returned a list of possible matches contained in a Return array that was sorted through to find best bit score/expect value pair (MaxScore). If the MaxScore was greater than MIN (10 À6 ), all returns that had this best Score were sorted into a new array Rank1. Detailed SubSeqs requirement is described in Supplementary Data. image of the database obtained on this date.) The default gap penalty and nucleotide match score were used. The nucleotide mismatch penalty, -q, parameter was set to À1 rather than the default. The results of any BLAST query with an expect value <0.0001 were returned in tabular format from the blastall program. The information about each return (bit score, expect value, mismatches, length of match) was placed in the Return{hash key}{info} hash using the SubSeq identity as the hash key for further analysis. The next subtask of ProSeq identification Task(I) carried out was the determination of SubSeq() states and is shown in Figure 3 . The BLAST algorithm gives a ranking score which can be reported as accounting for the size of the database (expect value) or not (bit score). The full taxonomic classifications of every return for a SubSeq were retrieved from the NCBI taxonomy database obtained on February 7, 2006. Using the scores and taxonomy relationships it was possible to find a reduced number of returns that had the best match with the HybSeq. These results were summarized by identifying the taxonomic class to which all the returns belonged to, 'identified organism', and a parameter that indicated how they are related to each other, 'organism uniqueness'. A detailed description of the steps is contained in Supplementary Data. After each SubSeq was examined, the algorithm moved to the next subtask, which was to determine the identified organism of the ProSeq from the SubSeq (Figure 4 ). The subsequences from the same ProSeq were only allowed to support a single 'identified organism' determination. The procedure shown in Figure 4 demonstrates the decision method used to arrive at this determination (detailed description in Supplementary Data). After the subtask covered in Figure 4 was completed, the ProSeq identification Task(I) loop continued until all ProSeqs were examined. A list of ProSeqs that had detected organisms was built up in the Result1 array. After the ProSeq identification Task(I) was completed, ProSeq grouping Task(II) (Figure 1 ) was used to examine the identified organism values listed in Result1 and grouped them together if they identified the same taxonomic class. Each entry in Result1 was examined and a new entry was created in Result2 if the identified organism did not appear in this list. The entries of Result2 represented the distinct individual organisms identified, but might still contain redundant information. When the ProSeqs were designed to detect the same organism and they all hybridized well, this grouping led to a reduction in redundant information being reported. But, when one ProSeq did not hybridize as well for a variety of possible reasons, multiple entries would appear in Result2 that actually represent hybridization from the same pathogen. This is because there is an alternative cause for the ProSeq hybridizing in this manner. This hybridization could be caused by two different but closely related organisms both being present in a sample and hybridizing to the microarray. Because we have not yet developed methods to distinguish these cases, no further reduction of the list of organisms is made for in ProSeq grouping Task(II) in cases where the level of identification varied on different ProSeq targeted for the same organism. Although it was difficult to relate results from separate Pro-Seqs to each other, it was desired to have a simple final detection decision be made in pathogen determination Task(III). The first task was specifically implemented so that information about what a ProSeq was intended to detect was not considered and the second task only minimal consideration of this was taken into account. This allowed these initial tasks to be capable of recognizing not only just positive and negative identifications of target pathogens but also cases that were indeterminate. In the final task, the algorithm considered whether the identified organisms belonged to the list of organisms the ProSeqs were designed to detect. The task would group organisms from ProSeq grouping Task(II) together that belonged to or were child classes of the taxonomic class of a target organism. The taxonomic class reported was the common taxonomic group of all the organisms. When all the ProSeqs for a pathogen hybridized well, a fine level discrimination was reported. But if one or more ProSeqs hybridized less well, the reported positive target pathogen was only identified at the level of the less detailed level. This is conservative because methods have not yet been developed to clearly discriminate mixtures of very closely related organisms causing different ProSeqs to hybridize from variable hybridization of a single organism on several ProSeqs. The results of all When there was a single SubSeq or one scored much better than the other, the ProSeq inherited the properties of that SubSeq. Detailed SubSeqs requirement is described in Supplementary Data. three tasks were reported and a more experienced user can view ProSeq grouping Task(II) results to clarify some cases. Note that organisms identified in ProSeq grouping Task(II) that only belonged to target pathogens were reported as positives. Clear negative ProSeqs were not mentioned in the output. ProSeqs that were indeterminate or that detected close genetic organisms were never reported as positives. These organisms were instead reported as being detected. A resequencing microarray (RPM v.1) was designed previously for detection and sequence typing of 20 common respiratory and 6 CDC category A biothreat pathogens known to cause febrile respiratory illness based on ProSeqs without relying on predetermined hybridization patterns (15, 16, 18 Purified Chlamydia pneumoniae nucleic acid samples with 10-1000 genome copies (via method in Lin et al., submitted for publication) were chosen to illustrate how pathogen detection and identification were done when multiple ProSeqs were targeted for the same pathogen. RPM v.1 has three highly conserved ProSeqs selected from the genes encoding for the major outer membrane proteins VD2 and VD4, and the DNA-directed RNA polymerase (rpoB) gene. The HybSeqs from the different samples differed only in the number of unique base calls as shown in Table 1 . The percentage of the ProSeq called varied from 80 to 100% except for one case at a concentration of 10 that had only 11% of the rpoB ProSeq producing unique base calls. Because the samples at this concentration are not reproducibly generating the same percentage of base calls, this is probably the detection limit of this ProSeq of the assay. Table 1 listed the determinations made for the SubSeq and at the end of each task for the various samples. The ProSeq from the different cases produced the same number of SubSeqs. These SubSeqs from different samples reported different bit scores for the same top ranked returns from BLAST. In fact VD2 and VD4 produced exactly the same results. The NCBI taxonomy database classified the returns into four distinct groups, which represented the C.pneumoniae taxonomic group and three child strain groups. AE001652, AE002167, AE017159 and BA000008 appeared in the returns of all the ProSeqs for each sample, since they represented database entries of completely sequenced genomes. One rpoB Sub-Seq produced for its organism uniqueness, SeqUniqu. All other SubSeqs were TaxAmbig as multiple returns from different taxonomic classes were returned. Since the VD2 and VD4 ProSeq each have a single SubSeq, Task(I) assigned the Pro-Seq the state of the SubSeq. For the rpoB ProSeq, the bit score of one SubSeq was large enough that the algorithm assigned that SubSeq's identification to the ProSeq. Task II of the algorithm grouped all three ProSeqs together since they all had the same identified organism and TaxAmbig was assigned. The result of Task(III) was positive for target pathogen C.pneumoniae and this decision was straightforward as all the ProSeqs agreed with each other and belonged to the same target pathogen taxonomic class. Although the rpoB ProSeq was SeqUniqu, this was not the final conclusion for Task(II) as the ProSeq that was SeqUniqu was not the child taxonomic group and other ProSeq were TaxAmbig. The three recognized strains scored the same, which indicated that the sequence selected for the ProSeqs was very conserved and would not allow discrimination between the strains. Influenza and Human Adenovirus (HAdV) were the only pathogens that had ProSeq selected that would permit detailed strain level discrimination as discussed in previous work (15, 16) . This previous work using manual analysis found that the microarray results were in excellent agreement with the conventional sequencing results for clinical samples. A few of the results of running the CIBSI 2.0 program using the updated NCBI database on the raw microarray results are presented in Table 2 (the results for all samples used in the previous work are presented in Supplementary Table A) . The identified organisms were not identical to the original findings due to the difference in database used and because all ProSeqs were considered rather than only the Flu A and B hemagglutinin. In fact, the conventional sequencing results that were submitted to NCBI from that work were found for every sample to be among the returns with the best score for the hemagglutinin ProSeq (Supplementary Table B ). It should be noted that the previous work based its analysis upon only the results of the hemagglutinin ProSeq. For 8 of 13 Influenza A and 3 of 12 Influenza B cases, the results of ProSeq identification Task(I) and ProSeq grouping Task(II) found that the conventional sequencing was the single best return for the hemagglutinin ProSeq. Owing to the large number of isolate sequences in the database for the hemagglutinin gene it was not surprising that in some cases a single unique entry was not found. In each of the remaining five Influenza A samples, the other sequences returned differed by <0.2% from the conventional sequence. The fewer samples with unique isolate identifications for Influenza B were due to an older reference sequence used for the ProSeq, which allowed less Note: Within a row the first listing of a specific strain was followed by a two-letter abbreviation used in the remaining columns of that row. hybridization to occur (18) . This also meant that when multiple sequences were returned for a sample they represented greater genetic variation, up to 2%. As a result of the current method of making pathogen determination Task(III) level identification, the final organism reported was less specific (H3N2 or Flu B) for every sample than what was reported as possible in ProSeq grouping Task(II). For HAdV samples, the algorithm also reproduced the finer scale discriminations that had been made previously by manual methods (data not shown). The next example of detection for the Mycoplasma pneumoniae pathogen demonstrated a case where there was only a single ProSeq for the target pathogen. A total of 48 test samples were performed using multiplex PCR (via method in Lin et al., submitted for publication) where for 46 of the samples M.pneumoniae organism was spiked into nasal wash with several other pathogens from 100 to 100 000 colony forming units per ml, the remaining 2 samples were purified with nucleic acid from culture stock at a concentration of 1000 genome copies per reaction volume. This ProSeq was also not optimal for fine discrimination because it was selected from a highly conserved region (345 bp) of the cytadhesin P1 gene. In every case taxonomic database entries for M.pneumoniae or its one recognized distinct strain tied for MaxScore (Supplementary Table 3 ). To better understand these returns, the database sequences were examined and subdivided into three groups of sequences, A, B and C, based on how well they matched the reference sequence used to make the ProSeq. The placement of the database entries into the three groups was determined from a CLUSTAL alignment of the sequences of this gene. This alignment confirmed that the database entries differed significantly more from each other in regions not represented by the ProSeq and contained sufficient variability that would have allowed finer discrimination. Members of Group A exactly matched the ProSeq and could not be distinguished between on the microarray. Similarly, members of group B matched the Pro-Seq except at the 199th position where the base called was C rather than T. Group C sequences contained a few database entries that were more variable and might be distinguished from other entries within the ProSeq. For the 48 experimental tests of M.pneumoniae, as much as 80% of the ProSeq hybridized for 19 samples, yet only 5 of these samples had an unambiguous base call at the 199th position. When it was unambiguous, it always matched group B sequences. In the cases where an N base call was made at the 199th location, both groups A and B sequences were returned with the same score. Regardless of this, the target pathogen positively identified was M.pneumoniae for every sample tested. These examples showed how decisions were made independent of whether single or multiple ProSeqs were dedicated to a target pathogen. They also illustrated that the level of discrimination possible was strongly determined by the quality of the selected ProSeq. It is possible that for some pathogens fine level discrimination is not required and the currently tested selections on RPM v.1 would provide satisfactory information. The CIBSI 2.0 algorithm demonstrated its capability to automatically report the maximum level of discrimination that could be supported by the HybSeq information. To demonstrate how the algorithm handled closely related genetic species, a sample of a non-targeted pathogen was considered using multiplex PCR (via method in Lin et al., submitted for publication). For Variola major virus, one of the biothreat pathogens on the RPMv.1, the validation runs demonstrated that Variola major virus purified DNA templates of plasmids were always positively identified when detected (Table 3) . Table 4 shows the results when purified Vaccinia genomic DNA was spiked into nasal washes and processed at various concentrations using multiplex PCR. The array has two ProSeqs from hemagglutinin (VMVHA, $500 bp) and cytokine response modifier B (VMVcrmB, $300 bp) genes for Variola major virus detection. The percentage of the ProSeq that hybridizes is sufficient that if hybridization patterns were only considered one might assume that this tile is identifying the presence of its target. This would indicate that reference sequence selected was not the best choice. However, when our algorithm was applied none of the samples is in fact identified as Variola major or minor virus. Vaccinia was always one of the Orthopoxvirus species listed with the highest scores for VMVcrmB ProSeq, but in only seven cases was it uniquely identified as the probable species detected. In only one sample at the lowest concentration and fraction of VMVcrmB hybridizing, did this ProSeq even identify Variola major and minor virus as one among the Orthopoxvirus species that could be the cause of the hybridization. The lower limit of detection for the amplification method used was between this concentration and the one above it for Variola major itself. Table 2 showed that HybSeqs split into multiple SubSeqs were capable of very specific identification. To illustrate the other filtering performed, when multiplex strategies rather than generic were used for amplification Figure 5 also contains the raw and mask filtered results of this region for the A/Puerto Rico/8/34 sample. It was necessary to perform additional filtering to remove potential biasing caused by the specific primers as described in the methods. In the case shown in Figure 5 , a sequence of 18 bases present in the raw result are made N after filtering since they are in a location that interacts with the primers. If these base calls were included in the subsequences constructed, even though the HybSeq would still be split into the same number of SubSeq, the query for the ProSeq would favor an incorrect strain. The algorithm we have developed successfully provided pathogen identification to the maximum level of detail possible (species or strain) depending on the quality of each Pro-Seq. This identification capability requires minimal input on the identity of the pathogens, making non-expert use feasible. The crucial feature incorporated that allowed complete automation was the use of taxonomic databases, which classify organisms into ordered groups and provide relationships between organism entries, allowing removal of redundancies, comparison of different related prototype sequences and simplification of data presentation. This allows databases, i.e. NCBI, that are redundant and subject to minimal curation but which constantly receive updated and new sequence information to be used with great success. Although we have demonstrated this using only the NCBI databases, other databases or custom made ones could have easily been used, which might improve performance. The algorithm is capable of providing accurate identifications at all analysis levels for pathogens that are less variable or are represented by highly conserved ProSeqs. For more variable or rapidly mutating pathogens, e.g. Influenza A virus, ProSeq identification Task(I) and ProSeq grouping Task(II) still provided accurate detailed identifications, but the pathogen determination Task(III) was unable to report fine scale discrimination. The comparison of the conventionally sequenced Influenza virus gene sequences illustrated that the algorithm is capable of automatically adjusting for updates in databases. The algorithm demonstrated its capability to properly distinguish hybridizations on a ProSeq caused by the specified pathogen from those caused by genetically close (near neighbor) strains and did not make incorrect identifications, eliminating one potential cause of false positives. Filtering the raw hybridization results served to reduce the computation time, accounted for potential primer interference and more importantly reduced potential biasing. This simple integrated algorithm provided sufficient and accurate identification, so that immediate use of the RPM v.1 or similar resequencing arrays and assay is possible. Although not discussed in this paper, the algorithm has successfully detected the presence of simulated multi-infections (Lin et al., submitted for publication). The algorithm as currently developed will detect mixtures when the organisms have sufficient variation; however, detection of a mixture of an organism and its mutation strain in a sample is uncertain in its present phase. In principle it may be possible to detect such mixtures as the resequencing microarray can detect and sequence diploid organisms. Besides demonstrating the success of the CIBSI 2.0 program, the work involved in developing the algorithm allowed insight into the importance of proper ProSeq selection. The RPM v.1 was the first resequencing array designed specifically for multiple pathogen detection using database similarity searching and served as a prototype for this application. We have demonstrated that a single ProSeq with as few as 100 bp, when designed correctly, can be sufficient to unambiguously identify an organism. However, it is clearly indicated that several longer ProSeqs provide better confirmation and more detailed information of a pathogen. Although the algorithm provides accuracy equivalent to manual analysis for determinations of individual ProSeqs, the current algorithm is only partially successful in integrating information from multiple ProSeqs. The emphasis of the design to this point has been on capabilities that are generally applicable to any pathogen. We are incorporating these insights in our newer more comprehensive resequencing array designs. Improving on level of detail reported in pathogen determination Task(III) will require more information about an individual pathogen and may have to be developed for each specific pathogen or class of pathogens. This information is also required for the algorithm to identify which differences between a sample and database entries represent significant mutations. Future work will involve improving the use of the current taxonomic database or potentially developing a new relational database that is specific to our needs and then incorporating more specific information of target pathogens. The hierarchal design of the data analysis makes it easy to incorporate analysis that build upon the analysis already performed. We have met with some success in the current version but want to have increased automated discrimination. We have a well-defined path to completing this aim. The use of properly designed resequencing microarrays and this automated detection algorithm provides a way forward to developing assays that can test for multiple organisms simultaneously while providing fine strain level discrimination giving access to information about detailed strain recognition, antibiotic resistance markers and pathogenicity. This is a capability that other approaches cannot currently provide. In addition, since the design of the original 30 kb RPM microarray, the possible sequence content of the current array has increased 10-fold to 300 kb and further increases in array density are still attainable. This, coupled with our identification algorithms, will allow the analysis of partial sequence information from even more organisms for applications such as differential diagnostics for illnesses with multiple potential causes (i.e. febrile respiratory illness), tracking of emergent pathogens, distinction of biological threats from harmless near genetic neighbors in surveillance applications and for tracking the impact of coinfections or super infections. The concept of categorizing and reporting different degrees of identification depending on the quality of samples and set of target sequences is not limited to resequencing microarrays but is more generally applicable to any platform that is capable of returning sequence level calls that can be used to query a reference DNA database. As the trend for assays that test for multiple pathogens increases, automated analysis tools, such as this one, become more crucial for rapid identification in simple formats useful to the non-expert on a day to day basis. The remaining hurdle to using resequencing microarrays as a routine assay method is now clearly the sample processing methods. Further automating these steps is an important area of future research and development. The program can be obtained free of charge for research purposes by contacting the authors.
67
Molecular dynamics simulations of human [Formula: see text]: the role of modified bases in mRNA recognition
Accuracy in translation of the genetic code into proteins depends upon correct tRNA–mRNA recognition in the context of the ribosome. In human [Formula: see text] three modified bases are present in the anticodon stem–loop—2-methylthio-N6-threonylcarbamoyladenosine at position 37 (ms(2)t(6)A37), 5-methoxycarbonylmethyl-2-thiouridine at position 34 (mcm(5)s(2)U34) and pseudouridine (ψ) at position 39—two of which, ms(2)t(6)A37 and mcm(5)s(2)U34, are required to achieve wild-type binding activity of wild-type human [Formula: see text] [C. Yarian, M. Marszalek, E. Sochacka, A. Malkiewicz, R. Guenther, A. Miskiewicz and P. F. Agris (2000) Biochemistry, 39, 13390–13395]. Molecular dynamics simulations of nine tRNA anticodon stem–loops with different combinations of nonstandard bases were performed. The wild-type simulation exhibited a canonical anticodon stair-stepped conformation. The ms(2)t(6) modification at position 37 is required for maintenance of this structure and reduces solvent accessibility of U36. Ms(2)t(6)A37 generally hydrogen bonds across the loop and may prevent U36 from rotating into solution. A water molecule does coordinate to ψ39 most of the simulation time but weakly, as most of the residence lifetimes are <40 ps.
Ditchfield, R., Hehre
68
A novel endonuclease IV post-PCR genotyping system
Here we describe a novel endonuclease IV (Endo IV) based assay utilizing a substrate that mimics the abasic lesions that normally occur in double-stranded DNA. The three component substrate is characterized by single-stranded DNA target, an oligonucleotide probe, separated from a helper oligonucleotide by a one base gap. The oligonucleotide probe contains a non-fluorescent quencher at the 5′ end and fluorophore attached to the 3′ end through a special rigid linker. Fluorescence of the oligonucleotide probe is efficiently quenched by the interaction of terminal dye and quencher when not hybridized. Upon hybridization of the oligonucleotide probe and helper probe to their complementary target, the phosphodiester linkage between the rigid linker and the 3′ end of the probe is efficiently cleaved, generating a fluorescent signal. In this study, the use of the Endo IV assay as a post-PCR amplification detection system is demonstrated. High sensitivity and specificity are illustrated using single nucleotide polymorphism detection.
Nucleic acid assays utilizing cleavage enzymes to generate a fluorescent signal from dual-dye-labeled oligonucleotide probes have been extensively used in diagnostics (1) (2) (3) . In these assays, the ability of fluorescent dyes to transfer energy absorbed from light to nearby molecules forms the basis of homogeneous nucleic acid based assays (1, 4, 5) . In the unhybridized state the fluorophore and quencher are within close proximity, which allows the quencher to absorb energy from the fluorophore to affect quenching. The use of fluorogenic probes in a 5 0 -nuclease polymerase assay, where the probe is enzymatically cleaved to release the fluorophore (2) , is well known. The signal amplification reaction utilizes invasive cleavage with structure-specific 5 0 -nucleases (6) . The method requires annealing of two oligonucleotides, called the upstream oligonucleotide and the probe, to a target sequence, which results in the formation of a unique substrate for the 5 0 -nuclease. In the DNAzyme-PCR strategy the primer contains a target-specific sequence and harbors the antisense sequence of a 10-23 DNAzyme. During amplification, amplicons are produced containing active sense copies of DNAzymes that cleave a reporter substrate included in the reaction mixture (7) . A target nucleic acid sequence can be amplified exponentially in vitro under isothermal conditions by using three enzymatic activities essential to retroviral replication: reverse transcriptase, RNase H (RNA cleavage) and a DNA-dependent RNA polymerase (8) . Cycling probe technology represents a simple method for the detection of DNA target sequences, utilizing a chimeric DNA-RNA-DNA probe which is cleaved by RNase H when hybridized with its complementary target (9) . Apurinic/apyrimidinic (AP) sites in DNA arise via spontaneous or mutagen-induced hydrolysis of the N-glycosylic bond, or through the repair activity of DNA glycosylases (10) . If left unrepaired, AP sites are potentially lethal or mutagenic (11). To cope with the deleterious consequences of AP sites, organisms possess AP endonucleases that initiate the repair of these DNA lesions (10) . Class II AP endonucleases initiate repair by catalyzing the hydrolysis of the 5 0 -phosphodiester of an abasic site to generate a 3 0 -OH group and a 5 0 -abasic phosphate residue (10) . The substrate specificity of human apurinic endonuclease on synthetic substrate analogs has been reported previously (12) . Here we describe a novel endonuclease IV (Endo IV) assay substrate ( Figure 1C ) that mimics the abasic lesions that normally occur in double-stranded DNA ( Figure 1A and B). The first lesion ( Figure 1A ) is a typical abasic (apurinic or apyrimidinic site generated by spontaneous or enzymatic loss of a nucleic acid base. The second lesion ( Figure 1B) is an atypical abasic site appearing as a result of inherent instability of the 3 0 -phosphodiester bond in abasic deoxyribose in lesion 1 or its cleavage by a Class I AP endonuclease. As shown by the arrow, AP endonucleases cleave the phosphodiester linkage at the abasic sites. In the novel Endo IV assay a short probe with a fluorophore linked through a phosphate at the 3 0 end and a short enhancer oligonucleotide generate an artificial, lesion 2 type Endo IV substrate. This arrangement allows the specific and efficient cleavage of the phosphodiester bond of the probe by Endo IV and release of fluorescent dye ( Figure 1C ). The addition of a quencher at the 5 0 end of the probe allows quenching of the fluorescence in the uncleaved probe. The specific cleavage of the phosphate bond and the generation of fluorescence is the basis of the new assay. In this study we used endonuclease IV from Escherichia coli (Endo IV). The enzyme functions as a Class II AP endonuclease and as a 3 0 terminal phosphodiesterase (13, 14) . Endo IV is enzymatically similar to exonuclease III, the major AP endonuclease of E.coli, but it differs from the latter enzyme mainly in lacking the associated exonuclease activity (14) . It was also found that E.coli Endo IV is fairly heat stable to $70 C (15, 16) . Full characterization for all the compounds described below can be found in the Supplementary Data. A total of 102 unrelated Centre Etude Polymorphism Humaine DNA samples were obtained from the Coriell Institute of Medical Research (http://locus.umdnj.edu/) after specifying that the DNA samples should be used for research purpose only. The list of templates used is available at http:// snp500cancer.nci.nih.gov. For the synthesis of the fluorophore-modified controlled pore glass(CPG) supports, please see the Supplementary Data. Fluorogenic probes were synthesized on fluorophore-linkerbased CPG supports (Scheme 1) using standard 3 0 -DNAphosphoramidites. An Eclipse Dark Quencher was introduced at the 5 0 end using the corresponding phosphoramidite (www. glenresearch.com). Probes were purified by reverse-phase high-performance liquid chromatography (HPLC), dried and re-dissolved in 1· TE buffer. A nearest-neighbor model was applied to calculate extinction coefficient (e 260 ) of oligonucleotides (17) . A 260 measurements were made in PBS (pH 7.2) at ambient temperature and assumed to be a random coil DNA structure in solution. For each Eclipse Quencher, fluorophore 1 (FL1), fluorophore 2 (FL2) or fluorophore 3 (FL3) substitution an e 260 correction of +6600, +18 500, or +25 100 or 28 600 M À1 cm À1 was used, respectively. Preparation of labeled probes with aminohexanoylextended hydroxyprolinol linker (FI 4, Table 1) 3 0 -Hydroxyprolinol modified probe precursor was synthesized using a modified CPG support (18) , and 3 0 -DNAphosphoramidites and Eclipse Quencher phosphoramidite on a 0.2 mmol scale. The probe was purified by reversephase HPLC using a gradient of acetonitrile in 0.1 M triethylammonium bicarbonate buffer (pH 8-9) followed by drying in a SpeedVac evaporator. The probe precursor was redissolved in 50 ml of dry DMSO and treated with 10 mmol of pentafluorophenyl N-FMOC-6-aminohexanoate (19) for 10 h. The FMOC-aminohexanoyl modified probe was isolated by reverse-phase HPLC and deprotected by treatment with concentrated ammonium hydroxide for 1 h at room temperature. The deprotected oligonucleotide was further purified by reverse-phase HPLC using the triethylammonium bicarbonate buffer as described above, dried and re-dissolved in 50 ml of dry DMSO. To this solution, 10 mmol compound 8c (Scheme 2) and 1 ml triethylamine were added. After being kept at room temperature for 10 h the reaction was combined with 1 ml of 2% solution of NaClO 4 in acetone. The precipitated crude conjugate was further purified by reverse-phase HPLC, dried and reconstituted in 1· TE buffer. An additional molar absorbance (e 260 ) of 28 500 M À1 cm À1 was used in the probe concentration calculation to correct for the presence of the dye. The synthesis of special CPG supports for the preparation of required oligonucleotide probes are shown in Schemes 1-3. Methyl 3-(3-chloro-2,4-dihydroxyphenyl) propanoate (4a) (Scheme 1) was synthesized starting from 3-chloro-2,4dimethoxybenzaldehyde (1) (20) , which was converted to substituted cinnamic acid 2 by Knoevenagel condensation. Intermediate 2 was hydrogenated in the presence of Pd/C to yield substituted phenyl propionic acid 3, which was first demethylated with HBr/HOAc and then converted to the methyl ester 4a. Synthesis of the pentafluorophenyl ester dye intermediates 11 required to synthesize the dye-modified CPG supports 12 is shown in Scheme 2. Compounds 4a and b were reacted with substituted phthalic anhydrides in the presence of AlCl 3 to give the benzophenones 5, which were reacted with the resorcinol analogs 6 (a-c) in trifluoroacetic acid/ methanesulfonic acid to yield the carboxyethyl substituted dyes 7. Treatment of compounds 7 with trifluoroacetic anhydride generated lactones 8, which were first reacted with the DMT-protected hydroxyprolinol 9 (21) and then with trimethylacetic anhydride to afford the protected dyes 10. Reaction of intermediates 10 first with succinic anhydride and then with pentafluorophenyl trifluoroacetate, yielded the PFP esters 11. The reaction of the PFP ester 11 with the long chain aminoalkyl CPG (LCAA-CPG) yielded the desired CPG supports 12 as shown in Scheme 1. Endo IV cleavage activity was measured in real-time as shown in Figure 2 . The model system in Figure 3A was used to determine optimum E.coli Endo IV (Trevigen, Gaithersburg, MD) concentration, magnesium concentration and pH; the latter two were determined to be at 20 mM Tris buffer, pH 8. 6 The endo IV reaction was prepared by adding 5 ml Endo IV Master Mixture to 5 ml PCR mixture. Reaction was incubated at 50 C for 50-60 min. Fluorescent detection was performed on ABI 7700 or 7900 in the two channels typically used to detect FAM (for FL1) and VIC (for FL2, FL3 and FL4) utilizing the ABI software for Allelic Discrimination. Alternatively, the fluorescence could also be detected in a Figure 3B and D, respectively. The reaction mixture contained Endo IV at 0.04 U/ml concentration, probe and enhancer at 150 nM, and the target at 5 nM. The experiment was performed on LightCycler. LightCycler Ò LC24 Real-Time PCR System. All assays unless otherwise indicated were performed at least in triplicate. Thermodynamic parameters of duplex formation were derived by the van't Hoff analysis methods. The shape of each melting curve was fitted to the two-state model with linear base lines (22) using a nonlinear least-square program (23) . Unless otherwise stated, T m of DNA duplexes were measured at a concentration of 5 · 10 À7 M in 1· PCR buffer containing 10 mM Tris-HCl (pH 8.5), 5 mM MgCl 2 and 40 mM NaCl, as described earlier (23) . The MGB Eclipse Ò Design Software 3.0 (http://www. nanogen.com/products/customassays/mgb_ecliplse/) was used to design the primers for PCR. It was also used to calculate the T m of the enhancer and probe. An example of the new Endo IV-based assay is shown in Figure 2 . A probe and an enhancer are hybridized to a complementary synthetic target with a one base gap between them. The signal is substantially diminished in the absence of the enhancer. A rigid hydroxyprolinol linker ( Figure 3B ) between the oligonucleotide and fluorophore FL1, critical for sensitivity and specificity, was used in this experiment. The results in Figure 2 illustrate the importance of all components used in the assay. The cleavage of the fluorophore is highly target dependent and its rate is amplified in the presence of the enhancer oligonucleotide. Very low non-target-dependent cleavage is an important feature of the Endo IV assay. Effects of pH and cations on E.coli Endo IV activity was determined using the probe, enhancer and target ( Figure 3A) . It was shown that Endo IV cleavage of the probe has a relatively flat pH optimum in the range between pH 8.5 and 9.5. For compatibility with post-PCR Endo IV amplification the assay buffer with pH 8.6 was chosen, similar to that of the PCR master mixture. Inclusion of the divalent magnesium cation in the buffer ranging from 0 to 20 mM, resulted in more than 2-fold increase in cleavage rate, the optimum range was $5-8 mM. A concentration of 5 mM was chosen for the Endo IV assay. It was determined that the monovalent cations, such as Li, Na, K and Rb, inhibit Endo IV cleavage activity for the concentrations studied in the range from 0 to 100 mM. An inhibition of $30% was observed at 20 mM monovalent cation concentration, which increased to $60-70% at 100 mM concentration. The optimum probe concentration was determined to be $600 nM, please see Supplementary Figures 1-3 . Figure 3A shows a model Endo IV assay that requires a target complementary to an enhancer and a probe labeled at the 5 0 end with an Eclipse Dark Quencher and a fluorescent dye at the 3 0 end. As shown, the probe (calculated T m ¼ 46.7 C) and enhancer (calculated T m ¼ 50.2 C) are separated by one base to mimic a natural abasic site. Four probes with different fluorescent dyes and linkers ( Figure 3B ) were evaluated for their performance in the new assay. Fluorophores FL1, FL2 and FL3 of the probes I, II and III (with similar quantum yields) contained a hydroxyprolinol linker between the 3 0 -phosphate of the probe sequence and the core of the dyes. Fluorophore FL4 of probe IV was analogous to FL3 of probe III, with an additional aminocaproic spacer introduced between the hydroxylprolinol moiety and the dye; the purpose of this spacer was to distance the fluorophore from the cleavage site. The probe specificity was evaluated by comparing the cleavage rates in the presence and absence of target. Targetdependent (specific) and target-independent (non-specific) cleavage rates of the probe were measured as the percent target cleaved per minute. The results are summarized in Table 1 . Probes I-III demonstrated similar specific cleavage. A 2-to 3-fold decrease in the rate of non-specific cleavage for probes II and III versus probe I was observed. These low, non-specific cleavage rates translate into high-specific to non-specific cleavage ratios. It appears that chloro-or chloro/ methyl substitutions in the fluorophores of probes II and III, respectively, are beneficial for the reduction of non-specific cleavage. One possible explanation for this reduction may be an effect of increased hydrophobicity of FL2 and FL3 on enzyme activity. In the case of probe IV, with the extended hydroxyprolinol linker, the rate of specific cleavage almost doubled. However, the corresponding increase in the rate of non-specific cleavage was $20 times higher than that observed in probe III, substantially compromising the ratio of specific to non-specific cleavage rates. Once again, this result implies the existence of inhibiting interactions between Endo IV enzymatic activity and fluorophore which can be adjusted by moving the fluorophore away from the active center. The rigid hydroxyprolinol linker with low non-specific cleavage rates utilized in probes I-III was identified as the linker of choice in the Endo IV assays. The use of a flexible, straight chain linker (C6) instead of hydroxyprolinol-based linker increased the rates of nonspecific cleavage to even higher degree than the extended hydroxyprolinol linker and reduced the specific to nonspecific ratios to an unacceptable level (data not shown). It appears that both the structure of the sterically constrained hydroxyprolinol linker and proximity of the bulky dye moiety to the cleavage site moderate cleavability of the 3 0 -terminal phosphodiester bond and practically eliminate non-template-dependent cleavage. Figure 4A shows the target, probe and the enhancer aligned with the target to indicate different gaps. The T m of the enhancer is typically chosen to be at least 5 C higher than that of the probe. The cleavage rates in the presence of enhancer are shown with 1-5 base gaps ( Figure 4B ). As expected, based on the natural abasic substrate requirement shown in Figure 1B , a one base gap gave the highest cleavage rate. The probe in the absence of enhancer showed a cleavage Table 1 . Effects of dyes and linkers on the rates and ratios of specific and non-specific cleavage Probe nos Dye Substitution on Figure 3B Cleavage rate Specific/non-specific rate ratios Target-specific %/min Target-non-specific %/min rate nearly equivalent to that of a 4 base gap in the presence of enhancer. As shown in Figure 4C , the maximum cleavage rate is achieved at a concentration of enhancer greater than $80 nM. In practice, the concentration of the enhancer is typically at least equivalent or larger than that of the probe. Some cleavage is observed with no enhancer present. I FL1 R 1 -R 7 ¼ H; n ¼ 0 0.97 0.0031 313 II FL2 R 1 ¼ R 3 ¼ R 4 ¼ R 5 ¼ R 6 ¼ R 7 ¼ Cl; R 2 ¼ H; n ¼ 0 0.93 0.0012 775 III FL3 R 1 ¼ R 3 ¼ R 5 ¼ R 6 ¼ Cl; R 4 ¼ R 7 ¼ H; R 2 ¼ CH 3 ; n ¼ 0 1.04 0.0010 1040 IV FL4 R 1 ¼ R 3 ¼ R 5 ¼ R 6 ¼ Cl; R 4 ¼ R 7 ¼ H; R 2 ¼ CH 3 ; n ¼ 1 1.90 0.0260 73 - - - - - - - - - - - - - - - - -4 - - - - - - - - - - - - - - - -2 - - - - - - -3 - -5 3'-GCTGAGCCGGGAACGGGCGGTGTAACCGTGGCTGACACTGA -5' |||||||||| 5'-Q-CTCGGCCCTT Dye Probe T m optimization Figure 5 shows how the cleavage rate of probes with different T m s changes with temperature. Probe lengths ranging from a 6mer to a 14mer were investigated and are shown in Figure 5 with calculated T m s of between 14 and 60 C. All the probes showed a bell-shaped relationship for cleavage rate with temperature. As expected, the cleavage rate for all probes increased with increased temperature. A relatively sharp cleavage rate optimum was observed for the probes a, b and c with calculated T m s of 60, 48 and 42, respectively. It appears that the best probe cleavage occurs close to the calculated T m or slightly above it. Performing the assay at a temperature slightly higher than the T m ( Figure 5 ) showed not only optimum activity but also appears to allow probe cleavage cycling, an important feature of this assay. The ability of the Endo IV assay to detect different target concentrations is shown in Figure 6 . It was observed that in the case of 125 nM target concentration, the probe-based fluorescence plateaued in $1 h, while in the presence of 25 nM target a fluorescence plateau was reached in $3 h. The limit of detection of this assay is 0.04 nM, clearly distinguishable from the stable background ( Figure 6 ). This stable background, a characteristic feature of the assay, is maintained over a period of 15 h. This property is also observed in the assay with other nucleic acid targets. The ability of the Endo IV assay to discriminate mismatches at different positions of a 14mer probe is illustrated in Figure 7 . Mismatches were introduced in the probe one at a time, from positions 1 to 8. The Endo IV assay shows excellent specificity for the different mismatches from base 1 to base 6 with large match/mismatch signal ratios. As shown in Figure 7 , the mismatch signal is often about the same as that of the NTC or slightly higher. The exceptions are the difficult 4-G/T-and 8-T/G-mismatches (24) where match/mismatch ratios of $16 and 7 were observed, respectively. These ratios are quite adequate to differentiate matched and mismatched sequences. It appears (data not shown) that the discrimination in positions 1 and 2 is largely determined by substrate requirements while discrimination in other positions is determined by thermodynamic considerations. Experiments with a large number of assays have also indicated that, with 10mer to 12mer probes, satisfactory discrimination is observed with the mismatch position in all except the last three bases at the 5 0 end. The exquisite specificity of the Endo IV assay appears to be the result of at least two factors. The first is the strict Relative Flourescence @ 60°C intrinsic substrate specificity requirement of the Endo IV enzyme which appears to require the rigid prolinol linker ( Figure 3B , n ¼ 0) used in the probes I-III. Second, the use of short probes (10mer to 12mer) gives the Endo IV a distinct advantage over most other assays where probes are typically much longer (1,2,6,25). The application of the Endo IV cleavage as a post-PCR detection system is shown in Figures 8 and 9 . Following an asymmetric PCR using primers to produce predominantly single-stranded amplicon, two allele-specific probes, an enhancer and Endo IV enzyme are added to the amplification mixture and incubated isothermally for about an hour. The cleaved fluorescence from the two allele-specific probes, each labeled with a different fluorescent dye, is then plotted in a scatter diagram as shown in Figures 8 and 9 for the genotyping analysis of two single nucleotide polymorphisms (SNPs) which are of immediate importance to in cancer (http:// snp500cancer.nci.nih.gov). The scatter diagrams of two polymorphisms, namely agouti signal molecular epidemiology studies protein (ASIP-01) and adenomatous polyposis coli (APC-03) (http://snp500cancer.nci.nih.gov) are shown in Figures 8 and 9 , respectively. In both cases, a probe specific to the wild-type allele is labeled with FL1 and a probe specific for the mutant allele labeled with FL3 is used. The probe sets, designed to analyze ASIP-01 and APC-03 polymorphisms, were used to assay 102 unrelated human DNA samples obtained from the Coriell Institute. In the case of the ASIP-01 target the mismatch was situated two bases from the 3 0 end of the probe, while in the case of APC-03, the mismatch was six bases from the 3 0 end of the probe. From our experience, the spacing of the allele clusters in Figures 8 and 9 are comparable or better than those scatter plots observed in other methods (2, 26, 27) . The exquisite specificity achieved with the Endo IV enzyme, combined with the shorter probes used and the high signal accumulation, gives this assay an advantage over other SNP assays. In comparison with other signal accumulating assays (2), the Endo IV assay has similar levels of high signal accumulation, but uses shorter probes with higher specificity than those assays. In the case of hybridization-based assays (5, 25) , the Endo IV assay has not only a signal accumulation advantage, but also uses shorter probes with a specificity advantage. In addition, the shorter probes allow much more flexibility in assay design compared to other methods with longer probes. We disclosed a novel nucleic acid signal detection system based on the cleavage of a phosphodiester linkage in a dualdye-labeled oligonucleotide probe by the Endo IV enzyme from E.coli. The parameters influencing the activity of the Endo IV enzyme were optimized. This optimized Endo IV detection assay requires, in addition to an oligonucleotide probe substrate, a fluorophore attached through a relatively rigid linker to the oligonucleotide and a helper oligonucleotide. The rigid linker conveys the exquisite specificity of Endo IV cleavage of the phosphodiester linkage in the probe and shows high-specific/non-specific rate ratios. The synthesis of fluorophore analogs coupled to a rigid linker and solid support, used in the synthesis of labeled oligonucleotides, was also disclosed. Excellent mismatch discrimination is observed from positions 1 to 6 in a 12mer probe. The ability to put the mismatch in any one of positions 1-6, allows for flexibility in probe design for genotyping assays where sequence and secondary structural constraints are considered. This method allows the use of probes shorter than those typically used in other methods (28) contributing to the improved mismatch discrimination. The combination of the Endo IV assay with a PCR as a post-amplification detection system has been demonstrated. Post-PCR Endo IV genotyping shows excellent spacing of scatter plot endpoint signals of no template controls, wildtype, mutant and heterozygous samples allowing clear allele differentiation. In contrast, other endpoint genotyping methods may suffer from overlapping endpoint signals due to PCR efficiency issues (25, 29) . This Endo IV post-PCR genotyping has been successfully used in an industry setting for >5000 genotype assays.
69
Pandemic influenza preparedness: an ethical framework to guide decision-making
BACKGROUND: Planning for the next pandemic influenza outbreak is underway in hospitals across the world. The global SARS experience has taught us that ethical frameworks to guide decision-making may help to reduce collateral damage and increase trust and solidarity within and between health care organisations. Good pandemic planning requires reflection on values because science alone cannot tell us how to prepare for a public health crisis. DISCUSSION: In this paper, we present an ethical framework for pandemic influenza planning. The ethical framework was developed with expertise from clinical, organisational and public health ethics and validated through a stakeholder engagement process. The ethical framework includes both substantive and procedural elements for ethical pandemic influenza planning. The incorporation of ethics into pandemic planning can be helped by senior hospital administrators sponsoring its use, by having stakeholders vet the framework, and by designing or identifying decision review processes. We discuss the merits and limits of an applied ethical framework for hospital decision-making, as well as the robustness of the framework. SUMMARY: The need for reflection on the ethical issues raised by the spectre of a pandemic influenza outbreak is great. Our efforts to address the normative aspects of pandemic planning in hospitals have generated interest from other hospitals and from the governmental sector. The framework will require re-evaluation and refinement and we hope that this paper will generate feedback on how to make it even more robust.
In this paper, we present an ethical framework for pandemic influenza planning. The ethical framework was developed with expertise from clinical, organisational and public health ethics and validated through a stakeholder engagement process. The ethical framework includes both substantive and procedural elements for ethical pandemic influenza planning. The incorporation of ethics into pandemic planning can be helped by senior hospital administrators sponsoring its use, by having stakeholders vet the framework, and by designing or identifying decision review processes. We discuss the merits and limits of an applied ethical framework for hospital decision-making, as well as the robustness of the framework. The need for reflection on the ethical issues raised by the spectre of a pandemic influenza outbreak is great. Our efforts to address the normative aspects of pandemic planning in hospitals have generated interest from other hospitals and from the governmental sector. The framework will require re-evaluation and refinement and we hope that this paper will generate feedback on how to make it even more robust. As the world prepares for the emergence of a pandemic strain of influenza, trans-national, national and local organisations and agencies are designing plans to manage community outbreaks. In addition, the medical commu-nity is identifying scientific research priorities and needs related to the anticipated pandemic [1] [2] [3] [4] [5] . There is also a need to examine the ethical issues that arise from planning for a public health crisis of this magnitude. Who should get the limited supply of antivirals? Are health care workers duty-bound to care for the ill in a pandemic when they may have competing familial obligations? Who will be prioritized for scarce ventilated hospital beds? When should hospitals cancel elective surgeries or restrict hospital visitation? To date, the bioethics community has been slow to respond to public health issues in general [6, 7] , and pandemic influenza planning in particular [8, 9] . In this paper we discuss the need for ethics in pandemic influenza planning and discuss the ethical framework we developed to guide pandemic planning in hospitals. In the only article we could find that has an in-depth analysis of the ethics of pandemic planning, Kotalik offers an ethical analysis of the pandemic plans of three countries. His arguments are primarily about the ethics of pandemic planning efforts, as opposed to the ethics in pandemic planning. For example, he argues persuasively that it is problematic that all three countries' plans accept particular conditions of resource scarcity as planning assumptions [10] . While Kotalik has raised important issues about the ethics of pandemic planning in his article, our ethical framework focuses specifically on providing guidance to decision-makers about ethical issues in pandemic planning. This includes providing guidance on how to design an ethical process for decision-making, and providing guiding ethical values for the consideration of substantive issues. The framework here proposed is an example of practical ethics that attempts to provide decision-makers with an introduction to and articulation of generally accepted ethical principles or values. The significance of this ethical framework is a) in the unique collaborative approach taken to its development that involved ethicists with different areas of expertise and a variety of health care stakeholders, and b) that it fills an important need in pandemic planning for an ethical framework to guide decision-making that has been unmet in most pandemic planning processes world wide. One of the characteristics of a public health crisis is that health needs overwhelm available human and material resources. Difficult decisions must be made about how, where and to whom resources should be allocated. Medical science provides valuable information to help make these decisions. However, science alone is insufficient. Now consider that resource allocation decisions are just one kind of decision decision-makers face in preparing for, and getting through an influenza pandemic [9] . As a few scholars have begun to point out, pandemic planning needs to take ethical considerations seriously, and not allow the urgency of logistical and scientific needs to sideline a discussion of ethical considerations [10, 11] . Kotalik argues that as "every discourse about health care has not only a scientific but also a moral dimension, [pandemic influenza] plans also presuppose certain ethical values, principles, norms, interests and preferences" [10] . It is important to make these presuppositions explicit, because, as the SARS experience in Toronto taught health care organisations, the costs of not addressing the ethical concerns are severe: loss of public trust, low hospital staff morale, confusion about roles and responsibilities, stigmatization of vulnerable communities, and misinformation [12] [13] [14] . Another key insight from SARS that we overlook at our peril was that in times of crisis, "where guidance is incomplete, consequences uncertain, and information constantly changing, where hour-by-hour decisions involve life and death, fairness is more important, rather than less [emphasis added]" [14] . As we shall argue, fairness considerations are both procedurally and substantively important: there is a need for fair decision-making processes, as well as equitable distributions of scarce human and material resources. Take the example of triaging ventilated beds in an ICU. In theory, decision-makers rely on scientific evidence to determine how best to maximise benefit in the allocation of ventilated beds, but science cannot tell us whether or not the initial decision to maximise benefit is just. Because the notion of maximising benefit is derived from a reflection on values, ethical analysis is required to determine why a utilitarian approach to triage though maximisation of benefit is preferable to the assignment of ventilated beds on a different basis, for example that of greatest need. Even if the utilitarian maximisation of benefit is thought to be ethically sound, how to implement a system based on this criterion is not ethically straightforward, and requires ethical reflection about what counts as good stewardship, and about the moral obligation to demonstrate transparency, accountability, fairness and trustworthiness in the allocation of scarce resources. The importance of ethics to pandemic planning is in the "the application of value judgements to science" [15] , especially as they are embedded in planning assumptions, and within the practice of medicine itself. For example, while ethics might have little to contribute to understanding the mechanism of influenza virus transmission, it can make a significant contribution to debates such as what levels of harm the public are prepared to accept, how the burdens of negative outcomes should be distributed across the population and whether or not more resources should be invested in stockpiling antiviral medications. The use of ethical frameworks to guide decision-making may help to mitigate some of the unintended and unavoidable collateral damage from an influenza pandemic. As Kotalik argues, the incorporation of ethics into pan-demic plans can help to make them "instruments for building mutual trust and solidarity at such time that will likely present a major challenge to our societies" [10] . Using ethical frameworks to help guide decisions can offer greater assurance that the values instantiated within them, such as accountability, transparency and trust, will be carefully thought about in decision-making and when reviewing decisions with stakeholders. One of the key lessons from the Toronto SARS experience was that health care institutions and their staff could benefit from the development of ethical frameworks for decision-making [12] . The intention of this section is not to systematically derivate from normative theory the values and principles in the framework. This paper has a more narrow focus -it is an example of applied/practical ethics that attempts to introduce and articulate values that are already commonly accepted. It is not our intention to comprehensively defend the values in the framework, but rather to show from which areas of scholarship they were drawn, articulate their relevance to pandemic planning, and to demonstrate their discursive legitimacy through a process of stakeholder engagement and vetting. To our knowledge, no other pandemic planning process has attempted to a) develop an ethical framework to guide pandemic influenza planning and b) assess an ethical framework's robustness and resonance in the community of its intended users. Thus, the significance of the procedural elements of the development of the framework is not to be minimized, nor are the insights we have gleaned from implementing the framework in health care organisations and in a governmental setting. Building on key lessons from SARS [12] [13] [14] and the "emergency ethics" literature and drawing on our expertise in clinical, organisational, and public health ethics, we identified key ethical processes and values that are relevant for health care organisations. These values were presented to and vetted by a variety of health care stakeholders. Thus, this framework is the product of an iterative and inclusive process. In Ontario the need for guidance on the ethical issues pertaining to an influenza pandemic has been widely acknowledged. As word of our work on an ethical framework for Sunnybrook and Women's College Health Science Centre (S & W) became known, we were invited to join other hospitals' pandemic planning efforts. There was also broader sectoral interest in ethics, and we were invited to join the Ontario Ministry of Health and Long Term Care's (MOHLTC) efforts to design a pandemic plan. Our working group was formed in response to the pandemic planning initiative that took place at S & W in early 2005. The hospital's Clinical Ethics Centre was invited to provide ethics support in this planning initiative. It soon became apparent that the scope of the issues went beyond the purview of clinical ethics to include organisational and public health ethics. Expertise in organisational and public health ethics was quickly procured through the University of Toronto Joint Centre for Bioethics which is a partnership between the University and sixteen affiliated healthcare organizations that includes S & W among its partners. S&W was subsequently de-amalgamated into Sunnybrook Health Sciences Centre and Women's College Hospital, thus the ethical framework is currently being implemented at Sunnybrook HSC. As the framework took shape, we were invited to join the MOHLTC planning efforts. We began to work with the Vaccine and Antiviral working group at the MOHLTC, and we adapted our work to meet the related but distinct challenges facing government. While our work with the MOHLTC began with the Vaccine and Antiviral working group, the ethical framework we developed for the MOHLTC was eventually included in the Ontario Health Pandemic Influenza Plan [16] not as an annex to the section on vaccines and antivirals as we had originally anticipated, but as an ethical framework for the plan as a whole. Expertise in clinical ethics was important to the development of this framework because of the knowledge, skills and experience clinical ethicists need to address dilemmas or challenges found in the daily clinical arena. An obvious challenge was how to integrate expertise in public health ethics into a framework designed to guide decision-making in clinical health care settings. A related challenge was to thoughtfully integrate generally accepted principles and values from clinical ethics with those in public health ethics. In order to meet this challenge, the authors turned not only to the respective ethics literature, but also to the SARS experiences of Toronto hospitals and health care providers. A review of the SARS literature, and that of public health ethics more generally, guided the integration of the public health and the clinical ethics perspectives [6, 9, 10, [12] [13] [14] [17] [18] [19] . The Toronto experience with SARS demonstrated that organisations faced unique ethical challenges when dealing with a public health crisis, and much of the ethics literature identified a need for greater forethought in how organisations can foster ethical decision-making in times of crisis [12] [13] [14] . We reasoned that the legitimacy of this framework would be enhanced by including insights from the analysis of a recent public health crisis like SARS. Not surprisingly, the literature on clinical ethics has little to say about disaster preparedness and how to make decisions about such things as triage under extraordinary circumstances. The ethics literature on bioterrorism and battle-field triage informed our thinking and called our attention to important issues such as the duty to care, reciprocity, equity and good stewardship [20] [21] [22] [23] [24] [25] . The importance of having ethically robust criteria and policies developed in advance of a pandemic influenza outbreak is underscored in this literature, for "critical decisions like these should not be made on an individual case-by-case basis" and "physicians should never be placed in a position of individually deciding to deny treatment to patients without the guidance of policy or protocol" [22] . Robust disaster preparedness requires practising preventive ethics. The ethical framework was vetted through S & W's Pandemic Planning Committee, the Joint Centre for Bioethics' Clinical Ethics Group (comprised of the affiliated health care organizations' clinical ethicists), the MOHLTC Vaccine and Antiviral Working Group, and the MOHLTC pandemic planning committee. Through this process, we refined the framework and we are grateful to these groups for their valuable insights. The ethical framework is intended to inform decisionmaking, not replace it. It is intended to encourage reflection on important values, discussion and review of ethical concerns arising from a public health crisis. It is intended also as a means to improve accountability for decisionmaking and may require revision as feedback and circumstances require. The framework is divided into two distinct parts, and begins with the premise that planning decisions for a pandemic influenza outbreak ought to be 1) guided by ethical decision-making processes and 2) informed by ethical values. Ethical processes can help to improve accountability and it is hoped that, to the extent that it is possible for ethical processes to produce ethical outcomes, the substantive ethical quality of decisions will be enhanced. Recognising, however, that ethical processes do not guarantee ethical outcomes, we have identified ten key ethical values to guide decision-making that address the substantive ethical dimensions of decision-making in this context. In planning for and throughout a pandemic influenza crisis, difficult decisions will be made that are fraught with ethical challenges. Our framework around ethical proc-esses is based upon the "accountability for reasonableness" model developed by Daniels & Sabin [26] and adapted by Gibson, Martin & Singer [27] . This model provides a useful means of identifying the key elements of ethical decision-making processes. An extensive literature has developed around Daniels' and Sabin's accountability for reasonableness framework. The Daniels and Sabin framework has broad applicability across institutional settings and priority setting situations [28] [29] [30] [31] [32] [33] [34] [35] . Because the Daniels and Sabin framework applies deliberative theories of democratic justice to the specific problem of health care priority setting, and because it is unique in this regard, we felt it promoted the kind of deliberative approach to pandemic planning that this ethical framework is intended to support. Table 1 outlines the characteristics of an ethical decision-making process. Stakeholders will be more able to accept difficult decisions during a pandemic influenza crisis if the decisionmaking process has, and is perceived to have, ethical legitimacy. The second part of the framework identifies ten key ethical values that should inform the pandemic influenza planning process and decision-making during an outbreak. These values are intended to provide guidance, and it is important to consider that more than one value may be relevant to a situation. Indeed, the hallmark of a challenging ethical decision is that one or more value(s) are in tension and that there is no clear answer about which one to privilege in making the decision. When values are in tension with one another, the importance of having ethical decision-making processes is reinforced (see above.) The values identified in our ethical framework were based initially on previous research findings on ethics and SARS at the University of Toronto Joint Centre for Bioethics (JCB). This work was funded by a Canadian Institutes of Health Research grant in 2004 through 2006 and has led to several key publications on the ethical dimensions of SARS [14, [36] [37] [38] [39] . In particular, Singer et. al., in their seminal British Medical Journal article begin to identify key ethical values that were of relevance during the SARS epidemic in Toronto. These values were then further articulated by our working group and adapted for the pandemic influenza planning context. Through a discursive process of stakeholder consultation with public health specialists, ministry officials, S & W's pandemic influenza committee, and the Clinical Ethics Group at the JCB, we augmented the values to include two new values (stewardship and trust [40, 41] ) and refined the definitions of each value in light of the anticipated demands of a pandemic influenza crisis compared to a hospital-based epidemic such as SARS. The substantive values identified and articulated in the framework are not intended to be an exhaustive set, and they may underdetermine how best to achieve the overall goals of pandemic planning, which generally include the minimization of morbidity, mortality, and societal disruption. Nevertheless, this is not to say that that a procedural engagement about the overall goals of a pandemic response would not benefit from using the ethical framework to guide and shape debate. A description of the values that should guide decision-making can be found in Table 2 . Included in the framework are "hot button" ethical issues that we identified through our work with Toronto hospitals and the MOHLTC. These issues were as follows: These "hot button" issues are not intended to be exhaustive, but rather they serve to illustrate how the values in the ethical framework can be used to identify key ethical aspects of decision-making. Let us take the issue of targeting and prioritizing populations for vaccine and antivirals to illustrate how the values in the ethical framework can help guide decision-making. The values of solidarity and protecting the public from harm would require that priorities be set to maximize the capacity to help society ensure that the ill are cared for during a pandemic. Furthermore proportionality would require that decision-makers consider who within the community are most vulnerable to the contagion as well as who are most likely to benefit from immunization. A well-informed public conversant with the values in the ethical framework and aware of the expertise that informed the ranking of priorities for immunisation would be consistent with value of trust and the principle of transparency. Lastly, while knowing how to use the framework to inform decision-making is vital, there is more to ensuring that the framework will be used or useful. We have identified three necessary, if not exhaustive elements to the successful integration of ethics into hospital pandemic planning processes. These elements are 1) sponsorship of the ethical framework by senior hospital administration; 2) vetting of the framework by key stakeholders and; 3) decision review processes. There should be mechanisms in place to ensure that ethical decision-making is sustained throughout the crisis. Decisions should be made explicitly with stakeholder views in mind and there should be opportunities for stakeholders to be engaged in the decision-making process. For example, decision-making related to staff deployment should include the input of affected staff. Decisions should be publicly defensible. This means that the process by which decisions were made must be open to scrutiny and the basis upon which decisions are made should be publicly accessible to affected stakeholders. For example, there should be a communication plan developed in advance to ensure that information can be effectively disseminated to affected stakeholders and that stakeholders know where to go for needed information. Decisions should be based on reasons (i.e., evidence, principles, values) that stakeholders can agree are relevant to meeting health needs in a pandemic influenza crisis and they should be made by people who are credible and accountable. For example, decision-makers should provide a rationale for prioritising particular groups for antiviral medication and for limiting access to elective surgeries and other services. There should be opportunities to revisit and revise decisions as new information emerges throughout the crisis as well as mechanisms to address disputes and complaints. For example, if elective surgeries are cancelled or postponed, there should a formal mechanism for stakeholders to voice any concerns they may have with the decision. Duty to Provide Care The duty to provide care and to respond to suffering is inherent to all health care professionals' codes of ethics. In an influenza pandemic, demands on health care providers and the institutions in which they work will overwhelm resources. Health care providers will have to weigh demands from their professional role with other competing obligations to their own health, to family and friends. Health care workers will face significant challenges related to resource allocation, scope of practice, professional liability, and workplace conditions. Decision makers should: • Work collaboratively with stakeholders and professional colleges in advance of an influenza pandemic to establish practice guidelines • Work collaboratively to develop fair and accountable processes to resolve disputes • Provide supports to ease this moral burden of those with the duty to care • Develop means through which institutions will handle appeals or complaints, especially with regards to work exemptions, or the vaccination/prophylaxis of staff Health care workers who are at increased risk because they are caring for patients with influenza must weigh familial obligations, and obligations to self with their professional duty to care. In addition, they may also have to comply with vaccination or antiviral regimens for prophylaxis which may conflict with their individual liberty. The principle of equity holds that, all things being equal, all patients have an equal claim to receive needed health care. During influenza pandemic, however, tough decisions will need to be made about which health services to maintain and which to defer because of extraordinary circumstances. Measures taken to contain the spread of a deadly disease will inevitably cause considerable collateral damage. In an influenza pandemic, this will extend beyond the cessation of elective surgeries and may limit the provision of emergent or necessary services. In allocating scarce resources, the value of equity could guide in developing fair criteria for allocation while consideration is given also to compensation for those who will not meet inclusion criteria yet are entitled to receive care. Individual liberty is a value enshrined in health care practice under the principle of respect for autonomy. Under usual circumstances, health care providers balance respect for individual autonomy with a duty to protect individual patients from harm. In a public health crisis, however, restrictions to individual liberty may be necessary to protect the public from serious harm. Patients, staff, and members of the public may all be affected by such restrictions. • Be proportional to the risk of public harm • Be necessary and relevant to protecting the public good • Employ the least restrictive means necessary to achieve public health goals • Be applied without discrimination Social distancing strategies that employ visitor restrictions in hospitals must be necessary for the protection of the public and must be proportionate to the threat being allayed. Individuals have a right to privacy in health care. In a public health crisis, it may be necessary to override this right to protect the public from serious harm. A proportionate response to the need for private information requires that it be released only if there are no less intrusive means to protect public health. • Disclose only private information that is relevant to achieve legitimate and necessary public health goals • Release private information only if there are no less intrusive means to protect public health • Determine whether the good that is intended is significant enough to justify the potential harm that can come from suspending privacy rights, (e.g. the harm from stigmatization of individuals or particular communities) • Provide public education to correct misconceptions about disease transmission and to offset misattribution of blame to particular communities The need to conduct contact tracing of possibly infected people might require that particular groups or even individuals are identified publicly. The need to do so must be weighed against the potential harm of exposing communities and individuals to stigmatization. Proportionality requires that restrictions to individual liberty and measures taken to protect the public from harm should not exceed what is necessary to address the actual level of risk to, or critical need of, the community. Decision makers should: • Use least restrictive or coercive measures in limiting or restricting liberties or entitlements • Use more coercive measures only in circumstances where less restrictive measures have failed to achieve appropriate public health ends. The decision to close an emergency room must consider if the potential harm in keeping the emergency room open is significant enough to warrant its closure. A foundational principle of public health ethics is the obligation to protect the public from serious harm. This principle requires that citizens comply with imposed restrictions in order to ensure public wellbeing or safety. To protect the public from harm, hospitals may be required to restrict public access to service areas (e.g. restricted visiting hours), to limit availability of some services (e.g. elective surgeries), or to impose infectious control practices (e.g. masks or quarantine). • Weigh the medical and moral imperative for compliance • Ensure stakeholders are made aware of the medical and moral reasons for public health measures • Ensure stakeholders are aware of the benefits of compliance & the consequences of non-compliance • Establish mechanisms to review these decisions as the public health situation changes and to address stakeholders concerns or complaints When making the decision to quarantine individuals, protection of the public from harm must be weighed against individual liberty. Note that while the ethical value of individual liberty is often in tension with the protection of the public from harm, it is also in individuals' interests to minimize harm to others. Reciprocity requires that society supports those who face a disproportionate burden in protecting the public good and takes steps to minimise their impact as far as possible. In an influenza pandemic, measures to protect the public good are likely to impose a disproportionate burden on health care workers, patients, and their families. Health care workers may face expanded duties, increased workplace risks, physical and emotional stress, isolation from peers and family, and in some cases, infection leading to hospitalization or even death. Similarly, quarantined individuals or families of ill patients may experience significant social, economic, and emotional burdens. • Easing the burdens of health care workers, patients, and patient's families in their hospitals and in coordination with other health care organizations • Ensuring the safety of their workers, especially when redeploying staff in areas beyond the usual scope of practice The provision of antiviral medication and/or vaccination to hospital staff for prophylaxis is one way hospitals can ensure the safety of their workers who may be exposed to greater than usual risks in discharging their duty to care. SARS heightened the global awareness of the interdependence of health systems and the need for solidarity across systemic and institutional boundaries in stemming a serious contagious disease. An influenza pandemic will not only require global solidarity, it will require a vision of solidarity within and between health care institutions. Solidarity requires: • Good, open and honest communication • Open collaboration, in a spirit of common purpose, within and between health care institutions • Sharing public health information • Coordinating health care delivery, transfer of patients, and deployment of human and material resources Territoriality between hospital departments and between health care institutions needs to be overcome with good communication and sense of common purpose in order to provide equitable care across jurisdictions In our society, both institutions and individuals will be entrusted with governance over scarce resources, such as vaccines, antivirals, ventilators, hospital beds and even health care workers. During a pandemic influenza outbreak, difficult decisions about how to allocate material and human resources will have to be made, and there will be collateral damage as a result of these allocation decisions. Those entrusted with governance roles should be guided by the notion of stewardship. Inherent in stewardship are the notions of trust, ethical behaviour, and good decision-making. Decision makers have a responsibility to: • Avoid and/or reduce collateral damage that may result from resource allocation decisions • Maximize benefits when allocating resources • Protect and develop resources where possible • Consider good outcomes (i.e. benefits to the public good) and equity (i.e., fair distribution of benefits & burdens) A hospital's decision to stock-pile antiviral medication must consider whether this is an effective way of protecting staff from infection, where the money for stockpiling will come from, and whether that money could be put to better use elsewhere. Trust is an essential component in the relationships between clinician and patient, between staff and the organization, between the public and health care providers, and between organizations within a health system. In a public health crisis, stakeholders may perceive public health measures as a betrayal of trust (e.g. when access to needed care is denied) or as abandonment at a time of greatest need. Decision-makers will be confronted with the challenge of maintaining stakeholders' trust while at the same time stemming an influenza pandemic through various control measures. It takes time to build trust. • Take steps to build trust with stakeholders before the crisis hits not while it is in full swing • Ensure decision making processes are ethical and transparent to those affected stakeholders Early engagement with stakeholders may go some distance to justify stakeholder confidence in decision-makers' trustworthiness. In part, the value of trust is respected and promoted by following the ethical processes outlined above. Whether or not an ethical framework is used to inform decision-making in a health care institution depends to a large extent on people in senior positions of an organisation seeing its relevance to the decision-making process. In part, this is dependant on how robust the framework is, but it also requires the willingness to frame (at least some) pandemic planning issues as normative in nature. Some may argue that the values in the framework are too stringent or impractical to implement under crisis conditions, especially those found in the Ethical Processes part of the framework (see Table 1 ). Certainly, crisis conditions may place constraints on the extent to which each principle can be acted upon. However, efforts should be made to put them into action to the fullest extent possible under the circumstances and in our experience this is only possible with the support of senior administrators. The senior administration at S & W (many of whom were part of the Pandemic Planning Committee) had previous experience with the accountability for reasonableness framework for decision-making, and thus their pandemic influenza planning committee was already familiar with the Ethical Processes part of the framework, and they were receptive to the idea of being guided by an ethical framework. Senior administrators may also have been receptive to the ethical framework because, as they learned from SARS, organisations that did not have decision-making processes that honoured the values for ethical process during SARS have been dealing with a legacy of collateral damage to staff and patients in the form of distrust and low morale [12] . For these reasons, the senior administrators at S & W played an important role in vetting the ethical framework. Ensuring that institutional "sponsors" are in favour of adopting an ethical framework is important for gaining widespread support for using an ethical framework in decision-making, and for ensuring that the ethical framework does not become something that looks good but remains unused. In order to obtain support for, or "buy in" to an ethical framework, it is important that key stakeholders in an institution vet the framework. This requires careful consideration of who the key stakeholders are in an institution. Not only should this include those with responsibility for decision-making, but also those who will be affected by decisions taken. For the vetting process is not just intended to create "buy in" but also to decrease the likelihood that interests and issues that are (morally) relevant to pandemic planning will be neglected or overlooked, thereby enhancing the moral legitimacy of the values in the framework. In addition, a process of stakeholder vetting increases the likelihood that the values instantiated in the framework resonate with the stakeholder community. It has been our experience that the values in the framework did resonate with the pandemic planners with whom we have shared this ethical framework. The primarily pragmatic justification for the selection of the values in the framework means that the framework is provisional so it ought to be subject to revision in light of compelling argument, empirical evidence and further stakeholder feedback. It is important to note, however, that the iterative and inclusive process through which the values in the framework were deliberated amongst the various stakeholder groups lends them a form of discursive ethical legitimacy and helps to justify their inclusion in the ethical framework. We intend that the framework invite further dialogue about its legitimacy and its adequacy. We will return to this issue in the final section of this paper. Ideally, the vetting process would include people who can represent the interests of patients, families and volunteers who are part of the hospital's constituency. Although patient relations, human resources and occupational health representatives from S & W provided guidance and feedback in the development of the framework, direct input from patients and family representatives was not obtained. One limitation of our framework is that is has yet to be vetted by these important stakeholders. The importance of solidarity to the management of a public health crisis would also suggest that the public and other health care organisations be considered stakeholders in hospital pandemic planning. While it may not be pragmatic for hospitals to undertake broad public consultation and vetting processes for their pandemic plans in general, and their ethical frameworks in particular, solidarity and equity suggest that these broader stakeholder interests are relevant to pandemic planning. Consequently, opportunities for broader ethical dialogue about pandemic planning need to be encouraged. In order to ensure that the support of key stakeholders is maintained through an outbreak, there need to be effective communication mechanisms in place. An important aspect of responsive decision-making processes is ensuring that there are formal opportunities to revisit and revise decisions as new information emerges. As part of our ethical framework, we formulated a template for decision review processes, (adapted from, Gibson, JL: Formal decision review process template. Unpublished; 2003) that aids organisations in identifying existing and establishing new mechanisms that can be used for the formal reviews of decisions. We believe decision review mechanisms are an essential part of ethical decision-making in a public health crisis, and are one way to put the values in the ethical framework in to action. Formal mechanisms for reviewing decisions are needed in order to capture feedback from stakeholders on key decisions, and to resolve disputes and challenges. These processes are important for ensuring that decisions are the best possible under the circumstances given changing information and for engaging stakeholders constructively around the difficult decisions that must be made. Given the unpredictable nature of public health emergencies and the difficulty this poses for those in charge of planning and decision-making, it is reasonable to assume that decisions will be revised throughout the pandemic influenza crisis. Disputes or challenges may arise from the restrictions or requirements imposed on staff, patients and families during a pandemic influenza outbreak. Thus, decision review processes are essential. Again, while some may argue that this is too stringent a measure for a time of crisis, we argue that reviews of decisions will be taking place regardless (most likely in an ad hoc manner), and that to formalize this process is to increase its fairness and moral legitimacy. Indeed, there may be existing mechanisms which can handle these kinds of reviews. It is important to distinguish between different types of ethical analyses in order to explain the approach that was taken to the development of the ethical framework discussed herein. Callahan and Jennings draw a useful distinction between applied ethics and critical ethics [7] . Our ethical framework is an example of applied ethics because the framework identifies and relies on "general principles that can be applied to real-world examples of professional conduct or decision-making" [7] and because it is "designed to give professionals guidance and to give clients and the general public standards to use in assessing professional conduct" [42] . While there is certainly a need for critical ethical analysis that pays attention to problems that are the "result of institutional arrangements and prevailing structures of cultural attitudes and social power" [7] , one would not expect a ethical framework designed to guide clinical decision-making to explicitly address these kinds of issues. This is not to say that this ethical framework cannot address the kinds of issues that a critical ethical analysis might address. For example, the framework promotes values and processes that seek to redress the power disparities within institutions. The section of the framework that deals with ethical processes in particular is a challenge to how institutional decisions are typically made. For example, the value of "inclusiveness" as a process principle is essential for redressing power differences amongst key stakeholders [27] . Thus, while the ethical framework is the product of applied ethical analysis, and should be evaluated in light of this, one of its strengths is that it can also redress what Callahan and Jennings would characterize as "critical" ethics problem of power disparities within institutions. Within pluralistic societies, there are many different ethical perspectives that exist simultaneously on issues about global, public and individual health. An ethical framework to guide decision-making is robust to the extent that it reflects the values and beliefs of the decision-makers who refer to it and the values and beliefs of those affected by the decisions being taken. Our framework relied heavily on the Toronto experience with SARS to surface and examine the ethical values that are important for a public health crisis. An influenza pandemic is likely to present us with particular ethical challenges that are different from SARS due to the predicted severity of the contagion and its spread to the community. It would therefore be important not to uncritically adopt such a framework but rather to use it as a basis for continued reflection and re-evaluation to ensure its relevance and responsiveness during the unfolding health crisis. It is also important to consider the extent to which an ethical framework is reflective of the community in which it is to be used. Lessons from SARS as it was experienced in China would likely surface some different ethical values, or emphasise different aspects of our framework. As Callahan and Jennings have argued: We submit that a rich discourse on ethics and public health cannot be advanced without relating it to the background values of the general society, and the particular communities, in which it will be carried out. [7] Indeed, as previously maintained, there are many issues related to pandemic influenza planning -particularly those raised by a critical ethical analysis -that require broad public debate. While these kinds of issues require public debate that takes place at the societal level, ethical pandemic planning requires that organisations and agencies foster internal dialogue about the values instantiated in an ethical framework. For it is imperative that the values outlined in a framework resonate with the members of an organisation, and the community it serves. The procedural aspects of the framework provide a means to ensuring that the values of the community are reflected in decision-making through the procedural principles of inclusiveness and responsiveness. It is important, too, to recognise that values are not static, and that circumstances will evolve rapidly during a pandemic influenza outbreak. Ethical frameworks will also require re-evaluation and revision. The challenge will be to continue to recognise the importance of moral reflection under circumstances that are not conducive to it and to encourage a process of re-evaluation that strives to assess whether resulting decisions are consistent with those values the framework is intended to promote. For this reason, it is imperative to start the ethical dialogue in advance, and to find ways to encourage consideration of ethical issues at all stages of decision-making. We hope that this paper will go some way towards advancing this objective, and that this paper stimulates discussion of the ethical issues and values that pervade pandemic planning. We believe that this framework is unique in its blending of clinical, public health, and organizational ethics. One of its strengths is that it draws on lessons from the recent public health crisis of SARS in Toronto, and it is to some extent empirically grounded. Another strength is that it is the product of an inclusive process of development that included stakeholder vetting. It is also unique in its attempt to provide guidance to decision-makers facing a public health crisis. We hope that the framework's acceptance by hospitals and the provincial government in Ontario signals a change in the way that decisions are taken by institutions that are charged with making decisions that have life and death consequences for the public. • Good pandemic planning requires reflection on values because scientific information alone cannot drive decision-making. • The development of an ethical framework for hospital pandemic planning calls for expertise in clinical, organisational and public health ethics. • Stakeholder engagement is essential for the ethical framework to be relevant and legitimate. • The ethical framework contains procedural and substantive ethical values to guide decision-making. • Three key elements of integration of ethics in to pandemic planning are 1) sponsorship from senior hospital administration; 2) vetting by stakeholders and; 3) decision review processes. • An ethical framework is robust to the extent that pandemic influenza planning decisions are seen to be ethically legitimate by those affected by them. • In order to increase the robustness of pandemic planning in general, timely public debate about the ethical issues is essential.
70
Open lung biopsy in early-stage acute respiratory distress syndrome
INTRODUCTION: Acute respiratory distress syndrome (ARDS) has heterogeneous etiologies, rapid progressive change and a high mortality rate. To improve the outcome of ARDS, accurate diagnosis is essential to the application of effective early treatment. The present study investigated the clinical effects and safety of open lung biopsy (OLB) in patients with early-stage ARDS of suspected non-infectious origin. METHODS: We undertook a retrospective study of 41 patients with early-stage ARDS (defined as one week or less after intubation) who underwent OLB in two medical intensive care units of a tertiary care hospital from 1999 to 2005. Data analyzed included baseline characteristics, complication rate, pathological diagnoses, treatment alterations, and hospital survival. RESULTS: The age of patients was 55 ± 17 years (mean ± SD). The average ratio of arterial partial pressure of oxygen (PaO(2)) to fraction of inspired oxygen (FiO(2)) was 116 ± 43 mmHg (mean ± SD) at biopsy. Seventeen patients (41%) were immunocompromised. Postoperative complications occurred in 20% of patients (8/41). All biopsies provided a pathological diagnosis with a diagnostic yield of 100%. Specific pathological diagnoses were made for 44% of patients (18/41). Biopsy findings led to an alteration of treatment modality in 73% of patients (30/41). The treatment alteration rate was higher in patients with nonspecific diagnoses than in patients with specific diagnoses (p = 0.0024). Overall mortality was 50% (21/41) and was not influenced by age, gender, pre-OLB oxygenation, complication rate, pathological results, and alteration of treatment. There was no surgery-related mortality. The survival rate for immunocompromised patients was better than that for immunocompetent patients (71% versus 33%; p = 0.0187) in this study. CONCLUSION: Our retrospective study suggests that OLB was a useful and acceptably safe diagnostic procedure in some selected patients with early-stage ARDS.
The clinical definition of acute respiratory distress syndrome (ARDS) includes the acute onset of bilateral pulmonary infiltrates, a ratio of arterial partial pressure of oxygen (PaO 2 ) to fraction of inspired oxygen (FiO 2 ) of 200 mmHg or less, and no evidence of left atrial hypertension [1] . Many risk factors, such as pneumonia, sepsis, and aspiration, are associated with the development of ARDS. However, other diseases and conditions, such as bronchiolitis obliterans organizing pneumonia (BOOP), adverse reaction to drugs, diffuse alveolar hemorrhage (DAH), and hypersensitivity pneumonitis (HP), can also cause ARDS; despite similar clinical presentations, etiological diagnosis can be difficult especially for early-stage ARDS. Although the mortality rate of patients with ARDS ALI = acute lung injury; ARDS = acute respiratory distress syndrome; BAL = bronchoalveolar lavage; DAD = diffuse alveolar damage; FiO 2 = fraction of inspired oxygen; HRCT = high-resolution computed tomography; ICU = intensive care unit; OLB = open lung biopsy; PaO 2 = arterial partial pressure of oxygen; PEEP = positive end-expiratory pressure. improves recently [2] , the rapid clinical deterioration of such patients, who often progress to multiple organ failure, remains a significant challenge for intensivists in the intensive care unit (ICU). To halt the disease progression of early-stage ARDS, accurate diagnosis is critical. It can be difficult to differentiate between infectious and noninfectious etiology as the cause of ARDS in its early stages. Current microbiological sampling techniques are insufficiently sensitive to determine the causes of ARDS in all patients [3] [4] [5] . In patients with negative microbiological cultures, separating a true infection from an inflammatory response with clinical data remains problematic. Empiric broad-spectrum antibiotics are typically prescribed to these critically ill patients immediately after admission. However, unnecessary antibiotic therapy for non-infectious patients can enhance the occurrence of antibiotic-resistant strains of bacteria and increase the potential for subsequent nosocomial infections. The therapeutic benefit of prolonged glucocorticoid therapy during the fibroproliferative stage of ARDS emphasizes the need for the elucidation of the underlying lung pathologies [6] . Additionally, the specific diseases such as BOOP, drug reaction, DAH and HP can cause an ARDS response to steroid therapy. However, inappropriate steroid therapy for patients with ARDS may be associated with complications such as gastrointestinal bleeding, hyperglycemia and increased susceptibility to infection. Some previous studies have demonstrated that open lung biopsy (OLB) is a useful and acceptably safe diagnostic technique for patients with ARDS [7] [8] [9] . In the study by Papazian and colleagues [7] , the results of OLB directly altered the therapeutic management for 34 of 36 patients with ARDS (94%), and the OLB complication of an air leak occurred in five patients (14%). The OLB results obtained by Patel and colleagues [8] led to a change in management in the majority of 57 patients with ARDS, the addition of specific therapy for 34 patients (60%), and the withdrawal of unnecessary therapy in 24 patients (37%); major complications occurred in four patients (7%). However, in both studies the duration from intubation to OLB was long: in Papazian and colleagues' study [7] the range was 5 to 89 days, and in Patel and colleagues' study [8] it was 0 to 25 days. This retrospective study attempted to evaluate the utility and safety of OLB in patients with clinically suspected non-infectious early-stage ARDS. The records of patients with ARDS who received OLB in two ICUs at a tertiary care referral center over a five year period between January 1999 and April 2005 were examined. Charts with a discharge diagnosis code 518.82 of the International Classification of Diseases, Ninth Revision, Clinical Modification, suggesting ARDS not related to surgery or trauma, were reviewed for possible inclusion in this study. A total of 819 patients with ARDS were identified and OLBs were performed in 68 patients (8.3%). Forty-one OLBs were performed during early-stage ARDS (one week or less after intubation). Patients supported with noninvasive positive-pressure ventilation or intubated for more than seven days at the time of biopsy were excluded. All patients met ARDS criteria defined by the American-European consensus conference [1] . Decisions to perform OLB were made by senior intensivists in charge of the respective ICUs. OLB was indicated when ARDS was suspected to be noninfectious in origin, with no obvious etiology and with a possible indication for corticosteroid treatment based on clinical presentations with rapid progression, relative symmetric distribution on chest X-ray, and predominant ground-glass attenuation in high-resolution computed tomography (HRCT) of the chest. Informed consent for OLB was obtained from each patient's family. Chest HRCT was performed before bronchoscopic sampling and OLB. The location for bronchoalveolar lavage (BAL) sampling was selected on the basis of HRCT findings, or on a chest X-ray when HRCT was unavailable. BAL was performed by introducing 200 ml of sterile warm (37°C) saline solution into a lung subsegment and aspirating it back in four 50-ml aliquots. The first aliquot returned (bronchial fraction) was discarded. Each specimen was sent for bacterial examination for Legionella, Mycoplasma pneumoniae, Pneumocystis carinii, and Mycobacteria, and for fungal and virological (cytomegalovirus, influenza virus, parainfluenza virus, adenovirus, herpes simplex virus, respiratory syncytial virus, and coxsackie virus) analyses. Specimens were also sent for cytology and iron stain analysis. BAL results were deemed positive when at minimum one microorganism grew to a concentration of more than 10 4 colony-forming units/ml. All procedures were performed within 24 hours of OLB. OLB was performed in an operating room or at the bedside in an ICU by an experienced thoracic surgeon. Bedside OLB was indicated when the FiO 2 used reached 1 with an applied positive end-expiratory pressure (PEEP) of at least 12 cmH 2 O. With regard to mechanical ventilator settings to prevent air leakage, PEEP was immediately reduced 2 cmH 2 O from the baseline level after surgery. Pulmonary tissue was harvested from a site considered new or from a progressive lesion identified by chest HRCT or chest X-ray. Each tissue specimen was cultured and examined by a pulmonary pathologist. Medical records from these 41 patients were reviewed and analyzed for the following data: age; gender; Acute Physiology and Chronic Health Evaluation (APACHE) II scores at admission to the ICU; acute lung injury (ALI) scores, PEEP, and PaO 2 /FiO 2 ratio at ARDS diagnosis; dates of ARDS onset, respiratory failure, intubation, and biopsy; underlying diseases; diagnostic tests before biopsy; and medications at time of biopsy. Results regarding complications of biopsy, pathological diagnosis, and postoperative therapeutic changes (addition or removal of drugs) were also analyzed. Outcome parameters, including ICU and hospital survival rates and cause of death, were also evaluated. For normally distributed data, values are reported as means ± SD. Student's t tests were used to compare normally distributed continuous variables. Differences between subgroups were compared by using the χ 2 test or Fisher's exact test when the expected number of events was less than five. The significance level (α) for all statistical tests was set at 0.05, and p < 0.05 was considered statistically significant. Sixty-eight patients underwent OLB for ARDS evaluation during the study period, of whom 27 were excluded because the duration between intubation and OLB exceeded seven days. A total of 41 patients were enrolled. Table 1 lists the baseline characteristics of the patients studied. Twenty-four patients (59%) were immunocompetent and 17 patients (41%) were immunocompromised. Causes of immunocompromise status were hematological malignancy in 10 patients and solid tumors in four patients (three had bronchogenic cancers and one had breast cancer), HIV infection in two patients and renal transplantation in one. The duration from intubation to OLB for these 41 patients was 3.0 ± 1.9 days (mean ± SD; range 1 to 7). BAL was performed 24 hours before OLB. Findings of BAL were compatible with pathological diagnosis for only four patients with diagnoses of bacterial pneumonia, mycobacterial tuberculosis, cytomegalovirus pneumonitis, and Pneumocystis carinii pneumonia. Twenty-two patients (54%) had chest HRCT before OLB to identify an appropriate biopsy site. For the remaining 19 patients who did not undergo chest HRCT, OLBs were performed from the right middle lobe in 12 patients and from the lingular lobe in seven patients. Of the 41 patients, 26 (63%) underwent OLB in an operating room and 15 (37%) received bedside OLB in an ICU. Videoassisted thoracotomy was performed in eight patients, and the remaining patients underwent limited anterior thoracotomy. No intra-operative complication occurred, and eight patients (20%) had postoperative complications (less than seven days after the operation). Two patients developed transient hypotension after OLB and regained normal status after fluid resuscitation and vasopressor treatment for 12 hours. Two patients had pneumothorax diagnosed by chest X-ray and required a chest tube with low-pressure suction (10 cmH 2 O) drainage for 24 hours after OLB. Two patients had subcutaneous emphysema localized in the chest area after OLB, which resolved spontaneously in two days. Additionally, two patients had bronchopleural fistula with persistent air leaking from the operative chest tube for at least one day and did not need further surgery. Although six of these eight patients (two with transient hypotension, one with pneumothorax, one with subcutaneous emphysema and two with bronchopleural fistula) died, no surgical complication resulted directly in death. The incidence of postoperative complication was 15% (4/26) and 27% (4/15) for patients undergoing OLB in an operating room or at the bedside in an ICU, respectively. Complication rates were not significantly different between these two groups (p = 0.3799). All biopsies provided sufficient data for pathological diagnosis (diagnostic yield 100%). The specimens obtained during OLB were sent for tissue culturing (for both bacteria and viruses); all culture results were negative. Pathological diagnoses were subdivided into specific and nonspecific categories. Eighteen patients (44%) had specific diagnoses established by OLB, and 23 (56%) had nonspecific diagnoses (Table 2) . Overall, OLB findings led to alteration therapy for 30 of 41 patients (73%). After OLB, 18 patients were administrated high-dose corticosteroid therapy (1 g/day methylprednisolone in divided doses for three days) and seven patients were treated with low-dose corticosteroid therapy (2-3 mg/kg per day methylprednisolone in divided doses). Three patients received co-trimoxazole for Pneumocystis carinii pneumonia. Antibiotics were changed in one patient and discontinued in one patient on the basis of pathological findings. Treatment was not changed in 11 of 41 patients (27%). Table 3 presents comparative results for patient characteristics, complication rates, alterations in treatment, and survival rates of patients with specific and nonspecific pathological diagnoses by OLB. The rate of treatment alteration was higher in the nonspecific pathological diagnosis group than in that with a specific diagnosis (56% versus 87%; p = 0.0243). No other significant differences between these two groups were noted. Twenty-one patients died in the ICU, resulting in an ICU survival rate of 49% (20/41). The hospital survival rate was the same as the ICU survival rate. Multiple organ dysfunction syndrome was the leading cause of death in 10 patients, followed by septic shock in nine patients, hypovolemic shock in one patient and acute myocardial infarction in one patient. Table 4 presents comparative results of patient characteristics and outcomes for survivors and nonsurvivors. No significant differences were observed between survivors and nonsurvivors for baseline data, such as age, gender, severity of illness, complication rate, and treatment alteration rate, between these two groups. Significantly more immunocompromised patients were in the survivor group than in the nonsurvivor group (60% vs 24%; p = 0.0187). Comparisons between immunocompromised and immunocompetent patients (Table 5 ) showed that immunocompromised patients were younger (p = 0.0004) and had lower ALI scores (p = 0.0045). Furthermore, immunocompromised patients had better hospital survival rates than immunocompetent patients (71% versus 33%; p = 0.0187). This study showed that OLB is an acceptably safe and useful procedure for some selected patients with early-stage ARDS. The treatment alteration rate was higher in patients with ARDS with nonspecific pathological diagnoses than in those with specific diagnoses. In recent studies of patients with ARDS [7, 8] , OLB was employed relatively late, and the time from intubation to OLB was considerable (5 to 89 days in the study by Papazian and colleagues, and 0 to 25 days in the study by Patel). In the present study, OLB was performed within one week of intubation (3.0 ± 1.9 days), substantially earlier than in the previous two studies. Patel Specific diagnosis rates based on OLB findings vary among studies of patients with different disease entities. The specific diagnostic rates in a review by Cheson and colleagues were 21 to 68% in immunocompetent patients and 37 to 95% in immunocompromised patients [11] [12] [13] [14] . In this study, specific and nonspecific diagnostic rates were 44% (18/41) and 56% (23/41), respectively, and specific diagnostic rates for immunocompetent and immunocompromised patients were 33% (8/24) and 59% (10/17), respectively. Although not statistically significant (p = 0.1052), the specific diagnostic rate between immunocompetent and immunocompromised patients was similar to that in previous studies, indicating that OLB obtains a high percentage of specific pathological diagnoses for immunocompromised patients. In this study, the rate of therapy alterations after OLB was 73% (30/41) and was not lower than those in previous reports (range 59 to 75%) [10, 14, 15] . For groups with nonspecific and specific pathological diagnoses, the rate of changed therapy was higher in the nonspecific group (87% versus 56%; p = 0.0243). This analytical finding resulted from a large number of patients with nonspecific pathological diagnoses undergoing corticosteroid treatment as a rescue or anti-inflammatory therapy after excluding potential active infection, such as the fibroproliferative stage of DAD [16] [17] [18] , interstitial pneumonitis, nonspecific interstitial pneumonitis, and organizing pneumonia. Early OLB can achieve diagnoses other than fibrosis that are potentially treatable with corticosteroid. Furthermore, the recent study by the ARDS Clinical Trials Network [19] did not support the routine use of methylprednisolone in patients with persistent ARDS (at least seven days after the onset) and suggested that methylprednisolone therapy might be harmful when initiated more than two weeks after the onset of ARDS. The duration of ARDS before corticosteroid treatment interacted significantly with survival. For immunocompromised patients, some studies [11, 20] suggested that OLB is advantageous for diagnosis and for treatment alteration but that its benefit to survival remains unclear. McKenna and colleagues [21] found that for immunocompromised patients, early OLB (average 3.6 days after admission) benefited the histological diagnosis of interstitial pneumonitis treated with steroids; however, OLB did not improve clinical outcome for all patients. The overall mortality rate was 51% (21/41) in the present study, which is similar to that obtained in previous reports (range 47 to 50%) [7, 8] . More immunocompromised patients were in the survivors group and had a better survival rate than the immunocompetent patients (60% versus 24%; p = 0.0187); the young age and low ALI scores of immunocompromised patients probably accounted in part for their better outcome. Furthermore, the enhanced survival rate of immunocompromised patients might be attributed to more immunocompromised patients (9/13; 69%) than immunocompetent patients (9/17; 53%) receiving high-dose corticosteroid therapy after active infection had been excluded by OLB. Various pulmonary conditions such as infection, disease progression, therapeutic reaction, new and unrelated pathologies, or a combination of these can be present in immunocompromised patients [21, 22] . For diagnostic yield and adequate treatment, early OLB has been considered to be a reliable diagnostic modality, providing an early and accurate etiological diagnosis in immunocompromised patients. Operative complication rates reported for OLB in patients with ARDS have ranged from 17 to 39% [7, 8, 10] . In this study, the overall rate of OLB postoperative complications was 20% (8/ 41). In the late fibrotic stage, lung parenchyma is stiffer than in the earlier exudative or fibroproliferative stages of ARDS. Although operative complications are multifactorial, early OLB in non-stiff lungs (less fibrosis in the present study than in other reports) may account for the low surgical complication rate in this study. Of the 41 patients in the present study, 15 could not be transported to an operating room because they were being administered 100% O 2 and a high PEEP; consequently, OLB was performed at the bedside in the ICU. No intra-operative complications or exacerbation of oxygenation and hemodynamics occurred, even in patients with ARDS with severe hypoxemia. Of these 15 patients, four developed postoperative complications of hypotension, pneumothorax, subcutaneous emphysema, and bronchopleural fistula, respectively. No death was attributable to OLB. The risk for complications due to OLB in early-stage ARDS was therefore acceptable, even for the most critically ill patients with severe hypoxemia. Several limitations of this study should be considered. First, because of its retrospective nature our study cannot directly address the question of whether early OLB has a survival benefit. However, understanding of a specific etiology would permit the initiation of specific therapy assuming that such a therapy is available. Many of the diagnoses found in this study (such as metastatic malignancy, infectious pneumonia and hypersensitivity pneumonitis) may have an established positive therapeutic effect on outcome. Second, the result of this study cannot be generally applied to all patients with ARDS. The decision to perform OLB was not made at random and the patients referred for OLB were unlikely to be a representative sample of our ARDS population. This selection bias of patients and intensivists would be expected to increase the possibility of an alternative intervention. A third limitation is that some specific diagnosis such as viral pneumonitis may be underdiagnosed because its identification depends on the availability of laboratory facilities. A standardized comprehensive microbiological examination of BAL before OLB should be established. This retrospective study demonstrates that OLB had a high diagnostic yield rate and an acceptable complication rate for some selected patients with early-stage ARDS. The rate of treatment alteration was higher in patients with nonspecific pathological diagnoses than in those with specific pathologically diagnosed ARDS. Further prospective, randomized and control studies should investigate the appropriate indication and effect of OLB on outcome in patients with ARDS. • Open lung biopsy is an acceptably safe diagnostic procedure for some selected early-stage patients with acute respiratory distress syndrome. • In patients with early-stage acute respiratory distress syndrome of suspected non-infectious origin, open lung biopsy may have a high diagnostic yield rate. • The role of open lung biopsy in patients with acute respiratory distress syndrome needs to be investigated in prospective, randomized and controlled clinical trials.
71
Association and Host Selectivity in Multi-Host Pathogens
The distribution of multi-host pathogens over their host range conditions their population dynamics and structure. Also, host co-infection by different pathogens may have important consequences for the evolution of hosts and pathogens, and host-pathogen co-evolution. Hence it is of interest to know if the distribution of pathogens over their host range is random, or if there are associations between hosts and pathogens, or between pathogens sharing a host. To analyse these issues we propose indices for the observed patterns of host infection by pathogens, and for the observed patterns of co-infection, and tests to analyse if these patterns conform to randomness or reflect associations. Applying these tests to the prevalence of five plant viruses on 21 wild plant species evidenced host-virus associations: most hosts and viruses were selective for viruses and hosts, respectively. Interestingly, the more host-selective viruses were the more prevalent ones, suggesting that host specialisation is a successful strategy for multi-host pathogens. Analyses also showed that viruses tended to associate positively in co-infected hosts. The developed indices and tests provide the tools to analyse how strong and common are these associations among different groups of pathogens, which will help to understand and model the population biology of multi-host pathogens.
Pathogens have highly variable host ranges: in natural conditions some infect only one or a few related species (i.e., specialist pathogens) while other can infect a wide range of hosts belonging to different taxonomic groups (i.e., multi-host or generalist pathogens). A large fraction of described pathogens of humans, animals and plants are generalists [1] [2] [3] . The ability to infect different hosts conditions the epidemiology and pathogenicity of generalist pathogens and, therefore, is highly relevant for pathogen management and disease control [1, 4] . The distribution of multihost pathogens over their host range, i.e. the frequency of infection in the various host species within an ecosystem, may vary largely, which could determine the population dynamics and structure of the pathogen. The distribution of a pathogen species over its host range may also determine important aspects of its biology in hosts significant from an anthropocentric viewpoint (i.e. target hosts), such as reservoirs and inoculum sources, emergence and reemergence, population thresholds for disease invasion or critical community size for disease persistence [e.g., 1, [4] [5] [6] [7] . Animal or plant species may be hosts for a range of pathogens, and most host populations encounter a large number of different pathogen species [8] . For significant host species, there is abundant evidence of differences in the infection frequency of the various pathogen species present in an ecosystem. The distribution of pathogens over their hosts, and the distribution of different pathogens within a host species, will affect the frequency of multiple infection of an individual host by different pathogens. Multiple infection may have important consequences for the infected hosts, for the pathogens, and for host-pathogen coevolution [8, 9] . In the host, frequent co-infections may lead to heterozygote superiority against multiple pathogens and contribute to the persistence in host populations of alleles conferring susceptibility to disease [10] . In multiple infected hosts, pathogens can cooperate or can compete for host resources, which will affect each other's fitness. Hence, multiple infections will be a factor in pathogen evolution. Theoretical analyses predict that the withinhost dynamics of microparasites in multiple infected hosts may have important consequences in the evolution of their virulence [11] [12] [13] [14] , and there is evidence that multiple infection may result in either increased or reduced virulence [e.g., [15] [16] [17] . Multiple infection of a host may also directly affect the genetic diversity of the pathogen population, as co-infection is a prerequisite for genetic exchange between different pathogen species or strains. Also, infection by one pathogen may result in an increased host susceptibility to a second pathogen, a common phenomenon named facilitation or predisposition by animal and plant pathologists, respectively [8, 18] . In spite of its potential impact on pathogenicity, evolution, epidemiology and control, the distribution of pathogens over their host range and the occurrence of co-infections have been largely overlooked, and most research on pathogen ecology and epidemiology has dealt with specific pathogen-host interactions [8] . To our knowledge, it has not been analysed whether the distribution of pathogens over their host range is random or, alternatively, associations between pathogens and hosts occur, neither has been addressed whether host co-infection by different pathogens is random or associations between pathogens occur in particular hosts. Here we address these issues. First, we propose indices for the observed patterns of host infection by different pathogens, and the observed patterns of coinfection, and tests to analyse if they conform to the null hypothesis of randomness or reflect associations. Second, we apply these tests to data on the prevalence of five insect-borne virus species in wild plant species within an agroecosystem in Central Spain. Results of these analyses uncover patterns that, if general, would be highly relevant to understand the ecology and evolution of pathogens. [19] . Except for TSWV, which has a single-stranded RNA genome of negative and ambisense polarity, all other viruses have single-stranded RNA genomes of messenger polarity. AMV, CMV and WMV are transmitted by aphids in a non-persistent manner, i.e. the virus is retained in the distal structures of the aphid mouth parts for short period of time. BWYV is transmitted in a circulative, non-propagative manner, i.e., the virus penetrates through the gut wall into the haemocoel of the insect vector, and circulates with the haemolymph to reach the salivary glands, from where it is inoculated into new plants. TSWV follows a similar path within the thrips body, but infects and multiplies in the insect cells [19] . All five viruses cause important diseases in vegetable crops worldwide, including the studied region in Central Spain, but infection in the analysed wild hosts was asymptomatic. Table 1 shows the number of samples analysed and the number of infected plants by each of these five virus species, in single or multiple infection, in the 21 most frequently found plant species in three monitored habitats (see Methods) for the analysed period. To this data set tests for association between hosts and pathogens (see Methods) were applied. The index of selectivity of pathogen (ISP), and its significance, is shown in Table 2 for the five viruses. The distribution of three of five analysed viruses over their hosts was significantly non-random, i.e. some of the available hosts were preferentially infected. Fig. 1 shows the relationship between prevalence and the ISP for the five viruses. A positive correlation was found for both parameters (r = 0.9347, P = 0.0189 in a Spearman rank correlation test), i.e., the more host-selective viruses were those with a highest prevalence in the analysed ecosystem. Similarly, the index of selectivity of the host (ISH), and its significance, was calculated for the 21 host plant species in Table 1 , and values are shown in Table 3 . For about half (9/21) of the analysed hosts (Amaranthus spp., Cirsium arvense, Convolvulus arvensis, Diplotaxis erucoides, Lactuca serriola, Medicago sativa, Portulaca oleracea, Solanum nigrum and Taraxacum spp.) differences in the prevalence of the five viruses departed significantly from random. Fig. 2 shows the relationship between virus prevalence and the ISH for the 21 host species. Again, a positive correlation between both parameters was found (r = 0.5161, P = 0.0166, in a Spearman rank correlation test), i.e., the more virus-selective hosts were those with a higher prevalence of virus infection. The relationships between prevalence and selectivity for viruses and hosts were not due to a coincidence in the frequency of infection among hosts by different viruses, as shown by a contingency analysis of counts of infected hosts by the different viruses (P,10 24 ). For 16 of the 21 plant species in Table 1 , co-infection with more than one of the five viruses occurred. For these 16 plant species, 102 plants were infected by at least one virus out of 1060 analysed plants ( Table 4 ). The above described test of association between pathogens was applied to this set. The data in Table 4 showed a tendency of the analysed viruses to associate positively: the distribution of the association index (AI) was skewed towards positive values (Fig. 3 ) so that out of 68 AIs computed for the five viruses in 16 plant species, 47/68 (more than two thirds) were positive and 21/68 were negative. Moreover, there was a conspicuous tendency of the positive AI values to have smaller probabilities (r = 20.6575, P,10 24 , in a Spearman rank correlation test). When the pooled sample from the sixteen plant species was considered, the AI was positive and significantly different from zero for each of the five viruses, i.e. each of the five viruses was found in co-infection with a frequency significantly higher than expected from the null hypothesis of independence of infection. However, this was not so when the data for each of the sixteen plant species were analyzed separately. Hence the association analysis uncovered two patterns that were not obvious: i) a general tendency of the analysed viruses to associate positively, ii) association depended on both the plant and the virus species. Most efforts to understand the population biology of pathogens have focussed on specialist pathogens, and population biologists have successfully developed a formal understanding of the dynamics and evolution of single-host pathogens. However, most pathogens of humans, animals and plants are multi-host pathogens [1] [2] [3] 20] . As stated by Woolhouse et al. [1] ''understanding the more complex population biology of multi-host pathogens will be one major challenge in the 21st century ''. There is evidence that within an ecosystem the prevalence of multi-host pathogens may differ largely for the different species of their host range [e.g., [21] [22] [23] [24] [25] ]. Similarly, there is evidence of large differences in the prevalence on a host species of the various pathogens that are able to infect it [e.g., [26] [27] [28] [29] ]. However, no attempt has been made, to our knowledge, to analyse if differences in the distribution of multihost pathogens over their hosts are random or if there are associations between hosts and pathogens. The uncovering of associations between hosts and pathogens would be highly relevant to understand and model the population biology of multi-host pathogens, and for understanding the phenomenon of generalism itself. We present here indices and tests to analyse if there is association between multi-host pathogens and their hosts. The proposed indices of selectivity for the pathogen and for the host measure the degree of association between hosts and pathogens. The tests analyse the homogeneity of distribution of a pathogen over different host species or populations, and of different pathogens on a host, and analyse how significantly the values of the indices departs from zero (i.e. no association). The literature on pathogen ecology does not abound with data on the prevalence of various pathogens on various hosts. Hence, we have applied these indices to our unpublished data on the prevalence of five insectborne plant viruses on 21 species of wild plants in an agroecosystem in central Spain over a three year period. The analysis of the prevalence of the different viruses in each host species by the homogeneity test that we propose, shows that half of the analysed plant species showed an index of selectivity of the host (ISH) significantly different from zero. The distribution of the host species showing virus selectivity was not related to taxonomy, habitat (fallow fields, edges or wastelands), seasonality or vegetative cycle (annual vs. perennial) (not shown). Interestingly, there was a positive correlation between the ISH and the average virus prevalence for these 21 host plant species, showing that the more selective hosts are more prone to be virus-infected, obviously by the virus(es) that better infects them. This phenomenon suggests that in spite that each host encounters a wide array of pathogens, mechanisms of escape and/or resistance [30] to some of them would operate, which could explain their selectivity. In fact, contingency analysis of counts of infected hosts by different viruses, suggest that different viruses specialise on different hosts. The analysis of the homogeneity of prevalence of a virus over its host species showed that for three of the five analysed viruses there was a significant host association, i.e., the value of the index of selectivity for the pathogen (ISP) significantly departed form zero. One major and unexpected finding of the analysis was that there was a positive and highly significant correlation between the value of the ISP and the prevalence of the viruses. The value of the ISP was not conditioned by the number of host plant species infected by each virus, as there was no correlation (r = 0.60, P = 0.173 in a Spearman rank correlation test) between ISP and the number of plant species that each virus infected in the analysed system i.e., the more selective viruses were not those infecting a smaller number of plant species. Thus, the more host-selective viruses were those that did best in the analysed ecosystem. This result could be highly relevant for understanding the evolution of generalism in pathogens. Although most described pathogens are generalists, the advantages of generalism are poorly understood. A generalist strategy provides the pathogen with more opportunities for transmission and survival, but it is predicted that evolution would favour specialism, because pathogen-host co-evolution could result in functional trade-offs that would limit the generalist fitness in any one host [1, [31] [32] [33] [34] . Our results are compatible with the hypothesis that specialism is advantageous for pathogens, as host selectivity is the rule for the analysed set of generalist viruses, and the more host selective is the virus, the more successful its strategy. Hence, our results could suggest that for generalist pathogens a degree of host specialisation, i.e. host-selectivity as defined here, is a successful strategy. Host specialisation in generalist pathogens would also be relevant for important issues of host and pathogen biology, as host specialisation will affect hostpathogen co-evolution and co-speciation, would reduce the opportunities for host switches and jumps, thus constraining the evolution of host expansion, and may result in spatial heterogeneity of hosts, thus favouring the stable maintenance of pathogen and host diversity [6, [35] [36] [37] . In addition, host specialisation may affect the opportunity for different pathogens of sharing a host and, thus, the consequences of multiple infection for pathogen and host evolution, as discussed below. We propose here also a simple procedure to estimate association among pathogens, which enables to compute an association index whose significance can be tested against the null assumption of independence of infections that follow a binomial distribution. The test was applied to the same data set as above, and the second major contribution of our analysis is the finding that co-infection was mostly non-random and that associations among the five analysed viruses were mostly positive. This result is relevant because co-infection of different pathogens may have important consequences for the pathogens, the infected hosts, and for hostpathogen co-evolution [8, 9, 14] . For viruses, co-infection of a host may result in the generation of new genotypes by recombination or by reassortment of genomic segments between different viral species or strains, often with dramatic changes in host range or pathogenicity. The classical example is the reassortment of avian and human strains of influenza A resulting in novel viruses with pandemic potential [38] [39] [40] [41] , but examples abound for both animal and plant viruses [e.g., [3, [42] [43] [44] [45] [46] [47] [48] ]. In the individual host, coinfection may lead to aggravated disease, often resulting from extracellular cooperativity of independently replicating viruses, by which one virus modulates the host response to infection to the benefit of the other [49, 50] . In addition, direct interactions of different viruses in co-infected cells may result in complementation of highly pathogenic defective genotypes, in increased virus replication or in modified cell and tissue tropisms [e.g., [51] [52] [53] [54] [55] [56] [57] ]. Alternatively, there is also evidence that mixed infections of pathogens result in reduced pathogenicity and less severe disease [17] . Examples from viruses include mixed infection with satellite or with defective interfering nucleic acids [58] . In our data set, association between viruses depended on each particular virus-host system. Hence, data suggest that in some hosts, but not in all, coinfection would be advantageous for some viruses, though the underlying mechanism remains to be analysed. The analysis here reported of plant virus infection on weeds has uncovered two major features that should be relevant to understand the population biology of viruses: i) the more hostselective viruses do better on the analysed ecosystem, ii) viruses tend to associate positively in co-infected hosts. It would be of high interest to know how general are these features and in which types of pathogens would they occur. The indices and tests that we propose here could be of general use in the analysis of the ecology of pathogens, and we hope that our results would prompt research on the ecology of pathogen-host and pathogen-pathogen associations, as these analyses might uncover pathogen properties relevant to the formal understanding of the population biology of multi-host pathogens. We study two factors relative of the distribution of pathogens in different hosts (i.e. different host populations, genotypes, species etc): if there are associations between pathogens and their hosts and if there are associations among pathogens. To analyse these two factors we propose the following tests and indices: Association between pathogens and their hosts Let us call N k the number of analysed individuals in host k (k~1, 2, :::: , n k ) and X ik the number of these individuals that are infected by pathogen i (i~1, 2, :::: , n i ). The prevalence of pathogen i in host k will be the ratio P ik = X ik /N k . The average prevalence of pathogen i over hosts will be Conversely, the average prevalence of the different pathogens in host k can be defined as Homogeneity of the prevalence of a pathogen among hosts can be tested by means of a 2xn i contingency table with elements X ik and (N k 2X ik ) [59] . Different proportions (i.e. lack of homogeneity) will indicate a property of the pathogen that we will call selectivity. Selectivity will be measured by the Cramer's coefficient of contingency [59] of the contingency table. If x 2 i is the chisquared of the 2xn i table, the index of selectivity of the pathogen will be: Both of these indices range from zero to one, with zero meaning equal prevalence of the pathogen over hosts, or of pathogens over the same host, i.e. no selectivity for the pathogen or the host. Association between different pathogens Let us call Xs ik and Xa ik the number of analysed individuals of host k that are infected only by pathogen i (single infections) and by pathogen i and at least another one (associated infections), respectively, (X ik = Xs ik +Xa ik ). The frequency of pathogen i in host k can be estimated as: which equals the above defined prevalence. Under the null hypothesis of independence of infection by different pathogens, the probability of a sampled host individual being infected only by pathogen i is: The conditional probability of non-infection by any other pathogen given the presence of pathogen i is: ps ik~P j=i (1{P jk ), and the conditional probability for the observed multiple infections given the presence of i is: So, under the hypothesis of independence of infection by different pathogens (non-association between pathogens), Xs ik will be distributed as a binomial with X ik trials and probability ps ik . We define the association index (AI) for pathogen i in host k as the difference between the proportion of samples that being infected by pathogen i are infected also by at least another pathogen (Xa ik /X ik ), minus the expectation of this proportion under the null hypothesis (pa ik ). This index has a range from one to minus one and an expected value, under the null hypothesis, of zero. The significance of the observation can be estimated as a onetail test from the binomial above. To test for association of different pathogens within a host, or for a given pathogen across different hosts, we follow the same process, as the expectation of a sum of observations will be equal to the sum of their expectations, and the corresponding sums of observations will be binomially distributed given the X ik . To single out significant tests in a group, raw significance probabilities were corrected by the sequential Bonferroni method for multiple independent tests as indicated in [60] . Plants were sampled monthly for three years in a horticultural area in central Spain within three habitats characterised by different degrees of human intervention: fallow fields, edges between fields, and wastelands. Plants were sampled systematically along fixed itineraries, with no consideration of symptom expression, as described in Sacristán et al. [21] . Infection by AMV, BWYV, CMV, WMV and TSWV in the sampled plants was analysed by double-antibody sandwich enzyme-linked immunosorbent assay (DAS-ELISA), using commercial antisera (Bio-Rad, Marnes-La-Coquette, France), according to the manufacturer's instructions. The distribution of the host species showing virus selectivity according to taxonomy, habitat (fallow fields, edges or wastelands), seasonality or vegetative cycle (annual vs. perennial) was analysed by chi-squared tests of 2x N contingency tables, and their significances assessed, as in the rest of tests of this work, by simulation following model III.
72
The Effectiveness of Contact Tracing in Emerging Epidemics
BACKGROUND: Contact tracing plays an important role in the control of emerging infectious diseases, but little is known yet about its effectiveness. Here we deduce from a generic mathematical model how effectiveness of tracing relates to various aspects of time, such as the course of individual infectivity, the (variability in) time between infection and symptom-based detection, and delays in the tracing process. In addition, the possibility of iteratively tracing of yet asymptomatic infecteds is considered. With these insights we explain why contact tracing was and will be effective for control of smallpox and SARS, only partially effective for foot-and-mouth disease, and likely not effective for influenza. METHODS AND FINDINGS: We investigate contact tracing in a model of an emerging epidemic that is flexible enough to use for most infections. We consider isolation of symptomatic infecteds as the basic scenario, and express effectiveness as the proportion of contacts that need to be traced for a reproduction ratio smaller than 1. We obtain general results for special cases, which are interpreted with respect to the likely success of tracing for influenza, smallpox, SARS, and foot-and-mouth disease epidemics. CONCLUSIONS: We conclude that (1) there is no general predictive formula for the proportion to be traced as there is for the proportion to be vaccinated; (2) variability in time to detection is favourable for effective tracing; (3) tracing effectiveness need not be sensitive to the duration of the latent period and tracing delays; (4) iterative tracing primarily improves effectiveness when single-step tracing is on the brink of being effective.
Control of epidemics of (emerging) infectious diseases, such as SARS, pandemic influenza, or foot-and-mouth disease, always faces the difficulty that some infectives are not yet observed. By concentrating control measures only on observed cases (treatment, isolation, culling), resources are used efficiently but control is often not effective enough. On the other hand, by directing control to the whole population (mass vaccination, prophylactic treatment, preventive culling), epidemics are most likely contained, but at high cost. Contact tracing of symptomatic infecteds works on an intermediate level: treatment or quarantine of contactees (by contact we mean the possible transmission event, by contactee the individual that is contacted) may be effective because unidentified infecteds are most likely to be found among contactees, and efficient because the resources can be directed towards individuals at risk only. Tracing and quarantine has been successfully applied for smallpox control, where the term 'Leicester method' refers to exactly this strategy, with the establishment of specific smallpox hospitals [1] . It was also successful during the more recent SARS epidemic [2] , but not during the British foot-and-mouth epidemic [3] . Here we wish to investigate how these differences can be accounted for by the tracing process, thereby distinguishing tracing and quarantine as having separate effects which can be explained as follows. Symptoms and detection divide the reproduction ratio R (the average number of secondary infecteds per primary infected in a susceptible population) into a part before detection and isolation and a part after [4] : where c is the reduction factor due to isolation, R 0 is the basic reproduction ratio, when no efforts are made to isolate, and R 0 pre is that part of R 0 occurring prior to detection. Contact tracing will lead to earlier prevention of transmission due to quarantine of traced infecteds, thereby reducing the uncontrolled transmission R 0 pre to R pre , and the reproduction ratio R to R q : The effect of decreasing R 0 pre to R pre is distinct from the effect of isolation and quarantine which reduce c. For tracing to be effective, R q should be smaller than 1, so R pre should be smaller than (12c)/(12cR 0 ), a threshold value determined by c and R 0 [4] [5] [6] . For comparison of contact tracing in different situations, we will choose c = 0 (as in [6] [7] [8] [9] [10] ), because it is easy and as arbitrary as any other value. The value c = 0 is valid for foot-and-mouth disease where traced herds are culled. Several tracing studies have been published, some focussing on specific infections [10, 11] and some with the objective to obtain more general insight, e.g. in a general Markov-type SIR model [7] , in a model incorporating symptom development [4] , or in simulation models with specific network contact structures [8, 9] . Different assumptions were made with regard to the possibility of only tracing contacts of symptomatic infecteds (single-step tracing [4, 10, 11] ), or of iteratively tracing the contacts of traced infecteds (iterative tracing [6] [7] [8] [9] ). However, no general insight has yet been obtained in reducing R pre in relation to time-related characteristics of the infection and the tracing process, such as the latent and infectious periods, the time of symptom-based detection, and delays in the tracing process. In this paper we consider detection and isolation as an autonomous process most likely governed by detection via clinical symptoms. This allows us to concentrate on contact tracing and quarantine, which are initiated by this autonomous process. Also we only regard transmission prior to control: we will consider tracing effective if R pre ,1. We study an epidemic in its initial phase and use a branching process for its description, which means that the epidemic can be regarded as a growing tree. Each branch (contact) connects two nodes (contactees) with a hierarchical relation, the infector having infected the infectee. If one of the contactees is notified as being infected, by becoming symptomatic or by being traced via a secondary contact, the contact is traced with a probability p c (all parameters and functions of the model are given in Table 1 ). Each contact may be traced only once, either from infector to infectee (forwards tracing) or from infectee to infector (backwards tracing), so its traceability can be determined at the time of transmission. Thus, infection trees with traceable and untraceable contacts arise (Figure 1 ). On these infection trees, two types of contact tracing are distinguished. The first type is single-step contact tracing, in which all traceable contactees of a symptomatic case are quarantined, but tracing only continues when an undiscovered infected, quarantined in this way, is detected itself ( Figure 1A ). The second type is iterative tracing, in which tracing of traceable contactees is continued, and a whole cluster of infecteds that is linked through traceable contacts is quarantined ( Figure 1B ). Such continuation would be possible if a test were available to determine the infection status of traced contactees. The underlying model for infection dynamics is based on the framework of [4] . In our model, t measures time since infection of an individual, which starts with a latent period until t = t lat without transmission of the pathogen. During the infectious period, lasting from t lat to t inf , infecteds give rise to b new infecteds per unit of time, as long as they are not detected. As we are interested in the effectiveness of tracing only, we do not model possible transmission that might occur while being isolated, so transmission 'ceases' after detection of the infected. This leaves us with the basic reproduction ratio prior to detection, defined as the expected number of secondary infections per infected in a susceptible population before detection, R 0 pre . Here we have adjusted the interpretation of the model in [4] , where R 0 is divided in an asymptomatic and a symptomatic part, by adding a detection delay after becoming symptomatic. The part R 0 pre is determined by t lat , t inf , b, and the distribution of the detection time, a Gammadistribution with mean 1 (so time is measured relative to the mean detection time) and shape parameter a. Throughout our analyses, b will be scaled accordingly to achieve a desired value of R 0 pre . This model construction allows a flexible way of exploring different assumptions about the time to detection, the infectious period, and their overlap, and it enables us to evaluate tracing effectiveness for most infections. For full understanding of contact tracing in our model, we analyzed four special cases regarding the infectious period and the detection time distribution. The infectious period was assumed to be either very short (all transmission occurs instantaneously, so t inf = t lat ) or very long (of infinite duration, so t inf = '). (In our model, infinite duration can be assumed because each infected will stop spreading the infection after detection. If some infecteds would never be detected, some large t inf ,' should be taken.) The detection time was assumed to be either fixed (a = ') or highly variable (a = 1, i.e. exponentially distributed). As a control we analyzed intermediate cases (results not shown) and four examples of real infections (influenza, SARS, smallpox, and foot-and-mouth disease), of which the parameter values are listed in Table 2 . These parameter values were obtained from literature [2, 3, [12] [13] [14] , assuming that the time to detection consists of the incubation period (time to symptom onset) plus a symptom-to-detection delay, which we assumed to be distributed as observed in the SARS epidemic (average 3.67 days, [2] ). In our analyses tracing effectiveness will be expressed as the critical tracing probability p c *, defined as the proportion of contacts that need to be traced to achieve R pre = 1. If at least that many contacts are traced, epidemics will certainly die out if transmission during isolation or quarantine is prevented or limited to a small number of health-care workers that do not re-introduce the infection into the general community (see also [5] ). If such reintroductions cannot be excluded, a lower R pre may be aimed for. Because of this threshold of 1, the R 0 pre values (without tracing) were assumed to have some value larger than 1 (otherwise tracing would not be needed at all) and less than published R 0 values for the specific cases (Table 2) . First we study p c * as a function of R 0 pre and t lat (with tracing delay d = 0), and second as a function of t lat and d (with R 0 pre = 1.5). In an epidemic where single-step tracing is applied, asymptomatic infecteds can spread the infection until they are detected and isolated. Detected infecteds are asked to name their contactees, and a proportion p c of all contactees will be reported and quarantined. Only when quarantined infecteds are detected themselves, tracing is continued ( Figure 1A) . We determined the critical tracing probability p c * as a function of the latent period for three values of R 0 pre (1.5, 2, and 3). The results are shown in Figure 2 . If the detection time is fixed (a = '), a too large latent period (larger than the detection time) results in a situation where every infected is detected before transmitting the infection, so tracing need not prevent any transmission; hence the maximum t lat of 1 in panels 2B and 2D. If we locate the approximate position of real infections (parameter values in Table 2 ) in Figure 2 , we observe that the long infectious period will be the best approximation for most infections, because we only regard pre-detection transmission and infecteds will often still be infectious at the time of detection. This does not entirely hold for influenza, which therefore fits best between the short and long infectious period. Because of the detection period distributions, smallpox and FMD are best described by panel 2D, whereas SARS and influenza fit best into panel 2C (influenza also into 2A). In three special cases (panels 2A,B,D), with fixed incubation period and/or short infectious period, the proportion of contacts to be traced is at least 121/R 0 pre . This lower limit 121/R 0 pre is due to forwards tracing, when all infecteds that are traceable via their infector, are quarantined before the end of their latent period. Then, the effective reproduction ratio will be equal to the untraceable proportion of R 0 pre , i.e. R q = (12p c )R 0 pre , resulting in p c * = 121/R 0 pre . This is likely to be case with smallpox (panel 2D). However, if the latent period is short, as seen for influenza and possibly FMD, quarantine will occur too late to prevent all infections and more contacts need to be traced. In the fourth panel (2C), with variable detection time and long infectious period, effective contact tracing requires a proportion of contacts smaller than 121/R 0 pre to be traced, if the latent period is large enough (like SARS). This can be explained as follows. If the detection time is variable and the latent and infectious periods are large, many infecteds will be detected before becoming infectious. If many infecteds are not infectious before being detected, the few that are should be very infectious (''superspreaders'') to attain a given R 0 pre (the average number of secondary infections before detection). Because many of the infectees of these superspreaders will be detected early (variable detection time), the superspreaders will be quarantined after backwards tracing, which adds to the effect of forwards tracing preventing infectees to reach their infectious period. Effectiveness may be very sensitive to the latent period, especially if there is little variation in the detection time. This is most apparent in the sharp transition in panel 2B, where tracing is only effective if all infectors are detected (at t = 12t lat ) before the infectious period (at t = t lat ), so t lat .0.5. The high sensitivity can be a problem for assessing the tracing effectiveness for a specific infection: the conclusion may largely depend on how correct the available estimates for the latent and incubation periods are, the incubation period determining the time to detection. Figure 2 indicates that this might be a problem for FMD. In the previous section tracing was an instantaneous process: detection and isolation of a given case were followed by quarantine of traceable contactees of that case without any delay. The effect of a delay of duration d is that contactees of detected infecteds can continue transmitting the infection for an extra d time units, unless they are detected themselves during this interval. The effect of a delay in each tracing step was studied by determining the critical tracing probability p c * for R 0 pre = 1.5, as a function of the latent period and tracing delay. Figure 3 shows the resulting contour plots for the four special cases with singlestep tracing, and it indicates the approximate positions of the four real infections. Along the t lat axis (d = 0) lie the R 0 pre = 1.5 curves of Figure 2 . It appears that the iso-p c * contours in Figure 3 are linear in three of the four special cases (panel 3A,B,D), and approximately linear in the fourth case, with long infectious period and variable incubation time (panel 3C: t inf = '; a = 1). The slopes of all (approximate) lines are always between 0.5 and 1.5. This means that the effect of tracing delays is comparable to the sensitivity to the latent period as observed in Figure 2 , so plots of the critical tracing probability as a function of the delay will resemble the plots in Figure 2 , only mirrored (as in Figure 4 ). In the contour plots in Figure 3 , tracing is ineffective in the dark grey areas (small latent period or large delay), so smallpox and SARS control are predicted to be able to cope with some delays, whereas it might be more of a problem with FMD and influenza. If forwards tracing is maximally effective and not affected by the delay, p c * = 121/R 0 pre = 0.33 as indicated by the light grey area (smallpox if d,0.5). If the detection time is variable, tracing may be effective even if a proportion less than 121/R 0 pre is traced (SARS if d,0.4, panel 3C). Because backwards tracing is needed to attain this result, it is only possible when the tracing delay is shorter than the infectious period (not shown in the Figures). The contour plots show that for some combinations of latent period and delay, the sensitivity to small changes in the delay is large (contours lie close to one another). Then, small changes in the delay can have a major impact when the time of quarantine is shifted from just before to just after the end of the latent period, similar to the sensitivity to t lat as observed in Figure 2 . This effect is most dramatic if there is little variation in the detection time (smallpox: d<0.7, and FMD: d<0.2). Figure 4 shows the effectiveness of single-step tracing for SARS, smallpox, influenza, and FMD using the parameter values listed in Figure 2 . The effectiveness of single-step contact tracing without tracing delays. Effectiveness is expressed as the minimum proportion of contacts that need to be traced for effective control (critical tracing probability p c *). The plots show p c * as a function of the latent period relative to the mean time to detection (t lat ). There are four special cases: A. Short infectious period and variable time to detection; B. Short infectious period and fixed detection time; C. Long infectious period and variable time to detection; and D. Long infectious period and fixed detection time. The three curves denote p c * for different values of the pre-isolation reproduction ratio R 0 pre . Indicated by dashed lines are the average t lat for four infections, in the panels with closest correspondence to the actual parameter values ( Table 2 ). Influenza appears in two panels with long and short infectious period, because it corresponds to both parameter sets equally. doi:10.1371/journal.pone.0000012.g002 Figure 3 . The effectiveness of single-step contact tracing with tracing delays, with the pre-detection reproduction ratio R 0 pre = 1.5. Effectiveness is expressed as the minimum proportion of contacts that need to be traced for effective control (critical tracing probability p c *). The contour plots show p c * as a function of the tracing delay d and the latent period t lat , measured relative to the mean detection time, for four special cases: A. Short infectious period and variable incubation period; B. Short infectious period and fixed incubation period; C. Long infectious period and variable incubation period; and D. Long infectious period and fixed incubation period. Dark grey shadows indicate areas where tracing is ineffective, light grey shadows indicate areas where p c * = 0.33. Indicated by dashed lines are the average t lat for four infections, in the panels with closest correspondence to the actual parameter values ( Table 2 ). Influenza appears in two panels with long and short infectious period, because it corresponds to both parameter sets equally. doi:10.1371/journal.pone.0000012.g003 Figure 4 . The effectiveness of single-step and iterative contact tracing for control of influenza, smallpox, SARS, and foot-and-mouth disease. Effectiveness is expressed as the minimum proportion of contacts that need to be traced for effective control (critical tracing probability p c *); p c * is plotted as a function of the relative delay (d, proportion of the incubation period) or the absolute delay (days). doi:10.1371/journal.pone.0000012.g004 Table 2 . Panel 4A shows the relation between p c * and the relative tracing delay d, from which it appears that the general cases shown in Figures 2 and 3 are good predictors for assessing tracing effectiveness. Influenza can be placed somewhere between long and short infectious periods (panels 2A,C, and 3A,C): it is hardly effective without delay, and ineffective already if d = 0.1. SARS control requires a tracing probability p c *,121/R 0 pre which is relatively insensitive to delays. Smallpox requires p c * = 121/R 0 pre , and is insensitive to delays up to some point (d<0.6) where tracing becomes quickly ineffective. Finally, FMD can be traced effectively only if d is small, but it is sensitive to delays already if d is small and requires a tracing probability p c *.121/R 0 pre . Measured in real time (panel 4B), effectiveness of tracing appears to be highly dependent on the actual generation time of the infection. Influenza control is hardly possible, FMD control will be difficult, whereas tracing is likely to be successful for smallpox and SARS. Suppose a test is available to determine the infection status of traced contactees. We can then continue tracing iteratively before infected contactees are detected due to symptoms, until no further infecteds are found ( Figure 1B) . We evaluated iterative tracing with and without delays for the same cases as single-step tracing, resulting in Figures similar to Figures 2 and 3 (see Supporting Information). As expected, the universal effect of iterative tracing is to lower p c *, although the lower limits 121/R 0 pre as observed in three of the four special cases remain intact. The largest difference between single-step and iterative tracing is observed when single-step tracing is on the brink of being effective, so that is the only situation where iterative tracing will make a difference (see also Figure 4 ). Without delays this difference is rather intuitive: single-step tracing is not effective if the latent period is small, but iterative tracing will always be effective, because if all contacts are traceable, the first detection is immediately followed by quarantine of every infected. This is the case with influenza, although the improvement will probably not be sufficient to making tracing effective for influenza control. On the other hand, when the latent period is large and single-step tracing is effective, iterative tracing only improves on this if there are significant delays, providing one or two more days to reach the required p c * (smallpox, SARS, and FMD). We studied the effectiveness of contact tracing in a model for the start of an epidemic, that is flexible enough to use for most infections. Effectiveness of contact tracing was expressed as the minimum proportion of contacts that need to be traced to obtain a reproduction ratio prior to control of 1 (p c *). Other threshold values may be chosen if more is known about transmission of the infection to the general community while infectious individuals are being isolated or quarantined. For instance, if isolated infecteds still cause an average of 0.3 new cases in the general community after isolation or quarantine, p c * should be redefined such that R pre = 0.7. The first conclusion from our model is that, for a given R 0 pre , the critical tracing probability can take any value depending on all infectious disease characteristics in the model: the latent period, the infectious period, and the detection time distribution. In contrast to some earlier publications on contact tracing [7, 9] , there exists no general expression for p c * as there is for the proportion to vaccinate for effective control in a well-mixed population. For smallpox the relation p c * = 121/R 0 pre holds reasonably well, but for SARS it is smaller, and for influenza and FMD it is larger. The second conclusion is that a variable detection time improves tracing effectiveness, possibly even resulting in p c *,121/R 0 pre . This does not mean, of course, that one should aim at late detection of some infecteds (which would increase R 0 pre ), but that apparent variability is an argument to use tracing. The reason is that the few infecteds that are detected late (or not at all, which is essentially the same) will be discovered by backwards tracing which is an additional effect to forwards tracing that in itself may already result in p c * = 121/R 0 pre . It was earlier found that p c * can be smaller than 121/R 0 pre [9] , but by a different mechanism, namely the presence of shared contacts in a network. The third conclusion is that the sensitivity of tracing effectiveness to the latent period and tracing delay may be large, especially in the case of single-step tracing. If this is the case already with a small delay (influenza and to a lesser extent FMD), reliability of parameter estimates will be crucial to establish whether tracing is advantageous. If it is only the case with larger delays (smallpox, SARS), tracing may be effective as long as it is done quick enough. The fourth conclusion is that in most situations single-step and iterative tracing are almost equally effective. A considerable difference can only be expected if single-step tracing is not or hardly effective, which is also when the sensitivity to the latent period and tracing delay is largest (see above). Thus, influenza control will benefit from iterative tracing already without delays, whereas with the other infections one or two days may be gained ( Figure 4C,D) , thus having more time to achieve the required p c *. In the real world, the choice between single-step and iterative tracing will be based on what is possible. Capacity problems may reduce the effectiveness of iterative tracing if effort is directed towards secondary contactees prior to primary contactees; on the other hand, if not quarantine but vaccination is applied, it might be worthwhile to traced contactees of contactees even without diagnostic tests [6] . For determining the cost-effectiveness of contact tracing, not only the critical tracing probability p c * is required, but also how easily that p c * can be achieved. A key aspect will be the possibility to distinguish relevant contacts: for control of sexually transmitted diseases relevant contacts are easily identified (that is, if people are willing to co-operate), but in case of respiratory pathogens like SARS or influenza the notion of relevant contacts is rather diffuse. Good insight will be obtained by not only regarding the sensitivity of tracing (p c ), but also the positive predictive value: what proportion of traced individuals is actually infected, and if we put more effort into increasing p c , how will it decrease the positive predictive value? A relatively easy way to increase the proportion of traced contacts for respiratory pathogens like influenza might be quarantine of households. This can be effective if households are the prime location for spread of the infection; specific models taking into account the contact structures within and between households will be better-suited to study this strategy. In our model we only regarded transmission of the infection before tracing or isolation, allowing us to focus on the characteristics of the contact tracing itself. Regarding the control of animal disease epidemics, e.g. foot-and-mouth disease or avian influenza, it is not unreasonable to assume that there will be no other transmission: all infected and suspected farms will be culled and they really will stop being infectious. However, with human infections it is very likely that isolation and quarantine will be less effective. This results in a complicated situation, because transmission will not only be reduced after detection, the contact structure will also change, with only a limited number of people (health-care workers) having contact with multiple isolated patients. For the case of SARS, a more detailed model on exactly this aspect was studied by Lloyd-Smith et al [5] , who indeed concluded that minimizing within-hospital transmission might be crucial, especially if tracing and isolation occur rather slowly. Further studies are required for this complex problem. The present paper provides a good understanding of the principles of contact tracing, and how infectious disease characteristics determine the effectiveness. The effect of single-step tracing can be measured by the effective reproduction ratio R q , the average number of new infections per infected. Whereas the rate of being traced backwards at time t since infection (due to detection of any infectee) is equal for all infecteds, the rate of being traced forwards (due to detection of the infector) depends on the generation-time distribution, which in turn depends on the number of traceable generations backwards in the transmission tree (traceable ancestors). Hence, infectives need to be typed according to the number of traceable ancestors j, and R q is the largest eigenvalue of the next-generation matrix K with entries k ij being the expected number of type-i infecteds (with i traceable ancestors) produced per type-j infected [15] . Because type-j infecteds only produce type-0 and type-j+1 infecteds, by untraceable and traceable contacts respectively, all entries other than k 0j and k j+1,j are equal to 0. Even though the matrix is infinitely large, we conjecture from all observed numerical calculations that the positive elements k j+1,j converge in the sense that |k j+1,j 2k j+2,j+1 |R0 as jR'. Therefore, for numerical evaluation we 'closed' the matrix to an (n+1) 2 matrix with k nn = k n+1,n . For the four special cases, R q could be calculated numerically in MathematicaH, but for other cases considered, the entries of the matrix had to be determined by numerical simulation. Analysis of single-step tracing with delay could be done by calculating R from the next-generation matrix, as for the model without delay (see Supporting Information for all details). With iterative tracing, isolation of a single infected results in quarantine of a cluster of infecteds, all mutually linked by traceable contacts ( Figure 1B) . By recognizing these clusters of infectives, an epidemic of infected individuals can be regarded as an epidemic of traceable clusters. Each untraceable contact in the transmission tree results in infection of a new cluster index case, so the average number of untraceable contact infections caused by a single cluster, the cluster reproduction ratio R c , determines the effectiveness of iterative contact tracing. If Y(p c ) denotes the expected cumulative infectiousness of a cluster at the time of cluster quarantine, then As with single-step tracing, effectiveness of iterative tracing is determined by considering the critical tracing probability (for achieving R c = 1): If p c = 0, then each infective will be a separate cluster, and R c = Y(0) = R 0 pre by definition. For p c .0 the cluster size will be larger than one, but cluster infectiousness need not be larger, as backwards tracing can reduce the infectious period of superspreaders with large detection time as it does in single-step tracing. For two of the four special cases (a = '), and for the case considered by Müller et al [7] (a = 1, t inf = ', t lat = 0), Y(p c ) can be calculated numerically in MathematicaH. For all other cases, including the real infections, stochastic simulations were needed. Incorporating delays into iterative tracing does not generally change the concept of R c , but it makes the calculation of Y(p c ) very complicated, because there is no longer a single time of cluster quarantine. Two complications arise: first, delays may cause contactees in the chain emanating from a symptomatic infected to become symptomatic themselves before the tracing process reaches them. They then initiate a new tracing process of their own within the same cluster. Second, if the infection process is faster than the tracing process, the cluster size grows infinitely large and iterative tracing becomes ineffective. We determined p c * by stochastic simulation of clusters until quarantine of the final infected (see Supporting Information for all details). Supporting Information Supporting information for the paper "The Effectiveness of Contact Tracing in Emerging Epidemics" Found at: doi:10.1371/journal.pone.0000012.s001 (0.60 MB PDF)
73
The Waiting Time for Inter-Country Spread of Pandemic Influenza
BACKGROUND: The time delay between the start of an influenza pandemic and its subsequent initiation in other countries is highly relevant to preparedness planning. We quantify the distribution of this random time in terms of the separate components of this delay, and assess how the delay may be extended by non-pharmaceutical interventions. METHODS AND FINDINGS: The model constructed for this time delay accounts for: (i) epidemic growth in the source region, (ii) the delay until an infected individual from the source region seeks to travel to an at-risk country, (iii) the chance that infected travelers are detected by screening at exit and entry borders, (iv) the possibility of in-flight transmission, (v) the chance that an infected arrival might not initiate an epidemic, and (vi) the delay until infection in the at-risk country gathers momentum. Efforts that reduce the disease reproduction number in the source region below two and severe travel restrictions are most effective for delaying a local epidemic, and under favourable circumstances, could add several months to the delay. On the other hand, the model predicts that border screening for symptomatic infection, wearing a protective mask during travel, promoting early presentation of cases arising among arriving passengers and moderate reduction in travel volumes increase the delay only by a matter of days or weeks. Elevated in-flight transmission reduces the delay only minimally. CONCLUSIONS: The delay until an epidemic of pandemic strain influenza is imported into an at-risk country is largely determined by the course of the epidemic in the source region and the number of travelers attempting to enter the at-risk country, and is little affected by non-pharmaceutical interventions targeting these travelers. Short of preventing international travel altogether, eradicating a nascent pandemic in the source region appears to be the only reliable method of preventing country-to-country spread of a pandemic strain of influenza.
The emergence of a pandemic strain of influenza is considered inevitable [1] . Provided the emerged strain is not too virulent, it may be possible to eliminate a nascent influenza pandemic in the source region via various combinations of targeted antiviral prophylaxis, pre-vaccination, social distancing and quarantine [2, 3] . If early elimination in the source region is not achieved, then any delay in a local epidemic that a country can effect will be highly valued. To this end, countries may consider introducing non-pharmaceutical interventions such as border screening, promoting early presentation of cases among arriving passengers, requiring the use of personal protective equipment during travels (e.g. the wearing of masks), and reducing traveler numbers. While the case for believing that measures such as these can not stop the importation of an epidemic from overseas has been argued strongly, whether it be SARS or influenza [4] [5] [6] , the extent to which such interventions delay a local epidemic is currently not well quantified, and hence of considerable interest. In this paper we demonstrate how the delay to importation of an epidemic of pandemic strain influenza may be quantified in terms of the growing infection incidence in the source region, traveler volumes, border screening measures, travel duration, inflight transmission and the delay until an infected arrival initiates a chain of transmission that gathers momentum. We also investigate how the delay is affected by the reproduction number of the emerged strain, early presentation of cases among arriving passengers, and reducing traveler numbers. As noted in previous simulation modeling [7] , many aspects of this delay have a significant chance component, making the delay a random variable. Therefore, the way to quantify the delay is to specify its probability distribution, which we call the delay-distribution. Some issues of the delay distribution, such as the natural delay arising in the absence of intervention and the effect that reducing traveler numbers has on this delay has been studied previously [6] [7] [8] . Specifically, if the originating source is not specified, and homogeneous mixing of the worlds population is assumed, then the most likely time to the initial cases arising in the United States is about 50 days assuming R 0 = 2.0 [7] . The additional delay arising from travel restrictions appears minimal until a.99% reduction in traveler numbers [6] [7] [8] . This paper adds to previous work [5] [6] [7] [8] by simultaneously including a wider range of epidemiological factors and possible interventions, such as elevated in-flight transmission, flight duration, the effect of wearing of mask during flight, early presentation of cases among travelers, and quarantining all passengers from a flight with a detected case at arrival. Consider a region in which a new pandemic strain of influenza has emerged, and a region currently free from the infection. We refer to these as the source region and the at-risk country, respectively. Travel between these countries is predominantly via commercial air travel and/or rapid transport which could potentially be subject to border screening and other interventions. We restrict our discussion to air travel. The aim is to assess the effects that a variety of non-pharmaceutical border control measures have, individually and in combination, on the time it takes before the epidemic takes off in the at-risk country. An epidemic is said to have ''taken off'' when it reaches 20 current infectious cases, after which its growth is highly predictable (i.e. nearly deterministic) and the probability of fade-out by chance is very low, if intervention is not enhanced. The source country of origin will undoubtedly have a large impact on the natural delay until importation of an epidemic, although this is difficult to quantify [7] . An alternative is to fix the originating city, for example a highly connected city such as Hong Kong [6] , with the obvious effect that results are highly dependent on the choice. We adopt no specific source region, but assume that the number of international travelers originating from it is reasonably small (see Methods), suggestive of a rural or semirural source region [2] . It is further assumed that the current heightened surveillance for pandemic influenza is continued and that a nascent pandemic with human-to-human transmission is identified and the pandemic is declared when there are 10 concurrent cases in the source region. For an epidemic to take off in an at-risk country, a series of events need to occur. First, the epidemic needs to get underway in the source region. Second, an intending traveler needs to be infected shortly before departure. Third, the infected traveler must actually travel and successfully disembark in the at-risk country. Fourth, the infected traveler, or fellow travelers infected during the flight, must initiate an epidemic in the at-risk country with the infectiousness that remains upon arrival. Finally, the epidemic needs to reach a sufficient number of cases to begin predictable exponential growth. International spread of the emerged pandemic strain of influenza may occur when a recently infected person travels. By 'recently infected' we mean that their travel is scheduled to occur within ten days of being infected. We assume that the number of individuals traveling from the source region to the at-risk country each day is known. The probability that a randomly selected traveler is a recently-infected person is taken to be equal to the prevalence of recently-infected people in the source region on that day. The incidence of infection in the source region is assumed to grow exponentially initially, with the rate of exponential growth determined by the disease reproduction number (the mean number of cases a single infective generates by direct contact) and the serial interval (the average interval from infection of one individual to when their contacts are infected) ( Figure 1A) . The time since infection of a recently-infected traveler is a key component of the calculations, because it affects the chance of positive border screening, the chance of in-flight transmission and the infectivity remaining upon arrival in the at-risk country. The time since infection at the time of scheduled departure is random and the dependence of its probability distribution on the exponential growth rate of infection is illustrated by Figure 1B (see also Supporting Information). The higher the epidemic growth rate in the source region, the greater the probability than an infected traveler will have been infected more recently. It is assumed that individuals detected by departure screening are prevented from traveling. To be detected by screening an infected traveler must be symptomatic and positively screened. An individual is assumed to become symptomatic 48 hours after being infected (cf. [3] who use 1.9 days). The probability of being symptomatic when presenting for departure screening is computed from the curve in Figure 1B . The distribution of the time since infection immediately after departure screening, given that the infected traveler was not detected, is given by the curve in Figure 1C . It contains an adjustment for the probability of being detected at departure. The instantaneous rate at which susceptible contacts are infected depends on the time since infection, and is described by an infectiousness function ( [9] , page 45). We use a peaked infectiousness function, motivated by viral shedding and household transmission data [2] , which has a serial interval of 2.6 days. The basic reproduction number (R 0 ), namely the reproduction number when there is no intervention in place and every contacted individual is susceptible, is given by the area under the infectiousness function. However, our concern is with the effective reproduction number R that holds when various interventions are in place. We obtain any R by simply multiplying the infectiousness function by the appropriate constant (to make the area under the curve equal to R). This keeps the serial interval the same. In the absence of suitable data we assume for most scenarios that the aircrafts ventilation and filtration systems are functioning properly, and that infected travelers transmit the infection at the same rate during a flight as they would while mixing in the community. We examine the sensitivity of this assumption by increasing the inflight transmission by as much as 10-fold (as could potentially happen if air-circulation and filtration systems malfunction, e.g. see [10] ). The in-flight transmission rate is set to zero under the optimistic scenario that all travelers wear 100% effective masks during transit. In terms of a sensitivity analysis this illustrates what would be achievable in a best-case scenario. The number of offspring that an infected traveler infects during a flight is a random variable, taken to have a Poisson distribution with a mean equal to the area under the infectiousness function over to the flight duration. Travelers infected during flights of less than 12 hours duration are asymptomatic at arrival and will not be detected by screening. The probability that an arriving traveler who was infected in the source region is detected on arrival is computed from the distribution of the time since infection on arrival. This distribution is obtained from the curve in Figure 1C by shifting it to the right by an amount equal to the duration of the flight. The distribution of the time since infection for an individual infected in the source region, who passes through arrival screening undetected has a further adjustment for the chance of being detected at arrival ( Figure 1D ). This curve shows that an infected traveler who escapes detection at departure and arrival is highly likely to enter the at-risk country with most, or all, of their infectious period remaining. Authorities are assumed to implement one of two control options when detecting an infected traveler by arrival screening. Under option one (individual-based removal), all passengers who test negative are released immediately and only passengers who test positive are isolated. Under the second option (flight-based quarantining), authorities prevent all passengers from dispersing into the community until the last person has been screened from that flight. Should any one passenger be detected as infected then all passengers will be quarantined, as previously recommended [5] . Transmission chains can be initiated in the at-risk country by infected travelers who mix within the community upon arrival. Suppose now that a flight arrives with one, or more, infected passengers who mix within the community. We classify these infected arrivals into those who are 'pre-symptomatic' and those who are 'symptomatic' at entry. It is assumed that the 'symptomatic' infected arrivals do not recognize their symptoms as pandemic influenza and will not present to medical authorities. In other words, they spend the remainder of their infectious period mixing in the community. On the other hand, the 'presymptomatic' infected arrivals, including all individuals infected during flight, are assumed to mix freely in the community only from entry until they present to medical authorities after some delay following the onset of symptoms. Not all infected travelers entering the community initiate a 'major' epidemic, even when the reproduction number (R) exceeds one. Quite generally, the distribution of the size of an epidemic initiated by an infected arrival is bimodal, with distinct peaks corresponding to a major epidemic and a minor outbreak ( Figure 1E ). In the latter event the outbreak simply fades out by chance despite there being ample susceptibles in the population for ongoing trans- In (B), the step illustrates the probabilistic removal of travelers who have completed their incubation period. In (D), the distribution of time since infection in (C) will have shifted to the right by an amount equal to the flight duration, and cases incubated in-flight may be detected by symptomatic screening, as will those symptomatic cases that were not detected previously. Screening sensitivity for this illustration is 60% on both departure and arrival. (E) Upon entering the community undetected, an infected traveler may initiate a minor (inconsequential) or major epidemic, depending on the characteristics of the disease and public health policy. doi:10.1371/journal.pone.0000143.g001 mission [11] . The number of cases in an outbreak that fades out is typically very small compared to an epidemic. The probability that a typical infective generates a local epidemic is computed by using a branching process approximation [12] for the initial stages of the epidemic, and equating 'epidemic' with the event that the branching process does not become extinct. This calculation is well known (e.g. [13] , page 473), but is modified here to allow for the fact that the process is initiated by a random number of infected arrivals and some of them have spent a random part of their infectious period before arriving in the at-risk country. The distribution for the random number of individuals infected by an infected individual when all their contacts are with susceptible individuals is needed for the calculation. The lack of data prevents a definitive conclusion for the most appropriate offspring distribution for influenza transmission [14] , and we use a Poisson distribution with a mean equal to R, discounted for individuals who spent only some of their infectious period mixing in the at-risk country. A Poisson offspring distribution is appropriate when the area under the infectiousness function is non-random (i.e. all individuals have the same infection 'potential'). We assume that R is the same in the source region and the at-risk country. For an undetected infected traveler and all their in-flight offspring to fail to initiate an epidemic on arrival, all of the chains of transmission they initiate must fail to become large epidemics (see Supporting Information). We calculate the probability distribution of D, the total delay until an epidemic gathers momentum by noting that it is given by D = D 1 +D 2 , where D 1 is the time until an epidemic is first initiated and D 2 is the time from initiation until the local epidemic gathers momentum. For an epidemic to be first initiated in the at-risk country on day d, it must have not been initiated on all previous days. Hence the probability distribution of the time delay (D 1 ) until the epidemic is first initiated in the at-risk country following identification in the source region is described by: Pr (D 1~d )~ p p 1 p p 2 p p 3 ::: p p d{1 p d where p d denotes the probability that the epidemic is initiated on day d , and p p d~1 {p d denotes the probability that the epidemic is not initiated on day d (see Supporting Information for calculation of p d ). Once successfully initiated, an epidemic may initially hover around a handful of cases before reaching a sufficient number of cases for its growth to become essentially predictable. As mentioned, 20 concurrent cases is our criterion for an epidemic to have gathered momentum. We determine the distribution of D 2 , the time to this occurrence, from 10,000 stochastic simulations and approximate this empirical distribution by a shifted gamma distribution. Our criterion of 20 concurrent cases is conservatively high, as results from the theory of branching processes shows that the probability of a minor epidemic (and hence no take-off) starting from 20 concurrent cases is about 3610 28 when R = 1.5, and even smaller for higher values of R. Finally, the distribution of the total delay (D = D 1 +D 2 ) from the pandemic being identified in the source region until 20 cases in the at-risk country was calculated by the convolution of the distributions of D 1 and D 2 . For the illustrative purposes, we chose values of 1.5, 2.5 and 3.5 for R, which encompass estimates proposed for previous pandemics [2, 3, 15] . The number of people within the infected source region was assumed reasonably small (5 million), and there was one flight per day traveling from the source region to the at-risk country carrying 400, 100 or 10 passengers. A higher number of travelers affects the delay only marginally, assuming the epidemic takes off in the source region (see Results). We assume a typical travel duration between attempted departure and possible arrival of 12 hours, but also examine the effect of varying this from 0-48 hours. The time to presentation following symptom onset is varied from 'immediately' to 'never presenting', with a time of 6 hours considered likely in the presence of an education campaign. The sensitivity of symptomatic screening is varied from 0-100%, with results presented for 0, 50 and 100% sensitivity. The probability that a recently infected traveler evades screening is substantial even if screening reliably detects symptomatic travelers (Figure 2A) , because the typical travel duration is shorter than the 2-day incubation period. In addition, during the early stages of the epidemic a high R in the source region acts to increase the probability that an infected traveler has been infected quite recently and hence will escape detection due to being asymptomatic during their travels (Figure 2A ). For example, assuming 100% sensitivity for detecting symptomatic infection, we calculate that during the early stages of the epidemic the proportion of infected travelers that evade both departure and arrival screening after 12 hours of travel is 0.26, 0.45 and 0.59 for disease reproduction numbers 1.5, 2.5 and 3.5, respectively. As the duration of travel approaches the disease incubation period, effective symptomatic screening substantially reduces the likelihood that a traveler evades screening and initiates an epidemic ( Figure 2B ). Reducing the time from the onset of symptoms to presentation (and subsequent isolation) for each infected arrival also reduces the probability that a major epidemic is initiated, however the best case scenario of infected travelers and all their in-flight offspring presenting immediately following the onset of symptoms still poses a substantial risk of epidemic initiation arising from pre-symptomatic transmission ( Figure 1C ). The delay contains a fairly substantial natural component, primarily due to the time it takes to increase the number of infectives in the source region sufficiently to make the chance of a recently infected traveler appreciable ( Figure 3A ), and the time (D 2 ) it takes for a local epidemic in the at-risk country to gather momentum following successful seeding ( Figure 4A ). In the absence of any interventions, the number of infected individuals who successfully enter the community of the at-risk country initially increases exponentially ( Figure 3A ). With individual-based removal of infected travelers, the number of individuals entering the at-risk country undetected by screening is proportionately reduced over the course of the epidemic ( Figure 3A ). With flightbased quarantining, the number of infected individuals entering the at-risk country undetected is dramatically reduced over the course of the epidemic, even for relatively insensitive screening ( Figure 3A ). With flight-based quarantining, the number of infected passengers slipping through undetected is bimodal, with the first peak occurring when the number of infected travelers attempting to travel is still in single figures. Without screening, the daily probability that an epidemic is initiated (p d ) increases, and becomes near certain once the number of infected travelers arriving undetected exceeds about 10 ( Figure 3B, solid line) . With screening and individual-based removal of infected individuals, p d follows a similar pattern only reduced somewhat. With screening in combination with flightbased quarantining, this probability is changed dramatically. After an initial rise it dips, to become essentially zero during the height of the epidemic in the source region ( Figure 3B , dotted line). This arises because once a flight has several infected travelers, the probability that at least one is detected approaches one (even if screening is imperfect), and all passengers on such a flight are quarantined. Once the epidemic starts to wane in the source region (assuming the unlikely event of the pandemic strain is restricted to the source region), the probability of initiation rises once again. The corresponding distribution of D 1 , the delay until The effects of R and the time from symptom onset to presentation on the probability that an infected traveler, having entered the wider community following arrival, will initiate an epidemic. There is no screening. doi:10.1371/journal.pone.0000143.g002 the epidemic is first initiated in the at-risk country, is bi-modal in the presence of screening ( Figure 3C) . Although flight-based quarantining is effective in preventing the entry of infected travelers during the height of the epidemic, a substantial cumulative risk of initiation has already occurred before this from the handful of infectives that have slipped through undetected ( Figure 3B ). Hence, whilst the effect of border screening, particularly in conjunction with flight-based quarantining, on the daily probability of initiation is dramatic, its effect on the delay to initiation is much less pronounced ( Figure 3C ). Border screening, even with perfect sensitivity for detecting symptomatic cases, tends to increase D 1 , the time to an epidemic being initiated, by a matter of days to weeks. The time (D 2 ) from initiation (the arrival of the index case) to an epidemic reaching 20 concurrent cases within the at-risk country is adequately modeled using a shifted Gamma distribution ( Figure 4A ). The convolution of this right-skewed Gamma distribution with the left-skewed delaydistribution of D 1 (Figure 3C ) yields the distribution for D, the total delay until the epidemic reaches 20 cases in the at-risk country ( Figure 4B ). The distribution of D is approximately symmetrical. The effect of border screening on the total delay D is quite modest, though sensitive to how screening is implemented. For example, with R = 1.5 and 400 travelers per day, 100% sensitive screening with individual-based removal increases the median delay from 57 to 60 days ( Figure 4B ). Flight-based quarantining would extend the median delay to 70 days. In general, the added delay arising from flight-based quarantining is about four-fold that arising from individual-based removal. The natural component of the delay is highly sensitive to the disease reproduction number ( Figure 5A ). For example, with 400 passengers per day departing the source country and in the absence of any interventions, the median delay ranges from a low of 17 days for R = 3.5 to 57 days for R = 1.5 ( Table 1 ). The delay is less sensitive to the number of intending travelers, with little appreciable increase in the median delay occurring until traveler numbers become very low ( Figure 5B ). For example, if R = 1.5, with no other border control measures, decreasing the number of intending travelers departing the source region from 400 to 100 per day increases the median total delay D from 57 to 66 days. A further decrease in the number of intending travelers to 10 per day increases the median delay to 83 days ( Table 1) . The delay is quite insensitive to the rate of transmission in-flight. For example, with R = 1.5, a 12-hour flight, 400 travelers per day and no other interventions, preventing in-flight transmission altogether increases the median delay from 57 to 58 days. Conversely, doubling the rate of in-flight transmission reduces the median delay from 57 to 56 days. A 10-fold increase in the rate of transmission in-flight only decreases the median delay from 57 to 53 days. Encouraging the early presentation of cases among travelers following the onset of symptoms has a limited effect on the delay distribution ( Figure 5C ). For example, for R = 1.5, 400 intending travelers per day and no other interventions, reducing the time to presentation from 'never presenting' to 6 hours increases the median delay from 57 to 61 days. Immediate presentation at symptom onset only increases the median delay a further day in this scenario. In general, the additional delay achieved by introducing nonpharmaceutical border control measures is generally small in comparison with the natural delay ( Figure 5D ). For the scenario with R = 1.5 and 400 intending travelers per day, a combination of 100% flight-based quarantining, 100% compliance with mask wearing during travel and immediate presentation at symptom onset extends the estimated median delay from 57 to 79 days ( Figure 5D ). This added delay diminishes in absolute terms as R increases. For example, if the same interventions are applied with R = 3.5, the median delay is extended from 17 to just 20 days ( Figure 5D ). The one exception to this generalisation is when travel numbers are reduced dramatically. The added delay achieved when a drastic reduction in travel numbers is combined with other border control measures appears to be greater than adding the delays each achieves on its own. For example, if R = 1.5, and we reduce the number of intending travelers from 400 to 10 per day, implement 100% flight-based quarantining, implement compulsory mask wearing during travel and presentation at 6 hours following symptom onset then there is a substantial probability (0.74) that the pandemic strain will never be imported (assuming the epidemic is confined to the source country). The estimated quartile delay (the median in this case is undefined) to the start of a major epidemic in an at-risk country is extended from 50 to 125 days. Again, the added delay decreases rapidly as R increases, and if the above interventions were applied with R = 3.5, the estimated median delay is extended from 17 to 26 days, and the importation of the epidemic is certain ( Figure 5D ). We have formulated a model of the importation of an infectious disease from a source region to an at-risk country that permits a comprehensive analysis of the effect of border control measures. Our results are most relevant to the early stage of a pandemic when most cases are contained within a single source region. Once the pandemic has spread to several countries, models with greater complexity and ability to more realistically model global mixing patterns [6] [7] [8] are required. Our model is developed with a pandemic-strain of influenza in mind, but could apply to any emerging infectious disease that is transmitted from person to person. We have assumed a Poisson distribution for the number of secondary infections, which a natural choice when each infected individual has the same infectivity profile. A distribution with a larger variance is appropriate when individuals vary substantially in their infectiousness. Our results are conservative in the sense that they give an upper bound for the probability that an infected traveler manages to initiate an epidemic, compared to an offspring distribution with a greater variance but the same reproduction number [14] . The nature of the next pandemic influenza virus, and particularly its reproduction number, is uncertain. If its reproduction number is low (R,2.0), our results indicate that at-risk countries receiving a reasonably small number of travelers (say 400 per day) from the infected source region can expect a natural delay until importing an epidemic of the order of 2 months. This is quite variable and under favourable conditions it could be 4 months. However, the natural delay decreases rapidly as R increases. The additional delay from isolating individuals detected by border screening is merely a few days under most plausible scenarios, even if both departure and arrival screening is introduced and screening detects every symptomatic traveler. While the extra delay is more than quadrupled if flights with a detected case(s) are quarantined, the effect remains modest (weeks at most) and it is questionable whether the extra delay achieved warrants the disruption created by such a large number of quarantined passengers. In-flight transmission is a commonly raised concern in discussions about the importation of an infection, so inclusion of in-flight transmission is an attractive feature of our model. Events of substantial in-flight transmission of influenza have been documented [10, 16] and modeling of indoor airborne infection risks in the absence of air filtration predicts that in-flight transmission risks are elevated [17] . However, it difficult to estimate the infectiousness of influenza in a confined cabin space, as there is undoubtedly substantial under-reporting of influenza cases who travel and fail to generate any offspring during flight. Provided the aircraft ventilation system (including filtration) is operational, it is considered that the actual risk of in-flight transmission is much lower than the perceived risk [18] . Our results indicate that the delay is relatively insensitive to the rate of in-flight transmission, making in-flight transmission less of an issue than commonly believed. A highly elevated transmission rate inflight will hasten the importation of an epidemic only marginally. Consistent with this, eliminating in-flight transmission by wearing protective masks increases the delay only marginally. Early presentation by infected arrivals not detected at the borders was found to add only a few days to the delay. To some extent this arises due to our assumption that pre-symptomatic transmission can occur, for which there is some evidence. In contrast, Ferguson et al. [2] assume that the incubation and latent periods are equal, with a mean of 1.5 days. In their model presymptomatic transmission is excluded and infectiousness is estimated to spike dramatically immediately following symptom onset and declining rapidly soon afterwards. Under their model assumptions, immediate presentation at onset of symptoms would reduce transmission effectively. However, as presentation occurs some time after onset of symptoms and the bulk of infectivity occurs immediately after onset of symptoms the results on the effect of early presentation of cases are likely, in practical terms, to be similar to those found here. Given the variable nature of influenza symptoms, there is likely to be a difference between the onset of the first symptoms as measured in a clinical trial (e.g. [19] ) and the time that a person in the field first suspects that they may be infected with influenza virus. To fully resolve the issue of how effective very early presentation of infected travelers is in delaying a local epidemic we need better knowledge about the infectiousness of individuals before and just after the onset of symptoms. It is assumed that the pandemic is identified and declared when there are 10 concurrent cases in the source region attributed to human-to-human transmission, and that screening is applied at both departure and arrival. The time between screening events is assumed to be 12 hours and infected travelers are not isolated following the onset of symptoms. Of the border control measures available, reducing traveler numbers has the biggest effect on the delay and even then it is necessary to get the number of travelers down to a very low number. An equivalent control measure is to quarantine all arriving passengers with near perfect compliance. Our results indicate that short of virtually eliminating international travel, border control measures add little to avoiding, or delaying, a local epidemic if an influenza pandemic takes off in a source region. All forms of border control are eventually overwhelmed by the cumulative number of infected travelers that attempt to enter the country. The only way to prevent a local epidemic is to rapidly implement local control measures that bring the effective reproduction number in the local area down below 1, or to achieve rapid elimination in the source region, in agreement with other recent studies [6] [7] [8] . Preventing the exponential growth phase of an epidemic in the source region appears to be the only method able to prevent a nascent influenza pandemic reaching atrisk countries. Text S1 Estimating the daily probability of epidemic initiation Found at: doi:10.1371/journal.pone.0000143.s001 (0.08 MB PDF)
74
Neutralizing Antibody Fails to Impact the Course of Ebola Virus Infection in Monkeys
Prophylaxis with high doses of neutralizing antibody typically offers protection against challenge with viruses producing acute infections. In this study, we have investigated the ability of the neutralizing human monoclonal antibody, KZ52, to protect against Ebola virus in rhesus macaques. This antibody was previously shown to fully protect guinea pigs from infection. Four rhesus macaques were given 50 mg/kg of neutralizing human monoclonal antibody KZ52 intravenously 1 d before challenge with 1,000 plaque-forming units of Ebola virus, followed by a second dose of 50 mg/kg antibody 4 d after challenge. A control animal was exposed to virus in the absence of antibody treatment. Passive transfer of the neutralizing human monoclonal antibody not only failed to protect macaques against challenge with Ebola virus but also had a minimal effect on the explosive viral replication following infection. We show that the inability of antibody to impact infection was not due to neutralization escape. It appears that Ebola virus has a mechanism of infection propagation in vivo in macaques that is uniquely insensitive even to high concentrations of neutralizing antibody.
Editor's note: The potential efficacy of pre-and post-exposure prophylaxis against Ebola virus infection, as well as the fundamentally important question of whether neutralizing bodies are important for Ebola virus resistance, is addressed by a related manuscript in this issue of PLoS Pathogens. Please see doi: 10 .1371/journal. ppat.0030002 by Feldmann et al. Passive transfer of relatively high concentrations of neutralizing antibodies can protect against challenge with a range of viruses in animal models and in humans [1] [2] [3] . Protection in some cases is in the form of sterilizing immunity, i.e., no viral replication is observed following challenge [2, 4, 5] . In other cases (e.g., [5, 6] ), some replication is observed but protection from disease is achieved, presumably because neutralizing antibody sufficiently blunts infection for T cell and innate immunity to resolve infection [7] . It might be expected that passive neutralizing antibody would be most effective against challenge with acute viruses. Many acute viral infections are resolved even in the absence of neutralizing antibody, and the blunting effect of passive antibody would provide more time for the development of effective cellular immune responses. In contrast, chronic viruses may present a greater challenge to passive antibody, since, in the absence of sterilizing immunity, there is a window of opportunity for the virus to establish a chronic infection before cellular immunity can be mobilized. Ebola virus (EBOV) causes a severe acute infection in humans [8] . Infection with the Ebola Zaire strain, Zaire ebolavirus (ZEBOV), produces mortality in the range of 60%-90% [9] with death generally occurring around 7-11 d following the appearance of symptoms [8] . There is a single report describing the use of convalescent sera to treat EBOV infection [10] . However, the patients in this report may have already been through the worst stages of the disease, and it is not clear that serum antibodies were responsible for their recovery [10] . Further, neutralizing antibody titers in survivors of EBOV infection tend to be rather low, although we have isolated a neutralizing human monoclonal antibody (mAb), KZ52, of good potency from a convalescent individual [11] . The ability of passive antibody to protect against EBOV has been investigated in a number of animal models. The guinea pig and mouse models use EBOVs that have been serially passaged to adapt to replication in the respective animals and are highly lethal. Protection has been demonstrated in the guinea pig model using neutralizing horse, sheep, and goat immunoglobulin G (IgG) against EBOV [12, 13] and the human anti-EBOV GP mAb, IgG KZ52. This antibody neutralizes ZEBOV (1995, Kikwit) with a 50% inhibitory concentration (IC 50 ) of 0.05-0.3 lg/ml and an IC 90 of 0.5-2.6 lg/ml in Vero cells [11, 14] and an IC 50 of approximately 0.05-1 lg/ml and a IC 90 of 0.5-2 lg/ml in primary human monocytes/macrophages [14] . We showed that when administered subcutaneously at a dosage of 25 mg/kg up to 1 h after challenge, the antibody protects against robust ZEBOV challenge (10,000 plaque-forming units [pfu] ) in the guinea pig model [6] . Macaques provide a model of EBOV infection that is likely closer to human infection. The human virus can be used directly in macaques without need for adaptation and the course of disease mirrors that seen in humans [8] . In cynomolgus macaques (Macaca fascicularis), ZEBOV infection produces a mortality rate of 100% with death occurring 6-8 d following infection with 1,000 pfu [15] , while in rhesus macaques (Macaca mulatta) ZEBOV produces about 100% mortality with death occurring 7-10 d after infection with 1,000 pfu [16] . In contrast to the guinea pig experiments, the passively transferred polyclonal equine neutralizing IgG described above provided only some minor benefit in the form of a slight delay in the onset of viremia from day 5 to day 7 [13] following ZEBOV challenge of cynomolgus monkeys. No significant reduction in mortality was observed. However, protection against EBOV in primates has been observed in a low dose challenge model. Thus, neutralizing equine IgG protected baboons from ,30 LD 50 (50% lethal dose) ZEBOV challenge when the IgG was given up to 1 h after infection and the serum contained high neutralizing antibody titers (1:128 to 1:512) [17, 18] , and, similarly, neutralizing ovine serum protected baboons against 0.6 LD 50 ZEBOV challenge [19] . Here, we studied the ability of passively transferred neutralizing human mAb KZ52 to protect against ZEBOV challenge in rhesus macaques. This passive transfer failed to protect the macaques against challenge with ZEBOV, and, furthermore, had a minimal effect on the explosive viral replication following infection. We showed by ELISA that antibody was present at high levels in serum of the monkeys and that neutralization escape was not responsible for the resistance of virus to antibody prophylaxis. To evaluate whether IgGl KZ52 could protect against Ebola virus infection in a nonhuman primate animal model, antibody was passively transferred to rhesus macaques followed by challenge with the 1995 ZEBOV (Kikwit) isolate 24 h later. Protection against virus challenge by neutralizing antibodies in naive animals often requires high doses of antibody [2] . Therefore, we used a high dose of 50 mg/kg KZ52, which was close to the maximum practically achievable. In addition, we gave a second bolus of 50 mg/kg of KZ52 on day 4 following infection. The results for the four antibodytreated animals show a steady increase in plasma viremia up to 10 5 -10 7 pfu/ml on day 7 ( Figure 1 ). These levels of plasma viremia closely parallel those seen in the control animal and typically seen in historical controls [15] . The second bolus of antibody given on day 4 did not appear to have any impact upon the rate of increase of plasma virus ( Figure 1 ). Three of the treated animals were euthanized when moribund at day 9 or 10 post infection. The fourth treated animal showed a decrease in plasma viral load after the peak and survived to day 28 before becoming moribund when it too was euthanized. Although monkey CH46 had less severe symptoms than the other animals in the study, it was concluded that this animal, too, was suffering from disease due to ZEBOV, as evidenced, for example, by copious Ebola virus antigen in the lungs (see below). Serum antibody (KZ52) loads were measured 1 d before virus challenge (day À1) and 4 d after challenge but before antibody boosting (day 4) by an ELISA designed to detect KZ52 as a human antibody that has bound to immobilized ZEBOV glycoprotein. The serum KZ52 antibody levels on day 4 were in the approximate range 200-400 lg/ml ( Table 1 ). The Figure 1 . Plasma Viremia in Macaques Challenged with ZEBOV Shown is the measured viremia, in log 10 pfu per ml, for four antibodytreated monkeys (CH46, CH56, CH57, and CH83) and one untreated control animal (EHD) at days 4, 7, 9, and 10 in plasma by plaque assay as described in Materials and Methods. 50 mg/kg of KZ52 IgG1 human antibody [11] was given intravenously to four rhesus macaques 1 d before and again 4 d after challenge with 1,000 pfu (intramusculary) of the 1995 ZEBOV (Kikwit) isolate. Ab, antibody. doi:10.1371/journal.ppat.0030009.g001 PLoS Pathogens | www.plospathogens.org January 2007 | Volume 3 | Issue 1 | e9 0063 Anti-Ebola Antibody Activity In Vivo Ebola virus is one of the most feared of human pathogens with a mortality that can approach 90% and an extremely rapid disease course that can lead to death within days of infection. Antibodies able to inhibit viral infection in culture, neutralizing antibodies, can typically prevent viral infection in animals and humans when present prior to infection, at sufficient concentration. Such neutralizing antibodies may be provided through passive administration or induced by vaccination. We have previously shown that a human neutralizing antibody can protect guinea pigs against Ebola virus. However, here we show that this antibody does not protect monkeys against Ebola virus and surprisingly appears to have very little impact upon the rapid course of infection, despite being present at very high levels in the blood of the monkeys. We conclude that administering antibody prior to or immediately following exposure to Ebola virus, for example, after an accident in a research setting or a bioterrorist attack, is unlikely to be effective in preventing disease. Recent successes in protecting monkeys against Ebola virus through vaccination may be independent of antibody, or, more likely, critically dependent on the cooperation of antibody and cellular immunity. two control monkeys, EHD (untreated and challenged), and a negative monkey (neither treated nor challenged), had very similar background levels of reactivity to the glycoprotein and anti-human antibody as each of the treated monkeys before treatment. A 50-mg/kg dose typically produces serum mAb concentrations in animals on the order of 500 lg/ml after injection [6] . Since the neutralization titer of KZ52 (IC 90 ) for ZEBOV is on the order of 0.5-2.5 lg/ml, depending upon the target cell and the presence of complement, the concentrations of KZ52 in the animals at the time of challenge and for the first few days were, as expected, greater than, or on the order of 100 3 IC 90 . These concentrations typically provide sterilizing immunity against challenge by a number of viruses [4, 20] . One formal possibility is that the neutralizing antibody has little effect on the course of infection in the treated monkeys because of the rapid emergence of neutralization escape mutants. Accordingly, virus was isolated from a selection of plasma from day 4 (monkeys CH56, CH57, and CH83) and day 7 (monkeys CH56 and CH57). All of these viruses were sensitive to KZ52 so that essentially 100% neutralization was observed in vitro at 40 and 400 lg/ml KZ52 in a plaque assay (see Materials and Methods) using Vero E6 target cells. In order to gain a better understanding of any differences in pathology between the control and antibody-treated animals, the levels of virus in different organs were surveyed postmortem. Viral levels in the liver, spleen, kidney, adrenal glands, lung, and mesenteric and inguinal lymph nodes were high (10 4 -10 6 pfu/g) in the control and three of the four treated animals (Table 2 ). However, monkey CH46, who survived much longer than the other animals (to day 28) showed some major histopathological differences from the other infected monkeys and from the norm for ZEBOV infection [15] . Relatively low viral levels were observed in most of the organs of CH46, and none in the liver, spleen, and adrenal glands. In addition, large immunoreactive monocytes were found in the blood of monkeys CH56, CH57, and CH83, but not in CH46 (Figure 2) . Typically, smaller immunoreactive monocytes are seen with ZEBOV infection [15] . The presence of large immunoreactive monocytes may simply reflect uptake of antibody-coated virions via Fc receptors and subsequent viral clearance. However, it is interesting to note that previous studies have implicated mononuclear phagocytes as vehicles for transport of filovirus particles to specific organs such as liver and spleen [21] [22] [23] [24] . If virus particles could remain infectious following Fc receptormediated uptake in a subset of cells (compare DC-SIGN mediated uptake of HIV-1 by dendritic cells [25] ), then the course of disease in monkeys CH56, CH57, and CH83 might represent the net result of inhibition by neutralization and enhancement by antibody-mediated cellular uptake. Interestingly, monkey CH46, who fared somewhat better than the others and lacked virus in the liver and spleen, did not show the presence of large immunoreactive monocytes. This is suggestive of lowered Fc receptor-mediated uptake of virus or reduced activity of the mononuclear phagocytic system. Here, we describe the case of a potent neutralizing human monoclonal antibody, administered to give a high serum concentration, which is shown to be unable to protect macaques against challenge with a lethal dose of ZEBOV. The antibody appeared to have very little effect on the course of virus replication or disease in three of four treated animals. Neutralization escape does not appear to explain the lack of protection observed by the antibody. In one animal, a more limited infection was observed, but this macaque did also eventually succumb to the effects of viral disease. The challenge dose of 1,000 pfu used corresponds to the amount of EBOV contained in a relatively small quantity of fluid (on the order of 1 ll) from an infected individual given the high titers of virus typically found in such individuals (on the order of 10 6 pfu/ml of blood, for example). Therefore, the challenge dose was not unreasonable in terms of a natural exposure to virus. The negative results with passive antibody contrast strongly with recent successes in preventing EBOV infection in macaques through vaccination [26, 27] . Does this mean that neutralizing antibody is unimportant in vaccine protection? The answer to this question must await further studies. However, a plausible hypothesis to explain all the data would still allow for an important contribution of antibody to vaccine protection. This hypothesis would argue that passive antibody is unable to completely block all EBOV entry to cells, and once a few cells are infected, virus replication is so explosive that it cannot be contained by a de novo generated cellular immune response. Vaccination, on the other hand, will provide CD8þ memory T cells that can be rapidly recruited to become effector cells and limit infection. Certainly, although mAb KZ52 was able to provide protection from disease following ZEBOV challenge in the guinea pig model, immunity was not sterilizing and some viral replication was noted [6] . Since 1 pfu of EBOV is a lethal dose for primates [8] , a failure of passive antibody to achieve sterilizing immunity may be critical. Immunohistochemistry, as discussed above, gives some intriguing hints that uptake of antibody-coated virions by monocytes may possibly have a role to play in the course of infection following antibody treatment. We note, however, that previous in vitro studies using isolated human monocytes/macrophages did not find evidence of infectivity-enhancing antibodies [14] . More detailed in vitro and in vivo investigations will be required before any firm conclusions can be drawn. We also note that our experiments were carried out with a single neutralizing monoclonal antibody. It is possible that a more favorable outcome may have been apparent for a combination of neutralizing antibodies or even a combination of neutralizing and nonneutralizing antibodies [2] . However, these possibilities should be weighed against the very high concentrations of neutralizing monoclonal antibody used in the experiments and the efficacy of the antibody in the guinea pig model. In summary, the inability of high concentrations of neutralizing antibody to even slow viral replication in infected macaques is remarkable and implies a mechanism of infection propagation that is virtually insensitive to antibody. Overall, the results suggest that monoclonal antibody prophylaxis or post-exposure prophylaxis alone are unlikely to be effective strategies in protecting against EBOV, for example, following a needle-stick accident in a research setting or a bioterrorist attack. Passive transfer experiment. 50 mg/kg of KZ52 IgG1 human antibody [11] was given intravenously to rhesus macaques (weight, 3.9-4.4 kg) 1 d before challenge (day 0) with 1,000 pfu intramuscularly of the 1995 ZEBOV (Kikwit) isolate and again 4 d later (day þ4). One monkey was not given any antibody treatment. The animals were carefully monitored for signs of disease, and Ebola virus plasma viremia was determined at days 4, 7, 9, and 10 in serum by plaque assay as described below. The investigators adhered to the Guide for the Care and Use of Laboratory Animals when conducting research, using animals [28] . The United States Army Medical Research Institute of Infectious Diseases (USAMRIID) animal facilities and animal care and use program are accredited by the Association for Assessment and Accreditation of Laboratory Animal Care International. All infectious material and animals were handled in a maximum-containment biosafety level 4 facility at USAMRIID under standard operating conditions. Antibody purification. IgGl KZ52 was produced and purified as described by Parren et al. [29] and was .98% pure, as determined by sodium dodecyl sulfate-polyacrylamide gel electrophoresis and contained ,1 IU of endotoxin/ml, as determined in a quantitative chromagenic Limulus amoebecyte lysate assay (BioWhittaker, Cambrex, http://www.cambrex.com). Viremia determined by plaque assay. Plasma viremia and viral loads in organs was determined by virus titration in a conventional plaque assay on Vero E6 cells, as described elsewhere [13, 21] . Neutralization assay. Samples were diluted into Eagle's minimal essential medium (EMEM; Invitrogen, http://www.invitrogen.com) with 5% heat-inactivated fetal bovine serum (FBS). In the presence and absence of a constant dilution of 1:10 or 1:100 of 4 mg/ml KZ52 (thus, 0.4 mg/ml or 0.04 mg/ml), plasma were titrated from 10 À1 to 10 À6 dilutions. Viremia was determined by counting pfu on Vero E6 cell monolayers. Cells grown to confluence in 6-well plates were given 0.2 ml of plasma with and without additional KZ52. The titrated samples were incubated in the presence of KZ52 for 1 h at 37 8C in 5% CO 2 . After absorption, the cells were overlaid with 2 ml of EMEM containing 5% FBS, 25 mM HEPES buffer, 50-lg gentamicin per ml, and 1% agarose. After 10 d, plaques were visible and the cells were removed from the humidified 37 8C incubator to visualize plaques with an inverted phase microscope. 2 ml of neutral red (1:6,000 final concentration; Sigma-Aldrich, http://www.sigmaaldrich.com) was added to each well, and after an additional 24-h incubation, the plaques were counted [11, 30] . ELISA. Nunc-Immuno Maxisorp enzyme-linked immunosorbent assay (ELISA) plates (Nunc, http://nuncbrand.com) were coated with 100 ll/well of 10 lg/ml lectin from Galanthus Nivalis (Sigma-Aldrich) in PBS and incubated overnight at 4 8C. The plates were then blocked in phosphate-buffered saline (PBS) containing 10% FBS for 2 h. The wells were then washed twice with wash buffer, PBS containing 0.2% Tween 20 (Sigma-Aldrich). 293 cells were transfected with a mammalian expression plasmid coding for transmembrane domaindeleted Ebola glycoprotein. Supernatant (100 ll) from these cells, which contains 0.8-1.3 mg/ml total protein, was used to coat each well for 1 h at room temperature (RT) after the blocking solution was removed from the ELISA plates. Plates were then washed six times with wash buffer. Monkey sera were added in 10-fold dilutions from 1:10 to 1:10 5 in dilution buffer (PBS containing 1% BSA and 0.02% Tween) and incubated at RT for 1 h. The wells were then washed six times, and a secondary antibody alkaline phosphatase-conjugated goat anti-human immunoglobulin G (IgG) against the F(ab9) 2 portion of the antibody (Pierce, http://www.piercenet.com) diluted 1:500 was added, and this was incubated for 1 h. Finally, the plates were washed again six times and developed by one tablet of phosphatase substrate Typically smaller immunoreactive monocytes are seen with ZEBOV infection [15] . Immunohistochemistry was performed as described in Materials and Methods. doi:10.1371/journal.ppat.0030009.g002 (Sigma-Aldrich) in 5 ml of alkaline phosphatase stain buffer (pH 9.8) per plate. The assay was performed as per manufacturer's directions. The plates were read at an optical density of 405 nm on a microplate reader (Molecular Devices, http://www.moleculardevices.com) at 30 min after adding substrate. A panel of normal sera was run each time the assay was performed. Immunohistochemistry. Sections were pretreated with Dako Ready to Use Proteinase K (Dako, http://www.dako.com) for 6 min at RT after deparaffinization and rehydration through a series of graded ethanols. Blocking was performed with Dako's Serum-Free Protein Block for 20 min pre-antibody exposure. The tissue sections were then incubated overnight at 4 8C in primary antibody using an equal mixture of mouse monoclonal antibodies to EBOV GP and VP40 (1:5,000). An alkaline phosphatase-labeled polymer (Dako Envision System, alkaline phosphatase) was incubated on the sections for 30 min, and then color was developed by exposing tissue to 6-bromo-2hydroxyl-3-naphtholic acid (HistoMark Red; Kikegaard and Perry Laboratories, http://www.kpl.com) substrate for 50 min in the dark. Counterstaining was done with hematoxylin. Positive controls included archived EBOV-infected cynomolgus tissue, and negative controls included replicate sections exposed to anti-Marburg virus antibodies and uninfected cynomolgus macaque tissue [15] .
75
Public health preparedness in Alberta: a systems-level study
BACKGROUND: Recent international and national events have brought critical attention to the Canadian public health system and how prepared the system is to respond to various types of contemporary public health threats. This article describes the study design and methods being used to conduct a systems-level analysis of public health preparedness in the province of Alberta, Canada. The project is being funded under the Health Research Fund, Alberta Heritage Foundation for Medical Research. METHODS/DESIGN: We use an embedded, multiple-case study design, integrating qualitative and quantitative methods to measure empirically the degree of inter-organizational coordination existing among public health agencies in Alberta, Canada. We situate our measures of inter-organizational network ties within a systems-level framework to assess the relative influence of inter-organizational ties, individual organizational attributes, and institutional environmental features on public health preparedness. The relative contribution of each component is examined for two potential public health threats: pandemic influenza and West Nile virus. DISCUSSION: The organizational dimensions of public health preparedness depend on a complex mix of individual organizational characteristics, inter-agency relationships, and institutional environmental factors. Our study is designed to discriminate among these different system components and assess the independent influence of each on the other, as well as the overall level of public health preparedness in Alberta. While all agree that competent organizations and functioning networks are important components of public health preparedness, this study is one of the first to use formal network analysis to study the role of inter-agency networks in the development of prepared public health systems.
International and national events have brought critical attention to the Canadian public health system and how prepared the system is to respond to various types of con-temporary public health threats. Whether those threats result from emerging infectious diseases or bioterrorism, the public health system is responsible for protecting the health of the population. Outbreak response occurs first at the local or regional levels and thus the potential impact that a threat will have on the overall Canadian population can differ significantly depending on the capacity of public health systems to respond [1] . For example, following the SARS outbreak in Toronto, the Canadian National Advisory Committee (NAC) on SARS and Public Health held it as fortunate that the SARS outbreak struck primarily in Toronto and not in other parts of Canada where the capacity to combat public health threats is limited [2] . While a "general renewal of the public health infrastructure" and its capacity to respond are in order the effective coordination of all agencies at regional, provincial, and federal levels is necessary to assure a comprehensive surveillance and management of public health threats. As Ralph Klein, the Premier of Alberta, commented: "Albertans and all Canadians understand a disease outbreak like SARS or West Nile virus in one region affects other parts of the country. Coordinated approaches will help deal with public health threats." Despite the commitment of federal and provincial agencies in Canada to improve public health system performance and develop overall public health preparedness, little is still known about the current systems-level state of public health preparedness in Canada. While coordination is seen as a necessary element for public health preparedness, few measures of inter-organizational collaboration relevant to public health preparedness have been developed and assessed in relation to other features of the organizational and inter-organizational environment. This study protocol presents the premises, conceptual model and methods used to measure and assess public health preparedness in Alberta, Canada. Although provincial public health agencies are themselves embedded within federal and global public health preparedness systems, we focus primarily on the organizational environment of public health preparedness in Alberta so as to develop contextually-relevant measures and evaluation techniques as well as address issues related to potential cross-Alberta variations in public health preparedness. In Alberta, public health and emergency response procedures are primarily governed under the Disaster Services Act [3] and the Public Health Act [4] . Provincial response to a public health disaster event may thus call into action organizations and units under the jurisdiction of local municipalities, regional health authorities, emergency management districts, and/or the provincial government. There are a number of organizational actors that may be involved in preparation, response, or recovery phases of a public health disaster, including departments or portfolios within agencies. Each of which are potential actors in the provincial public health and emergency response network. While we consider the overall public health and emergency management network to consist of all poten-tial organizational actors involved in responding to public health threats, the precise set of organizations that come together around a particular public health threat depends on the nature of that threat. For example, pandemic influenza and the West Nile virus are both emerging infectious diseases but the nature of their transmission differs and distinctive sets of organizational actors must be in place to confront them. In the case of West Nile Virus, Alberta's Department of the Environment and Sustainable Resource Development would be integrally involved for mosquito surveillance and control whereas it would not be as integral in the event of a pandemic influenza outbreak [5] . To assess potential differences by public health threat in Alberta public health preparedness, we examine and compare two potential public health threats: pandemic influenza and West Nile Virus. Using an embedded multiple-case study design[6] and integrated qualitative and quantitative methods, the project proposes to measure empirically the degree of seamless coordination existing among municipal, regional or district-level and provincial public health and emergency management agencies in Alberta. Although federal authorities and agencies play an important role in provincial-level responses to public health emergencies, for the purposes of this research, the focus will be on the inter-organizational linkages and organizational environments of provincial and sub-provincial actors; relevant federal-level organizations will be identified for future investigation. The seamless coordination of activities and tasks that should characterize the overall federal-provincial-regional linkages should also characterize the intraprovincial linkages. Fluid communication and resource flows among such diverse organizations and administrative authorities require an integrated and coordinated network of actors. Our research is premised on empirical research that has shown that effective and coordinated approaches to public health threats require a well-integrated and responsive inter-organizational system and that the most appropriate method in which to analyze inter-organizational relations and the degree of integration is one based on network analysis [7] . Yet, strong inter-organizational ties are not enough for effective preparedness if the organizations involved do not have the necessary individual capacity to respond to public health threats. In this regard, fluid communication and integrated organizational connections are but one element in an overall systems-level approach to assessing public health preparedness. As shown in Figure 1 , we view public health preparedness to be a product of several interrelated factors: 1) individual organizational attributes, 2) the inter-organizational networks existing among relevant organizational actors, 3) the institutional environments in which organizations and inter-organizational networks are located and in which threats first emerge, and 4) the integrated coordination of specific action-sets. An actionset refers, in this case, to a subgroup of organizations in the network that come together purposefully to address a specific public health threat. We see these four factors together influencing system-level public health preparedness. Our target population for the research is all public health and emergency management-related organizations in Alberta. The construction of the project's sampling frame is based on a stratified, multistage cluster design. The strata are four administrative divisions: (1) provinciallevel agencies, (2) sub-provincial administrative agencies, i.e., emergency-management districts (disaster services) and regional health authorities, (3) metropolitan areas, and (4) towns and local areas. Within strata 1-3, the sampling frame includes all public health or emergency management organizations, i.e., agencies, departments, or units, within those strata. The fourth stratum will be substratified according to the Alberta Emergency Management district divisions into 6 geographical areas: (1) northwestern Alberta, (2) northeastern Alberta, (3) northcentral Alberta, (4) central Alberta, (5) south-central Alberta, and (6) southern Alberta. From each of the 6 geographical regions, we will draw a proportionate, random sub-sample of towns. For stratum 4, the sampling frame consists of all public health or emergency managementrelated organizations in those towns randomly selected. The project's frame population will thus consist of those organizations listed in the strata 1-3 sampling frame and those listed in the stratum 4 sampling frame. This study will be conducted in 2 phases: phase 1 qualitative and phase 2 quantitative research. Phase 1 is used in part to help identify the organizations responsible for public health and emergency responses and the relevant relationships among them. Initial qualitative research will consist of the analysis of official documents and reports and interviews with key organizational representatives. Government documents and reports will be analyzed to 1) help identify agencies involved in the response system, 2) determine the formal tasks assigned to those agencies, and 3) initially assess the overall level of preparedness and responsiveness for the three case studies. Interviews with Attributes key informants and health officials will be used to determine 1) if there are additional organizations involved in the response system that were not noted in the official documents, 2) the types of relationships characterizing the inter-organizational ties within and among actionsets, and develop 3) relevant measures of the institutional environment and of public health preparedness and responsiveness. While laws mandate that certain organizations be involved in emergency responses to specific public health threats, key organizational representatives may identify agencies that play an important albeit informal role in local emergency response systems. In addition to helping identify organizations that may be absent from official documents, the interviews will also investigate the perceived characteristics of inter-organizational relationships. Mandated relationships, for example, tend to be characterized by lower levels of perceived cooperation among organizations [8] . While such characteristics will be investigated using quantitative methods in phase 2, the qualitative research will reveal the points of conflict and cooperation that emerge in the course of public health and emergency responses and thus provide a basis for the questions that will be used in the organizational questionnaire. The final focus of the phase one interviews will centre on the development of measures of the institutional environment, network effectiveness, and public health preparedness. What are the resources that key agency representatives themselves identify as critical and important to their organization and the network? What processes do the respondents use to evaluate if their organization or the system was prepared? In phase 1, we will draw a non-proportional sample from the project's sampling frame. We will interview 6 organizational representatives from the provincial level (stratum 1), 12 representatives from the sub-provincial level (strata 2 and 3), and 18 representatives from the local level (stratum 4). Phase 1 interview respondents will be organizational representatives who are randomly selected from each of these strata and choose to participate. Interviews will be tape-recorded and transcribed so that content analysis, thematic, and inter-textual analyses can be conducted on the interviews. Transcripts will be coded for such indicators as (1) organizations involved in the public health and emergency management preparedness, (2) types of information-sharing among organizations, (3) types of resource flows existing among organizations, and (4) perceptions of local and provincial public health preparedness. Content analysis will be used to identify relevant organizations that were not identified in the documentary research [9, 10] . Thematic analyses will be used to identify the critical themes emerging in the 36 interviews; inter-textual analyses will be used to compare the responses of interviewees across the three strata dimensions of jurisdiction, domain, and geographical location. Phase 2 involves the administration of an organizational/ inter-organizational questionnaire to Alberta-based public health and emergency management agencies. Information on organizations and inter-organizational networks will be collected by web-based or telephone questionnaires. For phase 2, we intend to administer questionnaires to representatives of all organizations listed in the project's sampling frame (described earlier). The questionnaire consists of three components: 1) an organizational attribute component, 2) an inter-organizational network component, and 3) an organizational environment and network assessment component. The organizational attribute component consists of questions on the general characteristics of organizations such as staff size, training background of specific occupational roles in the organization, sectoral operations, budget and more emergency-response related questions, such as whether an agency has formally assessed its epidemiology capacity. The inter-organizational network component consists of items that ask representatives to identify those organizations with which their own organization has ties along specific dimensions. The specific content and dimensions of the network to be mapped will be determined following phase one analysis but candidate dimensions used in other organizational network studies include (i) information-sharing, (ii) resource-sharing, including staff and (iii) joint-planning. For each of the critical dimensions identified in phase one, the respondent will be presented with a list of organizations within the relevant action-set and asked, for each one, "who do you share information with?" or "who do you share resources with." The third component of the questionnaire will consist of the respondent's subjective assessments of the institutional environment and inter-organizational network, and the impact that each has on perceived public health preparedness. For example, how do representatives assess the environment in which they operate, i.e., do they see resources as easily available or scarce? How prepared do they perceive their own organization and local system to be for a public health threat? Since mandated relations involve "sequential interdependence," an important question is whether organizations perceive the task transitions among organizations as functioning smoothly. The actual questions that will be used in this component of the questionnaire will be determined following phase one data analysis. Ethical approval of the research has been granted by the Conjoint Health Research Ethics Board of the Calgary Health Region and University of Calgary, Faculty of Medicine (ID#17873) and of the Capital Health Region and University of Alberta (#B-251005). The capacity of any organization to be prepared for a public health threat is influenced in part by the system in which the organization is located, just as the preparedness of the overall system can be influenced by the strengths and weaknesses of any single organization. The cross-case and intra-provincial comparative design allows our research to discriminate among different system elements and examine the independent influence of organizational attributes, institutional environments, and inter-organizational networks on public health preparedness. In addition, the mixed methods approach enables our research to analyse the role of both informal and formal relationships in the emergence of inter-organization networks and the development of prepared and responsive public health systems. As far as we are aware, our project is one of the first to assess public health preparedness at the systemslevel while accounting for the critical role that inter-organizational relationships play in the emergence of a cohesive and responsive system. Moreover, in deciphering the independent influence of inter-organizational ties within the overall system, the research will provide measures that will help researchers and policy-makers evaluate the degree of coordination among organizational actors. Where are the gaps in system linkages? Which organizations need to be made more central to preparedness activities? These are just a few of the questions that our research is designed to answer.
76
The immunoregulatory and allergy-associated cytokines in the aetiology of the otitis media with effusion.
Inflammation in the middle ear mucosa, which can be provoked by different primary factors such as bacterial and viral infection, local allergic reactions and reflux, is the crucial event in the pathogenesis of otitis media with effusion (OME). Unresolved acute inflammatory responses or defective immunoregulation of middle inflammation can promote chronic inflammatory processes and stimulate the chronic condition of OME. Cytokines are the central molecular regulators of middle ear inflammation and can switch the acute phase of inflammation in the chronic stage and induce molecular-pathological processes leading to the histopathological changes accompanying OME. In this review we present cytokines identified in otitis media, immunoregulatory [interleukin (IL)-2, IL-10, transforming growth factor-beta]) and allergy associated (IL-4, IL-5, granulocyte-macrophage colony-stimulating factor), as crucial molecular regulators, responsible for chronic inflammation in the middle ear and the chronic condition of OME.
The immunoregulatory cytokines [interleukin (IL)-2, IL-10, transforming growth factor (TGF)-b] and allergy-associated cytokines [IL-4, IL-5, granulocyte Á/ macrophage colony-stimulating factor (GM-CSF)] are mediators of the immune system, which are actively involved in regulation of molecular and cellular processes accompanying different types of inflammation. IL-2 is the up-regulating cytokine, which stimulates primarily the cell-mediated inflammatory response by promoting growth, proliferation and differentiation of T cells, B cells, natural killer (NK) cells, monocytes and macrophages. It is secreted mainly by activated T cells. IL-2 induces activation and rapid clonal expansion of mature T cells, and cytokine production in T cells, including interferon (IFN)-g and IL-4; 1 growth of B cells and immunoglobulin (Ig) J-chain switching and secretion of IgM in B cells; 2 proliferation, production of IFN-g and cytolytic activity of NK cells; 3 proliferation and differentiation of macrophage precursors; 4 and cytolytic activity of monocytes. 5 In contrast to IL-2, IL-10 (known as the cytokine synthesis inhibitory and macrophage deactivating factor) down-regulates the immune reactions accompanying acute inflammation and limits the duration of inflammatory responses. IL-10 can also promote and regulate chronic inflammatory processes. IL-10 is produced by a variety of cell types, including CD4 ' T cells, activated CD8 ' T cells and activated B cells. The main anti-inflammatory activities of IL-10 */ namely, inhibition of cytokine production in macrophages, neutrophils, T cells and NK cells, 6 Á 9 and inhibition of the macrophage Á/monocyte activation and the antigen presentation abilities of these cells 10 */lead to the resolution of inflammation. However, if the acute inflammatory process has not been resolved, IL-10 can induce humoral inflammatory reactions such as the immunoglobulin isotype switching in B cells 11, 12 and differentiation of B cells into plasma cells, 13 and thus promote switching of inflammation in the chronic stage. TGF-b, produced by T cells, platelets and monocytes, plays an important role in regulation of the inflammatory processes and in tissue reparation. The effects of TGF-b depend on the differentiation state of the responsive cells. 14 TGF-b, as well as IL-10, participates in resolution of acute inflammation by inhibiting antigen presentation and cytokine production in macrophages, 15 and suppressing proliferation of T cells and NK cells and cytokine production in activated T cells. 14 Simultaneously, TGF-b recruits monocytes to the site of inflammation and upregulates their pro-inflammatory activity; secretion of cytokines and growth factors, in particular. 14 The contradictory effects of TGF-b on monocyte/macrophage lineage, such as up-regulation of monocyte functions and deactivation of macrophages, demonstrate the highly coordinated role of TGF-b in regulation of inflammatory responses. The immunoregulatory and allergy-associated cytokine IL-4, produced by CD4 ' T cells, mast cells and basophils, is actively involved in regulation of different types of inflammation. IL-4 up-regulates humoral inflammatory responses by inducing growth of B cells, isotype switching in activated B cells and differentiation of B cells into antibody producing plasma cells. 16 IL-4 also stimulates the cell-mediated inflammatory processes by promoting growth of activated T cells 17 and enhancing the development of virus-specific cytotoxic T cells. 18 Simultaneously, IL-4 possesses powerful anti-inflammatory effects. In particular, IL-4 controls numerous molecular processes in monocytes, macrophages and neutrophils, which down-regulate production and secretion of the pro-inflammatory cytokines tumour necrosis factor (TNF)-a, IL-1 and IL-8, 19 Á 21 deactivate inflammatory macrophages and lead to suppression of acute inflammation. IL-4 can also induce switching of acute inflammation in the chronic stage owing to its ability to up-regulate expression of the mannose receptor on activated macrophages. 22 The mannose receptor promotes fusion of activated macrophages and formation of giant multinucleated cells in the zone of inflammation 23 that create the cellular background for manifestation of chronic inflammation. IL-5 is an important regulator of the humoral immune response and is produced by CD4 ' T cells as well as NK cells. 24 Being a late regulating factor in differentiation of B cells, IL-5 plays an essential role in cytokine-induced production and secretion of immunoglobulins and supports humoral inflammatory reactions such as production and secretion of IgA. 25 IL-5 is also the allergic cytokine, which by stimulating production, activation, differentiation and migration of eosinophils creates a strong molecular background for eosinophilic inflammation. 26, 27 GM-CSF regulates proliferation and maturation of granulocyte and macrophage precursors and activates mature neutrophils, eosinophils and macrophages, 28 and thus participates in different types of inflammation. 29 Monocytes, T cells, fibroblasts and endothelial cells activated by macrophage cytokines IL-1b or TNF-a produce GM-CSF. The GM-CSF enhances neutrophil phagocytosis, enhances release of chemotactic factors by neutrophils, 30 induces production of the pro-inflammatory cytokines by macrophages and increases their function as antigen-presenting cells, 31 and in this way promotes acute inflammatory reactions. In the presence of IL-4, GM-CSF can initiate a specific differentiation of monocytes into osteoclast-like multinucleated giant cells 32 that create a background for chronic inflammation. GM-CSF can also contribute to allergic inflammation by promoting growth, prolonged survival and activation of eosinophils and basophils. 33 Both groups of cytokines, the immunoregulatory IL-2, IL-10 and TGF-b and the allergy-associated IL-4, IL-5 and GM-CSF, have been identified in otitis media and their presence is the evidence of their participation in regulation of local inflammatory processes. These cytokines can switch the acute phase of middle ear inflammation into the chronic stage and lead to the chronic condition of otitis media with effusion (OME). OME: a brief description OME, the commonest cause of childhood deafness in the developed world, is a chronic inflammatory condition of the middle ear cleft with repetitive recurrences of acute middle ear inflammation. The disease is characterized by a middle ear effusion that persists for months to years, cannot be cleared by the normal mucociliary transport mechanisms and is must usually removed by surgical operation. A surgical procedure, which includes myringotomy, aspiration of fluid and the insertion of ventilation tubes (grommets) in the anterior tympanic membrane (Fig. 1) , is the most effective option in treatment of OME. The main problem with ventilation tubes is secondary infection and post-operative otorrhea, which can be managed using antibiotic eardrops. 34, 35 However, increasing antibiotic resistance of bacterial pathogens 36 and possible ototoxicity of topical antibiotics 37 can complicate the antimicrobial therapy in post-surgical management of otitis media. Other non-surgical approaches in treatment and prophylaxis of OME include traditional methods, such as combined steroid Á/antimicrobial therapy 38, 39 and polyvalent pneumococcal vaccination, 40, 41 and relatively new methods based on herbal medicine 42, 43 and homeopathy. 44, 45 The chronic condition of otitis media is associated with proliferative changes in the middle ear tissues, especially in the surface middle ear mucosa, which is presented in OME as a modified pseudostratified epithelium, 46 where basal cells are differentiating into goblet cells, 47 goblet cells are proliferating with enhanced secretory activity 48 and formation of mucus glands occurs 49, 50 (Fig. 1) . Goblet cells produce and secrete mucins, which are important glycoproteins in the mucociliary transport system of the middle ear and are the main component of middle ear effusions, responsible for the viscous properties of effusions. 51, 52 However, under disease conditions, alterations that occur in the middle ear and eustachian tube mucin metabolism, 53 in the structure of mucin glycoproteins 54, 55 and in the glycoconjugate expression in cilia and goblet cells 56 promote the dysfunction of the normal mucociliary transport system and the formation of effusion in the middle ear cleft. The expression of several mucins has been detected in OME. 53, 57 However, overproduction of only two mucins has been observed in otitis media; namely, the membrane-bound MUC4 53 and the secreted MUC5B. 53, 55, 58, 59 Another secreted mucin (MUC5AC) is always presented in effusions, but in varying amounts, 53, 55, 57, 60 and its levels maybe linked to levels of the pro-inflammatory cytokines TNF-a, IL-6 and IL-8, which can promote different levels of MUC5AC secretion. 61, 62 The initial step in the pathogenesis of OME is the inflammatory process in the middle ear mucosa. Bacteria, 63, 64 viruses, 65,66 allergic reactions 67,68 and even gastroesophageal reflux 69 Á 71 in tandem with predisposing factors such as eustachian tube dysfunction, 72 cleft palate 73 and obstructive adenoids 74 can stimulate the middle ear inflammation. Different groups of inflammatory mediators were identified in the human middle ear mucosa, fluids and effusions. A lot of different mediators participate in initiation and early stages of the middle ear inflammation, including arachidonic acid metabolites (prostaglandin E 2 and leukotrienes LT-B 4 , LT-C 4 ), 75, 76 histamine, 77,78 platelet-activating factor, 79 surface cell adhesion molecules (intercellular adhesion molecule-1, vascular cell adhesion molecule-1, endothelial leukocyte adhesion molecule-1, platelet endothelial cell adhesion molecule), 80,81 soluble cell adhesion molecules (soluble intercellular adhesion molecule-1 and soluble vascular cell adhesion molecule-1), 82, 83 chemokine RANTES, 84 complement C3a anaphylatoxin 85 and interferon-g. 86 However, cytokines are the key mediators of the middle ear inflammation. Cytokines regulate different stages of inflammation, are responsible for resolution of inflammation and can initiate local molecular processes leading to histopathological changes in the middle ear mucosa and submucosa, and the chronic condition of otitis media. Different groups of cytokines have been identified in the middle ear effusions and mucosa: the pro-inflammatory TNF-a, IL-1b, IL-6 and IL-8; the immunoregulatory IL-2, IL-10 and TGF-b; and the allergy-associated IL-4, IL-5 and GM-CSF. The pro-inflammatory cytokines and their participation in the pathogenesis of otitis media have been already discussed, 87 and are more related to the acute phase of middle ear inflammation. In the present review we have analysed the immunoregulatory and the allergy-associated cytokines in otitis media, which, by providing the cellular and molecular background for the chronic inflammation in middle ear, promote the chronic condition of OME. The immunoregulatory cytokines IL-2, IL-10, TGF-b and their cellular and molecular networks in OME IL-2 in otitis media IL-2 has been detected in middle ear effusions from children with chronic OME at high concentration (greater than 300 pg/ml). 88 The analysis of peripheral blood immunologic pattern in patients with persistent middle ear effusions 89,90 showed the following: (1) the number of CD4 ' and CD8 ' T lymphocytes was decreased and the T-cell subset ratio (CD4 ' / CD8 ' ) was reduced; (2) proliferative responses of blood T cells were diminished; and (3) generation of IL-2 by peripheral blood lymphocytes was decreased. Another immunologic investigation of children with recurrent otitis media revealed a consistent inability of adenoidal T cells to turn on B cells to mature into immunoglobulin-secreting plasma cells that has been explained by defective production of IL-2 in adenoids. 91 This assumption was confirmed by cytokine assay of IL-2, which has been done simultaneously in nasopharyngeal lymphoid tissues and in peripheral blood of children with recurrent otitis. The adenoidal lymphocytes produced significantly less IL-2 compared with the patient's peripheral blood lymphocytes. 92 It has been assumed that poor production of IL-2 in adenoidal tissues is probably linked to a deficiency of IL-2 production in the middle ear tissues, and that could be associated with the persistence of OME. Thus, general and/or local deficiency of the IL-2 production could induce defective immunoregulation of the middle ear inflammation and promote the persistence of OME. The investigation of IL-2 in experimental models of otitis media showed the involvement of IL-2 in regulation of acute inflammation because: (1) the IL-2 producing cells were predominant in acute OME compared with chronic OME, 93 and (2) recombinant IL-2 trans-tympanically injected into the middle ear of normal guinea pigs caused the inflammatory middle ear effusion within 24 h, which cleared by 72 h after inoculation. 94 However, differing degrees of responsiveness of effusion production were observed following the instillation of IL-2 in the middle ear, 95 varying from pronounced middle ear effusion causing rather severe mixed hearing loss to complete lack of effusion, and IL-2 was detected in all the middle ear effusions in experimental OME with signs of sensorineural hearing loss. 96 By analysis of clinical and experimental data we can propose the following conclusions about the role of IL-2 in OME: 1. IL-2 is involved in regulation of the middle ear inflammation and, as the main T-cell growth and differentiation factor, IL-2 can support acute and chronic inflammatory processes in the middle ear. 2. IL-2 can regulate local acute inflammatory response by inducing activation, differentiation of T cells and production of the proinflammatory (IL-1b, IL-6) and anti-inflammatory cytokines (IL-10, TGF-b) by activated T cells (Fig. 2 ). 3. The deficiency of local or general IL-2 production can suppress reactions of acute inflammation and promote persistence of OME, which can develop with time into the chronic OME (Fig. 3 ). 4. The excessive production of IL-2 can provoke chronic cell-mediated and (or) humoral inflammatory processes (proliferation of activated T cells, increased production of cytokines IL-4, IL-5 and GM-CSF by proliferated T cells, immunoglobulin-switching and IgM secretion in B cells), which will induce switching of the middle ear inflammation in the chronic stage with subsequent development of chronic OME (Fig. 3) . The clinical investigations demonstrated the involvement of IL-10 in chronic otitis media. 97,98 IL-10 was identified in middle ear effusions from children undergoing tympanostomy tube placements and its mean concentration was 569/58.7 pg/ml. 97 It has been assumed that the presence of IL-10 in middle ear effusions might be one of the causes of a lack of clinical features of acute inflammation and could lead to a chronic inflammatory state. Chronic otitis media has been also associated with dysregulated local production of IL-10. 98 Patients with chronic otitis media showed high levels of IL-10 ( /100 pg/ml) produced by sinus lavage cell culture and simultaneously low production of IL-10 by peripheral blood mononuclear cells. Investigations of cytokine profiles in different experimental models of OME showed a rapid appearance of IL-10 in the early stages of acute otitis media 99 Á 103 and that IL-10 in mucosa was produced predominantly by CD4 ' T cells. 104 Streptococcus pneumoniae induced expression of IL-10 mRNA in middle ear mucosa within 1Á/2 days after inoculation. 99, 100 Lipoteichoic acid, one of the components present within the cell wall layer of most Grampositive bacteria, 101 and non-typeable Haemophilus influenzae (NTHi) 102,103 induced the expression of IL-10 mRNA in middle ear mucosa even faster, within 6 h after inoculation. It is interesting to note that the expression of IL-10 mRNA followed the expression of TNF-a and IL-6 mRNAs, 99,100 and sometimes occurred simultaneously with the pro-inflammatory cytokines. 101, 102 Thus, clinical and experimental data about IL-10 in OME lead to the following conclusions: 1. IL-10 is involved in regulation of the middle ear inflammatory response induced by bacterial infection. 2. The expression of IL-10 in acute otitis media is up-regulated shortly after or even simul- taneously with the pro-inflammatory cytokines, TNF-a and IL-6. That is, the evidence of initiation of corrective immunoregulation in the middle ear mucosa, because IL-10 as the main anti-inflammatory cytokine, can down-regulate the local population of inflammatory cells, macrophages and neutrophils, and production of the proinflammatory cytokines TNF-a, IL-1b, IL-6, IL-8 and promote resolution of inflammation (Fig. 2 ). 3. If the acute inflammatory response is not resolved correctly, the prolonged presence of IL-10 in the zone of inflammation can provoke the amplification of chronic humoral inflammatory processes, such as increased secretion of immunoglobulins, especially IgA, by activated B cells and possible generation of plasma cells, in the presence of IL-4 ( Fig. 4) . In this way, IL-10 can contribute to switching of the disease into the chronic stage. 4. The deficiency of IL-10 in acute otitis media can also lead to chronic condition of OME, because shortage of IL-10 will promote overproduction of the inflammatory cytokines, especially IL-1b, TNF-a and IL-6, which can induce tissue damage and irre-versible changes in the middle ear mucosa (Fig. 4) . General conclusions about the role of IL-2 and IL-10 in an aetiology of OME Both cytokines, IL-2 and IL-10, participate in balanced immunoregulation of the middle ear inflammatory response. IL-2 up-regulates cellular and molecular events, accompanying acute inflammation, whereas IL-10 down-regulates acute inflammatory reactions and promotes resolution of inflammation. Any imbalance in production of these cytokines will induce chronic inflammatory processes (cellmediated or humoral, or both) or stimulate uncontrolled inflammatory-related damage of the middle ear tissues (Figs 3 and 4) . Thus, IL-2 and IL-10 can be considered as the key cytokine mediators, regulating switching of the acute phase of middle ear inflammation in the chronic stage, which will lead to the chronic condition of OME. Clinical investigations of TGF-b in OME showed its relation to chronic otitis media. 105, 106 The identification of TGF-b in serous and mucoid middle ear effusions was associated with a history of previous tympanostomy tube placements. 105 Investigation of growth factors in tympanic membrane perforations showed that TGF-b could promote development of the fibrotic scar at the perforation margin, explaining the deficient healing pattern of tympanic membranes in chronic otitis media. 106 However studies of cytokines in experimental OME demonstrated the presence of TGF-b also in acute otitis media. 100, 102, 103 The expression of TGF-b mRNA was up-regulated in acute otitis media caused by S. pneumoniae 100, 102 and NTHi infections, 102, 103 and was detected early during the course of acute inflammation, within 24 h after inoculation, together with cytokines IL-1, TNF-a and IL-10. However, the highest mRNA levels for TGF-b were recorded considerably later, on days 6 and 28, for the NTHiinfected and pneumococcus-infected animals, respectively, and its expression continued even after resolution of acute otitis media. 102 TGF-b regulates host response in infectious diseases through the TGF-b-activated kinase 1 107 and TGF-b-Smad signaling pathway. 108 It has been shown that NTHi, a major human bacterial pathogen of otitis media, strongly induced up-regulation of MUC5AC mucin in human epithelial cells. This induction occurred via activation of the Toll-like receptor 2-MyD88-dependent p38 mitogen-activated protein kinase (MAPK) pathway. However, activation of TGF-b-Smad signaling by TGF-b resulted in downregulation of p38 MAPK by inducing MAPK phophatase-1, and subsequent down-regulation of MUC5AC transcription. 108 TGF-b thereby acted as a negative regulator of MUC5AC mucin gene expression. This is a very significant fact in understanding the role of TGF-b in otitis media, because varying levels of the secreted mucin MUC5AC are presented in the middle ear mucosa and effusions in patients with OME. 53, 55, 57 By analysis of these data we can make the following conclusions about the role of TGF-b in OME: 1. TGF-b participates in regulation of the middle ear inflammation. 2. TGF-b is a negative regulator of acute inflammation in the middle ear. Early activation of TGF-b together with another antiinflammatory cytokine, IL-10, in the course of acute otitis media demonstrates the involvement of TGF-b in the correct resolution of acute inflammatory response. TGF-b as well as IL-10 can inactivate macrophages and inhibit pro-inflammatory cytokine production at the site of inflammation (Fig. 2 ). 3. TGF-b can be also involved in down-regulation of bacterial-induced MUC5AC mucin production, and that is the additional evi- Thus, TGF-b can contribute to the pathogenesis and persistence of chronic OME in many ways. The allergy-associated cytokines IL-4, IL-5 and GM-CSF and their cellular and molecular networks in OME IL-4 was identified in the middle ear effusions of children with persistent OME 109 and in atopic children with OME undergoing myringotomy and ventilation tube placement. 110 The cytokine analysis of effusions showed a higher mean level of IL-4 in the allergy-positive group compared with the allergynegative group 109 and a higher percentage of cells expressing IL-4 in atopic patients with OME compared with that seen in non-atopic patients. 110 A higher level of IL-4 in effusions correlated with predominance of T lymphocytes, which was the sign of chronic inflammation and was also related to the atopic background of patients with OME. 110 In effusions of non-atopic patients, the level of IL-4 was lower and the predominance of neutrophils 109 demonstrated signs of an acute inflammatory response. Investigation of adenoids in patients with recurrent otitis media 92,111 resulted in the following important findings: (1) expression and secretion of IL-4 was detected in adenoids; (2) adenoidal lymphocytes produced the same level of IL-4, or even slightly more, compared with the patient's peripheral blood lymphocytes; 92 and (3) the epsilon germline transcripts for IgE were detected in the adenoids and the level of IgE epsilon germline transcript expression was dependent on the level of IL-4 mRNA expression. 111 This local IL-4-induced immunoglobulin class switching to IgE in the adenoids was considered the essential molecular process contributing to chronic inflammation in the middle ear and promoting the pathogenesis and persistence of OME. Thus, IL-4 was presented in both groups of patients with OME, allergic and non-allergic. However, local overproduction of IL-4 was associated predominantly with an allergic background of patients. Studies on IL-4 in experimental OME 93,112 demonstrated the involvement of IL-4 in regulation of different types of otitis media. In trans-tympanically challenged animals the expression of IL-4 was detected during acute and chronic otitis media, with the predominance of IL-4 ' cells in the acute form. 93 The trans-tympanical injection of soluble receptors for IL-4 prevented the allergic Eustachian tube dysfunction and formation of effusion in the middle ear cleft. 112 The involvement of IL-4 in regulation of MUC5AC and MUC5B mucin gene expression and secretion has been shown in human airways. 113 Á 115 The secreted mucins, MUC5AC and MUC5B, are presented in the middle ear mucosa and effusions in patients with chronic OME. 53,55,57 Á 60 Although relationships between IL-4 and production of MUC5AC and MUC5B mucins in the course of otitis media have not been investigated, the ability of a soluble receptor for IL-4 to prevent Eustachian tube dysfunction and formation of effusions in the middle ear cleft 112 demonstrated the potential involvement of IL-4 in the regulation of the middle ear mucin metabolism. All these data allow us to make the following conclusions about the role of IL-4 in OME: 1. IL-4 is involved in regulation of different types of otitis media. 2. IL-4 is one of the cytokines responsible for an allergic type of inflammation owing its ability to induce local immunoglobulin class switching to IgE (Fig 6) . 3. IL-4 can support local humoral inflammatory responses by increasing production and secretion of immunoglobulins in activated B cells (Figs 4 and 6 ). 4. IL-4 can also support cell-mediated chronic inflammation in the middle ear by promoting generation of multinucleated giant cells (Figs 5 and 6 ), although it is necessary to note that giant cells in OME have not been investigated. 5. IL-4 can also participate in the regulation of mucin metabolism and the mucociliary transport system in middle ear inflammation. Thus, IL-4 is the important regulator of cellular and molecular processes, accompanying different types and different stages of the middle ear inflammation, and an important factor contributing to chronic condition and persistence of OME. IL-5 was detected in the middle ear mucosa and effusions in atopic patients with persistent OME. 110, 116, 117 The expression of IL-5 mRNA in mucosa was up-regulated and correlated with increased expression of major basic protein (cell marker for eosinophils) and CD3 (cell marker for activated T lymphocytes). 116 Examination of the cytokine profile in the middle-ear specimens showed an increased number of cells expressing IL-5 and its correlation with the percentage of eosinophils and T lymphocytes. 110 Correlation between the percentage of eosinophils and levels of IL-5 in effusions was particularly striking in OME patients with asthma, 117 suggesting that middle ear eosinophilia could be dependent not only on the local, but also on the general level of IL-5 production. The percentage of cells expressing IL-5 in effusions of atopic patients was the same or slightly higher compared with cells expressing IL-4; 110 this was evidence that both cytokines IL-5 and IL-4 contributed to development of allergic inflammation in the middle ear. Thus, overproduction of IL-5 in OME is associated with the atopic background of patients, and IL-5 together with IL-4 is responsible for an allergic type of inflammation in the middle ear (Fig. 6) . This assumption was confirmed in an experimental model of OME, 112 where the late-phase allergic inflammatory response was prevented by pretreatment with IL-5 antibodies or soluble receptors for IL-4. However, the involvement of IL-5 in regulation of the non-allergic inflammatory response in the middle ear should not be ignored. The production of IL-5 was detected in effusions of non-atopic patients with OME, 110, 117 although it was significantly less compared with atopic patients. Analysis of immunoregulatory cytokines in experimental models of OME 93 demonstrated the absence of IL-5 in acute otitis media and significant up-regulation of IL-5 production in chronic otitis media, where IL-5 ' cells were numerous. Thus, the presence of IL-5 in OME of nonallergic patients shows the involvement of IL-5 in regulation of non-allergic chronic inflammatory reactions in the middle ear, where IL-5 can support humoral processes such as increased production and secretion of immunoglobulins, IgA in particular (Fig. 6) . Thus, IL-5 can be considered the late mediator of middle ear inflammatory response, regulating allergic inflammation and increasing humoral inflammatory processes, which contribute to the chronic condition of OME. Our knowledge about GM-CSF in OME is insufficient. However, the involvement of GM-CSF in regulation of middle ear inflammation has been shown as GM-CSF was identified in effusions of acute purulent and chronic otitis media. 118 The level of GM-CSF in acute otitis media was higher compared with the chronictype otitis media. Despite a lack of experimental data about GM-CSF in OME, this cytokine could participate in regulation of different stages of the middle ear inflammation. GM-CSF was identified in acute otitis media, where it could activate matured neutrophils and macrophages and support the acute inflammatory response (Fig. 2) . GM-CSF in the presence of IL-4 can induce monocyte differentiation towards generation of mul-tinucleated giant cells, and in this way create the cellular background for chronic middle ear inflammation (Fig. 5) . Finally, GM-CSF being the eosinophil survivalpromoting cytokine together with IL-5 can up-regulate allergic eosinophil-mediated inflammation in the middle ear (Fig. 6) . However, in the past year of investigations GM-CSF has been shown as a potential down-regulator of allergic inflammation, because GM-CSF induced apoptosis of human eosinophils through the eosinophil receptor Siglec-8 (sialic acid binding immunoglobulin-like lectin). 119 Thus, GM-CSF can be a very important regulator of the middle ear inflammation and additional experimental work needs to be done to clarify the exact role of this cytokine in the pathogenesis of chronic OME. Different types of immunoglobulins, namely IgM, IgG, IgA, secretory IgA and IgE, have been identified in effusions and middle ear fluid of chronic OME. 120 Á 123 The effusion level of IgM in some patients with chronic otitis media was markedly elevated. 120 The mucoid type of effusions contained a high level of IgG, IgA 121 and IgE. 124 Comparison of the immunoglobulin levels measured in effusions and in sera showed that, in many cases, the effusion level of secretory IgA 122 and IgE 125 was significantly higher than the corresponding serum level, and this was the evidence of local overproduction of immunoglobulins in the middle ear. The immunologic investigation of effusions detected the immune complexes of IgG (IgG Á/ICs) and FIG. 6. Cellular and molecular networks of the allergy-associated cytokines leading to the chronic condition of OME. IgA (IgA Á/ICs) in both acute and chronic otitis media. 126, 127 The highest level of IgG Á/ICs was found in subacute cases, whereas IgA Á/ICs were predominant in chronic OME. The immunoglobulin immune complexes provided the prolongation of inflammatory process in the middle ear. The presence of immunoglobulins in chronic OME was associated mainly with the bacterial infection. 128 Á 131 IgG and IgA antibodies specific to Hemophilus influenzae and S. pneumoniae , 128 IgG, IgM, IgA and secretory IgA antibodies specific to outer membrane antigens of Moraxella catarrhalis, 129 and Staphylococcus aureus -harboured bacteria, intensely coated with secretory IgA and IgG antibodies, 130, 131 were identified in chronic effusions. Only one type of immunoglobulins, the secretory IgA, was identified in effusions infected with respiratory viruses. 132 The presence of IgE in chronic OME has not been always associated with local allergic inflammation. 124, 133 However, local overproduction of IgE was usually accompanied by local allergic reactions, such as degranulation of mast cells found in the middle ear biopsy specimens 124 and expression of IgE on mast cells detected in nasal mucosa specimens from patients with OME. 134 Thus, the activity of immunoglobulins in chronic OME is evidence of chronic humoral inflammatory processes in the middle ear, which is obviously controlled by cytokines. Correlation between the levels and types of immunoglobulins and the immunoregulatory and allergy-associated cytokines in chronic OME has not been investigated. However, the presence of the main types of immunoglobulins, IgM, IgG, IgA, secretory IgA and IgE, in effusions is indirect evidence that cytokines IL-2, IL-10, TGF-b, IL-4 and IL-5, which are involved in regulation of immunoglobulin production and secretion in general, 2, 11, 12, 16, 25 participate also in local production and secretion of immunoglobulins and regulate humoral immune reactions during the course of middle ear inflammation (Figs 3 Á/6 ). The immunoregulatory cytokines and the allergyassociated cytokines can be considered the key regulators of the middle ear inflammation responsible for the molecular and cellular background of chronic OME. The immunoregulatory cytokines IL-2, IL-10 and TGF-b initiate and support molecular switching of the acute phase of inflammation in the chronic stage, whereas the allergy-associated cytokines IL-4, IL-5 and GM-CSF very probably provide the molecular and cellular background for chronic humoral, cell-mediated and allergic inflammatory processes in the middle ear, which lead to the chronic condition of OME. However, it is necessary to note that additional experimental work needs to be done in order: (1) to elucidate the molecular mechanisms of cytokine regulation of the middle ear inflammation; and (2) to analyse the possibility of anti-cytokine therapy in clinical treatment of OME.
77
Bench-to-bedside review: Critical illness-associated cognitive dysfunction – mechanisms, markers, and emerging therapeutics
Cognitive dysfunction is common in critically ill patients, not only during the acute illness but also long after its resolution. A large number of pathophysiologic mechanisms are thought to underlie critical illness-associated cognitive dysfunction, including neuro-transmitter abnormalities and occult diffuse brain injury. Markers that could be used to evaluate the influence of specific mechanisms in individual patients include serum anticholinergic activity, certain brain proteins, and tissue sodium concentration determination via high-resolution three-dimensional magnetic resonance imaging. Although recent therapeutic advances in this area are exciting, they are still too immature to influence patient care. Additional research is needed if we are to understand better the relative contributions of specific mechanisms to the development of critical illness-associated cognitive dysfunction and to determine whether these mechanisms might be amenable to treatment or prevention.
Since its advent more than 40 years ago, the specialty of critical care has made remarkable advances in the care of severely ill patients. Mortality rates for many commonly encountered critical illnesses such as severe sepsis [1] and acute respiratory distress syndrome (ARDS) [2] have declined sharply over the past 2 decades. As greater numbers of patients survive intensive care, it is becoming increasingly evident that quality of life after critical illness is not always optimal. For instance, nearly half of ARDS survivors manifest neurocognitive sequelae 2 years after their illness, falling to below the 6th percentile of the normal distribution of cognitive function [3] . Considering that 89% of Americans would not wish to be kept alive if they had severe, irreversible neurologic damage [4] , these findings are quite concerning. Cognitive dysfunction (CD) is quite common in critically ill patients, not only during the acute illness but also long after the acute illness resolves [5] . Delirium, a form of acute CD that manifests as a fluctuating change in mental status, with inattention and altered level of consciousness, occurs in as many as 80% of mechanically ventilated intensive care unit (ICU) patients [6] . Most clinicians consider ICU delirium to be expected, iatrogenic, and without consequence. However, recent data associate delirium with increased duration of mechanical ventilation and ICU stay [7] , worse 6-month mortality [8] , and higher costs [9] . Chronically, critical illnessassociated CD manifests as difficulties with memory, attention, executive function, mental processing speed, spatial abilities, and general intelligence. Interestingly, patients who develop acute CD often go on to develop chronic CD after hospital discharge [10] [11] [12] [13] , suggesting that the two entities may share a common etiology. Although there are clearly defined risk factors for critical illness-associated CD, there is little understanding of the underlying pathophysiology. The precise mechanisms are unknown and there are likely to be multiple mechanisms at work in any given patient ( Figure 1 ) [5, 14, 15] . We have chosen to focus on two mechanisms that appear to have the greatest merit: neurotransmitter abnormalities and occult diffuse brain injury. In this bench-to-bedside review, we discuss the evidence supporting these mechanisms, potential markers that could be used to evaluate each mechanism in individual patients, and emerging therapies that may prevent or mitigate critical illness-associated CD. relative dopamine excess in the central nervous system (CNS). Antipsychotics such as haloperidol, which antagonize central dopamine receptors, can counteract the cognitive effects of anticholinergic medications, further supporting the anticholinergic hypothesis. Drugs with potent central anticholinergic effects, such as tricyclic antidepressants and antihistamines, are particularly likely to cause delirium. Many medications that are commonly used in the ICU yet not generally considered to be anticholinergic, such as H 2 blockers, opiates, furosemide, digoxin, glucocorticoids, and benzodiazepines, were recently shown to have central anticholinergic properties [16, 17] . Volatile anesthetics, such as sevoflurane, and intravenous anesthetics, such as propofol, also have anticholinergic effects and may be responsible not only for postoperative delirium but also for the more complex phenomena of postoperative cognitive dysfunction [18] . Acute illness itself may be associated with production of endogenous anticholinergic substances [19] . In one study, 8 out of 10 elderly medical inpatients had had detectable anticholinergic activity in their serum, even though no medication used by these individuals had anticholinergic activity. Characterization of such substances might improve our understanding of delirium and lead to useful intervention strategies. Considering that activation of specific cholinergic pathways can inhibit proinflammatory cytokine synthesis and protect against endotoxemia and ischemia-reperfusion injury [20] , it is tempting to speculate that inhibition of these pathways, whether exogenous or endogenous, might contribute not only to CD but also to other outcomes of critical illness. In assessing the overall risk for developing CD posed by medications with central anticholinergic activity in a given patient, individual differences in drug pharmacokinetics make the dose received a poor estimate of a patient's overall anticholinergic burden [21, 22] . However, we can objectively measure anticholinergic burden in individual patients using an assay referred to as serum anticholinergic activity (SAA) [16] . First described by Tune and Coyle [23] , SAA measures the ability of a individual's serum to block central muscarinic receptors using a rat forebrain preparation. Elevated SAA levels are associated with cognitive impairment in studies of medical ward inpatients and community dwelling seniors [16, [24] [25] [26] [27] . Only a single, small study has used this assay to investigate CD in ICU patients. Golinger and colleagues [28] examined SAA levels in surgical ICU patients and found the mean SAA level drawn 4 hours after mental status change was significantly greater in delirious patients (n = 9) than in those without delirium (n = 16; 4.67 ng/ml versus 0.81 ng/ml; P = 0.007). Whether these results apply to all critically ill patients is uncertain because no study has examined SAA across a broad range of ICU admitting diagnoses or in medical ICU settings. Furthermore, because SAA measurement requires fresh rat brain preparations, its use is likely to remain limited to research settings for the foreseeable future. Pathophysiologic mechanisms and predisposing factors thought to underlie critical illness-associated cognitive dysfunction [5, 14, 15] . Apo, apolipoprotein; HIV, human immunodeficiiency virus; 5-HT, serotonin (5-hydroxytryptamine); GABA, γ-aminobutyric acid; NE, norepinephrine (noradrenaline). Other neurotransmitter systems such as dopamine, serotonin, γ-aminobutyric acid (GABA), norepinephrine (noradrenaline), and glutamate are also thought to contribute to critical illnessassociated CD. Dopaminergic hyperfunction is thought to underlie the cognitive symptoms of schizophrenia, and dopamine administration itself may be a risk factor for delirium [29] . Serotonin syndrome, a consequence of excess serotonergic agonism, can be seen not only with selective serotonin reuptake inhibitors but also with a variety of drugs and drug combinations [30] . Even a single therapeutic dose of an selective serotonin reuptake inhibitor can cause the syndrome, which manifests as mental status changes, autonomic hyperactivity, and neuromuscular abnormalities. GABA abnormalities are thought to contribute to hepatic encephalopathy, perhaps mediated by branched chain and aromatic amino acids acting as false neurotransmitters [31] . Excess GABA activity, such as that which occurs after withdrawal from chronic ethanol or benzodiazepine use, is a well known and quite dangerous cause of delirium [32] . Acutely, sedatives that stimulate GABA receptors, such as benzodiazepines and (probably) propofol, impair cognitive function and are deliriogenic [8, [33] [34] [35] . This raises the possibility that strategies to minimize sedative drug accumulation, such as daily interruption of sedative infusions [36] , which have been shown to reduce duration of mechanical ventilation, and ICU and hospital length of stay, might also reduce the incidence or duration of delirium. Whether these sedative drugs lead to neurocognitive deficits long after their use is unknown, but this has been suggested in certain high-risk groups, such as the very old (>75 years) and those with pre-existing cognitive impairment [37, 38] . Noradrenergic hyperfunction, as part of the 'fight or flight' response, can lead to panic attacks and delusions. Glutamate has been implicated in the 'Chinese food syndrome', in which food with high amounts of monosodium glutamate interferes with normal neurotransmission causing confusion [39] . For a more complete review of the other neurotransmitter abnormalities that may underlie delirium, the reader is referred elsewhere [40, 41] . If critical illness-associated CD were solely due to acute medication effects, it would probably resolve after the exposure has ended. However, a significant percentage of individuals developing delirium in the hospital continue to demonstrate symptoms of CD after discharge [10] [11] [12] [13] . These patients manifest decreased cerebral activity and increased cognitive deterioration, and are more likely to develop dementia than patients without delirium. Also, patients who develop delirium have a greater rate of decline on cognitive tests than do nondelirious patients [10] [11] [12] [13] . Taken together, these observations raise the possibility that some degree of occult diffuse brain injury, as a consequence of the local hypoxia, hypoperfusion, cytokine-mediated inflammation and microvascular thrombosis that characterize the multisystem organ dysfunction of critical illness, might have occurred in these patients [42] . Given that every other organ system can be damaged by these forces, it seems implausible that the brain would be uniquely spared. Many of the data supporting occult diffuse brain injury as a cause of critical illness-associated CD come from studies of sepsis and septic encephalopathy, a form of delirium. In animal models of sepsis, oxidative damage occurs early in the hippocampus, cerebellum, and cortex [43] , and significant alterations in cerebral vascular hemodynamics and tissue acid-base balance indicate that cerebral ischemia and acidosis do occur [44] [45] [46] [47] [48] . Sharshar and colleagues completed several studies comparing brain pathology in small numbers of patients who died from septic shock with that in patients who died from other causes. Septic patients demonstrated diffuse severe ischemic and hemorrhagic CNS lesions [49] , which correlated with persistent hypotension and severe coagulation disorders. Multiple microscopic foci of necrosis involving the white matter of the pons [50] were seen, as well as ischemia and apoptosis within the cerebral autonomic centers [51] . The white matter lesions were associated with elevated levels of proinflammatory cytokines, suggesting a possible role of inflammation and microvascular thrombosis in the genesis of CNS injury [52] . Although those studies demonstrated that ischemic brain injury occurs in sepsis, they did not determine whether delirium occurred. Two studies attempted to examine the relationship of ischemic brain injury to delirium. In one study of 84 patients with severe sepsis and multiple organ dysfunction [53] , severe hypotension was the only factor in multivariable analyses that was associated with delirium, suggesting that sepsis-related encephalopathy may be caused by ischemic damage rather than metabolic abnormalities. Another study examined cerebral blood flow and cerebral oxygen metabolic rates in patients with septic encephalopathy and multiple organ dysfunction [54] , and it found that both were significantly lower than those in normal awake individuals. Although these studies support the idea of occult brain injury as a cause of delirium, the authors did not use a standardized Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV based tool to diagnose delirium, such as the Confusion Assessment Method for the ICU [6] . Lending support to the hypothesis that acute inflammation leads to brain injury and subsequent development of delirium, a recent study found that delirium in postoperative hipfractured patients was significantly associated with serum levels of C-reactive protein, an acute-phase protein that is a marker of acute inflammation [55] . Importantly, patients in the study were diagnosed with delirium using the Confusion Assessment Method (the ward-based predecessor to the Confusion Assessment Method for the ICU), providing the first DSM-IV based evidence that acute inflammation may be in the causative pathway of delirium. The brain is a target for free radical damage because of its large lipid content, high rate of metabolism, and low antioxidant capacity. Free radical induced oxidative stress may play a role in the delirium seen after cardiopulmonary bypass. Karlidag and colleagues [56] noted that patients with low preoperative levels of catalase, a erythrocyte-based antioxidant enzyme, were more susceptible to delirium postoperatively. They suggested that preoperative catalase levels might some day be used to identify at-risk patients who could then be put on antioxidant treatment preoperatively. Whether this would reduce the incidence of delirium remains speculative. Regional cerebral blood flow appears to be reduced in delirium. Using xenon-enhanced computed tomography (CT), Yakota and colleagues [57] demonstrated significant focal and global brain hypoperfusion in 10 ICU patients with hypoactive delirium. After recovery from delirium cerebral blood flow returned to normal, implying that cerebral hypoperfusion may contribute to the development of delirium. Studies of ARDS survivors suggest that a combination of acute hypoxia, hypoperfusion, and hyperglycemia plays an important role in the long-term cognitive sequelae of critical illness [3, 58, 59] . However, it has been difficult to demonstrate a clear relationship, given the lengthy interval between stimulus and effect and the great number of additional contributing variables that can obscure downstream effects. Among ARDS survivors, Hopkins and colleagues showed that the degree of CD at 1 year is significantly correlated with the durations of hypoxia [58] and mean arterial blood pressure less than 50 mmHg during the ICU stay [3] . In animals, hyperglycemia markedly enhances hypoxic-ischemic brain damage due to increased brain edema and disrupted cerebral metabolism [60] . In ARDS survivors, the duration of blood glucose greater than 180 mg/dl has been shown to correlate with worse visual spatial abilities, visual memory, processing speed, and executive function at 1 year [59] . Given the recent interest in maintaining tight glucose control during critical illness as a means of reducing mortality, it will be interesting to see whether patients managed using this technique have better cognitive outcomes. Clearly, such an approach will need to balance the benefits of tight glucose control with the known risks that hypoglycemia poses to the CNS. One of the perceived difficulties with looking for evidence of occult brain injury in humans is the apparent need for CNS tissue specimens to prove that brain injury actually occurred. However, studies of stroke, trauma, and cardiopulmonary bypass-associated brain injury show that serum markers of brain injury correlate well with the extent of CNS damage. S-100β, neuron-specific enolase (NSE), and myelin basic protein (MBP) are three such markers that could be used to look for evidence of occult brain injury in critical illnessassociated CD. S-100 is a dimeric calcium-binding protein consisting of two subunits (α and β) [61] . The β unit (S-100β) is highly brain specific, located mainly in astrocytes. Circulating levels of S-100β are elevated in patients with cerebral ischemia [62] , cardiopulmonary bypass-associated decline in explicit memory function [63, 64] , and traumatic brain injury (TBI) [65] [66] [67] . Even in mild head injury, serum levels of S-100β are correlated with clinical measures of injury severity, neuroradiologic findings, and outcomes, including postconcussion symptoms [68] . Elevated serum S-100β levels were recently demonstrated in critically ill patients with respiratory failure [69] and in porcine models of endotoxic shock [70] and acute lung injury [71] . In this latter group, elevated S-100β levels were associated with hippocampal histopathologic changes, including basophilic shrunken neurons in the pyramidal cell layer [71] . Interestingly, S-100β may have both beneficial and detrimental effects, in that lower levels may have protective neurotrophic effects, yet higher levels can lead to exacerbation of neuroinflammation and neuronal dysfunction [72] . Whereas S-100β is a marker of astrocyte damage, NSE and MPB are markers of neuron and white matter (myelin) damage, respectively. NSE is protein-based enzyme that is found primarily in neurons. Serum levels of NSE are elevated after TBI, exhibiting a close relationship with outcome in severe head injury [73, 74] and with volume of contusion in minor head injuries [75] . Interestingly, elevated NSE levels were recently shown to predict death in one small study (n = 29) of patients with severe sepsis [76] , even though these patients had no acute CNS disorders, such as stroke or neurotrauma. MBP is the major protein component of myelin. Serum levels of MBP are elevated in diseases in which there is myelin breakdown. Studies of patients with TBI have shown that MBP levels correlate with clinical measures of severity and may allow early prediction of outcomes [74, 77, 78] . New developments in neuroimaging, such as functional magnetic resonance imaging (MRI) and positron emission tomography, have revolutionized our understanding of abnormal brain function in many disease states, including schizophrenia, Parkinson's disease, and post-traumatic stress disorder. To study further whether critical illness-associated CD is associated with occult brain injury in humans, it would be useful to have an imaging test that can detect subtle evidence of brain injury. Unfortunately, traditional CT scans and MRI do not appear to be sensitive enough to pick up the microscopic cellular changes that may underlie CD [42] . Two small studies assessed brain CT findings in critically ill patients with sepsis [79, 80] . Neither study demonstrated any CT abnormalities, although brain pathology in nonsurvivors was consistent with the previously cited findings of Sharshar and colleagues [49] [50] [51] [52] . A recent study of ARDS survivors (n = 15) [81] found that many of these individuals exhibited signs of significant brain atrophy and ventricular enlargement on head CTs obtained during their acute illness, but there were no significant correlations between these abnormalities and subsequent neurocognitive scores. A new MRI technique may prove useful for identifying occult brain injury in critically ill patients. Specifically, highresolution, three-dimensional MRI can be used to assess noninvasively differences in brain tissue sodium concentration, which is a highly sensitive marker of tissue viability that highlights areas that traditional MRI can miss [82] [83] [84] [85] [86] . The method is based on sodium ion homeostasis, which is tightly regulated in the body and is a major energy consuming process. Any event that perturbs the energy level of the cell enough to disrupt the sodium ion gradient, such as ischemia, has an important impact on cell viability. Although tissue sodium concentration MRI has been successfully used to evaluate the CNS, including nonhuman primate studies and clinical studies of stroke and reversible focal brain ischemia [87] [88] [89] , it has not been used to assess patients with either acute or chronic critical illnessassociated CD. There are several recent developments that, although preliminary, are of interest because of their potential to prevent or mitigate critical illness-associated CD. Haloperidol has been used for many years to manage agitation in mechanically ventilated ICU patients, and it is the recommended drug for treatment of ICU delirium [90] . Kalisvaart and colleagues [91] compared the effect of haloperidol prophylaxis (1.5 mg/day preoperatively and up to 3 days postoperatively) with that of placebo in 430 elderly hip surgery patients at risk for delirium. Although there was no difference in the incidence of postoperative delirium between treatment and control groups, those in the haloperidol group had significantly reduced severity and duration of delirium (5.4 days versus 11.8 days; P < 0.001). Haloperiodol also appeared to reduce the length of hospital stay among those who developed delirium (17.1 days versus 22.6 days; P < 0.001). A recent retrospective cohort study examined haloperidol use in 989 patients who were mechanically ventilated for longer than 48 hours [92] . Despite similar baseline characteristics, patients treated with haloperidol had significantly lower hospital mortality than did those who never received the drug (20.5% versus 36.1%; P = 0.004), an association that persisted after adjusting for potential confounders. Because of the observational nature of the study and the potential risks associated with haloperidol use, these findings require confirmation in a randomized, controlled trial before they may be applied to routine patient care. Leung and colleagues [93] tested the hypothesis that using gabapentin as an add-on agent for treating postoperative pain reduces the occurrence of postoperative delirium. Patients aged 45 years or older undergoing spine surgery were randomly assigned to gabapentin 900 mg or placebo by mouth 1 to 2 hours before surgery and continued for the first 3 days postoperatively. Postoperative delirium occurred in 0% (0/9) of gabapentin-treated patients and 42% (5/12) of placebo patients (P = 0.045). Reduction in delirium appeared to be due to the opioid-sparing effect of gabapentin. Given the small size of the study, these results require confirmation. Donepezil, a cholinesterase inhibitor that increases synaptic availability of acetylcholine, improves cognitive function in Alzheimer's disease. Sampson and colleagues [94] randomly assigned 33 elderly patients undergoing elective total hip replacement to donepezil 5 mg or placebo immediately following surgery and every 24 hours for 3 days. Donepezil was well tolerated with no serious adverse events. Although the drug did not significantly reduce the incidence of delirium (9.5% versus 35.7%; P = 0.08) or length of hospital stay (mean ± standard error: 9.9 ± 0.73 days versus 12.1 ± 1.09 days; P = 0.09), both outcomes showed a consistent trend suggesting possible benefit. The authors project that a sample size of 95 patients would be required for a definitive trial. Dexmedetomidine's sedative effects are due to selective stimulation of α 2 -adrenoreceptors in the locus ceruleus of the CNS. Because it does not have anticholinergic or GABA-stimulating effects, it has the potential to be a delirium-sparing sedative. In preliminary results presented in abstract form [95] , cardiac surgery patients (n = 55) randomly assigned to dexmedetomidine for postoperative sedation had a nonsignificantly lower incidence of postoperative delirium as compared with those sedated with propofol or a combination of fentanyl and midazolam (5% versus 54% versus 46%). The authors of that report plan to enroll a total of 90 patient in the study; perhaps these impressive differences will be statistically significant with a greater number of patients. Recombinant human erythropoietin (rHuEPO) has received considerable attention as a potential transfusion sparing strategy in the ICU. Interestingly, EPO and its receptor are both expressed by the nervous system, and systemically administered rHuEPO can reach sites within the brain. In preclinical studies, rHuEPO reduced neuronal injury produced by focal ischemia, TBI, spinal cord injury, and subarachnoid hemorrhage [96] [97] [98] . Enthusiasm regarding its use as a general neuroprotectant in the ICU has been tempered by potential risks such as thromboembolism and the considerable cost of the drug. Concerns over safety may be at least partially addressed by the recent finding of erythropoietin derivatives with tissue protective but not hematopoietic properties [99] . Xenon is a chemically inert gas that has been used as an anesthetic agent and for contrast enhancement in CT scans. In rats xenon appears to protect the brain from the neurologic damage associated with the use of cardiopulmonary bypass, an effect that is potentially related to N-methyl-D-aspartate receptor antagonism [100] . However, its tendency to expand gaseous bubbles, such as bypass-associated cerebral air emboli, could abolish any beneficial effect or even worsen cerebral outcome [101] . In the setting of ischemic stroke or TBI, there are a variety of compounds with the potential to improve neurologic outcomes. For example, NXY-059, a free radical trapping agent, reduced disability at 90 days when given within 6 hours of stroke onset [102] . In a pilot randomized trial in 56 patients, simvastatin given up to 12 hours after stroke onset significantly improved neurologic functioning (National Institutes of Health Stroke Scale score) at 90 days [103] . Ethyl pyruvate, a pyruvate derivative that prevents mortality in murine sepsis models, reduced motor impairments, neurologic deficits, and infarct volume in a rat stroke model when given as late as 12 hours after middle cerebral artery occlusion [104] . In rodent models of TBI, cyclosporin A reduced acute motor deficits and improved cognitive performance, even when given after the traumatic insult [105] . A phase II dose escalation trial is currently underway in humans. Mounting evidence suggests that mild-to-moderate hypothermia can mitigate neurologic injury. Shankaran and colleagues [106] found that whole-body hypothermia (33.5°C for 72 hours) reduced the risk for death or disability in infants with moderate or severe hypoxic-ischemic encephalopathy. In adults successfully resuscitated after cardiac arrest, moderate hypothermia (32-34°C for 12 to 24 hours) increased rates of favorable neurologic outcomes and reduced mortality [107, 108] . A practical limitation of therapeutic hypothermia is that reaching target temperatures takes at least 2 hours using the fastest currently available cooling techniques. However, Polderman and colleagues [109] demonstrated that hypothermia could be induced safely and quickly (about 60 min) by means of ice-cold intravenous fluid combined with icewater cooling blankets. Cognitive rehabilitation involves the teaching of skills and strategies to target specific problems in perception, memory, thinking and problem solving, with the goal of improving function and compensating for deficits. The benefits of cognitive rehabilitation are well known to those that care for patients with stroke, anoxia, or TBI. Predicting who will benefit and how much has proven challenging, but even severely disabled patients sometimes make dramatic neurocognitive recoveries [110] . Although there are no studies evaluating the effectiveness of cognitive rehabilitation in patients recovering from non-neurologic critical illness, it stands to reason that such patients could benefit when they are found to be cognitively impaired. Because cognitive impairments in critically ill patients appear to be underrecognized by ICU and physical rehabilitation providers [111] , few patients are referred for cognitive rehabilitation therapy [3] . Education regarding the cognitive sequelae of critical illness is needed to enhance referrals for rehabilitation, not only for weakness and physical debilitation but also for cognitive impairments. Cognitive function is an important and relatively understudied outcome of critical illness. Evidence suggests that neurotransmitter abnormalities and occult diffuse brain injury are important pathophysiologic mechanisms that underlie critical illness-associated CD. Markers that could be used to evaluate the influence of these mechanisms in individual patients include the following: SAA, certain brain proteins (S-100β, NSE, and MPB), and MRI tissue sodium concentration. Although recent advances in this area are exciting, they are still too immature to influence patient care. Additional research is needed if we are to understand better the relative contributions of specific mechanisms to the development of critical illness-associated cognitive dysfunction and to determine whether these mechanisms might be amenable to treatment or prevention. This article is part of a thematic series on Translational research, edited by John Kellum. Other articles in the series can be found online at http://ccforum.com/articles/ theme-series.asp?series=CC_Trans
78
ProCAT: a data analysis approach for protein microarrays
Protein microarrays provide a versatile method for the analysis of many protein biochemical activities. Existing DNA microarray analytical methods do not translate to protein microarrays due to differences between the technologies. Here we report a new approach, ProCAT, which corrects for background bias and spatial artifacts, identifies significant signals, filters nonspecific spots, and normalizes the resulting signal to protein abundance. ProCAT provides a powerful and flexible new approach for analyzing many types of protein microarrays.
DNA microarray technologies have proven to be extremely valuable for probing biological processes by measuring mRNA expression profiles. However, studies at the protein level have the potential to provide more direct information since most genes function through their protein products. Traditional investigations focus on individual proteins in a system and then combine such individual analyses to provide a more global perspective. Recently, technologies to analyze proteins in a high throughput and unbiased fashion have become feasible [1] . One particular powerful technology is protein microarrays, which contain a high density of proteins and allow a systematic probing of biochemical activities [2, 3] . There are two types of protein microarrays [3] . A 'functional protein microarray' contains a set of proteins individually produced and positioned in an addressable format on a microarray surface. Functional protein microarrays are useful for identifying binding activities or targets of modification enzymes. The first version of a proteome microarray was reported in 2001 and contained 5,800 yeast proteins with amino-terminal glutathione S-transferase (GST) tags printed on the array [4] . A second version of yeast protein microarrays was generated recently and contained 5,600 proteins with carboxy-terminal 6His-HA-ZZ domain tags [5] . Proteins from both collections were overexpressed, purified and spotted onto the protein microarrays. Global proteome studies were performed on these chips to understand various biological mechanisms. For example, 87 yeast kinases were examined for their substrates using yeast protein microarrays and over 4,200 in vitro substrates representing 1,325 unique proteins were identified [6] . Compared with the approximately 150 known in vivo kinase-substrate interactions, this global study served as an important first step for dissecting yeast signaling networks. In addition to searching for kinase substrates, proteome chips can be probed with labeled proteins, DNA, lipids, antibodies and many other molecules to search for interacting proteins [4, 7, 8] . Large amounts of data have been generated using protein microarrays, presenting significant challenges in developing robust methods to process the raw data and building reasonable biological hypotheses from the datasets. The second type of protein microarray, the 'analytical protein microarray' or 'antibody microarray', shares similarities with immunoassays and uses antibodies to detect specific probes. Studies have shown that these antibody arrays can recognize specific targets and generate dose-dependent signal intensities, indicating that they can be used to quantify levels of various targets in a crude mixture [9, 10] . Because of the crossreactivity of certain antibodies with a variety of proteins, only highly specific antibodies are suitable for this type of study. This remains a limiting factor in preparing antibody microarrays. Both DNA and protein microarrays are prone to systematic errors that are usually generated from different sources, such as surface defects and spatial artifacts. Many studies have offered insight on noise subtraction in DNA microarrays [11] [12] [13] [14] , but little investigation has been done for protein microarrays. Functional protein microarrays differ in many respects from DNA microarrays. First, the goals of these two microarrays are different. DNA microarrays measure the relative DNA levels in a pool of probes, whereas functional protein arrays often aim at discovering global interactions of a single probe molecule. Second, a typical DNA microarray experiment measures signal ratios between two color channels, one for a tested mRNA sample and the other for a reference sample [15] . Signals in the second channel may serve as intrinsic controls that can help to decrease the effects of various amounts of reagent on the arrays and any local array nonuniformity. Furthermore, many current scaling methods are then based on the assumption that signal intensities should be balanced between the two color channels despite variation in slide location, intensity and other sources of systematic variation [16] [17] [18] . However, such controls are missing in onecolor-channel protein microarrays. Third, several scaling approaches in DNA microarrays are based on a set of 'housekeeping' genes that give constant signal intensities at different conditions [19, 20] . However, in protein microarrays, such a control group must be customized according to the type of activities that are assayed, and, therefore, a ubiquitous reference group does not exist. Fourth, unlike DNA microarrays, in which non-specific binding can often be addressed by signal comparison with mismatch probes [21] , cross-reactivities of protein microarrays can not be as directly corrected for. A separate slide is, therefore, often required to be probed in parallel as a negative control in protein microarray experiments. Finally, several protein-specific artifacts serve as common noise sources in protein microarrays. In the kinase assay, for example, the signal from strongly phosphorylated spots can bleed into neighboring spots, leading to incorrect background measurement. These differences are particularly applicable to functional protein microarrays in comparison to antibody arrays, and, therefore, the normalization techniques used for DNA microarrays are usually not directly applicable to functional protein microarrays. We have developed a new protein chip analysis tool (ProCAT) to deal with various artifacts specific to functional protein microarrays. The work started from a careful survey and characterization of all potential sources of systematic errors in protein microarrays. Specific approaches were then designed to deal with each type of noise. A correction approach is applied to reduce measurement errors in the background signals. In addition, spatial variations can be reduced efficiently through a novel two-parameter signal normalization approach and calling positive spots locally. After generating a list of positives, negative control slides are analyzed in the same approach and spots are subtracted from the list if they appear in the control slide. Slide features with poor signal qualities are also removed. Finally, signal intensities of the positives are normalized according to their protein amounts. All modules that account for the challenges in data processing specific to protein microarrays are built into ProCAT and tested. ProCAT contains a flexible modular design whose individual components can be adjusted according to the experimental designs and stringency level selected by the users. Six sequential modules are currently implemented in ProCAT before a final annotation report is assembled ( Figure 1 ). These modules carry out: background correction; signal normalization; positive spot identification; spot cross-reactivity filter; signal qualities inspection; and protein amount normalization. The performance of many of the steps was tested using several types of experiments as described below. A fundamental issue in all microarray experiments is background correction, which aims at reducing noise in background quantification. Signal intensities are generally quantified by subtracting the foreground intensities with the local background intensities, which are measured as the background signals immediately surrounding the spot of interest (termed here the 'adjacent background'; Figure 2b ). However, in protein microarrays local background regions can be easily skewed by artifacts such as small speckles. In addition, strong positive signals from on-chip kinase assays tend to produce signal smears on both film and phosphoimagers that exceed the normal feature size (Figure 2a) . In both cases, the measurement for that spot will be inaccurate. First, the background intensity will be arbitrarily high, which will diminish the real signal intensity for that spot. Second, the intensities will be affected by the alignment of the grid and extent of the smear, and, therefore, the variance of the same protein at replicate experiments will be increased. Two methods can reduce the artifacts in local background. The user can manually adjust the grid size to fit the circles to each individual spot. However, the aligning process requires considerable time and effort. The size of the smear may even prevent refitting the grid without adversely affecting neighboring spots. Additionally, a larger spot size can diminish the signal of the spot because the signal density decreases with increasing spot size. The second method for background correction, which is applied in ProCAT, replaces the background intensity of the central spot with the background from its local neighborhood. A three by three surrounding window is assigned to each protein spot, and the median background of the nine spots will be used as the 'neighborhood background' value for the central spot (see Materials and methods for more details). No additional time is needed for further alignment, yet this method will significantly reduce artifacts that can produce erroneous measurements on spots background. In the analysis of the phosphorylome dataset [6] , we applied the neighborhood background correction and observed a high sensitivity in identifying positive targets. To further characterize the effects of neighborhood background correction, we performed a test kinase assay with 100 nM protein kinase A (PKA) spotted at 96 locations on one slide (Figure 2a) . Each of the 48 blocks on the slide contains two PKA pairs with random yeast proteins spotted elsewhere (approximately 12,000 spots). After incubating the slide with 33 P-γ-ATP, all of the PKA spots autophosphorylated and showed strong signals, and in many cases the signal went beyond the grid circle boundaries (Figure 2b ). We then applied the neighborhood background correction to the PKA spots. As expected, the median for PKA signal intensities was enhanced by 53%. Furthermore, the PKA signals from different positions are more similar to each other; the variance within them is decreased by 41% (p value = 0.006; Fig 2d) . Therefore, the neighborhood method for accessing background provides more robust measurements than that of the adjacent background method. Spatial artifacts arise from uneven signal distribution across the slide, in part due to uneven probing conditions and smear artifacts [13] . Uneven probing can occur by several means, such as uneven mixing of the probe, exposure to the probe solution, or uneven washing and drying of the slides. Twocolor-channel experiments of DNA microarrays provide intrinsic controls that can be used to account for spatial artifacts. Functional protein microarrays often use only one color channel and, therefore, are especially prone to spatial artifacts. Spatial artifacts will cause inaccurate measurements of signal intensities and can hinder the identification of significant interactions. Adding more controls can help remove spatial artifacts since the signal of each spot can then be normalized according to its local controls. Due to the variable shape and size of spatial artifacts, ideally a large number of controls would be needed. However, space constraints of the protein chip and an inability to anticipate all the uses of the arrays usually prevent the necessary number of controls to fully account for spatial artifacts on the array. A scaling method that reduces signal variations among spots of the same proteins at different array locations decreases spatial artifacts. We developed a new normalization method to deal with the spatial artifacts specific to functional protein microarrays. By assuming that signal distribution in large windows is consistent across the slide, the foreground signal of each spot can be normalized according to signal intensities in its surrounding neighborhood. This assumption is usually valid in protein microarray experiments in which proteins are randomly printed on the array ( Figure 3 ). Two parameters, the median and the median absolute deviation (MAD), are calculated to represent the signal distribution in the local window ( Figure 4 ). To perform the normalization, the median and MAD of all sliding windows are averaged. The average values are then used to correct the signal of the central spot to To test the performance of this two-parameter scaling approach for signal normalization within one slide, we designed a test microarray containing multiple positive controls printed at different positions on the slide. The test array was organized in the same format as the commercially available protein microarrays (Invitrogen). Each protein was printed in duplicate, and the array contained 24 blocks of 16 by 16 printed proteins (Figure 5a ). Two GST-fusion proteins, Sla2p and Myo4p, were purified separately and a 1:1, 1:5, and 1:25 dilution of each protein was prepared. Sla2p and Myo4p at each concentration were printed at eight random positions on the array. Other spots were occupied with bovine serum albumin (BSA) as negative controls. In order to visualize the two fusion proteins, anti-GST antibody was used to probe the slide, and one probing with typical spatial artifacts is shown in Figure 5 . The artifact-containing slide showed different signal levels between the edges and the middle portion of the array. This produced blocks that had a variable signal distribution that ranged from high to low from one edge of the slide to the opposite edge; the variability occurred across blocks and simple block normalization methods adopted in DNA microarray normalization approaches [17] would not be suitable for dealing with this problem. We applied ProCAT to normalize the slide with several different parameters ( Figure 5) . Five window sizes were tested, termed windows 1, 3, 5, 7, and 9. These numbers correspond to the window size as a function of the number of spots on one edge of a block. For example, a block of 20 by 20 spots analyzed using window 1 would have a window size of 0.1 that of the block edge, or in this case 2 spots above, below, and to either side of the central spot, whereas a window size of 9 would contain a 37 by 37 area roughly as large as 4 blocks. Three observations were made from the analysis of different window sizes. First, as the window size increases, the computational time used for the normalization also increases. Second, no obvious spatial artifacts were left after the normalization with any of the window sizes tested ( Figure 5b ). Third, a small window size diminishes any signal inequality that exists between positive signals and background noise. Indeed, a small scaling window tends to introduce extreme changes to the original signals and, therefore, increases the discrepancy between the duplicate spots of the same protein. The variance of the signals for the same protein after normalization with different window size was calculated. In five out of the six cases (three dilutions of two proteins) the scaling window 9 can successfully reduce the signal variance in a range from 31% to 90% (Figure 5c ). Decrease of signal variation suggests that a large scaling window will help to reduce spatial artifacts. Although larger window sizes are possible, 9 was used as the default number for ProCAT because the analysis can be done in a reasonable time and minimal improvement has been achieved after window size 7 (Additional data file 1). In addition to providing accurate measurements of spot intensities, ProCAT has been developed to assign thresholds for identifying positive targets in one experiment. Traditionally, a global cutoff can be calculated from all spots and applied to the whole slide. Due to variable spatial artifacts, cutoffs were assigned locally in ProCAT. For each spot on the array the signal distribution within a nine by nine window was calculated and a cutoff defined as a number of standard deviations away from the mean; the default for ProCAT is two standard deviations. This cutoff corresponds to 5% significance level if the signal distribution within this local window is normal. When many spots with strong signals are included in the window, the cutoff will be arbitrarily high and thus decrease the sensitivity of detecting positive spots by the program. To avoid this loss in sensitivity, ProCAT has a built in function to identify possible outliers, to remove those outlier spots that have extremely strong signals, and then to calculate a cutoff for identifying positive spots using the remaining spots. A receiver operating characteristic (ROC) curve was used to compare the performance of local window cutoffs versus a global cutoff on the test slide [22] . Area under ROC curve (AUC) is a performance indicator that ranges from 0 to 1, with 1 for the best performing method. Using GST-Sla2p and GST-Myo4p as positive controls and BSA as negative controls, the sensitivity and specificity for both local and global cutoff methods was estimated. Five window sizes were tested and compared with the global cutoff ( Figure 6 ). Prediction performance is increased significantly when using local windows with nine or more spots on one edge. Thus, a nine by nine window is used as the default in ProCAT since a larger A representative protein microarray with high-quality data Figure 3 A representative protein microarray with high-quality data. The slide image was reconstructed from a protein microarray experiment with minimal noise in the data. Density plots of signals in local 37 by 37 windows (window size 9) for all spots were computationally combined, and they showed high similarities. window size results in increased computing time with only minimal improvement in sensitivity. The AUC value is much larger in local cutoffs (0.992) compared to global cutoffs (0.916) and the improvement is unlikely to be due to random chance (p value = 0.002) [20] . Therefore, we can conclude that the local cutoff is significantly better in identifying positive spots than a global cutoff. Scheme for the signal scaling method Figure 4 Scheme for the signal scaling method. The signal of one spot on the array is normalized according to the distribution in its local neighborhood. For each spot, a surrounding window is chosen and all spots in this window are defined as its neighborhood. The signal of a center spot will then be normalized by comparing the local median and MAD with the average values. Norm, normalized signals; Origin, original signals. Average distribution Two layers of filters are implemented in ProCAT. First, all positive spots from negative control experiments are removed. For example, in on-chip kinase assays, kinase dead alleles were probed on separate arrays using the same experimental conditions as used with wild-type kinases. Spots that produce signals in the absence of active kinase were identified by ProCAT and removed from the target lists of kinase probings. When probing tagged protein to detect protein-protein interaction, testing the epitope tag in the absence of the protein of interest is also an essential control. If proper negative control experiments are available, ProCAT will analyze them in the same way as regular experiments to construct experimental positive spot lists void of proteins producing positive signals under control conditions. The second filter checks the quality of each positive spot. All proteins are spotted in duplicates on protein microarrays, hence should have very similar signal intensities. ProCAT then uses the difference between duplicate signals as an indicator of the signal qualities. The difference between signals of two duplicate spots (s 1 , s 2 ) is calculated as (s 1 + s 2 )/(|s 1 | + |s 2 |) and then fitted to a normal distribution. Proteins with exceptionally large differences in their duplicate spots are more likely to be biased by certain artifacts, and thus are removed from the positive list. The default threshold for the duplicate spot difference in ProCAT is set at two standard deviations away from the mean. One of the goals for protein microarray experiments is to identify the affinity of a binding interaction (in a protein-protein interaction assay) or the extent of phosphorylation (in a kinase assay) so that one can compare the relative strength of the reaction for each positive protein. Ideally, the spot intensity would directly correspond to the strength of interaction. However, a number of other factors contribute to the array signal intensities, including the systematic noise from various artifacts, as was already discussed, and the amount of protein printed on the chip. Nonetheless, semi-quantitative estimates can be obtained. After background correction and signal normalization, the raw signals can be standardized by relative protein amounts before they can be used to estimate the interaction strength. Although proteins on the microarray can have very different amounts, they do share the same epitopes for the purpose of large-scale protein purification [4, 5] . Therefore, probing with anti-epitope antibodies will provide an estimate of the relative protein amounts in each spot on the array. After the protein amount is determined for one spot at row i and column j, ProCAT divides the raw signal intensities S i,j by the protein amount signals A i,j and uses the quotient as an approximation of the strengths of interactions: This approximation generally works well across the slide except for the following two situations. Less abundant proteins will be biased because the A i,j values estimated in antiepitope probings are more susceptible to background noise and slide artifacts. On the other hand, overpowering spots can also be biased if they have saturated signal intensities. A saturated S i,j value is an underestimate to the real signal. For these two reasons, only proteins with amounts more than a minimal cutoff and signal intensities lower than a saturation threshold will be normalized with protein amounts. Proteins that do not conform to these two requirements will be recorded with unnormalized signals and flagged for further inspection. An additional caveat is that the relative protein amount assessed using antibodies includes both native and denatured protein at a given spot. Therefore, the estimation of interaction strength will be an underestimate since the amount of functional protein may be an overestimate. ProCAT was designed as a flexible tool to analyze functional protein microarray data. The program was scripted in Perl (version 5.6.1) on top of a Tomcat (version 5.0.30) web server [23] . Each module discussed above was implemented independently and can be included or excluded depending on various experimental designs. To input a dataset, the user has to characterize the data in three aspects: experimental designs, data file formats and normalization parameters. Experimental design contains parameters such as the number of test arrays and negative control arrays for one particular assay. Data file format describes the layout in the uploaded dataset so that ProCAT can recognize and extract the useful information from it. Normalization parameters allow users to try different stringency levels. These three levels supply sufficient information to uniquely characterize an experiment while still allowing ample flexibility for the individual user to customize parameters to suit many different types of experimental designs. After inputting all three descriptions and uploading the dataset, ProCAT takes five minutes on average to complete all analysis modules for each array. The time may vary depending on the selected analysis modules and the size of the protein microarrays. Each task is assigned a unique ID and results are organized into a database for future queries. Processed data including analysis parameters, a list of positive spots with protein annotations, and normalized signal intensities will be available for the users to download from the server. Functional protein microarrays serve as an efficient platform for screening protein biochemical functions. Here we present ProCAT as a systematic approach to process and analyze data specific to functional protein microarrays. Calibrated by explicit test experiments, ProCAT has proven to be able to handle many types of functional protein microarray studies with three unique features. ProCAT includes novel scaling methods that provide robust and reproducible measurement for quantitative signals. This is crucial for protein microarrays as chip signal intensities often indicate strength of interactions. In addition, by calling positive candidates locally, ProCAT demonstrated excellent performance in identifying positives in comparison to global thresholds. Finally, each step has been integrated into a modular design to fit various experimental designs and stringency requirements. A major challenge in designing any automated data processing method is thinking of and anticipating all possible situations that may arise. ProCAT uses a local three by three window to correct background containing signal smears or dust speckles. This method assumes the artifacts are sparse enough so that the majority of the nine spots in the local window still provide correct measurements of the background signals. Since the median value of nine spots is used to correct the background, a few biased spots within the window will not severely affect the corrected background value. This assumption is usually valid since the percentage of spots that are either positive or whose signal is contaminated by artifacts in protein microarray experiments is generally quite low. In extreme cases where such spots are likely to be very close to each other, a larger window (five by five for example) can be used. Large artifacts such as bright speckles and incubation bubbles may affect many spots in a particular region. Since the shapes of these artifacts are variable, it is necessary to manually flag these spots initially and then remove them from future analysis. Many commercially available software packages for microarray experiments have a built in flagging function, and ProCAT will automatically discard flagged spots. A key aspect of ProCAT is the two-parameter approach for reducing spatial nonuniformity. Several factors can affect the performance of ProCAT's normalization. First, ProCAT normalizes the signal of a spot according to the signal distribution in its local neighborhood. It diminishes the signal intensity if the spot is located in a high signal neighborhood, while compensating the intensity if it is in a low signal neighborhood. This approach is based on the assumption that signal intensities across the slide share the same distribution, and it holds true if and only if the regional variations observed on the slide are due to technical artifacts and not from real biological differences. Since proteins are printed in a random order on most of the current protein microarrays, it is unlikely a particular region of the slide will gain high intensities as a result of biologically relevant reasons. Second, the size of the neighborhood window can also largely affect the performance of the normalization. Small window sizes tend to add biases to signals and diminish all local variations, whereas large window sizes increase the computational burden and tend to preserve local variations. We found that the optimal window size of ProCAT is 9 for protein-protein interactions; this figure corresponds to approximately four blocks on the chip and is used as the default. Other window sizes can also be chosen to fit various shapes of spatial artifacts. ProCAT can be applied to many experiments using protein microarrays, such as kinase assays, protein-protein interactions and protein-DNA interactions. Thus far, the twoparameter scaling approach has only been used in single chip normalization; however, a similar strategy can be extended to rescale multiple slides by assuming signals in neighborhood windows on different slides are similarly distributed. Overall, ProCAT provides a powerful and flexible new approach for optimal processing and analysis of functional protein microarrays. For the slide used for testing background correction, 100 nM PKA (Sigma, St. Louis, MO, USA) was spotted at 96 different places as positive control. The slide was incubated with 200 μl of kinase buffer (100 mM Tris pH 8.0, 100 mM NaCl, 10 mM MgCl 2 , 20 mM glutathione, 20% glycerol) plus 0.5 mg/ ml BSA, 0.1% Triton X-100, and 2 μl 33 P-γ-ATP in a humidified chamber at 30°C for 1 hour. The slide was then washed twice with 10 mM Tris pH 7.4, 0.5% SDS and once with double distilled H 2 O before being spun dry and exposed to X-ray film (Kodak, Rochester, NY, USA). For the anti-GST probing, slides were printed with Sla2p and Myo4p as positive controls and 150 nM BSA as a negative control. The array surface was blocked using SuperBlock (Pierce, Rockford, IL, USA) at 4°C for 1 hour. Rabbit polyclonal IgG (Santa Cruz Biotechnology, Santa Cruz, CA, USA) was incubated with the slides at 1,000-fold dilution. The array was then washed with PBST (Sigma) and incubated with a 1:1,000 dilution of Cy5-conjugated anti-rabbit IgG antibody (Jackson Laboratories, Bar Harbor, ME, USA). Slides were then washed with PBST five times and scanned in an Axon Testing experiment for the signal scaling approach Figure 5 (see previous page) Testing experiment for the signal scaling approach. (a) The design of the test slide with positive spots shown as red spots and the five tested normalization window, indicated by red squares, for a given spot on the array, shown in blue. (b) Comparison of signal intensity before and after normalization using window size 9 on the testing experiment. The two images were computationally reconstructed from the signal files, either without or with normalization. GenePix scanner (Molecular Devices, Sunnyvale, CA, USA). Raw signals were extracted with GenePix Pro 6.0 software (Molecular Devices). For one spot, let i be the row and j the column on a protein microarray. Thus, B i,j represents the adjacent background intensity and F i,j denotes the foreground intensity. The raw signal intensity S i,j is calculated as: In neighborhood background correction, we use neighborhood background to replace the adjacent background. A local three by three window around B i,j is chosen and the neighborhood background is defined as: In a protein slide with N rows and M columns, a local window W i,j around one spot (i, j) is defined as signals of a set of spots S i,j that satisfy: The size parameter k is dependent on window size factor f win and the block size f block : in which f block represents the number of spots on one edge of the block, and f win is chosen by users from five options: 1, 3, 5, 7 and 9. Different windows can overlap with each other and go beyond the block edges. Let s denote signal intensities of spots within the local window; ProCAT uses two parameters to characterize the signal distribution of s: median (MED) and median absolute deviation (MAD): After calculating MED i,j and MAD i,j for all the spots on the array, they are averaged to obtain the two parameters and for the reference distribution. For one spot (i, j), ProCAT normalizes its raw signal S i,j by comparing MED i,j and MAD i,j with the average values: For a given spot at row i and column j, its normalized signal is compared to surrounding spots in a nine by nine window W ij (4) . Signals within this window are fit to a normal distribution. The mean μ i,j and standard deviation σ i,j will be calculated and the default threshold is set at two standard deviations above the signal mean. A spot (i, j) will be called positive only if its signal is above the threshold: When positive spots are likely to be close to each other, Pro-CAT uses box plots to examine and remove possible outliers from the surrounding window [24] . Let Q 1 be the lower quartile (25th percentile) and Q 2 be the upper quartile (75th percentile); the difference between Q 1 and Q 2 is termed interquartile range ΔQ. A spot (i', j') is then defined as an outlier if its signal: ROC curve comparing the global cutoffs and local cutoffs in calling positive spots Figure 6 ROC curve comparing the global cutoffs and local cutoffs in calling positive spots. The test slide has six unique positive controls (Sla2p and Myo4p in three different titrations). The performance of identifying the positive controls is increased by using local cutoffs generated in relatively large surrounding windows. Five window sizes were tested and the best performance was achieved using nine by nine or larger windows. To obtain a robust threshold, the corrected mean and standard deviation are generated using the non-outlier spots. The following additional data files are available with the online version of this paper. Additional data file 1 is a figure illustrating the variance reduction in positive controls using different normalization window sizes. Additional data file 2 is a table listing the raw signals generated in the autophosphorylation experiment for testing the background correction method. Additional data file 3 is a table listing the raw signals generated in the anti-GST probing experiment calibrating the signal scaling approach. Additional File 1 Variance reduction in positive controls using different normaliza-tion window sizes Variance reduction in positive controls using different normaliza-tion window sizes. Click here for file Additional File 2 Raw signals generated in the autophosphorylation experiment for testing the background correction method Raw signals generated in the autophosphorylation experiment for testing the background correction method. Click here for file Additional File 3 Raw signals generated in the anti-GST probing experiment cali-brating the signal scaling approach Raw signals generated in the anti-GST probing experiment cali-brating the signal scaling approach. Click here for file
79
Positional clustering improves computational binding site detection and identifies novel cis-regulatory sites in mammalian GABA(A) receptor subunit genes
Understanding transcription factor (TF) mediated control of gene expression remains a major challenge at the interface of computational and experimental biology. Computational techniques predicting TF-binding site specificity are frequently unreliable. On the other hand, comprehensive experimental validation is difficult and time consuming. We introduce a simple strategy that dramatically improves robustness and accuracy of computational binding site prediction. First, we evaluate the rate of recurrence of computational TFBS predictions by commonly used sampling procedures. We find that the vast majority of results are biologically meaningless. However clustering results based on nucleotide position improves predictive power. Additionally, we find that positional clustering increases robustness to long or imperfectly selected input sequences. Positional clustering can also be used as a mechanism to integrate results from multiple sampling approaches for improvements in accuracy over each one alone. Finally, we predict and validate regulatory sequences partially responsible for transcriptional control of the mammalian type A γ-aminobutyric acid receptor (GABA(A)R) subunit genes. Positional clustering is useful for improving computational binding site predictions, with potential application to improving our understanding of mammalian gene expression. In particular, predicted regulatory mechanisms in the mammalian GABA(A)R subunit gene family may open new avenues of research towards understanding this pharmacologically important neurotransmitter receptor system.
Understanding transcription factor (TF) mediated control of gene expression remains a major challenge at the interface of computational and experimental biology. Computational techniques predicting TF-binding site specificity are frequently unreliable. On the other hand, comprehensive experimental validation is difficult and time consuming. We introduce a simple strategy that dramatically improves robustness and accuracy of computational binding site prediction. First, we evaluate the rate of recurrence of computational TFBS predictions by commonly used sampling procedures. We find that the vast majority of results are biologically meaningless. However clustering results based on nucleotide position improves predictive power. Additionally, we find that positional clustering increases robustness to long or imperfectly selected input sequences. Positional clustering can also be used as a mechanism to integrate results from multiple sampling approaches for improvements in accuracy over each one alone. Finally, we predict and validate regulatory sequences partially responsible for transcriptional control of the mammalian type A g-aminobutyric acid receptor (GABA A R) subunit genes. Positional clustering is useful for improving computational binding site predictions, with potential application to improving our understanding of mammalian gene expression. In particular, predicted regulatory mechanisms in the mammalian GABA A R subunit gene family may open new avenues of research towards understanding this pharmacologically important neurotransmitter receptor system. Co-regulation is a basic mechanism to coordinately control expression of genes in modules, biochemical pathways and protein complexes (1) (2) (3) . In eukaryotes, expression is most often mediated by transcription factors (TFs) that bind upstream of the transcription start site (TSS) and recruit the polymerase assembly (4) . TFs bind, with varying affinity, to a set of similar, short (6-20 nt) sequences (5) . Computational binding site discovery focuses on finding significantly overrepresented sequences in upstream regions of co-regulated genes (6) (7) (8) . Thus, computational TFBS prediction algorithms must begin with an input set of promoters from genes hypothetically co-regulated by a shared TF. The algorithms aim to predict the binding positions and consequently the nucleotide specificity of the TF (9) (10) (11) . The first part of transcription factor binding site (TFBS) discovery, the input set, can be identified using either computational or experimental methods. Experimental techniques, such as chromatin immunoprecipitation (ChIP) (12) , have been successfully used to generate a genome scale mapping of approximate TF-binding positions (10, 13, 14) . Computational techniques, such as phylogenetic profiling (15, 16) and artificial neural networks, can also be used to identify sets of co-regulated genes. Both experimental and computational approaches, however, suffer from a significant false positive (FP) prediction rate. Inclusion of extraneous promoters in the input sets dilutes the enrichment of the shared TFBS sequences making computational TFBS discovery significantly more challenging (17) . We term such erroneously included promoters decoy sequences (DSs). After receiving a set of upstream regions co-regulated by a shared TF as input, computational methods aim to predict the binding positions of that TF (6) (7) (8) 18) . Given a set of input promoters, motif detection algorithms identify a set of short, oligonucleotide segments hypothesized to bind to the TF of interest. The predicted sequences can be used to construct a position weight matrix (PWM) representing the average nucleotide frequencies for each position in the site (19) . Ideally, computational detection will return all sequences that bind to every TF with biologically relevant function in those upstream regions. However, since the source of binding specificity for TFs is not well understood (20) , heuristic approaches and ad hoc multiple alignment based scoring schemes are used to identify locally optimal solutions (17) . Each local optimum that exists in a given set of promoters may correspond to distinctly different motifs, and may score differently relative to each other according to different scoring schemes. Binding site prediction algorithms are generally confounded by several factors: degeneracy in the binding site; the unknown length of the binding site; the relatively large length of promoters; and the inclusion of DSs in the input sets (17, 21, 22) . As a result as few as 10% of predicted positions correspond to biologically functional binding sites (23) . Due, in part, to the low accuracy rate, computational binding site identification has been of limited use (23) . Problems identifying binding sites are further exacerbated in mammalian genomes by larger promoter regions (24) and scarcity of reliable information on co-regulation of genes. Thus, the most demanding test of efficacy for TFBS identification approaches lies in their application to mammalian systems and subsequent validation of predictions. Because of computational complexity of the problem, Gibbs sampling is often used to identify binding positions (18) . In this paper, we present a new strategy that clusters Gibbs sampling results at each input nucleotide-a technique we term positional clustering-to improve accuracy of predicted TF binding. We evaluate the efficacy of our approach using known examples of binding and regulation in yeast and experimentally testing predicted TF-binding sites upstream of the subunit genes coding for the heteromeric mammalian neurotransmitter receptor system, the type A g-aminobutyric acid receptor (GABA A R). The GABA A R is the major inhibitory neurotransmitter receptor in the central nervous system (CNS) (25, 26) with important roles in development (27, 28) and disease (29) (30) (31) . The receptor is believed to be a pentamer made up of multiple subunits that come from at least four different subunit classes (a, b, g and d) (32) . At least 19 genes code for the various subunits that differentially combine to form numerous pharmacologically distinct GABA A receptor isoforms (29, 30) . Isoform utilization depends in part on the relative abundance of the subunits, which may change under various conditions (33) (34) (35) . Understanding subunit regulatory mechanisms may provide insight into GABA A receptor isoform usage and related phenotypes (36) . In the current study, we test the ability of positional clustering to detect known TF-binding sites in a series of increasingly noisy sets of yeast promoters, and found marked improvement in the percentage of correct predictions over Gibbs sampling alone. We also present de novo predictions of TF-binding sites in promoter regions of GABA A receptor subunit genes (GABRs) whose expression is altered (either up-regulated or down-regulated) in an animal model of temporal lobe epilepsy (35) . Positional clustering identified a number of putative cis-regulatory sites, many of which correspond to known binding elements for TFs found in the CNS. Mobility shift assays showed several predicted GABR-binding sequences specifically bind nuclear proteins derived from primary neocortical neurons kept in culture. Furthermore, a particular non-consensus GABR putative regulatory sequence was shown to have a functional role in cultured cortical neurons demonstrating the efficacy of positional clustering in detecting functional regulatory elements in mammals. We identified S.cerevisiae genes predicted at high confidence (P < 0.001) to be regulated by the TF STE12 in YPD growth media, according to whole-genome TF location data (14) . For the 51 identified genes, we collected upstream intergenic promoters. Intergenic regions were truncated at 1 kb upstream of the gene's TSS. We selected for study a set of six GABRs: GABRA1, GABRA4, GABRB1, GABRB3, GABRD and GABRE. Promoters were extracted for each gene, including two alternative first exons of the GABRB3 (37), giving a set of seven promoters. The length of each promoter was: GABRA1, 3733 bp; GABRA4, 1546 bp; GABRB1, 1353 bp; GABRB3 (exon 1), 1310 bp; GABRB3 (exon 1A), 2080 bp; GABRD, 6625 bp; and GABRE, 5278 bp. We augmented the input set with orthologous promoters from rat, with the exception of GABRB3 for which an orthologous gene from mouse was used. In total, 14 promoters upstream of six GABRs were selected for analysis. For a given input set of promoters, we ran the Gibbs sampler BioProspector (8) 400-550 times, evenly distributed across all motifs widths from 6-15 bp. We used a third-order background model derived from appropriate genomic promoters. We collected the best three results from each BioProspector run. We counted the number of times BioProspector identified each nucleotide in the input set. For each promoter, we identified the maximally occurring nucleotide, and extracted all positions identified by BioProspector >35% of the maximum. We clustered together neighboring positions into putative TFBS. As a dust filter, we removed all putative TFBSs <6 bp long ( Figure 1 ). For sets of S.cerevisiae promoters, we used 1200 results from 400 BioProspector runs in our evaluation. For GABRps, we considered all 127 non-empty subsets of the seven promoters (orthologous sequences were always considered together). We used results from 70 000 BioProspector runs, evenly distributed across all promoter subsets, in our analysis. In addition to dust filtering, we required putative TFBSs to occur both in the human and in the orthologous rodent promoter. We used positive predictive value, to evaluate STE12-binding site predictions. We classified predictions as true positive (TP) or false positive (FP) by comparison to the STE12-binding motif, TGAAACA, as determined by (14) . For each sequence, we calculated distance from the known STE12 PWM using a modified local ungapped sequence alignment similar to that in (38) . Alignments were scored as the sum of Pearson's correlation coefficient, between prediction X and the STE12 PWM across all aligned positions. Thus, scores ranged from zero, with no positions aligned, to seven, the length of the STE12 PWM. We observed a bimodal distribution of scores (Supplementary Figure S1) , and chose the alignment score corresponding to the minima of the distribution (alignment score ¼ 4.5) as the threshold to classify predictions as TP or FP. We complemented the seed set of 51 STE12-bound promoters with 1-50 randomly chosen yeast promoters. We performed our motif detection procedure on each input set, and compared the PPV of putative TFBS with that of raw BioProspector results ( Figure 2 , solid lines). To evaluate the background rate of STE12-binding site recovery, we created a seed set of 51 randomly chosen S.cerevisiae promoters. We evaluated the percentage of STE12-like binding sites identified in the random seed set, as well as in versions of the seed set augmented with 1-50 randomly chosen yeast promoters (Figure 2, dashed lines) . For additional yeast evaluations (HAP4, TEC1, YAP1 and YDR026C), we substituted for BioProspector an in-house implementation of the BioProspector algorithm. Comparisons of results from each implementation show the two implementations to be approximately equivalent. We ran MotifScanner (39) to search GABR promoters for all vertebrate TF-binding motifs found in TRANSFAC (40) . For each promoter analyzed, we used a prior probability of 0.1 and the corresponding organism specific third-order promoter background model from Eukaryotic Promoter Database (EPD) (41) .We considered positional overlap between Motif-Scanner predictions and putative TFBSs indicative of known binding motifs in our predictions. Double-stranded oligonucleotides for EMSA contained the following sequences: Nuclear extracts were prepared (42) and used for gel shift analysis after concentration (Microcon no. 10 columns, Amicon, MA). Quantification was performed on EMSAs under conditions that yield a standard curve for band intensity. Single-stranded sense and antisense phosphorothioate oligonucleotides for the predicted GGCGGCGTGCACACACACGC-CCACCGCGG binding site are annealed by boiling sense and antisense oligonucleotides for 5 min at equal molar ratios in dH 2 O. Oligos are then cooled to room temperature and placed on ice. Transfections using DOTAP (Roche)/HEPES solutions are performed with oligonucleotides corresponding to wildtype, mutant or with DOTAP (Roche)/HEPES solution lacking oligonucleotides (MOCK) as described in (29) . Effects of oligonucleotide application to neurons are assessed by real-time RT-PCR. Since TFBS are predicted computationally by local optimization strategies, we evaluate the extent to which one of these strategies, Gibbs sampling, identifies the same set of segments in repeated runs using the same input data. Identifying stably recurring motifs requires clustering of related results which, in turn, requires definition of 'related'. Sequence similarity based clustering is impaired by the combination of sequence variation within motifs, the short length of TF-binding sites, and aligning motifs of different lengths. Instead of using sequence based clustering, we chose to cluster results by position, counting the number of times Gibbs sampling identifies each nucleotide in the promoter (Figure 1 ). We find that Gibbs sampling predictions, generated using BioProspector (8) are power-law distributed over nucleotide position (Supplementary Figure S2) . Gibbs sampling converges on the majority of nucleotides very infrequently, and a small number of nucleotides very frequently. Thus, the most frequently recurring nucleotides appear in as few as 10% of results. Moreover, we find the power-law distribution of results is robust to Gibbs sampling algorithm and scoring scheme (data not shown). We can hypothesize that the most frequently occurring positions are the most biologically significant. Thus, discarding the least frequent Gibbs sampling results may yield higher accuracy and robust identification of biologically insignificant positions. As a preliminary test of the above hypothesis, we applied repeated runs of Gibbs sampling to a set of 51 S.cerevisiae promoters enriched in STE12 binding as identified by wholegenome ChIP-chip experiments (10) . We used positional clustering of 1200 results to identify the most frequently recurring positions (see Methods). Incorporation of additional results did not significantly alter the distribution of results (data not shown). We chose STE12 because it is one of the best studied TFs, with a well known, highly conserved and experimentally well-defined binding motif (10, 43) . The most frequently recurring positions were compared with the known STE12binding motif (40) . We classified predictions into two categories: true positive (TP) if they resemble the experimentally identified STE12-binding motif, and false positive (FP) otherwise (see Methods). Finally, we calculated the positive predictive value PPV as PPV ¼ TP/(TP + FP). We find that positional clustering and subsequent selection of frequently recurring nucleotides improved the PPV of the STE12 binding site by at least 37% over Gibbs sampling alone (Figure 2 ). To validate that the above results were not specific to the number of input promoters, the STE12-binding motif, or the particular Gibbs sampling implementation, we repeated the above prediction process for promoters predicted to bind to YAP1, TEC1, HAP4 and YDR026C. We also repeated the analysis replacing the original Gibbs sampling procedure with our own implementation and MotifSampler (44) . In all cases, we found positional clustering significantly improves on results over local optimization procedures alone ( Figure 3 ). Computational discovery of TFBS can have two types of FP predictions. One type is the identification of an incorrect motif from a set of upstream regions known to bind to a TF of interest as described above (see Methods). The second type of FP error is the background discovery rate of the correct motif using upstream regions that do not bind to the TF. To simulate this rate for STE12-like binding site recovery we repeated the analysis as described above starting with 51 randomly chosen yeast promoters. We find that positional clustering identifies STE12-like sites in <5% of results, compared with 10-15% for Gibbs sampling alone. Thus, using positional clustering, the performance of computational motif discovery is enhanced not only by improving the positive predictive value in promoters of genes co-regulated by STE12, but also by decreasing the false discovery of STE12-like sites by 10%. Next, we evaluated the effect of adding DSs on the performance of Gibbs sampling with and without positional clustering. Addition of DSs dilutes enrichment of the TF-binding site in the input set, making motif detection more challenging (17, 22) . Modeling DSs, we repeated our estimate of PPV of TFBS detection with the addition of 1-50 random yeast promoters (DSs) to the original set of 51 STE12-bound promoters. We found that positional clustering improves the PPV of Gibbs sampling by >20% through the addition of up to 80% noise or 40 DSs (Figure 2, Supplementary Figure S3 ). Additionally, results of Gibbs sampling both with and without positional clustering decay linearly with the addition of decoys [R 2 ¼ 0.81 and 0.95, respectively (Supplementary Figure S4) ]. Extrapolating, we predict positional clustering will maintain an improved PPV through the addition of >100% noise or 70 DSs. To address issues of generality, we repeated the procedure on additional sets of S.cerevisiae promoters (YAP1, TEC1, HAP4 and YDR026C). An added benefit is that we can evaluate the effect of information content of the binding motif and number of promoters on the improvement from positional clustering (22) . Repeating the analysis, we again find that independently of the set or sampling procedure, positional clustering improves accuracy through a broad range of random DSs (Figure 3 ). Improvement appears to be limited and unreliable only when sampling alone correctly identifies the binding site in fewer than 20% of results. This result is consistent with our analysis of STE12-bound promoters (Figure 2) , and may correspond to a lower limit for the efficacy of positional clustering. Recently, researchers have noted that complementary motif detection approaches can be used together to predict binding sites more effectively than either method alone (23) . With this in mind, we evaluated positional clustering in terms of its ability to combine results from two different sampling implementations. For each dataset, an equal number of results from each approach were combined into a single dataset, and positional clustering was used to predict binding sites as described above ( Figure 3C ). We measured the average percent change in PPV for each TF on each dataset, and found positional clustering improved combined sampling by 94% compared with 25% and 27% improvement for BioProspector and MotifSampler, respectively. Additionally, clustering combined sampling improved 19 of the 22 datasets evaluated, whereas clustering of BioProspector and MotifSampler results improved 17 and 16 datasets, respectively. Thus, positional clustering is an effective mechanism to integrate results from multiple sampling procedures. Identification of GABR cis-regulatory sequences As described above in Introduction, identifying functional TFBS in mammals is difficult due in part to inclusion of decoy sequence from long upstream regions and lack of information on co-regulation of genes. Positional clustering, as shown above, is more robust to noisy input than Gibbs sampling alone, and thus may be better suited to identify de novo cis-regulatory elements in mammalian promoters that are coordinately regulated. To test this possibility, we chose seven mammalian GABR promoters (GABRps) whose activity is potentially altered in response to status epilepticus as identified through change in mRNA levels of the gene products (10) . For each set, the initial promoters were analyzed using Gibbs sampling with positional clustering (solid triangles) and without (open triangles). Two Gibbs sampling approaches were applied to each dataset: a Gibbs sampler procedure similar to BioProspector (8) (row A), and MotifSampler (39) (row B). Row C shows the combination of both sampling procedures, along with positional clustering of the combined results. x-axis counts over addition of DSs. We evaluated the positive predictive value of each technique on each dataset, and found positional clustering generally improved the PPV through addition of 100% random DSs. (31, 35) . We also included orthologous rodent promoters in the input sets (45) . Orthologous promoters were included to provide more instances of binding sites in the input set than would be expected by random, allowing for easier detection of the sites. Inclusion of orthologous promoters has the additional effect of selectively amplifying evolutionarily conserved binding sites. Such binding sites are more likely to have major functional roles in the regulation of the GABR receptor. Thus, sensitivity to such sequences is improved at the expense of sensitivity to species-specific binding sites. With this effect in mind, we require all GABR-binding site predictions to exist in orthologous promoters. Since the mechanisms of co-regulation for the seven GABRs are unknown, hypothetical co-regulation models were evaluated by querying all 2 7 possible subsets of the seven GABRps. Clustering results on nucleotide positions and selecting the most frequently occurring positions, we predicted 13 functional TF-binding sites. Predictions were compared with instances of known binding motifs from TRANSFAC (40) , and 8 of the 13 predictions (61.5%) resembled known binding sites for 10 TFs (Table 1) . Of the 10 TFs, 7 have been identified in the CNS of rodents: SP-1 (46); AP-2, TST-1 (POU3F1), OCT-1 (POU2F1), OLF-1 (47); CP-2 (48); and RREB-1 (49) . Furthermore, previous analyses of GABR promoter regions agree with our predictions that assign putative regulatory roles to SP-1, OCT-1, OLF-1 in the regulation of GABRs (29) . We chose to validate novel motif predictions with EMSAs and functional studies in primary cultured neurons. EMSA (50) was performed with an excess of cold competitors to define specificity of protein binding in nuclear extracts derived from primary neocortical neurons and fibroblasts (FIBs) kept in culture. As shown in Figures 4-6 , out of six predicted binding sites found upstream of the (a, b, g and d) subunit genes, four (GABRA4, GABRB1, GABRB3 and GABRD) displayed specific binding. In addition to specific binding of neuronal extracts to novel GABRA4 motifs, we have evidence for specific binding using FIB extracts ( Figure 5A and B) , of especial interest given that the expression of GABRs is restricted to the nervous system and repressors such as the RE1-silencing transcription factor (REST) (51, 52) expressed in non-neuronal cells have been implicated in the neural specificity of gene expression. Clearly, protein binding to DNA does not always necessitate regulatory function. To begin to address the functional Table 1 . Positional clustering based predictions of transcriptional regulatory sequences upstream of GABRs In total, we predict 15 orthologous pairs of regulatory sequences, representing 13 unique sequences. Comparing with known mammalian binding motifs, eight of the predictions show similarity to previously characterized TFBS, as indicated. Where no known binding motif exists, the corresponding in vitro EMSA and functional assay, if applicable, is indicated. Similar predictions are grouped together and aligned by hand. significance of our predicted regulatory motifs, we evaluated the effects of transfecting neurons with double-stranded oligonucleotides containing one of the GABRA4 novel binding motifs (dsA4O), as described above. GABRA4 is especially interesting given that it is regulated by brain derived neurotrophic factor (BDNF) after status epilepticus (31, 53) . Transfection with the dsA4O produced a significant downregulation of GABRA4 gene expression in neocortical neurons as monitored by quantitative real-time RT-PCR with no change after MOCK transfection or transfection with a dsO containing three copies of a cAMP regulatory element (CRE) (Figure 6 ). How reliable are the binding site predictions returned by Gibbs sampling based TFBS identification algorithms? We began by evaluating the stability of binding site predictions via repeated runs of Gibbs sampling. To quantify the robustness of predictions, we counted the number of Gibbs sampling results at each nucleotide position in the input set ( Figure 1 ) over a large number of repeated trials. We find that the most frequently returned positions better predict TF binding sites than the maximally scoring motifs from Gibbs sampling (Figures 2 and 3 ). Since scoring functions are empirically derived and do not necessarily represent biological reality, the result is not altogether unexpected (17) . However, we find that selecting frequently recurring positions allows filtering of up to 90% of spurious sampling results caused by convergence on biologically uninformative local minima. Positional clustering allows unbiased aggregation of results from different motif widths, thus approximating the width of the binding site 'for free' (54) . Next we show that positional clustering improves robustness to the addition of DSs (Figures 2 and 3) . Such sequences arise from inclusion of promoter regions in input sets without direct binding to the TF either due to experimental error or computational mis-annotation (17, 22) . In the STE12 example studied, linear regression models indicate our approach will maintain an advantage over traditional Gibbs sampling through addition of up to 150% noise to the original signal (Supplementary Figure S4) . Empirical data, however, show a sharp decrease in improvement close to the addition of 45 DSs, or roughly double the input set ( Figure 2 ). Moreover, evaluations using promoters co-regulated by other TFs Figure 6 . Double-stranded oligonucleotide functional assay for GABRA4 regulation. Primary cultures of rat neocortical neurons were treated with DOTAP (N-[1-(2,3-dioleoyloxy)propyl]-N,N,N-trimethylammonium methylsulfate) alone (Mock) or with DOTAP and phosphothioate oligonucleotides from either a cAMP response element (CRE Decoy) or a sequence from the GABA-A4 promoter predicted using positional clustering (GABA-A4 Decoy) (GTGCACACACACGCCCACCGCGGCTCGGG). mRNA was harvested after 24 h, and real-time RT-PCR was performed with GABA-A4 specific primers. Error bars refer to individual experiments; i.e. different platings of cells from different animals. Data was normalized to rRNA levels, and expressed as relative mRNA levels (GABA-A4/rRNA). Results are shown as mean ± SEM, N ¼ 3, asterisk indicates significantly different from control at the 95% confidence interval. Figure 5 . EMSA of three putative TF binding sites form DNA-protein complexes in neocortical and fibroblast nuclear extracts. Neocortical (NEO) and fibroblast (FIB) nuclear extracts from E18 rat embryos were incubated with three 32 P-radiolabeled probes from human A4 and D receptor subunits. Cold wild-type oligonucleotides were used to define specificity through competition. Cold oligonucleotides were added at 100-fold excess over probe. indicate positional clustering is less likely to improve predictions when Gibbs sampling identifies a correct site in <20% of repetitions ( Figure 3 ). Thus, it is possible the rather simplistic linear model overestimates improvement in robustness beyond what is practically achievable. Moreover, when multiple motifs exist in the input promoters, preliminary evidence suggests positional clustering will uniquely identify a single dominant motif (Supplementary Figure S5) . With further refinement, however, it may be possible to recover subordinate motifs, enabling identification of cis-regulatory modules. In spite of these limitations, using positional clustering of repeated runs, researchers can successfully apply sampling algorithms in identification of functional binding sites in datasets with a significant proportion of noise. Computational prediction of TF binding in mammalian genomes poses just such a challenge due to increased decoy sequence in large upstream regions (24) . Thus, having established increased robustness to DSs in yeast, we applied our approach to identify potentially unknown GABA A receptor subunit gene regulatory sequences that may participate in the response of the genome to seizure activity. We reasoned that GABA A receptor subunit genes either up-regulated or down-regulated in the animal model of epilepsy would share common binding motifs. Using positional clustering, we predicted 13 TF-binding sites upstream of GABA A receptor subunit genes ( Table 1) . Twelve of our predictions were verified by either comparison to known binding sites or experimental verification using in vitro binding assays. Initially positive experimental results highlight the ability of computational techniques to direct research into transcriptional regulation in mammalian models. As such, our approach may be applicable in the study of other protein complexes in the mammalian proteome. The reported predictions may enable pharmacologically important downstream research. For example the predicted sites can be used as a starting point for quantifying in vivo effect on downstream transcription; for identifying the TFs bound; and even for the more complex task of understanding the role of each site in determining the relative abundance of GABA A receptor isoforms. Research along these lines may dramatically improve our understanding of GABA A receptor regulation and its role in disease and development. Additionally, a more comprehensive evaluation of the remaining GABA A receptor subunit genes may reveal additional TFbinding sites that uncover the evolutionary significance of g-a-b GABR clusters in the human genome. Supplementary Data are available at NAR Online. Charles DeLisi is partially supported by NIH grants A08 POGM66401A and J50 01-130021. Daniel S. Roberts is supported by NIH training grant 2T32 GM00854. Shelley J Russek is supported by NIH/NINDS Grant NS050393. Funding to pay the Open Access publication charges for this article was provided by the Boston University Bioinformatics Program.
80
The Transmissibility of Highly Pathogenic Avian Influenza in Commercial Poultry in Industrialised Countries
BACKGROUND: With the increased occurrence of outbreaks of H5N1 worldwide there is concern that the virus could enter commercial poultry farms with severe economic consequences. METHODOLOGY/PRINCIPAL FINDINGS: We analyse data from four recent outbreaks of highly pathogenic avian influenza (HPAI) in commercial poultry to estimate the farm-to-farm reproductive number for HPAI. The reproductive number is a key measure of the transmissibility of HPAI at the farm level because it can be used to evaluate the effectiveness of the control measures. In these outbreaks the mean farm-to-farm reproductive number prior to controls ranged from 1.1 to 2.4, with the maximum farm-based reproductive number in the range 2.2 to 3.2. Enhanced bio-security, movement restrictions and prompt isolation of the infected farms in all four outbreaks substantially reduced the reproductive number, but it remained close to the threshold value 1 necessary to ensure the disease will be eradicated. CONCLUSIONS/SIGNIFICANCE: Our results show that depending on the particular situation in which an outbreak of avian influenza occurs, current controls might not be enough to eradicate the disease, and therefore a close monitoring of the outbreak is required. The method we used for estimating the reproductive number is straightforward to implement and can be used in real-time. It therefore can be a useful tool to inform policy decisions.
A new highly pathogenic strain of avian influenza, H5N1, emerged in the poultry markets of Hong Kong in 1997 and subsequently re-emerged in Vietnam in 2003. From this time onwards it has rapidly spread across the globe and is likely to be endemic in poultry in many parts of the world. Although onward transmission to humans at present remains limited, the high case fatality rate in those people that are infected has raised concerns about the impact of a potential human pandemic [1, 2] . Whilst much research and planning is currently underway to contain any outbreak in humans, relatively little is known about the extent of infection in poultry and, in particular, the transmissibility of highly pathogenic avian influenzas between poultry farms. Such understanding is vital if we are to limit the potential for a human pandemic by reducing the extent of infection in poultry, either through movement restrictions, culling or vaccination. Avian influenza occurs naturally in wild water fowl, usually in a low-pathogenic version (LPAI) causing no symptoms or only mild disease. However, in poultry some strains also occur in a highlypathogenic form (HPAI) and result in a devastating disease which can kill up to 100% of infected birds within 48 hours, and is highly transmissible between individual birds [3, 4] . Transmission between flocks kept at different farms is thought to occur via movement of infected birds, equipment or staff, with current evidence suggesting that air-borne transmission over long distances is rare [5] . There has been an increase in HPAI outbreaks over the past ten years [4] . In addition to their implications for human health, these outbreaks also have severe economic consequences for the affected countries. Typical control measures for HPAI in poultry comprise of swift isolation and culling of flocks on infected farms, the restriction of movements between farms, increased bio-security, and the culling of flocks in the vicinity of infected farms to deplete the susceptible poultry population. Vaccination, if coupled with a strict surveillance programme, has also been demonstrated to be effective in reducing the risk of further outbreaks [5, 6] . The reproductive number for infected poultry farms, defined as the average number of farms that each original infected farm infects at the start of an outbreak (i.e., when most farms are susceptible), is an important measure of the overall transmissibility of the virus in a population. It determines whether a self-sustaining epidemic will occur and, more importantly, yields a tool to assess the effectiveness of control measures. If, on average, at any point in time, each infected farm infects more than one further farm, the epidemic will continue. However, if on average, each infected farm infects less than one further farm, the epidemic will decline and the intervention measures applied at that point can be interpreted as being sufficient to control the outbreak. In this paper, we analyse published data from four outbreaks of HPAI in commercial poultry in industrialised countries to estimate the farm-to-farm reproductive number of HPAI to explore the extent to which different intervention measures implemented during these outbreaks reduce the reproductive number. The results from our analyses can be used to inform current planning for an outbreak of HPAI in similar commercial poultry sectors. We analyse data from three different outbreaks of HPAI that occurred in the past 8 years in industrialised countries: an outbreak of H7N1 in Italy in 1999/2000, an outbreak of H7N7 in the Netherlands in 2003, that will be treated as two distinct outbreaks due to geographic separation, and an outbreak of H7N3 in Canada in 2004. Figure 1 shows the time course of these outbreaks. Brief details of these outbreaks are given below. 2.1.1. Outbreak of H7N1 in Italy in 1999/2000 Northern Italy has experienced a number of avian influenza outbreaks from 1997 onwards [6] [7] [8] [9] [10] . These all occurred in an extremely dense poultry production area (up to 70 000 birds/km 2 ) and involved a significant number of farms keeping turkeys, a species known from experimental studies to be highly susceptible to avian influenza [11] . Furthermore, in this region there are many wetlands and resting sites for migratory waterfowl in close proximity to the poultry industry, which likely lead to multiple introductions from the wild bird host. In March 1999, H7N1 LPAI was detected in a farm keeping turkeys [6, 7, 9] . This outbreak was not controlled rigorously and so AI continued to circulate. In December, a case of H7N1 HPAI was found and strict control measures were implemented, including culling of affected flocks, movement restrictions and pre-emptive slaughter of flocks deemed at high risk. However, due to LPAI circulating at the time, the confirmation of HPAI was delayed, and so the disease had already infected a number of farms by the time control measures were enforced. This resulted in an HPAI epidemic affecting a total of 413 flocks. The LPAI/HPAI epidemic lasted until April 2000, and involved a total of over 13 million birds. The H7N7 epidemic in the Netherlands in 2003 affected a total of 255 commercial flocks in two distinct geographical and temporal clusters. The outbreak was situated in the Gelderse Vallei, the densest poultry production area in the Netherlands, in which over 10 million birds are kept in 984 flocks with a density of 4 flocks/ km 2 [12, 13] . Two months into the outbreak the infection passed to Limburg, another very dense poultry production area, where it continued to spread. In the Gelderse Vallei, HPAI was confirmed on 28 th February, 6 days after clinical signs appeared in the first infected farm, and between March and early April, a total of 212 farms were infected. In Limburg, a further 43 farms were infected between April and early May. A number of control policies were enforced in several stages. From 1 st March all movement of poultry and poultry products was banned, the tracing of dangerous contacts was initiated and reinforcement of strict bio-security measures was implemented. Two days later, from 3 rd March, culling of infected farms was initiated. On 5 th March the additional pre-emptive culling of farms within a 1 km radius of any infected farms was put in place. This was further extended to a 10 km radius for turkey flocks and 3 km radius for all other flocks on 7 th April [14] . However, these control measures were insufficient and it is hypothesized that the epidemics in both areas finally came to a halt due to depletion of susceptible flocks, after the culling of 30 million birds in 1,255 commercial and 17,421 hobby flocks [15] . During this epidemic, 89 human infections were also reported, most of whom presented with conjunctivitis or mild influenza-like illness. One person died from their infection. There was also evidence of limited human-to-human transmission [16, 17] . Following detection of the index case, a broiler breeder farm, a surveillance program was initiated, which led to the detection of the second case on 11 th March. The Fraser Valley south of the River Fraser was declared a Control Area, restricting movements of birds, bird products and equipment. Furthermore, active surveillance was undertaken in a High Risk Region (HRR, 5 km around the index case) and in flocks deemed dangerous contacts in a Surveillance Region (SR, 10km around index case). After the identification of 7 infected farms, all birds within the HRR were slaughtered from 24 th March onwards, but as this failed to stop transmission, on 5 th April it was decided to depopulate the whole Control Area, containing approximately 19 million birds. Infected farms were located mainly in three distinct local clusters within the Control Area. It is hypothesized that long distance spread between these clusters was due to bird, equipment or people movement, whereas once a farm in a densely populated area became infected, where sheds are sometimes within a few hundred metres from each other, the virus spread via dust or feather debris. Assuming homogeneous mixing, that all farms are equally infectious, and that the time-dependence of infectiousness from the point of infection is identical, we can estimate both the distribution of generation time intervals and the reproductive numbers of individual farms from the time-course of an epidemic using the following method [19] . Suppose there are N infected farms, labelled i = 1,…,N, and ordered so that the first k farms are those that contracted their infection from outside sources. The infection times of these farms are t = (t 1 ,…,t N ) such that t 1 = … = t k = 0. Under the simplest model that neglects any differences between farms, spatial locations, etc., the probability that farm j[fkz1,:::,Ng was infected by farm i[f1,:::Ng is where the generation time distribution has density w(T;h), which is defined to be 0 if T,0 and indexed by unknown parameter vector h. Under the above assumptions, the number of farms infected by farm i (i.e. the reproductive number of farm i) in the outbreak can be represented as an outcome from a random variable that is, a sum of Bernoulli random variables, which has expected value Now denote the 'infection tree' by v = (v k+1 ,…,v N ), defined such that v j = i if farm j was infected by farm i. Under (1), the likelihood for h when v and t are observed is But as v is unobserved we sum over all possible infection trees to obtain the 'integrated likelihood' where S j = {1,…,N}\{j} is the set of all indices other than j. The integrated likelihood is a genuine likelihood (up to a multiplicative constant) permitting valid inferences about h conditional on outbreak size N. time The maximum likelihood (ML) estimateĥ is obtained by minimizing twice the negative log-likelihood More details on how ML estimation is performed are given in Appendix S1. We assume the generation times T = t j 2t i are Weibull distributed, with density and so h = (k, g). Minimization was performed using the Downhill Simplex method [20] , the code used for these calculations is given in Code S1 and Code S2; to ensure the global minimum is reached, the procedure was run from 10000 different starting points. We further investigated whether the generation time distribution changed after control measures were introduced. To do this, we extended the above model to allow for distinct parameters for the generation times before and after controls, h pre and h post , see Code S3 and Code S4. The improvement in fit compared to the original model was assessed using a likelihood ratio test. 2.2.2. Estimation of the reproductive number Givenĥ we can estimate the mean and variance of the generation time distribution. Moreover, we can estimate the reproductive number for each infected farm via equation (3), and the mean reproduction number for any subset of infected farms. To calculate confidence intervals for the reproductive number we use an approximation of the parametric bootstrap percentile interval method [21] . To obtain proper parametric bootstrap intervals would involve generating infection times and trees according to the underlying epidemic model, which we do not wish to specify completely. Instead, the following two-step approximation is used, which we propose will be a good approximation for large N . These two steps approximate generating realisations from the underlying epidemic process. First, we take bootstrap samples of parameter values from the conventional approximation to the sampling distribution of the ML estimator, that is, from a bivariate normal distribution with mean (k k,ĝ g) and variance-covariance matrix V based on the inverse of the observed information matrix (see Appendix S1 for further details, the code used to generate the bootstrap sample is given in Code S5 and Code S6). This first stage can be loosely thought of as sampling the mean behaviour for a subgroup of possible outbreaks. To allow for variability within each subgroup, stage two involves fixing (k * ,g * )and independently generating reproductive numbers for each farm according to model (2) . Steps one and two together give R * = {R * i :i21,…,N}, an approximate bootstrap sample of the reproductive numbers for each farm. Here, 1000 samples of (k,g)pairs were drawn, and 500 sets of reproductive numbers generated for each. Finally, the approximate 95% CI for each R i is given by the 2.5 th and 97.5 th percentiles of the bootstrap distribution. The second stage of the calculation of the approximate CIs was done using Code S7 and Code S8. The generation time is defined as the time between the infection of a farm and the time at which the farm passes on infection to another farm. We have assumed the generation time distribution is Weibull. While this is a biologically plausible choice, we cannot verify it empirically. As such, we assessed robustness to this choice using other plausible choices such as the gamma distribution (results not shown). However, the following results under these alternatives did not differ substantively from those shown below. Figure 2 shows the estimated generation time distribution for the four different outbreaks; the parameter estimates are detailed in Table 1 . The estimates, and hence the distribution, differs substantially between the outbreaks. It could be hypothesised that the generation time would shorten after measures were put in place to isolate the infected farms. However, allowing for different generation time distributions for the pre-and post-control time periods did not significantly improve the model fit. Table 2 ) with upper 95% bounds in the range 1.5-3.6. The impact of control measures on the effective reproductive number can be clearly seen in all four outbreaks. For the outbreak in Italy (Figure 3a) , their introduction rapidly reduced the reproductive number, hovering around the threshold of 1 for the next few months before finally dying out. In British Columbia (Figure 3b ) controls were put in place after detection of the first IP. However our estimates of the reproductive number remain high until 24 th March when the decision was taken to cull the whole high risk region. Our estimates show that the control activities following this decision were effective in reducing the reproductive number to below one. Our results show that the situation in the Gelderse Vallei, The Netherlands (Figure 3c ) differed in that the initial control measures failed to bring the reproductive number reliably below 1, and the epidemic only died out at the end of March after the depletion of susceptible flocks in the affected area [15] . The same controls were applied to the Limburg epidemic but our estimates show in this case the reproductive number was reduced to just below 1 (Figure 3d) , and so potentially effective in controlling the outbreak. However, the end of the epidemic in late April coincided here too with the depletion of susceptible flocks and therefore it is possible that the epidemic would have taken substantially longer to control had there been a larger pool of susceptible flocks in the area. Our estimates of the farm-to-farm reproductive number prior to interventions for HPAI are in the range 1.1 to 2.4 and were remarkably consistent across the four datasets. However, these estimates are substantially lower than those previously reported for the Dutch epidemic. Prior to the implementation of control measures we obtained estimates of 1.1 (95% CI 0.9-1.5) in the Gelderse Vallei and 1.9, (95% CI 1.0-3.0) in Limburg which are significantly lower than those previously reported for the same outbreak prior to notification (6.5 (95% CI 3.1-9.9) for the epidemic in the Gelderse Vallei). However, as demonstrated in Figure 3c , there was substantial variation in our estimates of individual reproductive numbers prior to interventions. In addition, in the previous study, the generation time was not estimated directly from the data but based on observational and experimental data on the course of infection in the farms. Our estimate of the generation time for this region is of the order of 2 days, whereas the values previously assumed for the infectious period were defined per flock as the time between detection and culling plus an additional 4 days to cover the time before the infection was detected but during which birds were infectious. The previously published estimates therefore assumed a much longer mean generation time and this could also lead to a higher estimated reproductive number. Our results showed substantial differences between the estimated generation time distributions for the different outbreaks. Whilst much is known from experimental studies on the course of infection in individual birds [11, [22] [23] [24] , estimates of the generation time at the farm-level are more difficult to obtain. Although it is perhaps surprising that the generation time differs between the outbreaks, it is plausible that such differences could arise because of variation in farming practices or in the contact patterns between farms. In addition, the latent and infectious periods determining the generation time may differ by the strain of HPAI. Alternatively, the estimates may be biased because of assumptions made in the method. In particular, we assumed that the datasets were complete (and thus that all infected farms were detected) and that only the first farm in each outbreak was infected from an outside source. If, however, further undetected farms had played a role in transmission, this would substantially alter the estimates of the generation time and the reproductive number, particularly if these infections occurred towards the beginning or end of the epidemics where overall cases are sparser. All of the outbreaks investigated here occurred within dense poultry farming areas and hence were difficult to control. The control policies implemented in the different outbreaks were similar, comprising strict bio-security measures for movement of poultry and poultry products, swift culling of infected flocks, and if these failed to control the epidemics, additional pre-emptive culling of flocks in the neighbourhood of any infected farms. Our results demonstrate that the bio-security measures, movement restrictions and culling of infected farms, all of which were initiated early on in the outbreaks, did have an effect but for all four outbreaks only reduced the reproductive number to close to the threshold value of 1. The additional pre-emptive culling of flocks and de-population of the areas was needed to fully control the outbreaks. Current contingency plans for HPAI outbreaks in Europe focus on the former set of control measures to contain any outbreak [25] . Whilst differences in farming practices between countries mean that it is difficult to predict whether these measures will be sufficient for a new outbreak, our analyses suggest that additional interventions may well be required. Close monitoring of outbreaks, coupled with quantitative estimation of the reproductive number, is therefore needed to ensure that such additional measures, if required, are promptly implemented. The method used here to estimate the reproductive number and generation time parameters is an extension of that developed by Wallinga and Teunis [19] for the SARS-epidemic. This method requires only time-series data for an outbreak, and is therefore easily applied even in real-time. Technically, appropriate censoring terms should be added to the likelihood to account for infection times yet to occur, but a straightforward application of the method as described here will give estimates unbiased in an asymptotic sense. If data on the spatial location of infected farms are also available, this information can easily be incorporated to estimate the spatial transmission kernel and improve the estimation of the reproductive numbers. Such an approach was successfully applied to the Foot-and-Mouth epidemic in the UK in 2001 [26] . Further work is required to explicitly incorporate missing data, as this is likely to have a strong influence on the estimates of both the generation time and the reproductive number. Such methods are of particular importance to estimating the reproductive number for outbreaks of HPAI in Asia in which, with high general levels of poultry mortality, cases are likely to be less well documented. Appendix S1 Code S1 Source code for the maximum likelihood estimation of the parameters of the generation time distribution, shown in Figure 2 and Table 1 Code S7 Source code for estimating R0 and the confidence intervals based on the uncertainties inherent in the estimation procedure and the uncertainty of the generation time distribution, shown in Figure 3 .
81
The influenza pandemic preparedness planning tool InfluSim
BACKGROUND: Planning public health responses against pandemic influenza relies on predictive models by which the impact of different intervention strategies can be evaluated. Research has to date rather focused on producing predictions for certain localities or under specific conditions, than on designing a publicly available planning tool which can be applied by public health administrations. Here, we provide such a tool which is reproducible by an explicitly formulated structure and designed to operate with an optimal combination of the competing requirements of precision, realism and generality. RESULTS: InfluSim is a deterministic compartment model based on a system of over 1,000 differential equations which extend the classic SEIR model by clinical and demographic parameters relevant for pandemic preparedness planning. It allows for producing time courses and cumulative numbers of influenza cases, outpatient visits, applied antiviral treatment doses, hospitalizations, deaths and work days lost due to sickness, all of which may be associated with economic aspects. The software is programmed in Java, operates platform independent and can be executed on regular desktop computers. CONCLUSION: InfluSim is an online available software which efficiently assists public health planners in designing optimal interventions against pandemic influenza. It can reproduce the infection dynamics of pandemic influenza like complex computer simulations while offering at the same time reproducibility, higher computational performance and better operability.
Preparedness against pandemic influenza has become a high priority public health issue and many countries that have pandemic preparedness plans [1] . For the design of such plans, mathematical models and computer simulations play an essential role because they allow to predict and compare the effects of different intervention strategies [2] . The outstanding significance of the tools for purposes of intervention optimization is limited by the fact that they cannot maximize realism, generality and precision at the same time [3] . Public health planners, on the other hand, wish to have an optimal combination of these properties, because they need to formulate intervention strategies which can be generalized into recommendations, but are sufficiently realistic and precise to satisfy public health requirements. Published influenza models which came into application, are represented by two extremes: generalized but oversimplified models without dynamic structure which are publicly available (e.g. [4] ), and complex computer simulations which are specifically adjusted to real conditions and/or are not publicly available (e.g. [5, 6] ). The complexity of the latter simulations, however, is not necessary for a reliable description of infection dynamics in large populations [7] . A minimum requirement for a pandemic influenza planning tool is a dynamic modelling structure which allows investigation of time-dependent variables like incidence, height of the epidemic peak, antiviral availability etc. The tool should, on the other hand, be adjustable to local conditions to adequately support the pandemic preparedness plans of different countries which involve considerably different assumptions (Table 1) . Here we describe a publicly available influenza pandemic preparedness planning tool [8] which is designed to meet the requirements in preparedness planning. It is based on an explicitly formulated dynamic system which allows addressing time-dependent factors. It is sufficiently flexible to evaluate the impact of most candidate interventions and to consider local conditions like demographic and economic factors, contact patterns or constraints within the public health system. In subsequent papers we will also provide examples and applications of this model for various interventions, like antiviral treatment and social distancing measures. The model is based on a system of 1,081 differential equations which extend the classic SEIR model. Demographic parameters reflect the situation in Germany in 2005, but can be adjusted to other countries. Epidemiologic and clinic values were taken from the literature (see Tables 1, 2 , 3, 4, 5, 6 and the sources quoted there). Pre-set values can be varied by sliders and input fields to make different assumptions on the transmissibility and clinical severity of a new pandemic strain, to change the costs connected to medical treatment or work loss, or to simply apply the simulation to different demographic settings. Model properties can be summarized as follows. The mathematical formulation of this model is presented in detail in the online supporting material. The corresponding source code, programmed in Java, and further information can be downloaded from [8] . According to the German National Pandemic Preparedness Plan [9] , the total population is divided in age classes, each of which is subdivided into individuals of low and high risk ( Table 2) . Transmission between these age classes is based on a contact matrix (Table 3) which is scaled such that the model with standard parameter values yields a given basic reproduction number R0. Values for the R0 associated with an influenza strain with pandemic potential are suggested to lie between 2 and 3 [10] . This value is higher than the effective reproduction number which has been estimated to be slightly lower than 2 [11, 12] . As a standard parameter, we use R0 = 2.5 which means that cases infect on average 2.5 individuals if everybody is susceptible and if no interventions are performed. Susceptible individuals who become infected, incubate the infection, then become fully contagious and finally develop protective immunity (Table 4) . A fraction of cases remains asymptomatic; others become moderately sick or clinically ill (i.e. they need medical help). Depending on the combination of age and risk group, a fraction of the clinically ill cases needs to be hospitalized, and an agedependent fraction of hospitalized cases may die from the disease ( Table 5 ). This partitioning of the cases into four categories allows combining the realistic description of the transmission dynamics with an easy calculation of the resources consumed during an outbreak. The degree and duration of contagiousness of a patient depend on the course of the disease; the latter furthermore depends on the age of the patient (Table 5) . Passing through the incubation and contagious period is modelled in several stages which allows for realistic distributions of the sojourn times ( Table 4 ). The last two stages of the incubation period are used as early infectious period during which the patient can already spread the disease. Infectiousness is highest after onset of symptoms and thereafter declines geometrically (Table 6 ). Clinically ill patients seek medical help on average one day after onset of symptoms. Very sick patients are advised to withdraw to their home until their disease is over, whereas extremely sick patients need to be hospitalized and may die from the disease (Table 4) . After the end of their contagious period, clinically ill patients go through a convalescent period before they can resume their ordinary life and go back to work (Table 4) . We provide some examples of model output of InfluSim [8] , version 2.0, by means of four sensitivity analyses; further investigations will be presented elsewhere. Figure 1 shows the graphical user interface of the software which is divided into input and output windows. The user may set new values in the input fields or move sliders to almost simultaneously obtain new results for the course of an epidemic in a given population. Figures 2A and 2B show pandemic waves which result from varying the basic reproduction number from 1.5 to 4.0. Using the standard parameter values as given in Tables 2, 3 , 4, 5, 6 and omitting all interventions in a town of 100,000 inhabitants results in a pandemic wave which lasts for about ten weeks (Figure 2A , with R 0 = 2.5). The peak of the pandemic wave is reached after six to seven weeks, with a daily incidence of up to 2,340 influenza patients seeking medical help, with up to 280 hospital beds occupied by influenza cases and with up to 14,000 out of 60,000 working adults unable to go to work because of illness or convalescence. These results depend on the assumptions concerning the yet unknown contagiousness and pathogenicity of the virus. Figures 2C and 2D show how the shape of the curves depends on the course of contagiousness: the pandemic wave proceeds relative slowly if the contagiousness does not change during the infectious period (x 50 = 50%), but proceeds quickly if the contagiousness is highest after onset of symptoms and decreases thereafter (x 50 > 50%). The influenza pandemic preparedness planning tool InfluSim stands between simple spreadsheet models and sophisticated stochastic computer simulations. It describes a pandemic wave within a homogeneously mixing population like a town or city, but surprisingly produces the same dynamics as individual-based simulations which explicitly consider geographic spread through the US (cf. [6] and [5] with Figure 2 using R 0 = 2). Similar observations were made with a simple deterministic compartmental model [7] . Stochastic models are known to behave quasi-deterministically when the simulated population becomes very large. A further reason for the congruence of complex stochastic and simple deterministic models must lie in the incredi-bly quick way in which pandemic influenza spreads geographically. Unless being controlled at the place of origin [12, 13] , a pandemic starting in a far-off country will lead to multiple introductions [14] into the large industrialized nations where it can be expected to quickly spread to neighbouring towns and to rural areas. The large populations which have to be considered susceptible to a pandemic virus and the quick geographic spread tend to diminish the differences between the results of sophisticated individual-based and simple deterministic models. However, a deterministic model like InfluSim cannot reliably represent effects originating from stochasticity, from effects in small populations, or from heterogeneities. Examples are: (i) a geographically limited spread and fairly effective control measures can imply that the epidemic affects only a small population and thus, may be strongly influenced by stochastic events [15] [16] [17] ; (ii) transmission which predominantly occurs in households or hospitals, or which is driven by other substantial features of the contact network is not in agreement with the assumption of homogeneous mixing in the deterministic model cannot reliably predict the spread of infection [18] [19] [20] [21] [22] [23] . In particular, (iii) super-spreading events can substantially change the course of an epidemic compared to the deterministic prediction [24] [25] [26] [27] . Apart from such factors, the predictability of intervention success is generally subject to uncertainties in the choice of parameter values, Assumed scenarios and outcomes of pandemic preparedness plans. * Gross attack rate (i.e. clinically ill and moderately ill cases). A population of N = 100,000 inhabitants of Germany is subdivided according to age a and risk category r. We assume that all age groups are fully susceptible at begin of the outbreak. A fraction of F a = 6% of all children (age < 20 years) are regarded as being under high risk (r = r 1 ) after an influenza infection whereby the remaining 94% are under low risk (r = r 2 ). The high risk fractions of working adults (ages 20-59) and elderly (ages 60+) are F a = 14% and F a = 47%, respectively. Source: [9] demanding additional efforts like Bayesian approaches [28] to evaluate the reliability of predictions [29] . Pandemic preparedness plans must consider constraints and capacities of locally operating public health systems. The time-dependent solutions of InfluSim allow assessing peak values of the relevant variables, such as outpatients, hospitalizations and deaths. Various interventions may be combined to find optimal ways to reduce the total number of cases, to lower the peak values or to delay the peak, hoping that at least part of the population may benefit from a newly developed vaccine. Special care was taken when implementing a variety of pharmaceutical and non-pharmaceutical interventions which will be discussed in subsequent papers. Despite its comprehensible structure, the model does not suffer from over-simplifications common to usual compartment models. Instead of implicitly using exponentially distributed sojourn times, we have implemented realistically distributed delays. For example, the model considers that individuals may transmit infection before onset of symptoms, and that some cases may remain asymptomatic, but still infecting others. Such features have serious implications for the success of targeted control measures. InfluSim is freely accessible, runs on a regular desktop computer and produces results within a second after changing parameter values. The user-friendly interface and the ease at which results can be generated make this program a useful public health planning tool. Although we have taken care of providing a bug-free program, including the source code, the user is encouraged to treat results with due caution, to test it, and to participate in bug-reports and discussions on the open-source platform [30] which also provides regular updates of InfluSim. The author(s) declare that they have no competing interests. ME developed the model, MS designed the software, HPD wrote the manuscript and SOB formulated the public The who-acquires-infection-from-whom matrix shows the frequency of contacts (per week per person) between different age classes. Source: [38] . Distribution of sojourn times (the last two stages of the latent period are used as early infectious period with an average duration of D L = 0.5 days). Sources: A [11] , B [39, 40] , C assumed, D [41] health requirements of the software. All authors read and approved the final manuscript. Susceptible individuals S a, r are infected at a rate λ a (t) which depends on their age a and on time t. Infected individuals, E a, r , incubate the infection for a mean duration D E . To obtain a realistic distribution of this duration, the incubation period is modelled in n stages so that progression from one stage to the next one occurs at rate δ = n/D E . The last l incubation stages are regarded as early infectious period during which patients may already spread the infection (this accounts for an average time of lD E /n for the "early infectious period" which is about half a day for the standard set of parameters). After passing through the last incubation stage, infected individuals become fully contagious and a fraction of them develops clinical symptoms. The course of disease depends on the age a of the infected individual and on the risk category r to which he or she belongs: a fraction c a, r (A) becomes asymptomatic (A a ), a fraction c a, r (M) becomes moderately sick (M a ), a fraction c a, r (V) becomes very sick (V a ) and the remaining fraction c a, r (X) becomes extremely sick (X a ) and need hospitalization (i.e., c a, r (A) + c a, r (M) + c a, r (V) + c a, r (X) = 1 for each combination of a and r). ) . A fraction f V (t) of all severe and a fraction f X (t) of all extremely severe cases who visit the doctor within D T days after onset of symptoms are offered antiviral treatment, given that its supply has not yet been exhausted. As our model does not explicitly consider the age of the disease (which would demand partial differential equations), we use the contagious stages to measure time since onset and allow for treatment up to stage m a, T Sources: Contagiousness of asymptomatic cases: [11] ; degree of contagiousness during the early infectious period and equality of the contagiousness of moderately and severely sick cases: assumed. Independent of age a and risk group r, a fraction c a, r (A) = 33% of infections result in asymptomatic cases, a fraction c a, r (M) = 33.5% become moderately sick and the remaining fraction develops severe disease. An age-and risk-dependent fraction h a, r of untreated patients with severe disease needs hospitalization. An age-dependent fraction d a of hospitalized cases dies. Sources: fraction of asymptomatic cases: [11] ; 50% of symptomatic cases see a doctor: [9] ; hospitalizations per severe case: [9] ; case fatality of hospitalized, but untreated patients calculated from [4] . (see below for details). This imposes some variability to the maximum time until which treatment can be given, which may even improve the realism of the model with respect to real-life scenarios. Antiviral treatment reduces the patients' contagiousness by f I percent and it reduces hospitalization and death by f H percent. Extremely sick patients, whose hospitalization is prevented by treatment, are sent home and join the group of treated very sick patients(W a, T ). The remaining duration of disease and contagiousness of treated cases is reduced by f D percent so that their rate of progressing from one stage to the next has To obtain a realistic distribution of this sojourn time, convalescence is modelled in j stages so that progression from one stage to the next occurs at rate ρ = j/D C . Fully recovered patients who have passed through their last stage of convalescence join the group of healthy immunes I; working adults will go back to work. Further interventions, describing the reduction of contacts, will be discussed after the presentation of the differential equations. InfluSim user interface Figure 1 InfluSim user interface. x 50 = 95% means that 95% of the cumulative contagiousness is concentrated during the first half of the contagious period, see Table 6 ). D: Cumulative number of deaths for values of x 50 as in C. All other parameters as listed in Tables 2-6 . Hospitalized, but untreated cases Contact matrix For the mixing of the age classes, we employ a whoacquires-infection-from whom matrix which gives the relative frequency of contacts of infective individuals of age a i with other people of age a s . In this paper, we assume bi-directional contacts (e.g. children have the same total number of contacts with adults as adults with children). Multiplication of this matrix with an appropriate constant scaling factor κ (see below) results in the matrix of crude contact rates . In the absence of interventions, we have to multiply these contact rates with the contagiousness factors b L , b A , b M and b V to obtain the effective contact rates: during the early infectious period, of asymptomatic cases, of moderately sick cases, of (untreated) very sick cases. To assess the effect of day care centre and school closing on the transmission of an infectious disease, we have to first make an assumption on what fraction r sch of the contacts among healthy children who are in the same age class occurs in day care centres and schools. The contact rates between very sick or hospitalized children (who do not attend day care centre or school) and other children need, therefore, be reduced to (contact rate between healthy and very sick children in the same age class, i.e. a i = a s ). As very sick children have to be taken care of by adults at home or in hospital, their contact rate to adults increases by a factor F HC (contact rate between very sick children of age a i and adults of age a s ). Contacts between very sick children and other children in a higher or lower age class remain unchanged: (contact rate between healthy children of age a s and very sick children of a different age a i ). Closing day care centres and schools at time t will not necessarily prevent all the contacts that would have happened with other children. During the closing of schools and day care centres, the contact rates between susceptible children of age a s and infected children of age a i who are in their late incubation period ( ), who are asymptomatic ( ), or who are moderately sick ( ) are reduced by the factor r sch if the children are in the same age class: where 1 sch (t) is a function which indicates when schools and day care centres are opened or closed: ,..., While day care centres and schools are closed, children (age a i ) need adult supervision at home. Their contact with susceptible adults (age a s ) increases by the "child care factor" F CC : Child care at home also increases the exposure of healthy children (age a s ) to contagious adults (age a i ): Cancelling mass gathering events effects only the contacts of adults who are healthy enough to attend such events. Assuming that such an intervention at time t reduces contacts by a fraction r mass , we get for all contacts between susceptible adults of age a s and infectious adults of age a i the following contact rates: where 1 mass (t) is a function which indicates when mass gathering events are possible or when they are closed: As contacts with adults who are too sick to attend such mass gathering events cannot be prevented by this measure it is . During some time in the epidemic, the general population may effectively reduce contacts which can be a result of wearing facial masks, increasing "social distance", adopting improved measures of "respiratory hygiene" or simply of a general change in behaviour. This will be implemented in the program by reducing the contacts of susceptible individuals at that time t by factor r gen (t while mass gathering events are forbidden while m mass gathering events are allowed. while the population reduces their contacts while the population behaves as usual. The contact rates of cases in the late incubation period and that of asymptomatic cases remain unchanged: for infected individuals in the late incubation period, for asymptomatic cases. To allow for a contagiousness which changes over the course of disease, we multiply each contact rate with a weighting factor whereby k is the stage of contagiousness. This leads to the following contact rates: for asymptomatic cases in For x = 1, contagiousness is equally high in all stages; for x = 0, only the first stage is contagious; for 0 <x < 1, the contagiousness decreases in a geometric procession. We make the simplifying assumption that contagiousness does not change during the late incubation period for cases in stage k = n -l,..,n of the incubation period. At time t = 0 and in the absence of interventions, the next generation matrix has the following elements where is the fraction of untreated extremely severe cases who die from the disease (see below for details). The dominant eigenvalue of this matrix is called the basic reproduction number R 0 . If κ (which determines the value of the contact rates ) is given, the eigenvectors of this matrix can numerically be calculated. The user-specified value of R 0 is now used to determine numerically the scaling factor κ. Let be the eigenvector which has the largest eigenvalue R 0 . ) ) − Using the user-specified numbers of people N a in the age classes and the fractions F a of people under high risk within each age class (Table 2) , we obtain the initial population sizes according to age and risk class: Using these initial values, the set of differential equations is solved numerically with a Runge-Kutta method with step-size control. if and in treatment window otherwise 0 ⎩ ⎩
82
Immune pathways and defence mechanisms in honey bees Apis mellifera
Social insects are able to mount both group-level and individual defences against pathogens. Here we focus on individual defences, by presenting a genome-wide analysis of immunity in a social insect, the honey bee Apis mellifera. We present honey bee models for each of four signalling pathways associated with immunity, identifying plausible orthologues for nearly all predicted pathway members. When compared to the sequenced Drosophila and Anopheles genomes, honey bees possess roughly one-third as many genes in 17 gene families implicated in insect immunity. We suggest that an implied reduction in immune flexibility in bees reflects either the strength of social barriers to disease, or a tendency for bees to be attacked by a limited set of highly coevolved pathogens.
While evident in social organisms ranging from humans to birds (Brown & Brown, 2004; Masuda et al ., 2004) , the impacts of sociality on disease are especially vivid within social insect colonies. Here, typically thousands of individuals interact in close quarters, at densities far exceeding those of even the most crowded vertebrate social groups (Wilson, 1971) . This density, coupled with a relatively homeostatic nest environment and the presence of stored resources, makes social insects attractive targets for disease agents (Schmid-Hempel, 1998) . As expected based on their parasite and pathogen pressures, social insects have evolved both individual and group strategies to combat disease. Grooming, nest hygiene and other behavioural traits found throughout the social insects can reduce the impacts of pathogenic bacteria, fungi and parasitic mites. For example, 'hygienic behaviour' first described for honey bees (Rothenbuhler, 1964) is now a classical example of a social defence, whereby workers identify and remove infected larvae from among the healthy brood (Spivak & Reuter, 2001) . Other defences enabled by sociality include the construction of nests from antimicrobial materials (Christe et al ., 2003) , the raising of offspring in sterile nurseries (Burgett, 1997) , social 'fever' in response to disease (Starks et al ., 2000) , transference of immune traits (Traniello et al ., 2002; Sadd et al ., 2005) , and heightened risk-taking by infected individuals (Schmid-Hempel, 2005) . Like most eukaryotes, colony members also possess individual defences, including immune responses toward disease agents (Casteels-Josson et al ., 1994; Evans, 2004) . The recent sequencing of the honey bee genome (Honey Bee Genome Sequencing Consortium, 2006) allows the first global analysis of immune components in honey bees, and the second opportunity (after humans) to use genomic insights to better understand disease resistance in a highly social organism. Insects have diverse mechanisms to combat infection by pathogens. Many insects are protected by a layer of antimicrobial secretions on their exterior, and by a gut environment that is hostile to pathogens. When pathogens move beyond these defences, the epithelium is often sufficient to stop further progress. Should pathogens defeat the morphological defences of insects, they are often met by efficient cellular and humoral immune defences. Insect immunity shows many parallels to the innate immune responses of humans and other vertebrates, involving a diverse set of actions including the secretion of antimicrobial peptides, phagocytosis, melanization and the enzymatic degradation of pathogens (Hoffmann, 2003; Hultmark, 2003) . Further, insect immune pathways share both an overall architecture and specific orthologous components with the innate immune system of vertebrates (Beutler, 2004) . This suggests both a shared root for these immune pathways and selection to conserve many components over hundreds of millions of years. In the first part of this paper, we propose honey bee models for four non-autonomous pathways implicated in inducible host defence, Toll, Imd, Janus kinase (JAK)/STAT and JNK (Boutros et al ., 2002) , based primarily on extensive searches for orthologues to well-studied fruit fly, mosquito and moth species. While these pathways engage in cross-talk and can direct some of the same immune effectors, they have well-defined structures and interaction sets, and are best tackled as individual entities. Most honey bee components for these pathways remain to be validated by functional tests, yet we feel that the presented models serve two important purposes. First, they point toward the most likely orthologues involved with all stages of the immune response, thus setting the stage for postgenomic functional work on honey bee immunity. Second, the models themselves show intriguing differences between species for these canonical immune pathways with respect to gene losses and duplications. Next, we show that many immune-gene families in bees appear to be reduced in number, when compared to Drosophila and Anopheles . While genome-wide analyses of the honey bee have identified many gene families with reduced diversity (Honey Bee Genome Sequencing Consortium, 2006) , such reductions appear to be especially pervasive in the immune system. These reductions hold for each stage of immunity, from recognition and signalling to immune effectors. We couple gene-family data with data on specific orthologues to test five hypotheses: (1) missing genes are not represented in the current draft honey bee genome assembly but are present in the genome; (2) immune-related genes in the bee have diverged especially quickly at the sequence level and, as such, have escaped annotation based on sequence similarity to other species; (3) honey bees enact immune responses using pathways and/or components not currently identified as immune players in other insects; (4) honey bees are targeted by a small set of coevolved pathogens and their immune systems are thereby tuned to these pathogens at the expense of being responsive to a wider range of threats; and (5) 'social' defences and barriers in honey bee colonies are effective in reducing pathogen pressure and, as such, bees are not as reliant as other insects on individual immune responses. Honey bees possess apparent orthologues for the core members of each of the four pathways implicated in immunity (Figs 1 and 2, Supplementary Material Table S1 ) and precise 1 : 1 : 1 orthology between honey bees and the flies Drosophila melanogaster and Anopheles gambiae is evident for most pathway members, especially for the intracellular components. Of the dozens of described actors in four signalling pathways predicted to play a role in insect immunity, only one protein appears to be completely absent in the bee genome: the ligand unpaired from the JAK/STAT pathway. The presence of the JAK/STAT cytokine receptor domeless and all other members of this pathway (Fig. 2) suggest that JAK/STAT remains functional in honey bees and is triggered by a currently unrecognized ligand. Insect Toll and the Toll-like receptors (TLRs) are transmembrane signal transducing proteins that play critical roles in both immunity and development. They are orthologous to mammalian TLRs, all of which have been implicated in immunity (Beutler, 2004) . In Drosophila , the Toll signalling pathway is enacted when the cytokine-like molecule Spaetzle binds to the extracellular domain of the transmembrane receptor Toll. The Drosophila genome encodes a family of six Spaetzle-related molecules, that are believed to function as ligands for the nine Drosophila Toll receptors (Parker et al ., 2001) . Two plausible Spaetzle orthologues are evident in the bee genome (GB15688 and GB13503; Fig. 1 ), and functional tests will be needed to determine which act as Toll-binding cytokines. Following conformational changes of the activated receptor, several intracellular death-domain (DD) containing proteins are recruited to form a receptor complex. Activation of this complex leads to the degradation of the NF kappa B inhibitor (I κ B) Cactus and subsequent nuclear translocation of the NF-κ B transcription factor Dorsal (or the Dorsal-related immune factor, Dif, in Drosophila (Royet et al ., 2005) . Two homologues of Dorsal were found in the honey bee genome (Fig. 1) , neither of which was orthologous with Dif. This lends support to the view that Dif is a highly derived branch found in brachyceran flies but absent from other insects. In mosquitoes (Shin et al ., 2005) , and arguably honey bees, Dorsal (called REL1 in mosquitoes) is a functional alternate for Dif. Functional tests can help determine which of the two dorsal paralogues is the key transcription factor for this pathway. The intracellular components Tollip, Pellino, Cactin and TNF receptor associated factor-2 (TRAF-2) are believed to aid the main players of this pathway, and all appear to be present in both fly species as well as the honey bee. Candidate effectors for the immune-related Toll pathways in honey bees include a compliment of antimicrobial peptides, the melanizing agent phenoloxidase and three lysozymes. While it has not been confirmed that these effectors are triggered by the Toll pathway as opposed to other pathways described below, it is evident that some of the bee effectors are responsive to pathogens and/or mechanical wounding of bees (Fig. 3) . Candidate honey bee members for the Toll pathway. Names are given for the Drosophila pathway components, along with vertebrate orthologues (in parentheses). Honey bee matches given as named during the genome project. Honey bee names in italics refer to genes with close paralogues which cannot readily be distinguished with respect to pathway components from Drosophila. Underlining indicates genes shown to be transcriptionally up-regulated after immune challenge. Candidate honey bee members for the Imd, JNK and JAK/STAT pathways, below names for Drosophila pathway components along with vertebrate orthologues (in parentheses). Honey bee matches presented as named during the genome project. Honey bee names in italics refer to genes with close paralogues which cannot readily be distinguished with respect to pathway components from Drosophila. Underlining indicates genes shown to be transcriptionally up-regulated after immune challenge. While Toll signalling in flies serves a dual purpose in development and immunity, the signalling process activated by peptidoglycan recognition protein (PGRP)-LC and Imd is specific for antimicrobial defence and is dispensable for normal development (Hultmark, 2003) . Via the NF-κ B-like transcription factor Relish, this signalling induces transcription of all major antimicrobial effector peptides in Drosophila . In Drosophila , Imd signalling is often said to be specific for Gram-negative bacteria, although Gram-positive bacteria with diaminopimelic acid-type peptidoglycans are at least as strong as elicitors. A weaker response is also seen to other types of peptidoglycan and even to fungi (Hultmark, 2003; Werner et al ., 2003; Stenbak et al ., 2004) . This broad specificity is caused by the three alternative splice forms of Drosophila PGRP-LC, which carry different peptidoglycan recognition domains (Werner et al ., 2003; Mellroth et al ., 2005 . Chang et al ., 2006 . Interestingly, the Imd signalling pathway is highly conserved in the honey bee, with plausible orthologues for all components (Fig. 2) . While this strongly suggests that Imd signalling is similar in flies and bees, it does not necessary imply similar biological roles. Besides the activation of Relish, Imd signalling also leads to activation of components of the JNK signalling pathway (Boutros et al ., 2002) , and recent evidence indicates that this pathway can provide both positive and negative feedback for the expression of the antimicrobial peptides (Wojda et al ., 2004) . Plausible orthologues for each of the major components of the JNK signalling were also identified in the honey bee genome (Fig. 2) . The JAK/STAT signalling pathway may also contribute to innate immunity by induction of complement-like factors and the overproliferation of haemocytes. JAK/STAT appears to be initiated via cytokine-like molecules in blood cells (Agaisse & Perrimon, 2004) . In flies, the extracellular glycosylated protein Upd acts as a ligand that activates the JAK/STAT pathway, which in turn promotes phagocytic activity of haemocytes. The JAK/STAT pathway has also recently been shown to participate in an antiviral response in Drosophila (Dostert et al ., 2005) . Honey bee homologues for the Drosophila JAK/STAT signalling pathway (Fig. 2 ) comprise the cytokine receptor domeless (Dom), JAK tyrosine kinase (Hopscotch), the STAT92E transcription factor 3. Transcript abundances for immune candidate genes in adult workers 24 h after injections of Escherichia coli (Ec), saline buffer, or the bee pathogen Paenibacillus larvae , and controls (left four columns). Two columns on right show transcript abundances in 2nd-instar larvae challenged orally with an infective dose of P. larvae or unchallenged controls. Cluster A = genes strongly up-regulated by adult injection or wounding, Cluster B = genes up-regulated in infected larvae, Cluster C = genes down-regulated or minimally changed in challenged bees. and two negative pathway regulators SOCS (suppressor of cytokine signalling) and PIAS (protein inhibitor of activated STAT). Orthologues of two recently identified components of this pathway (Baeg et al ., 2005; Muller et al ., 2005) , the tyrosine phosphatase Ptp61F (XP392429) and the WD40and bromo-domain-containing protein BRWD3 (XP395263), are also present in the honey bee. Although the key ligand (Upd) for the JAK/STAT pathway was not found in the honey bee genome, the presence of the gp130 cytokine receptor homologue Domeless and all other members of the signalling pathway indicates that this mechanism may be common across insects and is intact in honey bees as well as in flies. In addition to up-regulating the complement-like thiolestercontaining proteins (TEPs; Lagueux et al. , 2000; Boutros et al ., 2002) , the JAK/STAT pathway in Drosophila regulates expression of the Turandot (Tot) genes that encode humoral factors induced by severe stress Agaisse et al ., 2003) . None of the Tot factors (Tot A-Z) are apparent in honey bees. While honey bees appear to have maintained each of the known insect immune-related pathways, they appear to do so with a reduced number of paralogous members. When comparing a set of 17 gene families and functional groups implicated in immune responsiveness (Christophides et al ., 2002) , honey bees have substantially lower paralogue counts than either Drosophila or Anopheles (Table 1 ). The 71 genes placed into these groups for honey bees are in sharp contrast to the 196 and 209 found in Drosophila or Anopheles , respectively. Bees have the lowest gene counts for 12 of the 17 families and are tied for the lowest count two more times. Drosophila and Anopheles were lowest for only one family each (defensins and dorsal, respectively). In contrast, Drosophila and Anopheles show the highest paralogue counts for this triad seven and eight times, respectively, versus once in bees (for the Toll-pathway candidate cactus, with three copies). These rankings are significantly different under an ordinal contingency-table analysis ( P < 1.0 × 10 -4 ), and reflect differences for genes involved with pathogen recognition and signalling, as well as effectors. PGRPs, major players in pathogen recognition (Hultmark, 2003; Steiner, 2004; Royet et al ., 2005) are less diverse in honey bees versus flies and other insects for which genomic data exist (e.g. the moth Bombyx mori ). There are only four PGRPs in the honey bee genome, compared to 13 and seven in Drosophila and Anopheles , respectively (Fig. 4A , Table 1 ). Further, bees show no capacity for the splice variation that contributions to diversify peptidoglycan recognition specificity in flies and mosquitoes. Specifically, the single membrane-bound PGRP in bees (PGRP-LC, GB17188) is similar to fly and mosquito PGRP-LC but lacks the potential to insert alternative peptidoglycan recognition domains by alternative splicing, a factor in recognition breadth for flies. PGRP-LC and PGRP-S2 are both up-regulated in honey bees after disease challenge (as in flies), suggesting that the products of these genes indeed play a defensive role (Fig. 3) . As in Anopheles , there is only one Class C Scavenger Receptor (SR-C) in the honey bee. This group has diversified into four members in Drosophila , three of which show selective signs suggestive of an immune role (Lazzaro, 2005) . There are 10 Class B scavenger receptors in the bee, a number roughly similar to that in the fly and mosquito (Fig. 4B , Table 1 ). Several other recognition classes also seem to be reduced in honey bees, including β -glucan recognition proteins ( β GRPs), galectins and fibrinogen-related proteins (Table 1) . Of these, the fibrinogen-domain genes are especially striking, due to the absence in bees of high lineage-specific diversification found in mosquitoes and Drosophila [resulting in 57 and 13 domain-family members, respectively (Christophides et al ., 2004) ]. Along with PGRP-LC and SR-C , two other receptors were recently found to participate in the phagocytosis of infectious non-self in Drosophila , DSCAM and Eater. DSCAM, long implicated in neuronal development, was shown to have a likely role in the binding of bacteria by Drosophila haemocytes (Watson et al ., 2005) . This gene, which has > 12 000 potential splice variants in honey bees thanks to three sets of highly interchangeable exons (Graveley et al ., 2004) , is a very interesting candidate for determining the extent to which bees might better tune their immune response to specific pathogens. Another Table 1 . Gene counts for a subset of gene families implicated in insect immunity. Anopheles gambiae and Drosophila melanogaster counts based on Christophides et al. (2002 Christophides et al. ( , 2004 recently identified protein involved in Drosophila cellular immunity is Eater, a phagocytic receptor characterized by several repeats of an EGF motif in its extracellular domain (Kocks et al., 2005) . Many genes with EGF motifs are present in the bee genome (e.g. GB14654, Supplementary Material Table S1 ), as in flies, although their orthology with Eater is unclear. Bees are also likely to engage in the encapsulation of endoparasites and pathogens, and show typical integrins (Honey Bee Genome Sequencing Consortium, 2006) implicated in lamellocyte encapsulation . None of these cellular immunity components appear to be more diverse in bees than in fly species (Supplementary Material Table S1 ). In addition to their function in digestion of food, serine proteases (SPs) in insects participate in regulatory cascade pathways in embryonic development and in immune responses (Kanost & Clarke, 2005) . Many haemolymph SPs and serine-protease homologues (SPHs) implicated in immunity contain one or more clip domains at their amino terminus, which may regulate or localize the immune responses stimulated by protease cascades (Jiang & Kanost, 2000) . Among the 57 SP-related proteins in the honey bee genome, 12 SPs and six SPHs contain at least one clip domain, significantly fewer than in Drosophila (24 SPs and 13 SPHs) (Ross et al., 2003) or Anopheles (26 SPs and 15 SPHs), and smaller than the number of clip domain SPs identified to date (n = 14) from expression data in Manduca sexta (Jiang et al., 2005) . Additional phylogenetic and functional relationships among insect SPs and SPHs are discussed in Zou et al. (2006) . Serine protease inhibitors from the serpin superfamily regulate protease cascades in mammals and in arthropods (Reichhart, 2005) . In insect haemolymph, serpins inhibit activated proteases to maintain homeostasis and prevent unregulated activation of immune responses such as melanization or Toll-mediated antimicrobial protein synthesis (Kanost & Clarke, 2005) . In the honey bee genome, there are seven annotated genes encoding five serpins and two proteins with serpin-like regions (GB10078 and GB15070). The number of serpin genes in the honey bee is much lower than in Drosophila (28) or Anopheles (14), mirroring the reduced size of the protease gene family in bees. Nine Toll-related receptor genes are known from the Drosophila genome and 10 from Anopheles (Tauszig et al., 2000; Christophides et al., 2004) . In Drosophila, Toll is the primary family member implicated in immune-related function (Lemaitre et al., 1996) , although it appears that Drosophila Toll-5 and Toll-9 (the paralogue that is structurally most similar to mammalian TLRs), are also involved in immune-related signalling (Ooi et al., 2002; Bilak et al., 2003) . We have identified only five Toll-related genes in the honey bee: Toll1, Toll2/18w, Toll6, Toll8/Trex/Tollo and Toll10. Additional Toll members in Drosophila (Toll-3, -4, -5, -7 and -9) apparently reflect gene duplication events in the fly lineage ( Fig. 4) or, in the case of Toll-9, arguably a loss in the honey bee and lepidopteran lineages. For instance, the ancestral Toll appears to have diverged into two different groups, Toll-1 and Toll-5, in flies. In contrast, Toll-1 remains the only member of this clade in honey bees. Similarly, the Toll-7/2 clade is represented only by a single honey bee homologue Am18w (Aronstein & Saldivar, 2005) . Presumably, the ancestral Toll 7/2 gene was duplicated in flies following their divergence 300 Mya from the lineage leading to honey bees. A protein named Apis mellifera toll7 (Kanzok et al., 2004) , seems more likely to be an orthologue of Toll-10. The five Toll receptors present in the A. mellifera genome (Toll-1, -6, -2/7, -8, -10) are also present in the sequenced genomes of other insects that belong to the orders Diptera, Lepidoptera and Coleoptera, with few exceptions. For example, whereas orthologues of Toll-6, -7 and -8 are found in D. melanogaster and A. gambiae (Diptera), Bombyx mori (Lepidoptera), and Tribolium castaneum (Coleoptera), Toll-1 appears to be absent from the genome of the lepidopteran insect B. mori, whereas Toll-10 appears to have been lost in the fruit fly. This suggests that these five genes encode the basic set of Toll receptors that was present in the common ancestor of these insects. These five receptors are highly expressed, in a dynamic and tissue-specific manner, during Drosophila embryogenesis (Kambris et al., 2002) . Of note, Toll-6 and Toll-8, are adjacent in both the D. melanogaster and A. mellifera genomes. Among the immune effectors, the total of six honey bee antimicrobial peptides contrasts with the 20 and nine found in Drosophila and Anopheles, respectively (Table 1) . Two of these (abaecin, and apidaecin) are in the class of prolinerich antimicrobial peptides, two are conventional defensins and two (apisimin and hymenoptaecin) are distinct from all other recognized antimicrobial peptides. Genomic analysis reveals that the gene encoding apidaecin consists of a conserved N-terminus followed by several exons, each of which encodes a complete 28-amino acid peptide. Peptide and cDNA evidence for this gene was used to predict a mechanism for ratcheting up expression of apidaecin in response to bacterial challenge (Casteels-Josson et al., 1993) . The genomic structure of apidaecin raises the possibility of a mechanism for generating specific responses to pathogens by splice variation. As each exon is a functional and distinct antimicrobial peptide, it is conceivable that splice variation at this locus can further refine this gene as an immune effector. Apidaecin exons differ greatly across individual honey bees in both number and in their encoded amino acid sequences. The two bee haplotypes sequenced in this project differed in sequence and exon number and also differed from three previously described apidaecin cDNAs (Casteels et al., 1993) . Collectively, the two haplotypes sequenced here, and the three described by Casteels et al. encode a range of 4-11 secreted peptides each. While some peptides are shared between the various haplotypes for this gene, there is a surprisingly high level of sequence variation, such that the 35 peptides expressed by these five haplotypes reflect 23 different amino acid variants. With the exception of defensin-2 (identified from genomic sequences generated during this project; Klaudiny et al., 2005) all of the honey bee antimicrobial peptides were first characterized by protein sequencing (Casteels-Josson et al., 1994) , a fact that belies the difficulty in discovering such genes by sequence similarity across the millions of years separating insect species. Still, there is no evidence for close paralogues for any of the honey bee antimicrobial peptides, in contrast to such patterns in other insects (e.g. the gene-rich cecropin family). Five of the six bee antimicrobial peptides are up-regulated across diverse immune challenges ( Fig. 3; Evans, 2004) . Honey bees possess only one prophenoloxidase (proPO) gene, versus three and nine in Drosophila and Anopheles, respectively. Like most proPO, the honey bee proPO lacks a signal peptide and has the consensus sequence of NRFG around the activation site. The gene encoding proPO is expressed more strongly in older honey bee larvae and pupae (Lourenço et al., 2005) . This gene was not up-regulated in our challenge experiments, but a gene identified as a proPO activator was up-regulated during natural infection (Fig. 3) . There are only three lysozymes in the honey bee genome, two c-(chicken) type and one i-(invertebrate) type. One of the c-class lysozymes is up-regulated by challenged honey bees (Fig. 3) . There were fewer thiolester-containing proteins (TEPs) in the bee genome (four) than expected based on flies ( Table 1 ). The Anopheles genome encodes 15 TEPs, most of them originating from species-specific expansion (Christophides et al., 2004) versus six members of this gene family in Drosophila (Agaisse & Perrimon, 2004) . TEPs are induced after septic injury and promote phagocytosis in mosquitoes (Blandin & Levashina, 2004; Moita et al., 2005) . They also play a central role in vertebrate innate immunity as the complement factors. In Anopheles, TEP1 was found to promote phagocytosis of Gram-negative bacteria and is also a major player in the host response to plasmodium infection. Members of this group are implicated as both recognition proteins and effectors (opsonins) in insects (Levashina et al., 2001; Moita et al., 2005; Stroschein-Stevenson et al., 2006) . As social animals, honey bees are at considerable risk from parasites and pathogens. Specifically, increased genetic relatedness and the high population densities that typify honey bee societies can strongly favour pathogen spread and epizootic outbreak. However, such social costs to honey bee immunity might be offset by social defences including mating strategy (e.g. multiple mating by queens: Tarpy, 2003) , mutual grooming, and the maintenance of a sheltered environment for colony members. Given their well-studied natural pathogens, immune pathway models from the current annotated draft genome and unique genetic traits, honey bees can join with several fly and moth species as important systems with which to understand the genetic causes of immunity and disease. They also join humans as organisms for which there is great interest in understanding the social drivers of disease, and in using this information to improve host survival. Our analyses indicate that the basic set of molecules defining the insect host-defence system is present in honey bees, including intact pathways for the key processes implicated in immunity and development. Single orthologues can be assigned for many pathway members, while others show several potential bee genes for which further work is needed to confirm roles. Interestingly however, whereas in Drosophila and Anopheles the host appears to have diversified its molecular arsenal through species-biased gene duplication (Table 1) , we have not found examples of such gene expansions in honey bees. Thus, despite having wideranging parasites and pathogens, and tremendous losses to these pests at least in domesticated settings (Morse & Flottum, 1997) , bees appear to have relatively diminished capacities to respond to and defend against pathogens. Our first hypothesis to help explain this observation, that bee gene families are systematically smaller than in other insects, is not supported because there is no evidence for a systematic downward bias in paralogue counts across the bee genome at the level seen for immune-gene families (Honey Bee Genome Sequencing Consortium, 2006) . Hypothesis two, that bee genes were simply missed due to sequence divergence, should apply particularly for immune genes that are short and subject only to limited sequence constraints, such as those encoding antimicrobial peptides. Indeed, few such peptides have been identified from bees (n = 6) and the means with which these few were discovered (directly at the protein level in all but one case) support the idea that sequence-level searches might have missed additional family members. Nevertheless, members of the remaining gene-poor families do not comprise especially short genes, nor genes divergent enough to be missed completely by comparisons at the level of insect orders. In fact, at least one significant bee:Drosophila:Anopheles orthologue is present in each of the remaining discussed families, indicating sufficient sequence-level conservation for identification of bee counterparts. Our third hypothesis, that bees simply have an undescribed mechanism for broadening their immune efficacy, is not testable at this point but would be very surprising given similarities in immune actors across the diverse insect orders studied to date. Two of our initial hypotheses, that bees face a less diverse set of successful parasites and pathogens, and that societal defences by bees lessen pathogen pressures, therefore seem best supported by the data in hand and can now be compared. On one level, bee pathogens are diverse, ranging from Gram-positive and Gram-negative bacteria to fungi, RNA viruses, microsporidia and amoebae (Morse & Flottum, 1997) . Bees are also parasitized by mites and other arthropods, raising risks of both pathogen infection (Kanbar & Engels, 2003; Chen et al., 2004) and lower abilities to combat disease (Gregory et al., 2005; Yang & Cox-Foster, 2005) . Still, despite having pests that range over several kingdoms, common disease agents in bees are in fact restricted to several pathogens, two of which (the bacterium Paenibacillus larvae and the fungus Ascosphaera apis) are predominant. While functional data on the specificity of responses toward these and other pathogens are not yet available, gene-expression changes after challenge do not appear to be especially precise with respect to honey bee pathogens vs. exotics (e.g. Escherichia coli) or stress generally (Fig. 3) . Thus, there is no compelling evidence at this point that the bee immune response is channelled just toward a small set of 'true' pathogens. With respect to the final hypothesis, bees, like many social insects, are relentlessly hygienic, removing alien organisms from their nests, and secreting antimicrobial substances that can reduce the viability and growth of pathogens in the colonies. Bees also raise their young in individual cells using, as a food source, substances with strongly antimicrobial properties (e.g. royal jelly; Albert & Klaudiny, 2004) . A testament to this hygiene is the fact that, even when facing severe colony-level infections by bacterial pathogens such as Paenibacillus larvae (for which < 10 spores are normally fatal to young larvae (Brodsgaard et al., 1998) , the vast majority of larvae show no signs of exposure (Evans & Pettis, 2005) . 'Social' barriers might also reduce exposure to minor, opportunistic, pathogens or saprophytes that have been proposed as generalized targets of insect immune defences (Hultmark, 2003) . While bees do carry an assemblage of microbes, and bacteria in particular (Gilliam, 1997) , exposure to these microbes is arguably lower than in free-living Drosophila (decaying plant material as larvae and adults) or Anopheles (septic aqueous environments as larvae). Further, bacteria found in bee colonies have only rarely been associated with disease pathologies, despite extensive study. In fact, some resident bacteria in colonies appear to add to external defences through their inhibition of bee pathogens (Evans & Armstrong, 2006) . Future genomic work can help reveal whether other species of highly social insects, including ants, wasps, bees and termites, also appear to have more simplistic innate defence systems. More generally, social and solitary insects with more 'exposed' life histories are predicted to have a greater number and higher functional diversity of immunepathway genes and end products, when compared to sister taxa that are more sheltered. Data on parasite and pathogen abundance across social (e.g. Boomsma et al., 2005) and solitary insects could be used as a surrogate for disease loads in different taxa, although field and epidemiological data are most needed to assess the relative fitness impacts of disease and the efficacies of different lines of defence. Through this analysis, we present the first plausible models for immune pathways in a social insect, the honey bee. We show nearly complete conservation of candidate genes for these pathways yet show that bees have consistently undercut numbers of genes that embellish these pathways in other insects. Genome-wide expression studies newly available for bees, and the proven success of gene knockdown techniques such as RNA inactivation (Lourenço et al., 2005) , allow for more refinement of the roles played by pathway members as well as the discovery completely novel players in honey bee immunity. The latter discoveries, combined with analyses across more species of social and solitary bees, will help determine whether the observations described here are unique to the highly social honey bees. Longstanding agricultural interest has helped generate a wealth of data on honey bee pathologies (Morse & Flottum, 1997) , and it is now possible to connect these data with immune traits that help limit pathogen efficacy. Through these connections, honey bees will provide a valuable and tractable model for disease transmission, immunity, 'socialized medicine' and pathology. Immune-gene candidates from other insects were used in several ways to query the honey bee genome, primarily using the BLAST family of search functions (www.ncbi.nlm.nih.gov). Most searches were initiated by BLASTP queries against the consensus protein list (GLEAN3, derived from HBGP assembly 2.0) using BLASTP and algorithms (BLOSUM and PAM variants) appropriate to gene size and structure. Honey bee orthologues were also identified by searching honey bee genome assemblies 2.0 and 3.0 directly using TBLASTN and either local databases or the BeeBase BLAST server (http://racerx00.tamu.edu/blast/blast.html). Searches for missing genes were also carried out a on smaller coverage set of honey bee contigs that were too short to be included in the assembly, as well as the unassembled reads from the project (http://www.hgsc.bcm.tmc.edu/projects/honeybee/). Given honey bee candidates, searches were repeated in the hope of identifying paralogues missed by interspecific comparisons. PSI-BLAST was used to identify honey bee genes on the basis of conserved domains, followed by RPS-BLAST to confirm the significance of these domain matches. Putative matches were aligned and, in the case of serine proteases, scavenger receptors and C-type lectins, screened for additional motifs using pfam categories (http://sanger.ac.uk/software/pfam). Tentative matches were aligned and checked for gene-prediction errors (in the case of genes from the official protein list) as part of the annotation of immune candidate genes for the Honey Bee Genome Project (Honey Bee Genome Sequencing Consortium, 2006). All protein matches were ported to Apis mellifera assembly 3.0 using the alignment program BLAT (Jim Kent, University California, Santa Cruz) to establish scaffold locations (Supplementary Material Table S1 ). Best-match honey bee sequences were then aligned with counterparts from D. melanogaster and A. gambiae, along with other insects where sufficient genome-level data were available. Amino acid sequence alignments were carried out using GONNET series weight matrices, with the program CLUSTAL_X (Chenna et al., 2003) . Alignments were used to propose phylogenetic relationships using maximum-parsimony and neighbour-joining algorithms, with the programs PAUP* (Sinauer, Sunderland, MA) or PHYLIP (http:// evolution.genetics. washington.edu/phylip.html). The PGRP tree is based on the conserved domain region only. Other alignments were edited manually to reduce or remove ambiguous regions. For scavenger receptor class B proteins, human (NP_005497) and mouse (NP_031669 CD36) proteins were added to alignments. Honey bee scavenger receptor AmelSCRB8 was not included because its relatively short predicted length (apparently the result of an intercontig gap) precluded an unambiguous alignment. All alignments are available on request. Two experiments were carried out to screen for immune-related transcript changes. In the first, adult worker bees from a single local A. mellifera ligustica colony were removed, then injected abdominally with either dilute phosphate-buffered saline or saline solution containing 103 live cells of E. coli or 103 vegetative spores of the honey bee bacterial pathogen Paenibacillus larvae. These bees, along with uninjected controls, were maintained for 24 h at high humidity and 34 °C and then were immediately frozen at −70 °C prior to RNA extraction. To assess immune responses following natural infection, eight 1st-instar larval bees from the same stock were given per os challenges of P. larvae in their food [(50 spores/µl as described in Evans (2004) ], then maintained 24 h at 34 °C and high humidity. Parallel control larvae were given the same food without bacterial spores. All samples were frozen at −80 °C following incubation. RNA was extracted from whole abdomens of the adult bees using a standard TRIzol (Invitrogen, Carlsbad, CA, USA) procedure while RNA was extracted from individual larvae using the RNAqueous kit (Ambion, Austin, TX). RNAs were pooled by sample duration for the eight larvae challenged with the bacterial pathogen P. larvae, and the eight controls prior to cDNA synthesis, giving six RNA pools. DNA was removed from all extracts, then first-strand cDNA was synthesized as described by Evans (2004) . Transcript abundances for these cDNAs were assayed by quantitative real-time PCR with an Icycler real-time PCR machine (Bio-Rad, Hercules, CA, USA). Primer pairs were designed to amplify 120-300 bp sections of 39 honey bee immune-related genes derived from Supplementary Material Table S1 and ribosomal protein S5 as a control gene (primers in Supplementary Material Table S2 ). Primer sequences were modified, where necessary, to run in duplicate on 96-well plates using a fixed thermal protocol consisting of 5 min at 95 °C, then 40 cycles of a four-step protocol consisting of 94 °C for 20 s, 60 °C for 30 s, 72 °C for 1 min, and 78 °C for 20 s was used (Evans, 2006) . Reactions were carried out on 0.5-2 µg cDNA along with 1 U Taq, the provided PCR buffer (Roche Applied Sciences, Indianapolis, IN, USA), 1 mM dNTP mix, 2 mM added MgCl 2 , 0.2 µM each primer, 1× concentration SYBR-Green I dye (Applied Biosystems, Foster City, CA, USA), and 10 nM fluorescein in a 25 µl reaction volume. Amplification was followed by a meltcurve dissociation program in order to confirm expected product size. Thresholds were calculated individually for each target gene on the 96-well plate. For adult bee samples, data were pooled for the three replicates in each single-bee injection treatment (or controls). Results were screened for the appropriate dissociation (melt-curve) values, and by 1% agarose gels, in order to ensure against primer artefacts and the presence of DNA contamination (which would have been evident for numerous primers spanning two exons). Immune-gene transcripts were normalized relative to expression levels for the gene encoding ribosomal protein S5, a gene with consistent expression across honey bee life stages and disease status (Evans & Wheeler, 2000; Evans, 2004) . For display purposes, transcript abundance values (CTcontrol-CTtarget) for each gene were median-normalized across each panel of genes and clustered by average linkage clustering (using Cluster 3.0, M. Eisen, www.rana.lbl.gov/EisenSoftware.htm) and presented as relative grey-scale values (using Treeview, M. Eisen).
83
ElaD, a Deubiquitinating Protease Expressed by E. coli
BACKGROUND: Ubiquitin and ubiquitin-like proteins (Ubl) are designed to modify polypeptides in eukaryotes. Covalent binding of ubiquitin or Ubls to substrate proteins can be reversed by specific hydrolases. One particular set of cysteine proteases, the CE clan, which targets ubiquitin and Ubls, has homologs in eukaryotes, prokaryotes, and viruses. FINDINGS: We have cloned and analyzed the E. coli protein elaD, which is distantly related to eukaryotic CE clan members of the ULP/SENP protease family that are specific for SUMO and Nedd8. Previously misannotated as a putative sulfatase/phosphatase, elaD is an efficient and specific deubiquitinating enzyme in vitro. Interestingly, elaD is present in all intestinal pathogenic E. coli strains, but conspicuously absent from extraintestinal pathogenic strains (ExPECs). Further homologs of this protease can be found in Acanthamoeba Polyphaga Mimivirus, and in Alpha-, Beta-and Gammaproteobacteria. CONCLUSION: The expression of ULP/SENP-related hydrolases in bacteria therefore extends to plant pathogens and medically relevant strains of Escherichia coli, Legionella pneumophila, Rickettsiae, Chlamydiae, and Salmonellae, in which the elaD ortholog sseL has recently been identified as a virulence factor with deubiquitinating activity. As a counterpoint, our phylogenetic and functional examination reveals that ancient eukaryotic ULP/SENP proteases also have the potential of ubiquitin-specific hydrolysis, suggesting an early common origin of this peptidase clan.
Ubiquitin, as well as Ubls such as Nedd8 and SUMO, are proteins (almost) exclusively expressed by eukaryotes [1] . Ubiquitination controls many cellular processes, including degradation of proteins by the proteasome and intracellular trafficking. Nedd8 is a ubiquitin-like modifier that regulates the rate or extent of ubiquitination, whereas SUMO1 is involved mostly in regulation of transcription factors and in nuclear import. The attachment of ubiquitin or ubiquitin-like modifiers to substrate proteins is covalent, yet reversible [2] . A large family of eukaryotic cysteine proteases is involved not only in generation of ubiquitin(-like) proteins from their precursors, but also in their removal from modified substrates [3, 4] . Pathogens can tamper with the ubiquitin-proteasome system to cripple the cell's defenses [5] . For instance, ubiquitination and proteasomal degradation of p53, initiated by a Human Papillomavirus protein [6] , or stabilization of Ikb-a by Yersinia deubiquitinases [7] have been described. The continuing discovery of new deubiquitinating proteases in viruses broadly hints at how important it is for these pathogens to seize control of posttranslational modifications in host cells [8] [9] [10] . Here, we focus on cysteine proteases of the CE clan [11] , as defined by the MEROPS database [12] . CE peptidases are expressed by viruses, bacteria and eukaryotes. In eukaryotes, this clan represents the family of Ubl-specific proteases (ULP/SENP), which remove SUMO or Nedd8 from substrate proteins [13, 14] . Viral homologs of ULP/SENPs can act as deubiquitinases, but they also cleave unrelated proteins, as long as a glycine motif is present at the C-terminus of the substrate [15, 16] . Examples of bacterial deubiquitinases include YopJ and ChlaDUBs. First, YopJ is a protease that is secreted into host cells by Y. pestis [17] , and homologs to this peptidase can be found in other bacteria, too [18, 19] . Injection of YopJ eventually suppresses the inflammatory response in affected cells. The precise molecular mechanism of YopJ family peptidases has not yet been solved, as they lack a hallmark tryptophan following the active-site histidine [20] . Yet, in vitro and in vivo data strongly suggest that these proteins are indeed proteases with specificity for ubiquitin or SUMO [7, 18, 19] , although their effectiveness as virulence factors might depend on additional functions such as acetylation [21, 22] . A second example of bacterial CE peptidases can be found in pathogenic Chlamydiae. We have shown that pathogenic Chlamydiae, but not a nonpathogenic environmental strain, express proteases that specifically recognize ubiquitin and Nedd8. They do so presumably to remove both modifiers from target proteins of the host cell, as Chlamydiaelike most other bacteria-possess neither a ubiquitin nor a Nedd8 homolog [23] . Intrigued by the finding of CE peptidases in bacteria and considering their functional similarity to eukaryotic ULP/SENP proteases, we explored whether additional bacterial homologs with deubiquitinating activity might exist. To search the sequenced genomes of bacteria for new members of the CE protease clan, we employed PHI-BLAST [24] to find proteins with the typical catalytic triad of histidine, aspartate (or glutamate or asparagine), and the active-site cysteine [25] . Candidate proteins were subjected to a second round of analysis, in which we excluded proteins without the hallmark oxyanionstabilizing group, consisting of at least one glutamine (or asparagine) close to the active-site cysteine. Lastly, the predicted secondary structure of candidate proteases was compared to the solved structure of eukaryotic CE clan homologs [26, 27] . Apart from YopJ and the ULP/SENP homologs in Chlamydiae, we found potential CE peptidases in Alpha-, Beta-, and Gammaproteobacteria ( Figure 1 ), but not in bacteria from other branches. We also found a peptidase homolog in the giant Acanthamoeba Polyphaga Mimivirus [28] and in African Swine Fever Virus, the first virus in which a ubiquitin conjugating enzyme was discovered and for which deubiquitinating activity has been suggested [29] [30] [31] [32] . The bacteria we identified share a close (symbiotic or pathogenic) relationship with eukaryotes, and all viruses with CE proteases possess a dsDNA genome. The sequence variations among these peptidases are too extensive to allow for significant bootstrap values and it is therefore not possible to faithfully infer phylogeny from this dataset. Yet, this is also true for the eukaryotic ULP/SENPs, which fall into three functional classes, specific for either SUMO (ULP1 and ULP2 group) or Nedd8 (SENP8 group). While best reciprocal BLAST hits and consensus phylogram trees tend to correctly predict to which functional class a particular ULP/SENP homolog belongs, bootstrap support is generally not significant [33] . We were especially interested in the protein elaD (belonging to ''group II'', see Figure 1 ), expressed by E. coli and with orthologs in Legionella pneumophila and in all currently sequenced strains of Salmonella ( Figure 2 ). E. coli is an abundant commensal in the human gut and also a relevant pathogen. Nonetheless, little is known about genes that define pathogenicity of various E. coli strains, apart from those that encode obvious toxins [34] [35] [36] . We furthermore chose elaD, because it represents one of the more distantly related hits in our bioinformatics screen and we aimed to test the robustness of our prediction by examining this protein's function. Figure 1 . Phylogram representation of CE clan proteases in viruses, bacteria, and eukaryotes. Eukaryotic peptidases (in blue) belong to the C48 subfamily and can be separated into three groups: ULP1 (including the mammalian proteases SENP1, 2, 3, and 5), ULP2 (including SENP6 and 7), and the SENP8 group with proposed specificity for SUMO (ULP1 and ULP2 group) and Nedd8 (SENP8 group), respectively. Bacterial proteins are indicated with a preceding ''B'', viral proteins with a ''V''. We have further divided microbial protease homologs by color: green for biochemically tested proteases, red indicating the absence of published data on the function of these putative proteases, and yellow for the group representing elaD and its orthologs. The C5 family contains Adenovirus proteases with deubiquitinating activity, C55 comprises the bacterial YopJ homologs, and C57 the Vacciniavirus I7 peptidases. Based on sequence similarity, two bacterial C48 family groups can be distinguished: a group of Proteobacteria (located at one o'clock) which appear to be closely related to fungal SENP8 homologs (common node indicated with a circle, bootstrap support.60%), and Chlamydiae, for which we had previously shown the presence of deubiquitinating and deneddylating activity. Three additional groups have not yet been assigned to specific CE clan subfamilies in the MEROPS database, including Mimivirus (''group I''), Gammaproteobacteria (''group II''), and Rickettsiae (''group III''). The African Swine Fever Virus protease and the I7 Vacciniavirus protease have not been tested for deubiquitinating or Ublspecific activity, but they both require a glycine-based motif at the C-terminus of the substrate, as found in ubiquitin or Ubls. The unrelated CD clan peptidase Clostripain is used as outgroup in this phylogram. For clarity, this tree does not contain all orthologs and paralogs of the different groups or families. Sequence information is provided in Table 2 . doi:10.1371/journal.pone.0000381.g001 The goal of our first experiment was to confirm protease activity and to determine the substrate specificity of elaD. Because of its relationship to ULP/SENPs, we hypothesized that elaD might recognize Ubls or ubiquitin. We expressed elaD by in vitro transcription/translation in rabbit reticulocyte lysate and incubated the metabolically labeled polypeptide with electrophilic probes, in which a Michael acceptor was added to the C-terminus of ubiquitin or the Ubls SUMO1, Nedd8, and ISG15 [14] . As shown in Figure 3 , elaD readily forms a covalent adduct with the ubiquitin probe and to a much lesser extent also with the Nedd8 probe, but not detectably with SUMO1 or ISG15. Moreover, when mutating the putative active-site cysteine at position 313 to serine, covalent binding to the electrophile is abolished. This indicates that the cysteine residue in elaD is essential for catalytic activity, as has been observed for YopJ [7] , for the Chlamydia protease CT868 [23] and for the eukaryotic ULP/SENPs [2] . To date, specific labeling of putative proteases with activity-based probes has shown excellent correlation with enzymatic activity [14, 37, 38] . Next, we assessed enzyme kinetics by measuring hydrolysis of fluorogenic substrates derived from ubiquitin, SUMO1 and Nedd8. For these experiments, we expressed elaD in E. coli. The growth rate of the bacteria was unaffected when overexpressing elaD, but we only recovered about 50% of the wildtype protein when compared to the amounts of C313S mutant ( Figure 4A ). As demonstrated by release of the fluorophore 7-amino-4-methylcoumarin (AMC), elaD cleaves ubiquitin-AMC, but not SUMO1-AMC or Nedd8-AMC, and the C313S mutant of elaD fails to cleave either substrate ( Figure 4B and data not shown). The initial rate of hydrolysis with 100 nM ubiquitin-AMC and 50 nM elaD is in the order of 0.3-0.6 per minute, defining elaD as a moderately active deubiquitinase, compared to the rapid Isopeptidase T with a rate of ca. 8 per minute [39] or to the much slower ubiquitinprotease USP14 with a rate of,0.01 per minute (data not shown). It should be noted that we could not assess V max , because the enzymatic rate increased linearly with substrate concentration (we tested up to 20 mM ubiquitin-AMC; 50 nM elaD then hydrolyzed ubiquitin-AMC at an initial rate of 5 per minute). Similar observations have been made with the SARS virus deubiquitinase [8] . Overall, our functional analyses confirm the prediction of our bioinformatics screen and define elaD as a deubiquitinating protease. Why do eukaryotic CE peptidases show specificity distinct from their prokaryotic counterparts? Could this indicate a profound discrepancy, hinting towards a separate origin of these two protease groups? We set out to challenge the presently held notion that ULP/ SENP proteases do not exhibit ubiquitin-specific activity [13] , while most tested bacterial homologs apparently do. To this end, we chose to biochemically define CE clan members of a more deeply branched class of eukaryotes. Pezizomycotina, a subgroup of fungi that includes A. fumigatus, A. nidulans, M. grisea, N. crassa, and G. zeae encode putative SENP8 proteases that are related in amino acid sequence to a group of yet uncharacterized bacterial C48 homologs ( Figure 1 ). In our phylogram, the common node between the respective prokaryote C48 group and SENP8 homologs of Pezizomycotina replicates with a bootstrap support of.60%. We cloned, expressed and tested the putative SENP8 protease of G. zeae, as a representative of Pezizomycotina. Unlike mammalian SENP8, the G. zeae ortholog displays dual activity towards Nedd8 and ubiquitin, similar to the previously defined CE clan protease CT868 Figure 2 . Sequence comparison between SENP8 and its homologs in human pathogenic bacteria. Multiple sequence alignment of the catalytic core region of human SENP8 (NCBI protein sequence identifier GI: 33942066, shown are residues 100-165) with the homologs in C. trachomatis (CT868, GI: 76789615, residues 199-284) [23] , E. coli (elaD, GI: 15832411, residues 228-319), L. pneumophila (GI: 52843101, residues 189-265), and S. typhi (sseL, GI: 29141091, residues 197-264) [42] . The arrows indicate active-site histidine, aspartate (or asparagine), the catalytic cysteine, and the oxyanion-stabilizing group. Predicted secondary structures are shown at the bottom and have been confirmed with the solved structure of SENP8 [27] . doi:10.1371/journal.pone.0000381.g002 Figure 3 . Biochemical assay for substrate specificity of elaD. 35 S-methionine-labeled in vitro translated wildtype elaD forms covalent adducts with suicide inhibitors based on ubiquitin (ubiquitin-vinylmethylester, VME) and Nedd8 (Nedd8-vinylsulfone, VS), but not with probes based on SUMO1 and ISG15. All probes were tested for activity with bona fide substrates (not shown) [14] . Mutation of the active-site cysteine at position 313 to serine abolishes adduct formation of elaD to electrophilic probes. Samples were resolved by reducing SDS-PAGE and visualized by fluorography. Indicated at the right is the molecular mass in kDa. doi:10.1371/journal.pone.0000381.g003 of the prokaryote C. trachomatis ( Figure 5 ) [23] . This result indicates that ubiquitin-specificity is not restricted to bacterial or viral CE peptidases, but also exists in some ancient eukaryotic members of this protease clan. We have found previously undescribed CE peptidase homologs in several bacterial and in one viral species (Figure 1) . We have furthermore proven that one of the more distantly related homologs-the protein elaD in E. coli-can act as a deubiquitinating enzyme in vitro. Earlier, we had shown that related Chlamydia proteases also have deubiquitinating activity and similar observations had been made regarding the YopJ protease family. This suggests that the yet undefined bacterial CE peptidases could also display ubiquitin-or Ubl-specificity, especially considering their even closer sequence relationship to eukaryotic ULP/SENP proteases. These findings raise two important questions. First, are these proteases really specific for ubiquitin in vivo, or are we simply measuring an offtarget artifact, with the true substrates merely resembling ubiquitin? Second, why are these proteases-which are present in every eukaryote-so widely distributed in bacteria and viruses as well, and what can we learn about their genetic origin? One might argue that the bacterial CE clan peptidases have distinct specificity for bacterial substrates. For instance, it has been proposed that the origin of the ubiquitin system predates the split between eukaryotes and prokaryotes [40, 41] . Factors that distantly resemble Ubls and their conjugating and deconjugating enzymes can be found in bacteria. In this manner, the cleavage of ubiquitin by elaD could just be an artificial byproduct of our assays, while the true substrate is of bacterial origin. Similarly, the Adenovirus CE protease has relaxed specificity for a consensus site that is present in ubiquitin, but also in certain viral substrates [15, 16] . However, we have extended our analysis of elaD's specificity to the ubiquitin homologs Nedd8 and ISG15. Both share significant sequence similarity to ubiquitin, and ISG15 is even identical at the critical C-terminal region. We observed no reactivity between elaD and ISG15-vinylsulfone, and the binding of elaD to Nedd8vinylsulfone was significantly weaker than to the ubiquitin probe. Furthermore, we detected hydrolysis of the C-terminal peptide bond in ubiquitin-AMC, but not in Nedd8-AMC. These features clearly distinguish elaD from the more promiscuous viral CE peptidases. Given the similarity of primary, secondary and tertiary structure among these Ubls, we conclude that hydrolysis of ubiquitin by elaD reflects a highly specific interaction. With the exception of A. avenae, no bacterial strain in our dataset encodes a homolog of ubiquitin, making it likely that the substrate of elaD is indeed eukaryotic ubiquitin. Moreover, the ortholog of elaD in Salmonella-sseL-has recently been shown to be a virulence factor and to display deubiquitinating activity in vitro and in vivo [42] . This enzyme is encoded by all currently sequenced Salmonellae, but only present as a pseudogene in Shigellae [43] . Likewise, elaD is not essential for E. coli under laboratory conditions [44] . A comparison of the genomes of all 16 sequenced E. coli strains reveals that elaD is present in the commensal E. coli strain K12, and in all intestinal pathogenic strains (EAEC, EHEC, EIEC, EPEC, ETEC), but absent from all ExPEC strains (APEC, NMEC, UPEC) ( Table 1 ). To answer the second question raised above, the fact that all CE proteases show the same signature motifs at the catalytic domain and that they have substrate specificity for ubiquitin, Ubls, or related products, suggests a common genetic and functional origin of these proteases. The viral and bacterial organisms that express these CE proteases all share an intimate relationship with eukaryotes, either as commensals, symbionts or as pathogens. Additionally, lateral gene transfer has been proposed between eukaryotes and dsDNA viruses [28] , as well as between eukaryotes and bacteria such as Chlamydiae, Rickettsiae, and L. pneumophila [45, 46] . The notion of a dynamic horizontal gene transfer involving deubiquitinases is further underscored by the distribution of ubiquitin-specific proteases in Chlamydiae: all pathogenic strains express CE peptidases, except for C. pneumoniae, which instead encodes a homolog of the eukaryotic otubain-type deubiquitinating enzymes [47] . Also, one group of bacterial C48 peptidases in particular clusters close to the SENP8 homologs of Pezizomycotina ( Figure 1 ). The sequence relationship and the shared habitat of these organisms-plant symbionts and phytopathogens vs. environmental fungi-raises the possibility of gene transfer between them [48] . From a functional perspective, we could show that a homolog of SENP8 in Pezizomycotina does exert ubiquitin-specific hydrolase activity, like previously characterized bacterial CE proteases ( Figure 5 ). The dual specificity of SENP8 from the fungus G. zeae towards ubiquitin and Nedd8 is not trivial. Although Nedd8 is arguably the closest relative to ubiquitin, there are sequence differences that require distinct conjugating machineries and deconjugating proteases [27, 49] . In particular position 72-an arginine in ubiquitin, and an alanine in Nedd8-can act as a compatibility switch and a single replacement at this side chain can cause ubiquitin to mimic Nedd8 and vice versa [50, 51] . In this respect, the recognition of both ubiquitin and Nedd8 by the SENP8 homolog of G. zeae is in contrast to what has been observed in mammalian SENP8 [27] . One possible explanation for these observations is that CE proteases originally derived from a deubiquitinase, a protease specific for this most conserved eukaryotic protein. As CE peptidases in eukaryotes structurally diversified to accommodate the evolving family of Ubls, protease counterparts in bacteria, viruses, and some deeply branched eukaryotes retained their specificity for the ''ur-substrate'' ubiquitin. Together with the published literature, our data supports the notion that the clan of CE proteases was acquired by bacteria and viruses via horizontal gene transfer from eukaryotes. Why this family of enzymes forms such a particularly attractive substrate for genetic exchange is an intriguing question. The distribution of CE proteases in symbiotic and pathogenic prokaryotes and viruses is suggestive of a general role in host-microbe interactions, as exemplified by the Salmonella protease sseL [42] . Protein and DNA sequence data were obtained from the National Center for Biotechnology Information (www.ncbi.nlm.nih.gov), the Institute for Genomic Research (www.tigr.org), and the University of Wisconsin E. coli Genome Project (www.genome.wisc.edu). Protein sequence identifiers are listed in Table 2 . Sequences containing and surrounding the catalytic core of the proteases were aligned with the ClustalX algorithm (default parameters) (bips.u-strasbg.fr/fr/Documentation/ClustalX) [52] , manually edited with Genedoc (www.psc. [53] . Secondary structures were predicted with JPred (www.compbio.dundee.ac.uk/,www-jpred/) [54] . Phylograms were constructed with the MEGA software package [55] , using the Neighbor-Joining Method with Poisson correction (all substitutions, homogeneous pattern, c-distribution 2.0) and pairwise deletion of gaps. Figure 1 shows a consensus tree based on 100 bootstrap replications. Cloning, expression, and biochemical analysis of elaD and G. zeae SENP8 The full-length elaD gene (NCBI protein sequence identifier GI: 16130204) was amplified by PCR from the K12 strain BL21 (Novagen) and cloned into pET28a (Novagen The putative proteases are separated by groups (as shown in Figure 1 ), by species, and by subdivision of Proteobacteria. This list is not complete in terms of orthologs/paralogs or bacterial species.
84
Mutational analysis of human CEACAM1: the potential of receptor polymorphism in increasing host susceptibility to bacterial infection
A common overlapping site on the N-terminal IgV-like domain of human carcinoembryonic antigen (CEA)-related cell adhesion molecules (CEACAMs) is targeted by several important human respiratory pathogens. These include Neisseria meningitidis (Nm) and Haemophilus influenzae (Hi) that can cause disseminated or persistent localized infections. To define the precise structural features that determine the binding of distinct pathogens with CEACAMs, we have undertaken molecular modelling and mutation of the receptor molecules at previously implicated key target residues required for bacterial binding. These include Ser-32, Tyr-34, Val-39, Gln-44 and Gln-89, in addition to Ile-91, the primary docking site for the pathogens. Most, but not all, of these residues located adjacent to each other in a previous N-domain model of human CEACAM1, which was based on REI, CD2 and CD4. In the current studies, we have refined this model based on the mouse CEACAM1 crystal structure, and observe that all of the above residues form an exposed continuous binding region on the N-domain. Examination of the model also suggested that substitution of two of these residues 34 and 89 could affect the accessibility of Ile-91 for ligand binding. By introducing selected mutations at the positions 91, 34 and 89, we confirmed the primary importance of Ile-91 in all bacterial binding to CEACAM1 despite the inter- and intraspecies structural differences between the bacterial CEACAM-binding ligands. The studies further indicated that the efficiency of binding was significantly enhanced for specific strains by mutations such as Y34F and Q89N, which also altered the hierarchy of Nm versus Hi strain binding. These studies imply that distinct polymorphisms in human epithelial CEACAMs have the potential to decrease or increase the risk of infection by the receptor-targeting pathogens.
The bacterial pathogens Neisseria meningitidis (Nm) and Haemophilus influenzae (Hi) are frequently found in the nasopharynx of a substantial proportion of the healthy population but are capable of causing serious infections in susceptible individuals (Turk, 1984; Foxwell et al., 1998) . Nm and typable Hi (THi) can invade the nasopharyngeal epithelial barrier to cause septicaemia and meningitis, which in the case of Nm, may rapidly become life threatening (van Deuren et al., 2000) . Non-typable Hi (NTHi), which lack a polysaccharide capsule, are associated with localized respiratory tract and conjunctival infections (Foxwell et al., 1998) . Strains belonging to Hi-biogroup aegyptius (Hi-aeg) are also associated with Brazilian purpuric fever (Foxwell et al., 1998) . The factors that determine susceptibility to infection by these frequent colonizers are not entirely clear. For both colonization and pathogenesis, the first essential step is adherence to mucosal epithelial cells. Many investigations have shown bacterial targeting of specific human signalling molecules such as integrins, sialic acid binding Ig like lectins (Siglecs) and carcinoembryonic antigen (CEA)-related cell adhesion molecules (CEACAMs) can lead to cellular invasion (Virji et al., 1995; 1999; Hauck and Meyer, 2003; Jones et al., 2003) . Of these, CEACAMs have emerged as common targets of several respiratory mucosal pathogens and include Nm, Hi, Moraxella catarrhalis, as well as the urogenital pathogen Neisseria gonorrhoeae and enteric pathogens Escherichia coli and Salmonella (Leusch et al., 1991; Virji et al., 1996a; Chen et al., 1997; Gray-Owen et al., 1997; Hill and Virji, 2003) . Carcinoembryonic antigen-related cell adhesion molecules belong to the immunoglobulin (Ig) superfamily. Several members of the CEACAM subgroup are expressed on human epithelial cells and include the widely expressed transmembrane CEACAM1 as well as GPI-anchored CEA and CEACAM6. All CEACAMs have an N-terminal IgV-like domain and variable numbers of IgC2-like A and B domains. CEACAM1 comprises up to four extracellular domains: N, A1, B and A2 and either a long or a short cytoplasmic tail (Tsutsumi et al., 1990; Prall et al., 1996; Hammarstrom, 1999) . Various functions have been attributed to CEACAM1 including cell-cell adhesion, insulin regulation and angiogenesis (Obrink, 1997; Hammarstrom, 1999; Wagener and Ergun, 2000; Najjar, 2002) . Targeting of CEACAMs by N. gonorrhoeae, as well as Nm and Hi leads to cellular invasion and passage across polarized monolayers (Virji et al., 1999; Gray-Owen, 2003; M. Soriani, K. Setchfield, D.J. Hill, and M. Virji, unpublished data) . A wide range of bacterial adhesins are involved in targeting the CEACAM N-terminal domains and includes Opa proteins, a major adhesin family of pathogenic Neisseria and P5 proteins of Hi (Chen and Gotschlich, 1996; Virji et al., 1996a; Hill et al., 2001) . Nm and N. gonorrhoeae contain multiple copies of Opa genes that encode conserved domains which form b-barrel structures in bacterial membranes and variable domains that form surface exposed loops. In spite of the surface diversity afforded by the hyper-variable domains of the loops, the majority of the Opa proteins are capable of targeting CEACAMs (Virji et al., 1996a; Hauck and Meyer, 2003) . The P5 proteins of Hi are similar b-barrel forming proteins, also with surface variable loops (Webb and Cripps, 1998; Vandeputte-Rutten et al., 2003) . The interactions between these bacterial ligands and CEACAMs are complex and the binding domain appears to involve more than one variable loop of the bacterial adhesins (Virji et al., 1999; Bos et al., 2002; de Jonge et al., 2003) . Interestingly, antibody inhibition studies have shown that the diverse ligands of neisseria and haemophilus bind to an overlapping site on the N-domain. In addition, mutational analysis of the N-domain of CEACAM1 has identified several critical residues particularly Ile-91. Alanine substitutions at these sites abrogated binding of most Opa-and P5-expressing bacteria to CEACAM1. Additional residues such as Tyr-34, Ser-32, Val-39, Gln-44 and Gln-89, most of which located in the vicinity of Ile-91, appear to determine the efficiency of interactions of various Opa and P5 molecules (Virji et al., 1999; . The bacterial binding surface on the CEACAM1 N-domain is the protein face composed of the beta strands C′′, C′, C, F and G (CFG for brevity). Despite the extensive investigations in a number of laboratories particularly on neisserial Opa proteins, it remains unclear as to precisely how CEACAM-binding ligands engage with the receptors. A three dimensional structural model of the human CEACAM1 N-domain has been previously generated based on other Ig family molecules (Virji et al., 1999) . In the current investigation, we have refined our previous model based on murine CEACAM1 crystal structure (Tan et al., 2002) . Examination of this model showed a better confluence of the above residues into a continuous binding site and suggested that substitutions at positions 34 and 89 at the core of the binding region may affect bacterial access to the implicated key binding residue Ile-91. By introducing conservative and non-conservative substitutions at positions 34, 89 and 91, we have examined the binding of a variety of bacterial strains to the receptor constructs. The data suggest that single nucleotide polymorphisms (SNPs) in individuals or populations that may introduce substitutions in CEACAM sequence particularly at the bacterial binding site, could not only decrease but also significantly increase the functional affinity of pathogen interactions. Increased binding affinity may result in increased cellular invasion and thus may lead to increased host susceptibility to infection by CEACAM-targeting bacteria. A three-dimensional model of CEACAM1 N-domain was previously produced based on the Ig family molecules REI, CD2 and CD4 (Fig. 1A ) (Virji et al., 1999) . In this model, whilst most mutations affecting binding of Nm and Hi were located centrally on the CFG face of the protein, Val-39 involved in Opa binding, was located at a distance towards the bottom of the CFG face (Fig. 1A) . Subsequently, a crystal structure of murine soluble CEACAM1a domains 1 and 4 was reported (Tan et al., 2002) . Based on this, we have remodelled the human CEACAM1 Ndomain. The CC′ loop that contains Val-39 folds back against the CFG face of this model such that it lies in close proximity to the other critical residues involved in bacterial adhesion (Fig. 1B) . Further examination of this model suggested that the bacterial binding pocket forms a rather flat surface in the centre of the CFG face. It also appears that substituting neighbouring residues Tyr-34 and Gln-89 by Phe and Asn, respectively, could result in a further flattening of the bacterial binding surface (Fig. 1C-E) , providing increased access to the key Ile-91. These residue changes also increase the size of the hydrophobic patch centred on I91 (Fig. 1C-E) . Both of these effects might facilitate binding of some bacterial ligands to the target receptor. In order to assess the importance of these residues in interactions with mucosal pathogens, substitutions were introduced by site-directed mutagenesis at three sites (91, 34 and 89) on the N-domain. The substitutions introduced in CEACAM1 [NA1B]-Fc are shown in Fig. 2 . Chimeric receptors proteins with the native sequence or with sub- (Virji et al., 1999) . (B) New model based on the murine CEACAM1a N-domain crystal structure (Tan et al., 2002) . The models are presented as Van der Waals surface representation. In the latter case V39 locates more centrally with the other critical residues for bacterial binding. (C-E, left) Stereo pairs presented are ribbon diagrams for CEACAM1 N-domain in the native form (C), with Y34F (D) or Q89N (E) substitutions show a flattening of the bacterial binding region on CEACAM1 and improving accessibility of the primary binding residue I91. The side-chain atom colouring (C-green, O-red, N-blue) show the increased hydrophobic nature of the binding site around I91 in the mutant structures. (C-E, right) Surface presentations of the binding site (views corresponding to a 90°rotation of the stereo images in the horizontal axis) coloured according to hydrophobicity. stitutions were produced by transient transfection of COS cells for functional studies described below. The N-domain specific monoclonal antibody (mAb) YTH71.3 has been shown to require Gln-89 and Ile-91 for receptor recognition and substitution of these amino acids had no effect on the binding of the polyclonal anti-CEACAM antibody A0115 (Virji et al., 1999) . The novel receptor constructs produced in the current studies were first examined for their ability to bind to YTH71.3, A0115 as well as Kat4c. The latter mAb recognizes the A and B domains of the receptor (Jones et al., 1995) . AO115 and Kat4c bound to various receptor constructs and to the native NA1B-Fc molecule with equal efficiency. For YTH71.3, Leu but not Thr at position 91 supported binding to the same extent as Ile of the native molecule. Similarly at position 34, only Phe could be effectively substituted for Tyr. Finally Q89A and Q89N completely abrogated the antibody binding. Opa-expressing phenotypes of two Nm strains (C751 and MC58) were used to investigate their binding to the modified NA1B-Fc receptors. Three C751 derivatives express-ing distinct Opa proteins (OpaA, OpaB and OpaD) and the MC58 derivative expressing an Opa protein designated OpaX (Virji et al., 1999) were assessed by receptor overlay experiments (Fig. 3) . All the Nm isolates showed reduced binding to the soluble receptor with Y34A and I91A substitutions whilst Q89A affected OpaD binding most significantly confirming previous studies (Virji et al., 1999) . Introduction of a Leu or Thr residue at position 91 was generally less disruptive to Opa interactions. I91T substitution either did not affect binding (OpaB and OpaX) or reduced binding by 50-60% (OpaD and OpaA). I91L substitution had no effect on OpaD or OpaB whilst having opposite effects on OpaA (~50% reduction) and OpaX (twofold increase) binding. Substitutions at position 34 other than Ala also supported binding of some but not all Opa proteins. Y34S substitution was unsuitable for the three Opa proteins of strain C751 but was tolerated by OpaX of MC58. Interestingly, Phe proved to be a more favoured residue in at least two cases with OpaX as well as OpaB binding to Y34F construct at threefold higher levels than the native receptor. Substitutions of Asn at Gln-89 revealed distinct patterns of interactions and binding reduced in the order OpaA > OpaX > OpaB/D (Fig. 3) . Overall, the data suggest that Opa binding to the receptor requires an extended aliphatic chain at position 91 and certain arrangements of the chain may facilitate binding of most Opa proteins (Ile = leu > Thr). At position Α,Β D, E β−strand: (Virji et al., 1999) . Side chains of the residues introduced at positions 34, 89 and 91 in this study are shown in B for comparison 34, the removal of the hydroxyl residue (Y43F) is tolerated or preferred whereas the absence of aromatic ring (Y34S) has an overall deleterious effect on Opa binding. Finally, reduction of the side chain extension (Q89N) reduces binding of three of the four Opa proteins. Comparison of C751 Opa-A, -B and -D in which the differences exist only between their surface exposed loop structures (HV1 and HV2, shown in Fig. 9 ), suggests further that a combination of the two loop structures must be involved in presenting the appropriate binding partners for the distinct residues of the receptor. One non-typable (A950002) and two THi strains (Rd and Eagan) were used in receptor overlay experiments as above (Fig. 4) . Alanine substitutions at Ile-91 confirmed previous results (Virji et al., 2000) . Taken individually, substitutions of Ile-91 for Ala or Thr abrogated binding of Fig. 3 . Relative binding of CEACAM1-Fc constructs to N. meningitidis isolates expressing distinct Opa proteins. Bacterial lysates were dotted on to nitrocellulose and overlaid with NA1B-Fc constructs as indicated. Binding relative to the native receptor was determined by densitometric analysis of immno-blots using NIH Scion Image programme. One hundred per cent binding level is indicated by the horizontal line allowing comparison to native receptor binding. Mean values and SE of > 3 replicates are shown in each case, except Q89A (n = 1) for OpaA, B and X. However, alanine substitution at all three positions confirm previous observations (Virji et al., 1999) . Black inserts in C show the levels of binding of the various receptor constructs to Opaisolate of strain C751; experiments were conducted simultaneously with C751OpaD in the presence of the receptors shown. all three strains. However, substitution I91L reduced binding only of the THi strain Eagan, suggesting a requirement for the extended hydrophobic arm of Ile in Eagan binding. Alanine substitution of Tyr-34 abrogated the binding of Rd and Eagan, whereas no effect on NTHi A950002 was observed. In contrast, Y34F substitution led to an increased binding of all three strains varying from 45 to 80% above native receptor binding levels, whereas Y34S had a differential effect on the three strains tested, ranging from no binding of Rd to a slight increase in the binding of A950002. These data broadly reflect results with neisserial Opa binding. Finally substitutions at Gln-89 show that reduction or abrogation of the side chain (Q89N and Q89A) has no deleterious effect, rather its effect is often that of enhanced binding (Fig. 4) . In summary, Hi strains primarily require Ile-91 to enable receptor targeting. Tyr-34 influences binding of THi and as for Nm, removal of the OH (Y34F) provides a more favourable environment. Gln-89 side chain also limits bacterial interaction and its substitution to shorter side chains (Q89A and Q89N) is more favourable especially for Rd-CEACAM binding. In previous studies, interactions of Hi-aeg strains were shown to be more analogous to Nm (Virji et al., 2000) . The current studies accordingly demonstrated a requirement both for Ile-91 and Tyr-34 for all Hi-aeg (Fig. 5) . In addition, as with some Nm strains, the requirement for hydrophobic and aliphatic chains was partly fulfilled by Leu and to a lesser extent by Thr. Substitutions of residue Tyr-34 with Phe or Ser produced a complex profile. Phe generally created a suitable or better environment but the loss of the aromatic ring was also tolerated (Y34S). At position 89, Q > N substitution enhanced binding dramatically for Ha3. Overall, the three strains tested exhibited similar binding to all receptor constructs that had substitutions at residue Ile-91 as well as those with alanine substitution at residue Tyr-34 (Fig. 5) . However, strains exhibited differences in binding to the receptors with substitutions Y34F and Q89N demonstrating the structural diversity of the CEACAM1-binding ligands of distinct Hi-aeg strains also. From the above data, it is clear that receptor constructs with Q89N substitution have the capacity to increase the binding of some Hi over that of Nm. In previous investigations, we have shown that certain isolates of the two bacteria are capable of competing with the native receptor and that Nm C751 derivatives can displace Ha3 as well as Eagan from the receptor (Virji et al., 2000) . Here we investigated how changes in receptor structure at position 89 may influence this phenomenon. We used receptor dot blot overlay in the presence of varying amounts of the soluble receptor constructs to estimate the ability of C751OpaB and Ha3 to selectively adsorb the receptor. Consistent with the lower affinity of Ha3 compared with C751OpaB/D for CEACAM1 (Virji et al., 2000) , the proportion of the native CEACAM1 that interacted with Ha3 compared with C751OpaB was lower and declined further at limiting concentrations of the receptor (Fig. 6A ). In contrast, Ha3 and C751OpaB had similar affinities for the receptor when the Q89N construct was present in excess (> 0.13 mg ml -1 ; Fig. 6B ). Further, at limiting concentrations of Q89N, the greater affinity of Ha3 was apparent and bound this construct threefold more than the Nm derivative (Fig. 6B ). The data demonstrate the potential influence of structural modulations at the bacterial binding site in changing the colonization profile of the target tissue. To assess whether CEACAM mutations that either significantly increase or decrease bacterial binding to the soluble constructs also affect bacterial binding to cellexpressed receptor constructs, we analysed COS cells transiently transfected with CEACAM1-4L containing Q89A, Q89N and I91A substitutions. Initially, binding of the anti-CEACAM antibodies to transfected CHO cells was examined and their binding was as observed for the soluble receptors (Fig. 7) . However, bacterially expressed ligands may overcome decreased affinity for specific mutant proteins due to multiple ligand-receptor engagement at the target cell surface. Examination of adherent bacteria and levels of receptor expression by microscopy revealed that bacteria bound to all transfectants other than sham-transfected cells (Fig. 8) . However, receptor expression levels had to be considerably high in the case of I91A for significant bacterial numbers to bind the transfectants. Whereas for Q89A and Q89N, bacteria could bind to cells with barely detectable levels of receptors (Fig. 8) . Using CHO cells expressing full length CC1 or CC1(Q89N), the ability of the soluble native (CC1-Fc) or CC1(Q89N)-Fc to inhibit bacterial binding was assessed using C751 OpaB bacteria (Fig. 9 ). Data show the following: (i) Binding to CHO cells as reported previously (Virji et al., 1996b) was observed only with Opa+ and not Opa-bacteria and no binding was seen with untransfected cells (not shown), (ii) Despite the low levels of binding of the isolate to the soluble CC1 construct carrying the Q89N mutation in the dot immunoblot analysis (Fig. 3B) , bacterial binding to the receptor expressed on the cells was clearly visible (Fig. 9D) , (iii) As in the case of THi Rd described above, association was only significantly high with cells expressing high levels of the receptor (assessed in parallel experiments, not shown), (iv) CC1-Fc could compete with bacterial binding to the homologous native structure expressed on CHO cells (Fig. 9B) , confirming previous results (Virji et al., 1996b) , (v) In comparison, CC1-Fc could more efficiently compete out C751 OpaB binding to the CC1(Q89N) receptor expressed on CHO cells (Fig. 9E) , consistent with its higher affinity for the isolate than CC1(Q89N)-Fc (Fig. 3B) , (vi) CC1(Q89N)-Fc construct was largely ineffective in competing with either cell-expressed receptor ( Fig. 9C and F) ent bacterial numbers are thus also low (cf. Fig. 9D and F: lack of low level binding in F compared with D). The data emphasize the importance of cell presentation of receptor in addition to receptor structure in bacterial interactions. The final outcome of this depends on the balance between the effects exerted by residue substitutions and receptor density. Both can affect functional affinity of bacteria-host interactions. To assess the relative importance and the degree of influence of each substitution on bacterial interactions, degrees of receptor recognition were assigned several categories as shown in Fig. 10A . The data show that substitution I91A has a profound effect on the interactions of all strains and whilst the side chain of Ile-91 is best suited to all, Leu can effectively substitute for Ile for several strains, whilst the polar Thr is less well tolerated. However, some Nm Opa proteins can also tolerate I91T substitution. Tyr-34 is also required in most cases. Interestingly, Phe-34 is preferred in general at this position over the native Tyr. Substitutions at Gln-89 have a delete- For the native receptor, expression levels and bacterial binding correlated more frequently than for the mutated molecules. In the case of substitutions at position 89, bacterial binding could be seen at very low levels of receptor expression (arrows in Q89A and Q89N panels). The reverse was the case with I91A, where even at high receptor levels (arrowhead, left panel), bacterial binding (arrowhead middle panel) was relatively low. In sampling of 100 cells, c. 50% exhibited this phenomenon. Note that the receptor recognition by Kat4C was not affected by mutations in the N-domain as determined by immunoblotting or by immunofluorescence microscopy (Fig. 7) . rious effect on binding of some Nm Opa derivatives and the smaller Q89N apparently is better suited overall for binding by Hi and especially Hi-aeg strains. The data emphasize the extensive inter-and intraspecies ligand variations in this receptor targeting, perhaps with greater interstrain differences in Nm. Structural aspects of bacterial ligands that affect the receptor recognition can be considered for strain C751 Opa proteins whose structures are known (Fig. 9B ) (Hobbs et al., 1994) . The hypervariable loops HV1 and HV2 of Opa proteins have been implicated in the binding of CEACAMs (Virji et al., 1999; Bos et al., 2002; de Jonge et al., 2003) . Given that the combinations of HV regions of OpaA, B and D of strain C751 provide three distinct combinations (Fig. 10C) , this is consistent with the distinct patterns of targeting of the variant receptor molecules observed in the study. However, the proteins must all contain sufficient similarity to bind the hydrophobic region around I91. The N-terminal domains of the cell-expressed CEA family of molecules are highly homologous and the majority of CEACAMs are targeted by one or more pathogenic neisserial adhesins belonging to the Opa family of proteins (Virji et al., 1996a; Chen et al., 1997; Gray-Owen et al., 1997) . Despite the structural variability, all Opa proteins target a common site on the receptors whose centre of binding appears to be Ile-91 on the CFG face of the N-domains. Ile-91 is conserved throughout the CEA members. Most of the other important residues on CEACAM1 identified by alanine scanning mutagenesis located in a close proximity of Ile-91 (Virji et al., 1999) . Precisely how variant Opa proteins can bind to a common receptor site is not entirely clear but a complementary set of sequences of more than one variable domains of Opa proteins may be involved. This variability of the ligands may determine the binding preference for distinct members of the CEA family which represents one mechanism that may determine tissue tropism (Virji et al., 1999; Bos et al., 2002; Gray-Owen, 2003; Hauck and Meyer, 2003; de Jonge et al., 2003) . Diverse strains of typable and NTHi lineages including the biogroup aegyptius also bind to the CFG face of the N-domain of CEACAM1 (Virji et al., 2000) . In our previous analysis of receptor binding, THi isolates were shown to behave similarly: each strain tested having a primary requirement for Ile-91 and in addition, was affected by Y34A and Q44A substitutions. In the current study also, the THi strain Rd and Eagan demonstrate similar overall binding patterns. However, in all cases, intraspecies dif- Fig. 9 . Binding of N. meningitidis isolate C751OpaB to cell-expressed receptors: competition between cell-expressed and soluble receptors. Bacterial binding to cell-expressed CC1 or Q89N construct was detected using anti-Nm antisera and TRITC-conjugated secondary antibodies. The interactions of bacteria with the cell-expressed receptors were investigated in the absence (A, D) or presence of competing soluble receptors. Although binding to the soluble receptor Q89N is much lower than the native CC1 in dot-blots (Fig. 3B) , it is relatively high on certain cells (presumably, on those expressing high levels of the receptor) (D). In competitive experiments, the native CC1-Fc (B, E) and the CC1(Q89N)-Fc (C, F) were preincubated at 10 ug ml -1 with bacteria for 15 min, prior to infection of target cells. CC1-Fc inhibited bacterial binding to the homologous receptor significantly (B), and almost abrogated binding to cell-expressed Q89N (E); whereas the soluble Q89N was inefficient at inhibiting bacterial binding to CHO-CC1 (C). However, a level of homologous inhibition was apparent when examining the adhesion of bacteria to cells expressing low levels of the receptor (e.g. 'peppered' areas shown in D are less evident in F). CHO-CC1(Q89N) ferences in response to different amino acid substitutions were apparent. One consistency between all species and strains tested was the dramatic loss of binding following I91A substitution which, as observed previously, appears to be central in an overlapping bacterial binding footprint on CEACAM1 (Virji et al., 1999; . A similar situation occurs on the mouse CEACAM1a (MHVR1a) N-terminal domain, in which Ile-41 appears to engage with the mouse hepatitis virus spike protein. Replacement of Ile-41 by Thr in the MHVR1b allele reduces virus binding significantly (Tan et al., 2002) . Cell surface receptor interactions with their ligands often involve hydrophobic contact points which provides the major binding energy. Hydrophobic residues surrounding these contribute to the specificity of binding (Clackson and Wells, 1995; Kwong et al., 1998; Kim et al., 2001) . Thus in murine CEACAM1, the protruding Ile-41, which is surrounded by a number of surface exposed charged residues, e.g. Asp-42, Glu-44, Arg-47, Asp89, Glu-93 and Arg-97, might form such a binding area (Tan et al., 2002) . Accordingly from current studies also, mutations introduced at position 91 in the human receptor support the requirement for a hydrophobic pocket at this site. In addition, an extended aliphatic chain is preferred as Ala disrupted binding of all bacteria and the polar Thr reduced binding of all Hi and several Nm strains. Only I91L was more frequently tolerated. This binding pocket is flanked by several polar residues (Fig. 1 ) whose contribution to bacterial ligand binding is apparent and variable (Virji et al., 1999; . The importance of the Ile-91 and the surrounding residues on human CEACAM1 in bacterial binding is also supported by the observation that the mAb YTH71.3 directed against the N-domain also requires several residues in common with bacteria and blocks binding of all CEACAM1-binding bacteria we have investigated (Virji et al., 1999; Hill and Virji, 2003) . The murine N-domain strand arrangement derived form crystal structure depicts the CC′ loop to assume a convoluted conformation. The previous model of human CEACAM1 contained a flat CC′ loop like the Ig-folds on Fig. 10 . Binding of bacterial ligands to CEACAM1-Fc constructs as determined in the current study. A. Receptor recognition categories are colour coded according to percent binding relative to the native molecule*. B and C. Diagrams showing CEACAM-binding meningococcal ligand structures. As only the Opa protein structure has been analysed in details so far with respect to CEACAM binding (de Jonge et al., 2003) , for clarity and ease of discussion, a 2D Opa protein structure (B) and relationship of the three variable domains of strain C751 Opa proteins are shown (C). SV, semivariable domains of strain C751 Opa-A, -B and -D are identical as are the hypervariable structures HV1 of OpaA and OpaB and the HV2 of OpaB and OpaD. which it was based. This resulted in Val-39 and Gly-41 being located at a distance from the binding focus Ile-91. Both Val-39 and Gly-41 have been implicated in bacterial binding from alanine scanning and homologue scanning mutagenesis (Bos et al., 1999; Virji et al., 1999) . The remodelling shown here of CEACAM1 produces the convoluted structure of the CC′ loop relocating Val-39 close to Ile-91. Further, the aromatic ring of Tyr-34 suggested to be required to maintain the convoluted structure of the CC′ loop (Tan et al., 2002) when substituted with Phe almost always supported bacterial binding. Indeed, Y34F provides a better environment for most bacterial strains tested. On the other hand, Y34A frequently abrogated receptor recognition. However and surprisingly, Y34S is tolerated by a substantial number of bacterial ligands. Whether this is due to the flexible variable loop domains of the bacterial ligands which may produce an induced fit around the receptor needs consideration. Interestingly, Tyr-34 is conserved in the majority of the human CEACAMs with the exception of CEACAM4 which contains His at this site. However, the importance of Tyr-34 in human CEACAM1 maintaining the three dimensional structure requires human receptor crystallographic data. Substitutions Y34F and Q89N also produced interesting data from the point of view of pathogenesis. Whilst Q89A and Q89N appear to affect some Opa proteins by reducing receptor recognition, suggesting its potential contribution in determining tissue tropism (Virji et al., 1999) , Q89N substitution occasionally caused dramatic increase in bacterial adhesion, especially of Hi isolates. Y34F, as observed above, increased binding of strains within all species examined. As it is possible that multiple receptors presented on the target cells may overcome the reduced binding affinity of mutated receptors, we examined strain Rd binding to I91A and Q89A/N-substituted receptors expressed in transiently transfected COS cells. Whilst the latter receptors were targeted on cells with low levels of receptor expression, only a proportion (c. 50%) of cells with very high levels of I91A receptors had significant numbers of bacteria attached. Thus point mutations of CEACAMs can both decrease and increase bacterial load and additional factors that dictate bacterial binding include receptor levels on the target cells. The ligands of Neisseria (i.e. Opa proteins) and of Hi so far identified (i.e. P5 proteins) share similar beta-barrel structure with surface-exposed variable loops. The regions of P5 that may engage with the receptor have not been identified but those of Opa proteins were studied by mutagenesis of strain H44/76 (de Jonge et al., 2003) . This strain is related to strain MC58, one of the strains used in the current study. The studies of de Jonge et al. implicated G ¥ (I/V/l) ¥ (S/E/Q) as the key motifs of HV2 regions (Fig. 9 ) of meningococcal Opa proteins in receptor targeting. Together with this, an 99 ELK motif of the Opa HV1 region might be involved in the three dimensional presentation of the receptor-engaging residues of the bacterial ligand (de Jonge et al., 2003) . Within the strain C751, OpaB and OpaD proteins contain the motif GxLxS at positions 172-176 and 167-171, respectively, whereas Opa A contains a 168 PxIxN motif. In the HV1 region, OpaA and B contain 99 DLK whereas OpaD contains EDK. Studies in our laboratories are in progress to assess the precise variant C751 Opa and receptor residue pairs involved in mutual recognition. Polymorphisms that affect host susceptibility may be found at various sites in the genes encoding host receptors targeted by pathogens and may result in loss or gain of receptor-associated functions. Some SNPs may lead to multiple and diverse downstream effects, e.g. altered transcriptional response and manifestation of disease (Sakuntabhai et al., 2005) . Extracellular domain polymorphisms may have a more direct effect via altered binding of pathogen ligands to their receptors. SNPs of the innate immune system especially those affecting pathogen associated pattern recognition receptors and cytokines have been studied extensively. Changes such as Asp299Gly and Thr399Ile in the extracellular domains of LPS-binding Toll-like receptor 4 (TLR4) have been implicated in increased risk to bacterial infections (Schroder and Schumann, 2005) . These SNPs have also been associated with severe respiratory syncytial virus (RSV) bronchiolitis in infants. In this case, altered interaction of the viral fusion (F) protein, implicated as a ligand for TLR4, is regarded as the primary mechanism (Tal et al., 2004; Schroder and Schumann, 2005) . However, these TLR4 SNPs could not be correlated with meningococcal disease (Read et al., 2001) . In contrast, rare SNPs were found more commonly in the TLR4 genes of patients with meningococcal disease (Smirnova et al., 2003) . This supports the notion that rare rather than common variants of TLR4 may be associated with infectious disease susceptibility. Studies presented here suggest some possible polymorphisms that can increase bacterial load. Several SNPs in CEACAMs have been identified and are listed in the NCBI SNP database and three have been identified in the N-domain of CEACAM1 including one at Gln-89. In this case, a Gln-89 to His substitution is observed. Such a residue difference occurs within the members of CEACAM family. For example, CEA contains H at position 89 (Fig. 2) . This is the only major difference between CC1 and CEA that could affect bacterial binding (Fig. 2) . As such, the mutation Q89H in CC1 would be expected to produce CEA-like binding pattern and could affect tropism of the bacteria as observed for CEA (Virji et al., 1999; de Jonge et al., 2003) . Further studies are required to assess whether other CEACAM polymorphisms, for example, substitutions such as Y34F and Q89N occur in human Molecular analysis of bacterial ligand-CEACAM1 interactions 341 populations and their frequencies in susceptible populations. SNP substitutions may also change the colonization profile of the nasopharynx, because in our competition studies we could demonstrate that increased binding afforded by Q89N to Hi-aeg isolate Ha3 increases the binding of Ha3 such that it out-competes Nm isolate C751OpaB in an in vitro competition assay for this receptor. The situation with the native receptor was the reverse. As shown in Figs 8 and 9, additional factors that affect bacterial binding to cell-expressed receptor include receptor density. Whilst certain residue substitutions, e.g. I91A, reduce functional affinity of bacterial interactions, high receptor densities increase such affinity. The final outcome must depend on the interplay between these two parameters. In recent studies, the role of receptor density on enhancement of bacterial attachment and invasion have been evaluated in detail (Bradley et al., 2005; Rowe et al., 2006) . It would be interesting to analyse bacterial invasion in cell lines expressing variant CEACAM1 carrying the above mutations by employing cell lines in which the receptor expression levels can be controlled, which are under development. Besides receptor polymorphisms, several other scenarios may lead to increased bacterial ligand binding to CEACAMs. Both Nm and Hi CEACAM-binding ligands (Opa and P5) are known to undergo antigenic variation. Thus, in any population, antigenic/structural variants are present and may be selected for during the course of host colonization and subsequent pathogenesis (Virji et al., 1996a; Duim et al., 1997; Meyers et al., 2003) . The receptor repertoire and subtypes may select bacteria capable of binding with high affinity. As mentioned above, upregulation of receptor expression on target cells may also increase bacterial binding affinity. In such cases, high affinity interactions result in cellular invasion, whereas lower affinity or load of bacteria may not proceed beyond surface adhesion (Tran Van Nhieu and Isberg, 1993; Bradley et al., 2005) . The expression of CEACAMs on normal epithelia of the respiratory tract has been reported, which would allow bacterial attachment and possible subsequent penetration into these tissues (Tsutsumi et al., 1990; Virji, 2001) . Following exposure to cytokines such as IFN-g or TNF-a, CEACAM expression by colonic carcinoma cells has been shown to increase (Fahlgren et al., 2003) . In addition, certain viral infections have also been shown to upregulate CEACAM1 expression in several epithelial cell lines (Avadhanula et al., 2006) . Increased cytokine levels following viral infection could lead to increased CEACAM expression and bacterial association with respiratory epithelia and subsequent invasion of deeper tissue by these organisms. Such a situation may explain the epidemiological association of increased incidence of Nm and Hi infections following certain viral infections (Cartwright et al., 1991; Takala et al., 1993) . In summary, little is known about why certain people are more susceptible to infection by some of the frequent colonizers of the human nasopharynx. Interestingly, opportunistic pathogens such as Nm and Hi as well as M. catarrhalis (not investigated here) target CEACAM1. As specific substitutions such as Y > F and Q > N produce more favourable targets for distinct mucosal isolates, it is possible that occurrence of such receptor polymorphisms in the human population could lead to greater bacterial binding thus increasing the chances of cellular invasion. Given the colonization rate of these organisms (generally > 10% of the population) and the frequency of invasive infection (up to 3:100 000 population), a combination of events may be required to increase host susceptibility. Inflammatory conditions that increase receptor density in populations carrying specific polymorphisms could provide the worst scenario. The sequence of mature human CEACAM1 N-domain was aligned with the corresponding domain of murine CEACAM1, giving a gapless alignment for residues 1-109 with a residue identity of 42%. The crystal structure co-ordinates of mouse CEACAM1 (residues 1-109) were taken from the structure file containing domains 1 and 2 (PDB code 1L6Z) and a homology model of human CEACAM1 N-domain was built using standard methods. Final refinement of the model was performed by soaking it with a 5 Å thick layer of water and energy minimizing while constraining the backbone atoms to their original positions in the template structure. The final round of minimization was for 2000 conjugate-gradient steps, constraining the backbone heavy atoms with a force constant of 0.5 kcal/Å. A stereochemical analysis of the structure was performed using Procheck and found to be of similar quality to the template crystal structure. Production of mutants was carried out by site-directed mutagenesis of the pIG construct containing the DNA encoding the CEACAM1 NA1B domains described previously (Watt et al., 1994) . The oligonucleotide primers used to create amino acid substitutions at positions 91, 34 and 89 of the N-domain are shown in Table 1 . Some primer sequences have been published previously (Watt et al., 1994) . For introducing mutations, CEACAM1 was amplified by polymerase chain reaction from the pIG-NA1B construct using either the common forward primer and a reverse primer containing the desired mutation, or a complementary forward primer containing the mutation and a common reverse primer. CEACAM1 with the appropriate mutation was amplified using the common forward and common reverse primers. The gene was then cloned into pIG using the restriction sites HindIII and EcoRI. Chimeric soluble receptor proteins containing the appropriate amino acid substitutions were prepared as previously described by transient transfection of COS cells (Teixeira et al., 1994; Virji et al., 1999) . The CEACAM1-Q89A-Fc used in overlay experiments was kindly donated by Dr S. Watt (Virji et al., 1999; Watt et al., 2001) . The strains used in this study have been described previously (Virji et al., 1999; . Nm strain C751 (serogroup A) variants used were C751OpaA, C751OpaB and C751OpaD. The strain MC58 (serogroup B) variant used expressed an Opa previously designated OpaX, which is encoded by the opaB locus. Opa -C751 isolate, which has been shown not to bind to CHO-CC1 (Virji, 1999) and RdCC-, a derivative of THi Rd, known not to bind to CHO-CC1 (M. Virji, unpublished) were used as controls. THi strain Rd is an acapsulate serotype d isolate, Eagan is serotype b isolate and A950002 is a NTHi strain. Hi-aeg strains Ha3, Ha30 and F2087 are all conjunctiva isolates. Nm was grown on brainheart infusion (BHI) agar supplemented with 10% heated horse blood (HBHI). Hi strains were grown on HBHI agar further supplemented with Levinthal base (10 mg ml -1 each of NAD and haemin). All strains were cultured at 37°C in 5% CO2. COS-1 cells (African green monkey kidney cells) used for transient transfection were cultured in Dulbecco's modified Eagle's medium (DMEM) containing 2-10% heat-inactivated Foetal Calf Serum (FCS, Gibco™), 2 mM glutamine, 50 mg ml -1 penicillin and 50 mg ml -1 streptomycin in a humidified atmosphere of 5% CO2 at 37°C. Antibody binding to CEACAM1 constructs. NA1B-Fc proteins were dotted at 0.2 mg ml -1 on to nitrocellulose and non-specific binding sites blocked using 3% (w/v) BSA in Dulbecco's PBS containing 0.05% Tween-20 (PBST) for 1 h at room temperature. Receptor was detected using the following antibodies, rabbit polyclonal AO115, rat monoclonal YTH71.3 both directed against the N-domain and mouse monoclonal Kat4c which recognizes the A and B domains (Jones et al., 1995) . Bound antibody was subsequently detected using an appropriate secondary antibody conjugated to alkaline phosphatase and developed using nitroblue tetrazolium and 5-bromo-4-chloro-3-indolylphosphate. Bacterial interactions with receptor constructs. Bacterial lysates (c. 2 ¥ 10 7 bacteria) were applied to nitrocellulose strips, air-dried and non-specific binding sites blocked using 3% BSA in PBST for 1 h at room temperature. Strips were overlaid with either native or mutated soluble NA1B-Fc diluted in 1% BSA in PBST at required concentrations for 1 h at room temperature. In most experiments, excess (1-3 mg ml -1 ) of the receptor was used. In competition studies, a range of concentrations (0.008-0.5 mg ml -1 ) was employed. Following washing to remove unbound NA1B-Fc, receptor binding was detected using anti-human-Fc alkaline phosphatase conjugate and substrates as described above. For quantification, densitometric analyses of the developed immunoblots were carried out using NIH Scion Image software. In most cases, multiple estimations were carried out and means and SE of each determination have been reported. Site directed mutagenesis of CEACAM1-4L receptor gene was performed using the QuickChange® Site Directed Mutagenesis Kit (Stratagene, La Jolla, CA, USA) according to the manufacturer's instructions. Primers (Table 1) were used to introduce the desired mutations into the pRc/CMV-CEACAM1-4L construct (kindly provided by Professor Wolfgang Zimmermann). Following sequencing to ensure the desired substitution had been obtained, the pRc/CMV-CEACAM1-4L construct was transiently transfected into COS-1 cells for functional analysis using DEAE dextran method described previously (Teixeira et al., 1994; Virji et al., 1996a) .
85
Immune reconstitution inflammatory syndrome (IRIS): review of common infectious manifestations and treatment options
The immune reconstitution inflammatory syndrome (IRIS) in HIV-infected patients initiating antiretroviral therapy (ART) results from restored immunity to specific infectious or non-infectious antigens. A paradoxical clinical worsening of a known condition or the appearance of a new condition after initiating therapy characterizes the syndrome. Potential mechanisms for the syndrome include a partial recovery of the immune system or exuberant host immunological responses to antigenic stimuli. The overall incidence of IRIS is unknown, but is dependent on the population studied and its underlying opportunistic infectious burden. The infectious pathogens most frequently implicated in the syndrome are mycobacteria, varicella zoster, herpesviruses, and cytomegalovirus (CMV). No single treatment option exists and depends on the underlying infectious agent and its clinical presentation. Prospective cohort studies addressing the optimal screening and treatment of opportunistic infections in patients eligible for ART are currently being conducted. These studies will provide evidence for the development of treatment guidelines in order to reduce the burden of IRIS. We review the available literature on the pathogenesis and epidemiology of IRIS, and present treatment options for the more common infectious manifestations of this diverse syndrome and for manifestations associated with a high morbidity.
Since its introduction, ART has led to significant declines in AIDS-associated morbidity and mortality [1] . These benefits are, in part, a result of partial recovery of the immune system, manifested by increases in CD4 + T-lymphocyte counts and decreases in plasma HIV-1 viral loads [2] . After initiation of ART, opportunistic infections (OI) and other HIV-related events still occur secondary to a delayed recovery of adequate immunity [3] . Some patients initiating ART experience unique symptoms during immune system recovery. In these patients, clinical deterioration occurs despite increased CD4 + Tlymphocyte counts and decreased plasma HIV-1 viral loads [4] . This clinical deterioration is a result of an inflammatory response or "dysregulation" of the immune system to both intact subclinical pathogens and residual antigens [5] [6] [7] [8] [9] . Resulting clinical manifestations of this syndrome are diverse and depend on the infectious or noninfectious agent involved. These manifestations include mycobacterial-induced lymphadenitis [5] , paradoxical tuberculosis reactions [6, 7, 10, 11] , worsening of progressive multifocal leukoencephalopathy (PML) [12] , recurrence of cryptococcosis and Pneumocystis jirovecii pneumonia (PCP) [8, [13] [14] [15] [16] , Cytomegalovirus (CMV) retinitis [17] , shingles [18] , and viral hepatitis [19] , as well as noninfectious phenomena [20] . Because clinical deterioration occurs during immune recovery, this phenomenon has been described as immune restoration disease (IRD), immune reconstitution syndrome (IRS), and paradoxical reactions. Given the role of the host inflammatory response in this syndrome, the term immune reconstitution inflammatory syndrome (IRIS) has been proposed [21] and has become the most widely used and accepted term to describe the clinical entity. Possible infectious and noninfectious etiologies of IRIS are summarized in Table 1 . To date, no prospective therapeutic trials concerning the management of IRIS have been conducted. All evidence regarding the management of IRIS in the literature relates to case reports and small case series reporting on management practice. This does not provide reliable evidence regarding either the safety or efficacy of these approaches, but merely guidance regarding the practice of others in managing this difficult condition. In severe cases where the discontinuation of ART is a possibility, the potential disadvantages of therapy cessation, such as the development of viral resistance or AIDS progression, should be considered. Despite numerous descriptions of the manifestations of IRIS, its pathogenesis remains largely speculative. Current theories concerning the pathogenesis of the syndrome involve a combination of underlying antigenic burden, the degree of immune restoration following HAART, and host genetic susceptibility. These pathogenic mechanisms may interact and likely depend on the underlying burden of infectious or noninfectious agent. Whether elicited by an infectious or noninfectious agent, the presence of an antigenic stimulus for development of the syndrome appears necessary. This antigenic stimulus can be intact, "clinically silent" organisms or dead or dying organisms and their residual antigens. IRIS that occurs as a result of "unmasking" of clinically silent infection is characterized by atypical exuberant inflammation and/or an accelerated clinical presentation suggesting a restoration of antigen-specific immunity. These characteristics differentiate IRIS from incident opportunistic infections that occur on ART as a result of delayed adequate immunity. Examples of IRIS in response to intact organisms include, but are not limited to, the unmasking of latent cryptococcal infection [22] and infection with Mycobacterium avium complex (MAC) [4, 5, 23, 24] . The most frequently reported IRIS symptoms in response to previously treated or partially treated infections include reports of clinical worsening and recurrence of clinical manifestations of Mycobacterium tuberculosis (TB) and cryptococcal meningitis following initiation of ART [6, 7, 10, 13, 16, [25] [26] [27] [28] . In noninfectious causes of IRIS, autoimmunity to innate antigens plays a likely role in the syndrome. Examples include exacerbation of rheumatoid arthritis and other autoimmune diseases [29] . Given the role of this antigenic stimulus, the frequency and manifestations of IRIS in a given population may be determined by the prevalence of opportunistic and non-opportunistic infections to initiation of ART. The mechanism receiving the most attention involves the theory that the syndrome is precipitated by the degree of immune restoration following ART. In assessing this theory, investigators have examined the association between CD4 cell counts and viral loads and the risk of IRIS. Some studies suggest differences in the baseline CD4 profiles or quantitative viral load at ART initiation or their rate of change during HAART between IRIS and non-IRIS patients [4, [30] [31] [32] [33] [34] , while other studies demonstrate only trends or no significant difference between IRIS and non-IRIS patients [7, 35] . These immunological differences between groups have been difficult to verify due to small numbers of IRIS cases and lack of control groups. An alternative immunological mechanism may involve qualitative changes in lymphocyte function or lymphocyte phenotypic expression. For instance, following ART an increase in memory CD4 cell types is observed [36] possibly as a result of redistribution from peripheral lymphoid tissue [37] . This CD4 phenotype is primed to recognize previous antigenic stimuli, and thus may be responsible for manifestations of IRIS seen soon after ART initiation. After this redistribution, naïve T cells increase and are thought to be responsible for the later quantitative increase in CD4 cell counts [38] . These data suggest IRIS may be due to a combination of both quantitative restoration of immunity as well as qualitative function and phenotypic expression observed soon after the initiation of ART. The third purported pathogenic mechanism for IRIS involves host genetic susceptibility to an exuberant immune response to the infectious or noninfectious anti-genic stimulus upon immune restoration. Although evidence is limited, carriage of specific HLA alleles suggest associations with the development of IRIS and specific pathogens [39] . Increased levels of interleukin-6 (IL-6) in IRIS patients may explain the exuberant Th1 response to mycobacterial antigens in subjects with clinical IRIS [9, 40] . Such genetic predispositions may partially explain why manifestations of IRIS differ in patients with similar antigenic burden and immunological responses to ART. Despite numerous descriptions of the infectious and noninfectious causes of IRIS, the overall incidence of the syndrome itself remains largely unknown. Studies to date are often retrospective and focus on specific manifestations of IRIS, such as tuberculosis-associated IRIS (TB-IRIS). In a large retrospective analysis examining all forms of IRIS, 33/132 (25%) of patients exhibited one or more disease episodes after initiation of ART [4] . Other cohort analyses examining all manifestations of IRIS estimate that 17-23% of patients initiating ART will develop the syndrome [32] [33] [34] . Another large retrospective study reported 32% of patients with M. tuberculosis, M. avium complex, or Cryptococcus neoformans coinfection developed IRIS after initiating ART. Risk factors identified for the development of IRIS in one cohort included male sex, a shorter interval between initiating treatment for OI and starting ART, a rapid fall in HIV-1 RNA after ART, and being ART-naïve at the time of OI diagnosis [31] . Other significant predictors have also included younger age, a lower baseline CD4 cell percentage, a lower CD4 cell count at ART initiation, and a lower CD4 to CD8 cell ratio at baseline [4, 32] . It should be noted cohorts differ substantially in study populations and the type of IRIS (i.e. TB-IRIS only) examined, making conclusions regarding risk factors for IRIS difficult. Clinical factors associated with the development of IRIS are presented in Table 2 . Case reports describing different clinical manifestations of IRIS continue to appear, expanding the clinical spectrum of the syndrome. Because the definition of IRIS is one of clinical suspicion and disease-specific criteria have yet to be developed, determining the true incidence will be difficult. Taken together, these studies suggest IRIS may affect a substantial proportion of HIV patients initiating ART. Future epidemiologic and genetic studies conducted within diverse cohorts will be important in determining the importance of host susceptibility and underlying opportunistic infections on the risk of developing IRIS. In order to aid clinicians in the management of IRIS, we review the epidemiology, clinical features, and treatment options for the common infectious manifestations of IRIS. Additionally, manifestations associated with significant morbidity and mortality, such as CMV-associated immune recovery vitritis (IRV) or immune recovery uveitis (IRU), are also reviewed. Treatment options and their evidence are presented. Until disease specific guidelines are developed for IRIS, therapy should be based on exist- [4, 6, 7, 10, 11, 26, 30-32, 41, 43, 45] Rheumatoid arthritis [29] Systemic lupus erythematosus (SLE) [91] Graves disease [92] , Autoimmune thyroid disease [93] Mycobacterium avium complex [4, 5, 23, 31, [94] [95] [96] Sarcoidosis & granulomatous reactions [20, 97] Other mycobacteria [4, 56, 57, 98, 99] Tattoo ink [100] Cytomegalovirus [4, 33, 61, 63] AIDS-related lymphoma [ [112] Molluscum contagiosum & genital warts [32] Sinusitis [113] Folliculitis [114, 115] ing evidence and individualized according to the severity of presentation. Mycobacterium tuberculosis (TB) is among the most frequently reported pathogen associated with IRIS. Narita et al performed the first prospective study to evaluate the incidence of paradoxical responses in patients on TB therapy and subsequently initiated on ART. Of 33 HIV/TB coinfected patients undergoing dual therapy, 12 (36%) developed paradoxical symptoms [7] . The frequency of symptoms in this group were greater than those observed in HIV-infected controls receiving TB therapy alone, supporting the role of an exaggerated immune system response in the pathogenesis of the syndrome. Retrospective studies corroborate the finding that a significant proportion of HIV/TB coinfected patients undergoing HAART have symptoms consistent with IRIS, with estimates ranging from 7-45% [10,26, 30, 35, [41] [42] [43] . The association between a shorter delay between TB treatment initiation and ART initiation is an area of debate. While some investigators have found no difference in time from TB therapy to initiation of ART between IRIS and non-IRIS subjects [30] , others have reported a significant differences between groups [31, 35] . In general, IRIS occurred in subjects initiated on ART within two months of TB therapy initiation [35] . Based on these and other data, a decision analysis on ART initiation timing in TB patients found the highest rates of IRIS occurred in patients initiated on ART within two months of TB therapy initiation [44] . However, withholding or deferring ART until two to six months of TB therapy was associated with higher mortality in scenarios where IRIS-related mortality was less than 4.6%. Future reports from large, prospective observational cohorts may aid in resolving this difficult issue. Although consisting primarily of case reports [45, 46] , TB-IRIS affecting the central nervous system (CNS) poses a unique problem. As the availability of ART increases in endemic countries, the incidence of CNS TB-IRIS may increase. Thus, clinicians should be vigilant in its diagnosis. The commonest clinical manifestations of TB-IRIS are fever, lymphadenopathy and worsening respiratory symptoms [47] . Pulmonary disorders, such as new pulmonary infiltrates, mediastinal lymphadenopathy, and pleural effusions are also common [7] . Extrapulmonary presentations are also possible, including disseminated tuberculosis with associated acute renal failure [6] , systemic inflammatory responses (SIRS) [48] , and intracranial tuberculomas [45] . Pulmonary TB-IRIS can be diagnosed by transient worsening of chest radiographs, especially if old radiographs are available for comparison. Other symptoms are nonspecific, and include persistent fever, weight loss, and worsening respiratory symptoms. Abdominal TB-IRIS can present with nonspecific abdominal pain and obstructive jaundice. In most studies, TB-IRIS occurs within two months of ART initiation [6, 7, 10, 11, 25, 35, 45, 48] . Among 43 cases of MTB-associated IRIS, the median onset of IRIS was 12-15 days (range 2-114 days), with only four of these cases occurring more than four weeks after the initiation of antiretroviral therapy [7, 10, 25, 26, 30] . These studies suggest the onset of mycobacterial-associated IRIS is relatively soon after initiation of ART, and clinicians should maintain a high level of vigilance during this period. Paradoxical CNS TB reactions are well described in HIVnegative patients, and include expanding intracranial tuberculomas, tuberculous meningitis, and spinal cord lesions [49] [50] [51] . TB-associated CNS IRIS has also been reported in HIV-positive patients [45, 46, 52] . Compared to non-CNS TB-IRIS, symptoms tend to occur later, usually 5-10 months after ART initiation [45, 50, 52] . Crump et al [45] described an HIV-seropositive patient in who developed cervical lymphadenopathy after five weeks of ART. Five months later, CNS symptoms associated with an expanding intracranial tuberculoma appeared after initiation of antituberculous therapy. The significant morbidity in this case illustrates the importance of maintaining a high clinical suspicion for the disease, particularly in endemic areas. Treatment for mycobacterial-associated IRIS depends on the presentation and disease severity. Most patients present with non-life threatening presentations which respond to the institution of appropriate antituberculous therapy. However a range of life threatening presentations, such as acute renal failure [6] and acute respiratory distress syndrome (ARDS) [11] , are described and have significant morbidity and mortality. Morbidity and mortality might also be greater in resource-limited settings where limited management options exist. Since the pathogenesis of the syndrome is an inflammatory one, systemic corticosteroids or nonsteroidal anti-inflammatory drugs (NSAIDS) may alleviate symptoms. In studies where therapy for IRIS was mentioned, the use of corticosteroids was variable [7, 24, 25, 31, 41, 43] and anecdotally effective. Therapies ranged from intravenous methylprednisolone 40 mg every 12 hours to prednisone 20-70 mg/day for 5-12 weeks. These practices reflect the lack of evidence from controlled trials for the use of anti-inflammatory agents in IRIS. A randomized, placebo controlled trial examining doses of prednisone 1.5 mg/kg/day for two weeks followed by 0.75 mg/kg/day for two weeks in mild to moderate TB-IRIS is currently underway in South Africa. Until data become available, it is reasonable to administer corticosteroids for severe cases of IRIS such as tracheal compression due to lymphadenopathy, refractory or debilitating lymphadenitis, or severe respiratory symptoms, such as stridor and ARDS. Interruption of ART is rarely necessary but could be considered in life-threatening situations. In HIV-negative patients, adjuvant corticosteroid use in tuberculous meningitis provides evidence of improved survival and decreased neurologic sequelae over standard therapy alone [53, 54] . Once other infectious etiologies, have been excluded, standard antituberculous therapy should be initiated or continued as the clinical situation dictates, and a course of corticosteroid therapy should be considered for CNS TB-IRIS. Continuation of ART is desirable, although its discontinuation may be necessary in unresponsive cases or in those presenting with advanced neurological symptoms. In addition to TB, atypical mycobacteria are also frequently reported as causative pathogens in IRIS. Early observations involving atypical presentations of Mycobacterium avium-intracellulare (MAC) were first noted with zidovudine therapy [55] . Reports of atypical presentations of both Mycobacterium tuberculosis (MTB) and MAC increased in frequency with the introduction of protease inhibitors and ART. In larger cohorts, MAC remains the most frequently reported atypical mycobacterium [4, 5, 24] . Other atypical mycobacteria rarely associated with IRIS are referenced in Table 1 . In general, MAC-associated IRIS typically presents with lymphadenitis, with or without abscess formation and suppuration [5] . Other less common presentations include respiratory failure secondary to acute respiratory distress syndrome (ARDS) [56] , leprosy [57] , pyomyositis with cutaneous abscesses [23], intra-abdominal disease [58] , and involvement of joints, skin, soft tissues, and spine [58, 59] . Several studies have characterized the time of onset of Mycobacterium-associated IRIS. In one study of MAC lymphadenitis, the onset of a febrile illness was the first sign of IRIS and occurred between 6 and 20 days after initiation of antiretroviral therapy [5] . In another study, the median time interval from the start of antiretroviral therapy to the development of mycobacterial lymphadenitis was 17 days (range 7-85 days) [24]. As with TB-IRIS, evidence for treatment of IRIS due to atypical mycobacteria are scarce. Occasionally, surgical excision of profoundly enlarged nodes or debridement of necrotic areas is anecdotally reported [23, 59] . However, healing is often poor leaving large, persistent sinuses. Needle aspiration is another option for enlarged, fluctuant and symptomatic nodes. Otherwise, treatment is similar to TB-IRIS (see Mycobacterium tuberculosis IRIS -Treatment). In the pre-ART era, CMV retinitis, a vision-threatening disease, carried a high annual incidence and was one of the most significant AIDS-associated morbidities [60] . After the introduction of HAART, Jacobson et al described five patients diagnosed with CMV retinitis 4-7 weeks after ART initiation. They speculated that an HAART-induced inflammatory response may be responsible for unmasking a subclinical infection [17] . In addition to classical CMV retinitis, ART led to new clinical manifestations of the infection, termed immune recovery vitritis (IRV) or immune recovery uveitis (IRU), in patients previously diagnosed with inactive AIDS-related CMV retinitis [61] . Distinct from the minimal intraocular inflammation of classic CMV retinitis, these manifestations exhibit significant posterior segment ocular inflammation thought to be due to the presence of residual CMV antigens or proteins which serve as the antigenic stimulus for the syndrome [62] . Clinical manifestations include vision impairment and floaters. In a retrospective cohort, CMV-related IRIS was common (6/33 of IRIS cases, or 18%) [4] . In prospective cohorts, symptomatic vitritis occurred in 63% (incidence rate 83 per 100 p-yr) of ART responders who carried a previous diagnosis of CMV retinitis but had inactive disease at the onset of antiretroviral therapy. The median time from ART initiation to IRV was (43 weeks) [63] . Another large prospective surveillance study [64] identified 374 patients with a history of CMV retinitis involving 539 eyes. Thirtyone of 176 ART responders (17.6%) were diagnosed with IRU. Male gender, use of ART, higher CD4 cell counts, and involvement of the posterior retinal pole as factors associated with a reduced risk of developing IRU, whereas prior use of intravitreous injections of cidofovir, large retinal lesions, and adequate immune recovery on ART were associated with increased risk. The diagnosis of ocular manifestations of IRIS requires a high level of suspicion. In addition to signs of retinitis, inflammatory symptoms include vitritis, papillitis, and macular edema, resulting in symptoms of loss of visual acuity and floaters in affected eyes. Treatment of IRIS associated CMV retinitis and IRV may involve anti-CMV therapy with gancyclovir or valgancyclovir [17, 65] . However, the occurrence of IRU in patients receiving anti-CMV therapy draws its use into question [64, 66, 67] . The use of systemic corticosteroids has been successful, and IRV may require periocular corticosteroid injections [61, [68] [69] [70] . Due to its significant morbidity and varying temporal presentations, clinicians should maintain a high level of vigilance for ocular manifestations of CMV-associated IRIS. With the introduction of protease inhibitors, increasing rates of herpes zoster were noted in HIV-infected patients. Two studies comparing ART and non-ART patients reported increased incident cases of zoster and rates estimated at 6.2-9.0 cases per 100 person-years, three to five times higher than rates observed in the pre-HAART era [18, 71] . While another study [72] reported no difference in overall incidence between HAART eras (3.2 cases per 100 person-years), the use of HAART was associated with increased odds of developing an incident zoster outbreak (OR = 2.19, 95% confidence interval: 1.49 to 3.20). These studies suggest that ART may play a role in increasing the risk of zoster, which is reflected in large observational IRIS cohorts, where dermatomal varicella zoster comprises 9-40% of IRIS cases [4, 32, 33] . Mean onset of disease from ART initiation was 5 weeks (range 1-17 weeks) [71] , and no cases occurred before 4 weeks of therapy [18] . Both studies identified significant increases in CD8 T cells as a risk factor for developing dermatomal zoster. Although complications such as encephalitis, myelitis, cranial and peripheral nerve palsies, and acute retinal necrosis can occur in immunocompromised HIV patients, the vast majority of patients exhibit typical or atypical dermatomal involvement without dissemination or systemic symptoms [18, 71, 73] . A randomized, controlled trial demonstrated oral acyclovir to be effective for dermatomal zoster in HIV-infected patients, facilitating healing and shortening the time of zoster-associated pain [74] . Its use in cases of varicella zoster IRIS appears to be of clinical benefit [18] . The benefit of corticosteroids in combination with acyclovir in acute varicella zoster has been demonstrated in two large randomized, controlled trials. The combination of corticosteroids and acyclovir decreased healing times, improved acute pain, and quality of life, but did not affect the incidence or duration of postherpetic neuralgia [75, 76] . The incidence of postherpetic neuralgia in immunocompetent individuals does not differ significantly from HIV-infected patients, but increases with increasing patient age [77] . Successful symptomatic management involving opioids, tricyclic antidepressants, gabapentin, and topical lidocaine patches individually or in combination has been shown to be beneficial [78] [79] [80] [81] [82] and should be attempted in HIV patients with postherpetic neuralgia as a complication of herpes zoster IRIS. Accurate incidence of C. neoformans-associated IRIS is unknown. It is infrequently reported in overall IRIS cohorts, and many cases appear as single case reports. A recent study [90] evaluated antifungal combination therapies in the treatment of C. neoformans meningitis in HIV patients. Although significant log reductions in colony forming units were observed with all combinations, substantial numbers of patients remained culture positive 2 weeks after therapy. It may be important to delay ART until CSF sterility can be achieved with effective antifungal combinations such as amphotericin B and flucytosine. However, the exact timing of ART and whether attaining CSF culture sterility is important in avoiding IRIS is unknown. This is illustrated by cases of reactivation cryptococcal meningitis described in four patients who had received at least four weeks of antifungal therapy prior to ART [13, 22, 83] . It is reasonable to administer systemic corticosteroids to alleviate unresponsive inflammatory effects, as anecdotal benefits have been observed in these patients [21, 84] . Furthermore, serial lumbar punctures may be required to manage persistent CSF pressure elevations in these patients [85, 86] . Although continuation of ART has been performed safely [13, 84] , interruption of antiviral therapy may be necessary in severe or unresponsive cases. Other less common infectious etiologies, as well as noninfectious etiologies, are listed in Table 1 . Because these other infectious and non-infectious etiologies are rare, no recommendations exist for their management. While exact estimates of incidence are not yet available, IRIS in patients initiating ART has been firmly established as a significant problem in both high and low income countries. Because of wide variation in clinical presentation and the still increasing spectrum of symptoms and etiologies reported, diagnosis remains problematic. Fur-thermore, no test is currently available to establish an IRIS diagnosis. Standardized disease-specific clinical criteria for common infectious manifestations of the disease should be developed to: 1) identify risk factors for developing the syndrome and 2) optimize the prevention, management of opportunistic infections. Results of trials addressing the optimal timing and duration of treatment of opportunistic infections will assist in developing guidelines for the prevention and management of IRIS. Treatment of IRIS will remain a clinical challenge due to the variety of clinical presentations and the presence of multiple pathogens capable of causing the syndrome. Until a greater understanding of the syndrome is achieved in different regions of the world, clinicians need to remain vigilant when initiating ART and individualize therapy according to known treatment options for the specific infectious agent.
86
Global Surveillance of Emerging Influenza Virus Genotypes by Mass Spectrometry
BACKGROUND: Effective influenza surveillance requires new methods capable of rapid and inexpensive genomic analysis of evolving viral species for pandemic preparedness, to understand the evolution of circulating viral species, and for vaccine strain selection. We have developed one such approach based on previously described broad-range reverse transcription PCR/electrospray ionization mass spectrometry (RT-PCR/ESI-MS) technology. METHODS AND PRINCIPAL FINDINGS: Analysis of base compositions of RT-PCR amplicons from influenza core gene segments (PB1, PB2, PA, M, NS, NP) are used to provide sub-species identification and infer influenza virus H and N subtypes. Using this approach, we detected and correctly identified 92 mammalian and avian influenza isolates, representing 30 different H and N types, including 29 avian H5N1 isolates. Further, direct analysis of 656 human clinical respiratory specimens collected over a seven-year period (1999–2006) showed correct identification of the viral species and subtypes with >97% sensitivity and specificity. Base composition derived clusters inferred from this analysis showed 100% concordance to previously established clades. Ongoing surveillance of samples from the recent influenza virus seasons (2005–2006) showed evidence for emergence and establishment of new genotypes of circulating H3N2 strains worldwide. Mixed viral quasispecies were found in approximately 1% of these recent samples providing a view into viral evolution. CONCLUSION/SIGNIFICANCE: Thus, rapid RT-PCR/ESI-MS analysis can be used to simultaneously identify all species of influenza viruses with clade-level resolution, identify mixed viral populations and monitor global spread and emergence of novel viral genotypes. This high-throughput method promises to become an integral component of influenza surveillance.
Influenza viruses cause serious global economic and public health burdens. Annual influenza epidemics resulted in more than 30,000 deaths a year in the United States during 1990-1999 [1, 2] . Periodic pandemics result in significantly higher death tolls. Emergence of new influenza A virus strains can be caused by ''antigenic shift,'' resulting from reassortment of gene segments, including H and/or N types [3, 4] , ''antigenic drift'' resulting from the continuing accumulation of mutations in the H and N genes [5] , or a pathogenic virus jumping species and acquiring the ability to infect and be transmitted among humans, as in the 1918 pandemic [6] . The recent outbreak of highly pathogenic H5N1 avian influenza virus (HPAI), which originated in Southeast Asia and has since spread globally, has resulted in 166 deaths (272 confirmed human cases) as of February 6, 2007 (http://www.who. int/en/). The global emergence of this virus has brought renewed urgency to the effort to track the spread and the evolution of influenza viruses. Currently, rapid methods for influenza virus diagnosis rely on antigen-specific antibody probes [7] , or real-time reverse transcription PCR (RT-PCR) analysis of the matrix (M) gene for identification of the viral species [8, 9] followed by H and N subtype specific RT-PCR assays for determination of the viral subtypes [10, 11] . Since there are many H and N subtypes with significant intra-and inter-subtype sequence variations, these methods do not identify all H and N subtypes, nor are they likely to identify reassortants or newly emerging genetic variants. Further, none of the current surveillance methods provide information relevant to tracking antigenically novel strains that emerge each year or distinguish amongst multiple lineages of influenza viruses that can co-circulate and persist in a popula-tion [12] . Secondary genome sequence comparisons and phylogenetic analyses are necessary to fully understand the multiple lineages of viruses, recognize newly emergent influenza variants, and monitor global spread of these viruses [12, 13] . For instance, analysis of human influenza virus H3N2 sequences from 1999-2004 revealed that at least three major clades of influenza viruses were in circulation after the 2002-2003 influenza season [12] . The differences were due to multiple reassortment events though all shared a common H-gene lineage. Several similar wholegenome studies with avian influenza viruses have revealed the presence of multiple, region-specific sub-lineages of the HPAI H5N1 virus in Southeast Asia that are spreading to Europe and Africa [14] [15] [16] [17] [18] . We have developed a method based on broad-range RT-PCR followed by electrospray ionization mass spectrometry (RT-PCR/ ESI-MS) for rapid and accurate detection of influenza virus, subspecies characterization, and early identification of genetic changes in circulating viruses. This method has previously been applied to detection of other pathogens in human clinical samples [19, 20, 21, 22] , but it has unique capabilities and advantages for influenza surveillance. Here, we show how a highthroughput assay incorporating eight parallel RT-PCR reactions followed by ESI-MS analysis can be used to simultaneously survey for all species of influenza viruses, provide clade-level resolution, identify mixed viral populations in the same sample, detect reassortants, and facilitate monitoring of viral evolution, all integral components of broad influenza surveillance. To measure the breadth of coverage and resolution offered by the panel of primers described in Methods (details in Table S1 ), we tested 92 well-characterized influenza virus isolates collected from human, avian, and animal species. Despite the extensive genetic diversity of this sample set, the broad-range primers generated amplicons from all isolates, and base composition signatures distinguished the isolates (Figure 1 ). Most isolates showed base compositions consistent with expected signatures for the corresponding H/N sub-types based on bioinformatic analysis of existing sequence data. Two of the isolates, however, showed previously unknown base compositions at several primer loci suggesting these might be novel influenza virus types; these are noted as ''Unknown'' in Figure 1 . Base composition signatures provide a multidimensional fingerprint of the genomes of the various viruses, which can be used to determine clusters of related species/sub-types. One such representation ( Figure 2 ) shows base composition data derived from the PA, PB1, and NP gene segments analyzed on individual axes. Importantly, only three of the six influenza A primer pairs from Figure 1 are visualized on this three-dimensional plot. There was strong agreement between the bioinformatic analysis of sequence data from GenBank and experimental measurements of base composition signatures from Figure 1 . Human H3N2 and H1N1 viruses clustered independently from each other and from the avian/human H5N1 and H1N1 viruses. To assess the utility of the RT-PCR/ESI-MS assay for surveillance of influenza virus in human populations, we analyzed 656 blinded clinical samples collected over a seven-year period (1999) (2000) (2001) (2002) (2003) (2004) (2005) (2006) . The results were compared with conventional analysis of the same samples by virus culture/serology and real-time RT-PCR methods. Two hundred forty-three samples were influenza positive both by RT-PCR/ESI-MS and conventional assays. Ten samples were positive only by RT-PCR/ESI-MS while eight samples were positive only by culture/real-time RT-PCR, corresponding to approximately 97% sensitivity and 98% specificity. Of the influenza-positive samples, RT-PCR/ESI-MS analysis identified 186 as influenza A virus and 67 as influenza B virus, in complete agreement with conventional typing methods. Base composition analysis of multiple RNA segments enables further categorization of isolates into previously established clades determined by sequencing (details shown in Table S2 ). Of the 186 influenza A samples, we determined that 149 of were H3N2 subtype and 34 were H1N1. The subtype of three samples could not be distinguished between H3N2 and H1N2 because these viruses probably arose from a recent reassortment of the H gene from an H1N1 virus with gene segments from an H3N2 virus [12] . Importantly, our predictions agreed completely with serology and direct RT-PCR analysis of the H and N segments from these samples. Nonetheless, although base composition analysis of the PB1, NP, M1, PA, and NS gene segments can be highly predictive of the H and N types, direct analysis of the H and N segments is necessary for unambiguous subtyping due to the potential for viral reassortment. In addition to identification and species typing, RT-PCR/ESI-MS provided a quantitative estimate of the number of viral genome copies in the original patient sample. This was achieved by including a fixed amount (300 copies/well) of an internal RNA calibration standard in each PCR reaction [19] [20] [21] [22] . The genome copy numbers in the influenza samples ranged from low (,100 genome copies per well) to intermediate (100-2,500 copies/well) to high (.3,000 copies per well). Based on the samples within the linear range of the calibration standard, we estimated an average genome load of 750 copies/PCR reaction, representing an average viral load of ,1.5610 4 genomes from material extracted from the original swab or 200 mL of a nasal aspirate. Six of the samples required serial dilutions to obtain viral concentrations in the quantifiable spectrum and ranged from 1.5610 6 to 8.0610 7 genomes/swab. Thus, for influenza-positive patients with respiratory symptoms, we observed a range of five orders of magnitude in viral RNA shedding. To demonstrate the capabilities of RT-PCR/ESI-MS to track the evolution of circulating influenza viruses, we created a tree representation of the H3N2 influenza virus sequences from Genbank ( Figure 3 , black) as described in Methods. The 104 experimentally determined H3N2 base compositions were mapped onto this tree (Figure 3 , blue). Analysis revealed a distribution very similar to the sequence-derived clades of Holmes et al. [12] . The branch terminating in clade A represents the dominant branch between years 1999-2004, with a clade B branch that co-circulates in the same time period. Strikingly, however, clade-A isolates were not detected after this time and Table S3 ) and confirmed the findings described above. A visual display of the most likely relationships among the isolates is shown in Figure 4 . Collectively, these results demonstrate the richness of the genetic information provided by direct RT-PCR/ESI-MS analysis of human clinical materials. Figures 5A and B show the spectra from samples containing a mixture of the ''parent'' BCtype AADFAA and single-nucleotide variations AAHFAA and AADFBB, respectively, providing a snapshot of enduring ''fit'' quasispecies contributors and of potential viral evolution in action. The dynamic range for mixed RT-PCR/ESI-MS detections has previously been determined to be approximately 100:1 [20] , which allows for detection of viral variants with as low as 1% abundance in a mixed population. Figures 5C and 5D show results obtained with a clinical sample containing a mixture of viruses at the limit of detection for mixed populations. To demonstrate that these peaks truly represented mixed viral populations, the PCR amplicons were cloned and 450 independent colonies were sequenced. Nine (2%) of these 450 clones had the predicted mutations, correlating well with the measured amplitude of the low-abundance peaks. A total of 293 non-overlapping nucleotides, excluding the primer regions, were analyzed using the genetic loci targeted by the influenza A primers. This corresponds to 2.15% of the influenza A virus genome. Out of 174 human samples analyzed from the 2005-2006 season, only two showed evidence for mixed viral populations at one of the six loci, corresponding to 1.1% of the samples. Thus, assuming the same mutation rate for the broader viral genome as for the region analyzed by PCR/ESI-MS, about 50% of the human H3N2 virus samples would have a mixed population of viruses. Choosing amongst the various molecular methods available for pandemic influenza surveillance requires consideration of both practical issues (e.g., broad availability, convenience, cost, and throughput) and scientific issues relevant to public health (e.g., sensitivity, breath of coverage, and the depth and value of the information provided). At one end of the spectrum, a conventional RT-PCR test with specific primers and probes provides a highlyspecific, sensitive, rapid, convenient, quantitative, relatively inexpensive, and high-throughput format that can provide valuable surveillance information. However, these tests are not optimal for surveillance when the exact nature of the pandemic virus is not known. Moreover, without supplemental nucleic acid sequencing, conventional RT-PCR-based tests are not capable of signaling the appearance of new genetic variants, except by potentially demonstrating a loss of sensitivity. Further, a single RT-PCR test can achieve only a single presence/absence analysis limited to the specific target for which it was designed. Discrimination of all known variants of influenza at the level of resolution described here would require hundreds of independent RT-PCR reactions. At the other end of the spectrum, virus isolation using culture methods followed by complete genome sequencing does not require prior knowledge of the virus' sequence and provides cladelevel resolution and highly detailed information regarding virus evolution. Unfortunately, this method is slow, labor intensive, expensive, and low throughput, rendering it ineffective in public health arenas requiring rapid response. In this work we describe a novel method that employs some of the best properties of each of the discussed techniques, and also supplies additional valuable information not provided by those techniques. For example, our method may identify mixed Black font: types determined through sequence analysis; blue font: experimentally determined base composition types; red font: experimentally determined base composition types for season 2005-06. Ten rare sequence types (,1.5%) were not uniquely discernable by the base composition analysis of the eight amplicons used in this analysis, as more than one subtype produced the same BC-type. These BC-types are indicated by asterisks. doi:10.1371/journal.pone.0000489.g003 populations of viruses, either as viral quasispecies as previously illustrated (i.e., development of ''drift'' strains) or co-infections with circulating strains (i.e., potential for development of ''shift'' strains). Recent advances in ESI-MS using bench-top mass spectrometers have enabled analysis of PCR amplicons with sufficient mass accuracy that the nucleotide base composition (the A, G, C, and T count) of the PCR amplicon can be unambiguously determined [23] . The approach outlined here provides two important advantages for surveillance. First, broad-range primers targeted to highly-conserved sites within the influenza virus family that flank highly variable, information-rich regions can be used to amplify sequences from highly diverse viruses in the same assay. The measured base compositions allow identification of the viral species with a high degree of resolution. For this approach, we selected primer pairs targeting the core gene segments (PB1, PB2, PA, M, NS, and NP) that have conserved regions spanning all known influenza viruses and used the resultant base compositions to infer the influenza virus H/N types. In silico analysis of influenza virus sequences from the GenBank database showed that this approach would detect all known influenza viruses and distinguish .90% of all species and types. This overcomes the limitations of directly targeting the H and N segments, which requires specific primer pairs for each known H/N type. Further, since the H/N segments evolve rapidly, newly emerging strains might not be readily detected by the traditional approach. Second, mixed viral populations in the same sample that differ only by a single mutation and is present in as low as 1-2% of the virus population can be identified, providing early insights into viral evolution as an integral component of surveillance. This information-rich result is provided with the same throughput and consumable costs as conventional, sequence-specific RT-PCR assays. The results from 174 influenza A H3N2 samples from the 2005-2006 season in the northern hemisphere were particularly interesting because they revealed viral evolution during a single season. The viruses detected appeared to have been seeded from two of the more abundant BC types circulating during the previous season in the southern hemisphere. The majority (97) of the samples had identical BC types and probably arose from a single founder from the previous season in the southern hemisphere. Most of the remaining samples differed from this founder BC type by one or two additional point mutations within the target regions described here. Surprisingly, when mutations occurred, they became fixed rapidly in the viral population, since only two of the 174 samples from the 2005-2006 season showed evidence of mixed populations. Sequencing revealed that both mutations were silent transitions in third codon positions. In summary, the RT-PCR/ESI-MS method has the capacity to provide rapid diagnosis of human influenza in individual patients with respiratory symptoms, as well as public health surveillance of emerging, potentially pandemic strains, including novel reassortants. The use of RT-PCR/ESI-MS for human as well as avian/ animal surveillance offers the potential for new insights into viral evolution on a scale and at a cost previously not possible. Benchtop mass spectrometers are capable of analyzing complex PCR products at a rate of approximately one reaction product/minute, making the RT-PCR/ESI-MS technology practical for large-scale analysis of clinical specimens or for animal surveillance. Further, as we have demonstrated with influenza viruses, this method provides sensitive detection directly from patient specimens with specificity approximating sequence-level resolution. The ability to quantitate virus shedding, detect low-abundance, mixed infections, and identify new genetic variants without prior knowledge of viral sequence also make this technology ideally suited for monitoring the emergence of drug-resistance mutations during therapy or for identifying newly emerging antigenic variants. Sample processing Viral stock samples consisting of cultured virus or stocks obtained from ATCC were prepared for analysis using the Qiagen QiaAmp Virus kit (Valencia, CA). Both manual (mini spin) kits and BioRobot kits were used. Robotic-based isolations were done on both the Qiagen MDx and Qiagen BioRobot 8000 platforms. Clinical swab samples were stored in Viral Transport Media (VTM). VTM solution (1 ml) was passed over a 0.2 micron filter, which was then subjected to bead beating in a small amount of lysis buffer. The resulting viral lysate was then used following the same protocol as above. Based upon analysis of multiple influenza sequence alignments, pan-influenza virus RT-PCR primer sets were developed that were capable of amplifying all three influenza virus species (A, B, and C) and subtypes (HxNy) from different animal hosts (e.g., human, avian, and swine) and distinguishing essential molecular features using base composition signatures. Additional primer pairs were designed that broadly amplified all known members of a particular species, but that did not cross-amplify members of different species (e.g., pan-influenza A and pan-influenza B primers). A surveillance panel of eight primer pairs (Table S1 ) was selected comprising one pan-influenza primer pair targeting the PB1 segment, five pan-influenza A primer pairs targeting NP, M1, PA, and the NS segments, and two pan-influenza B primer pairs targeting NP and PB2 segments. All primers used in this study had a thymine nucleotide at the 59 end to minimize addition of non-templated adenosines during amplification using Taq polymerase [24] . The sensitivity of each RT-PCR primer pair was determined using known quantities of a synthetic calibrant RNA template as described previously [20] . Each of the primer pairs was sensitive to as few as twenty copies of the calibrant RNA and several primers were sensitive to five copies (Table S1 ). One The following RT-PCR cycling conditions were used: 60uC for 5 min, 4uC for 10 min, 55uC for 45 min, 95uC for 10 min, followed by 8 cycles of 95uC for 30 seconds, 48uC for 30 seconds, and 72uC for 30 seconds, with the 48uC annealing temperature increasing 0.9uC each cycle. The PCR was then continued for 37 additional cycles of 95uC for 15 seconds, 56uC for 20 seconds, and 72uC for 20 seconds. The RT-PCR cycle ended with a final extension of 2 min at 72uC followed by a 4uC hold. The RT-PCR products were analyzed using the Ibis T5000 universal biosensor platform (Ibis Biosciences, Inc., Carlsbad, CA; http://www.ibisbiosciences.com), which performs automated post-PCR desalting, ESI-MS signal acquisition, spectral analysis, and data reporting as described previously [23] . Briefly, the steps were as follows: 15 mL aliquots of each PCR reaction were desalted and purified using a weak anion exchange protocol as described elsewhere [25] . Accurate mass (61 ppm), high-resolution (M/dM.100,000 FWHM) mass spectra were acquired for each sample using high-throughput ESI-MS protocols described previously [20] . For each sample, approximately 1.5 mL of analyte solution was consumed during the 74-second spectral acquisition. Raw mass spectra were post-calibrated with an internal mass standard and deconvolved to monoisotopic molecular masses. Unambiguous base compositions were derived from the exact mass measurements of the complementary single-stranded oligonucleotides [26] . Quantitative results are obtained by comparing the peak heights with an internal PCR calibration standard present in every PCR well at 100 molecules [20] . To demonstrate the capabilities of RT-PCR/ESI-MS to track the evolution of circulating influenza viruses, we used a bioinformatic approach to develop a framework on which to display the RT-PCR/ ESI-MS results obtained with H3N2 viruses. Complete genome sequences of all H3N2 human influenza viruses available in GenBank were analyzed. A total of 731 genomes were included, from which we inferred the phylogeny of H3N2 influenza virus since 1996. Using the 565-nucleotide concatenated sequence of the six loci that we analyzed by RT-PCR/ESI-MS, we constructed a nonredundant alignment of 105 sequence types. Base compositions were determined for each genome segment (i.e., locus) analyzed and each unique base composition at each of these loci was assigned a letter according to decreasing number of occurrences (therefore, the letter A represents the most common allele identified at each locus). The concatenation of the six base-composition letters from the PB1, NP, M1, PA, NS1, and NS2 loci from the sequences of 731 H3N2 viruses (1996) (1997) (1998) (1999) (2000) (2001) (2002) (2003) (2004) (2005) available in GenBank yielded 95 six-letter codes referred to as base composition types (BC types). The predominant type is labeled AAAAAA. The topology of the tree was then deduced from the alignment of non-redundant sequences using the programs dnadist and neighbor from the Phylip package (http://evolution. genetics.washington.edu/phylip.html). Since the sequence types differ mostly by discrete single mutations, the original computergenerated tree was then extensively edited using graphics tools to place the labels of intermediate types within the tree itself in lieu of zero-length branches. Readability was further improved by sorting parallel branches chronologically. Table S1 Influenza virus primer pairs used in this study. Genbank reference sequence for each segment is indicated; however, the primer sequences are not identical to the reference sequence as described in the Methods. The limits of detection for each RT-PCR primer pair were determined using known quantities of a synthetic calibrant RNA template. Found at: doi:10.1371/journal.pone.0000489.s001 (0.06 MB PDF) Table S2 Distribution of BC-types observed in influenza A H3N2 positive human respiratory samples. Unique base compositions at each genome segment locus analyzed were assigned letter codes and concatenation of letter codes across the six loci analyzed yielded BC-types. H1N1 samples were not assigned a BC-type. Experimentally determined BC-types (marked RT-PCR/ESI-MS Analysis results) were compared to BC-type signature information of sequences currently available in GenBank and the closest matching strain is shown (right pane). Last column shows comparison to clade designation described in Holmes et al [12] . Found at: doi:10.1371/journal.pone.0000489.s002 (0.24 MB PDF)
87
Transmission Parameters of the 2001 Foot and Mouth Epidemic in Great Britain
Despite intensive ongoing research, key aspects of the spatial-temporal evolution of the 2001 foot and mouth disease (FMD) epidemic in Great Britain (GB) remain unexplained. Here we develop a Markov Chain Monte Carlo (MCMC) method for estimating epidemiological parameters of the 2001 outbreak for a range of simple transmission models. We make the simplifying assumption that infectious farms were completely observed in 2001, equivalent to assuming that farms that were proactively culled but not diagnosed with FMD were not infectious, even if some were infected. We estimate how transmission parameters varied through time, highlighting the impact of the control measures on the progression of the epidemic. We demonstrate statistically significant evidence for assortative contact patterns between animals of the same species. Predictive risk maps of the transmission potential in different geographic areas of GB are presented for the fitted models.
The 2001 FMD epidemic in the UK had a substantial cost in human, animal health and economic terms (Alexandersen et al. [1] , Kao [2] ). Understanding the risk factors underlying the transmission dynamics of that epidemic and evaluating the effectiveness of the control measures are essential to minimise the scale and cost of any future outbreak. Epidemic modelling [3, 4, [5] [6] [7] proved critical to decision making about control policies which were (in some cases controversially) adopted to control the 2001 epidemic [8] [9] [10] . Modelling now has a 'peace-time' contingency planning role. One weakness of the modelling studies undertaken in 2001 was the relatively ad-hoc nature of the parameter estimation methods employed. In their first paper, Ferguson et al. [4] used maximum likelihood methods to fit to the observed incidence time series, but did not attempt to fit to the spatio-temporal pattern of spread. In their later work, the same authors developed a more robust method for estimating species-specific susceptibility and infectiousness parameters and spatial kernel parameters (see Supplementary Information to [3] ), but at the time the statistical basis for the methods developed was lacking. In retrospect, the methods developed turned out to be closely related to those developed during the SARS epidemic by Wallinga and Teunis [11] , although the earlier work incorporated population denominator data to allow for spatial-and species-based heterogeneity in disease transmission. Nevertheless, the methods employed had the limitation of not being fully parametric, meaning they could not be extended to fit arbitrary transmission models to the observed data. Keeling et al. [5] used maximum likelihood methods to estimate transmission parameters, but it was also supplemented by more ad hoc least-squares matching to regional incidence time series. Therefore there remains a need to develop rigorous modern statistical approaches for parameter estimation of non-linear models for the 2001 FMD outbreak. Bayesian Markov Chain Monte Carlo (MCMC) techniques are the best established such methods and have been successfully employed in the analysis of a range of spatiotemporal outbreak data in the past [12] [13] [14] , as well as to purely temporal incidence data [15, 16] . Here we develop MCMC-based inference models for the 2001 FMD epidemic in GB. The models examine: the extent to which transmission was spatially localised and the temporal variation in transmission, species-specific variation in susceptibility, infectious-ness and heterogeneity in contact rates between and within species. We take the farm as the unit of our study and ignore the possible impact of within-farm epidemic dynamics. Thus we implicitly assume disease spread within a farm is so rapid as to be practically instantaneous, with all animals on a farm becoming infectious at the same time. Our data consists of information on all the farms in the UK listed in the 2000 agricultural census [see http://www.defra.gov. uk/footandmouth/cases/index.htm ]. There were a total of 134,986 farms listed in that dataset and uniquely identified by their County/Parish/Holding (CPH) number. Their spatial coordinates are provided together with the number of animals by species within each farm. A partition of all GB farms according to the animal types represented is shown in Figure 1a . Their geographical distribution is represented in Figure 1b as the number of farms per 565 km. Notice the high density areas in the North West (Cumbria), South West (Devon), Wales and Scotland where the main epidemic foci developed. There is also an area of high density in the Shetland Islands corresponding to very small crofter smallholdings. Figure 1c and d show the numbers of sheep and cattle kept per 565 km square. During the 2001 FMD outbreak, a total of 2026 infected premises (IPs) were recorded -farms where FMD was diagnosed, and which were subsequently culled. The IP dataset contains, for each farm, the estimated date of infection (determined by a clinical evaluation of the age of lesions on affected animals), and the dates of disease reporting, confirmation and culling. A total of 7457 other (non-IP) farms were also culled -mostly as contiguous premises (CPs, about 3103) or dangerous contacts (DCs, about 1287), but some under other local culling policies used in Cumbria and Scotland. For instance about 1846 (79%) out of a total of 2342 sheep farms in Cumbria had all sheep culled under the ''local 3 km radial sheep cull'' policy adopted there. Some of the farms (about 30) were recorded both as DCs and CPs. Multiple records per farm were often found in the disease control management system dataset, and it was often unclear whether this was due to data entry errors or as a result of sequential species-specific culls on the same farm. In our analysis we therefore considered the whole farm to be culled at the last recorded date of culling. The most frequent species are cattle and sheep (see Figure 1a ). There are less than 3% farms with pigs only and only 10 farms with just pigs were diagnosed as IPs in 2001 (less than 1% of all the IPs). This indicates a-priori that pigs contributed far less to the 2001 outbreak than many other FMD outbreaks (despite their high levels of shedding [1, 17] ), and we therefore decided to discard pigs-only farms from the current study to simplify the analysis. The Sensitivity Analysis section shows that this simplification does not significantly affect estimates of other epidemiological parameters. We discarded another three IPs due to missing information or possible mistakes regarding their location or number of animals, leaving a total of 2013 IPs in our analysed dataset. We model the epidemic as a space-time survival process [18] . The total observation time T is the 240 days between 7 th February and 5 th October 2001. Each farm i at the location (x i , y i ) is associated with an infection time t i (if infected), a removal time r i (if slaughtered) and two integers n c i and n s i representing, respectively, the number of cattle and sheep on the farm. S c and S s represent per-capita cattle and sheep susceptibility, respectively, while I c and I s represent per capita cattle and sheep infectivity. The susceptibility is a relative measure of animal sensitivity to the disease whereas infectivity represents the infectious risk posed by an animal to others. We use a continuous kernel to describe how the probability of contact between farms scaled with distance. Transmission is naturally assumed to decrease with the distance between farms according to the power law where d ij represents the Euclidian distance between the infected farm i and the susceptible farm j. The parameters a (kernel offset) and c (kernel power) are to be estimated. The kernel captures all forms of movement and contact between farms and as such, the use of a simple 2 parameter function is inevitably a highly simplified representation of the true complexity of inter-farm contacts. We examined other functional forms for the kernel (such as those used in some other analyses [19] ) but the resulting model fits were much poorer than found using the power-law kernel above. Given the susceptibility and infectiousness parameters and the kernel, the infection hazard from an infected farm j to a susceptible farm i is then quantified by This model is over-specified as stated, so we arbitrarily assume S s = 1 throughout, meaning S c represents the ratio of cattle-tosheep susceptibility. For a constant (distance-independent) kernel this is just a mass-action closed epidemic model with heterogeneous susceptibility and infectiousness. This model assumes susceptibility and infectiousness parameters scale linearly with the number of animals of different species on the farm, a relatively strong assumption imposed for model parsimony reasons. The mixing matrix embedded in (2) quantifies the 4 species-specific mixing rates between animals on different farms: cattle-to-cattle (S c I c ) sheep-to-cattle (S c I s ), cattle-to-sheep (S s I c ) and sheep-to-sheep (S s I s ). This model formulation is identical to that used by Keeling et al. [5] , except for the functional form of kernel used. The force of infection on a susceptible farm i at time t depends on the whole history of events and is just where , if the farm i is susceptible and the farm j is infectious at the time t 0, otherwise By default, we assume a latent period of 1 day (latency is represented within the function L); i.e. farms are infectious the day after they are infected. However, we test the sensitivity of our estimates to the assumption by also examining latent periods of 2 and 3 days. The probability density function that farm i is infected at time t is then given by Hence, the contribution that a farm i, observed to be infected at time t, makes to the log likelihood is just: A farm which is not infected contributes to the overall likelihood the probability that it escapes infection during the observation period, i.e. until the time it is culled (r i ) or for the duration of the epidemic T, whichever is shorter. Its contribution to the log likelihood is therefore The total log likelihood of the model can be written as We then extend the simple model above by introducing an additional parameter to understand to what extent the transmission within species is altered by between species transmission. The parameter r quantifies the degree to which mixing between species is assortative -with r,1 representing assortative mixing and r.1 disassortative mixing. The interaction model still assumes constant parameters with respect to time along the whole observation period T. The mixing matrix defined in equation (2) becomes S c I c rS c I s rS s I c S s I s ð8Þ where we again fix S s to be 1 to avoid model over specification. The force of infection (3) and model log likelihood equation (7) change accordingly. Assuming transmission parameters were constant in time throughout the epidemic is obviously a crude simplification. However, allowing infectivity to vary continuously in time results in an over-specified model and problems of parameter identification and confounding. We therefore examined two sets of models in which changes in transmission parameter were restricted to 2 significant points in time denoted by T cut , namely 23 rd February (when the national ban on animal movements was introduced) and 31 st March (when control measures were intensified and the so called 24/48 hour IP/CP culling policy was introduced). Models were respectively fitted to the individual case data from the start of the epidemic (conditioning on the first infection) or from after 23rd February (conditioning on the 54 farms that were already infected by that date). A detailed history of the epidemic is given by Kao [9] . We separately fitted model variants which assumed a discrete change in parameters on 23 rd February and on 31 st March. Confounding meant that only a very limited number of parameters could be varied in time, so we examined the effect of varying infectiousness and kernel parameters separately. We fitted four separate time-varying model variants: (i) varying the cattle infectivity by a factor and keeping sheep infectivity constant through time (Cattle Infectivity model); (ii) varying sheep infectivity by a factor but not cattle infectivity (Sheep Infectivity model); (iii) varying both cattle and sheep infectivity by the same ratio (Cattle & Sheep Infectivity model); (iv) varying the kernel parameters (Time Varying Kernel model). For the last model variant we also fitted a version which includes non-assortative mixing between species (see equation (8)). Hence the most general mathematical expression of the transmission model is: where The scripts pre and post are self-explanatory for time varying parameters. When fitting models with time varying infectivity parameters we actually fit I post and the ratio m = I pre /I post we called infectivity factor. This is a within species ratio, a parameter directly fitted by the models, unlike the between species infectivity ratio additionally calculated as explained later in the text (see Parameter estimates section). Note that all models above treat the epidemic as fully observed, i.e. infection times are assumed to be known (when in fact only estimated infection times are known -see Sensitivity Analysis section), and only IPs are assumed to be infectious. We adopt a Bayesian framework for statistical inference and use MCMC methods for fitting the model to individual case data. This is not strictly necessary, given our simplifying assumption that the epidemic was completely observed, but it provides a more consistent and robust framework within which to relax that assumption in future work. We obtained parameter estimates and equal-tailed 95% credible intervals from the marginal posterior distributions of the fitted parameters. For the basic model for instance we estimated the relative cattle susceptibility, S c , two infectivity parameters (I c (t);I c and I s (t);I s for all t) and two kernel parameters (c(t);c post ;c pre ;c and a(t);a post ;a pre ;a for all t). We used the posterior mean deviance as a Bayesian measure of fit or model adequacy as defined by Spiegelhalter et al. [20] . The posterior density deviance is defined as: where log{P(y|h)} is the log-likelihood function for the observed data vector y given the parameter vector h and C is a constant which does not need to be known for model-comparison purposes (being a function of the data alone). The smaller the mean posterior deviance, the better the corresponding model fits the data. If the posterior deviance distributions for two different models overlap significantly, it is necessary to use additional criteria to compare model fit -namely a comparison of the relative complexity of the models. The Deviance Information Criterion (DIC) is perhaps the most general of such methods, being a generalisation of the Akaike information criterion for Bayesian hierarchical models [20] . We define the complexity of a model by its effective number of parameters, p D , defined as where E[ ] represents taking expectations (the posterior average). The DIC is then defined as A lower value of DIC corresponds to a better model. This criterion offers flexibility for comparing non-nested models [20] and it is straightforwardly computed within an MCMC algorithm. We applied the classic random walk Metropolis Hastings algorithm [21, 22] and a block-sampling of parameters due to the computationally expensive form of the likelihood [23, 24] . A log scale has been used for sampling as the parameters were all positive definite and were expected to potentially vary by orders of magnitude. However, linear scale sampling yielded similar results. The convergence of the chains was also very much improved (see Robert [25] for more on perfect sampling and reparameterization issues) compared with sampling on a linear scale. The model was coded in C and parallelized using OpenMP 2.0. The MCMC sampler was allowed to equilibrate with convergence being evaluated visually from the likelihood and parameter traces. For the simpler models, 5,000 iterations were sufficient for equilibration, while this increased to 20,000 for the most complex models. Also, using log scale sampling, we verified that the chains were able to converge even if started with initial parameter values far from the final posterior mean values. Posterior distributions were estimated from 100,000 iterations. The rate of the acceptance varies from model to model. For the baseline model we achieved a 25% rate of acceptance and for the most complex model (8 parameters), a rate of approx 10%. These values compare well with the ''golden'' acceptance rate for Random Walk Metropolis Hastings of 23% (Roberts [26] ). We did not encounter common problems in MCMC estimation like slow convergence and slow mixing (O'Neill [27] ). There were some correlations between parameters, mostly having biological explanations (cattle and sheep infectivity for instance), but a careful parameterization lowers them. We verified parameter estimates were not dependent on parameterization choices -e.g. no difference was seen whether we fitted species infectivity individually, or just fitted sheep infectivity and then the ratio of cattle-to-sheep infectivity. Table 1 lists the parameter estimates we obtained for a set of fitted models conditioned only on the first infection whereas Table 2 presents the estimates for models conditioned on infections occurring up to 23 rd February. The posterior deviances for each set of models are plotted in Figure 2a and Figure 2b , respectively. Figure 2a illustrates some clear conclusions. Of the two models without time variation in parameters, the interaction model fits significantly better than the baseline model without heterogeneous mixing between species. However, fitting the interaction model broadened the credible intervals of the infectivity parameter estimates (Table 1) , indicating (unsurprisingly) slight confounding between the 4 infectivity and susceptibility parameters. Of the models which allowed infectivity to vary on 23 rd February, allowing only cattle infectivity variation gave a slightly better fit than varying sheep infectivity or both. However, of the models with parameters which vary on 23 rd February, the model variants which allow the 2 kernel parameters to vary at that time point fit substantially better (by both deviance and DIC criteria, see Table 1 ) than those which just allow a species-specific variation in infectivity. This is encouraging for the inference procedure, as the main control measure initiated on that date was the banning of (Figure 3a and Figure 3b ). The parameter estimates are less precise before 23 rd February (Table 1) due to the relatively small number of IPs (about 57) before that date. Looking at the most complex model (namely the interaction model with time varying kernel), cattle were estimated to be 5.7fold (4.6, 6.8) more susceptible than sheep (see Figure 3c and Table 1 ). Rather than mentioning animals' specific infectivity (see Figure 3d and Table 1 ), it is more informative to comment on the cattle:sheep infectivity ratio parameter for the most complex fit (this ratio does dot appear in the tables as it is not a model parameter). We calculated it within the MCMC algorithm as the ratio of the two species infectiousness for each sampled parameter point. The most complex model suggests that cattle are 5.95-fold (4.54, 7.63) more infectious than sheep (Figure 3e) . The parameter quantifying assortativity in mixing was estimated at r = 0.45 (0.31, 0.61) -well below 1, the level at which mixing between species is random (Figure 3f ). By comparison with the model with a time varying kernel but random mixing between species, the effect of heterogeneous mixing between species modified the between-species transmission as given by matrix (1.9) as indicated below. Cattle-to-cattle and sheep-to-sheep transmission is higher (by 19% and 54% respectively) for the model with non-random mixing, whereas the sheep-to-cattle and cattle-to-sheep transmissions dropped by 41% and 37 % respectively. Conditioned on 23 rd February, 7 model variants have been considered (Table 2 and Figure 2b) . We examined the baseline and interaction models (no change in parameters over time), allowing cattle infectivity to vary on 31 st March and both cattle and sheep infectivity to vary by the same factor after 31 st March (with and without heterogeneity in mixing) and allowing both kernel parameters to vary on 31 st March. Unsurprisingly, the kernel parameters were not significantly different if allowed to be different before and after 31 st March, neither did this model prove to be the best fit. Overall, while the variations in mean deviance (Figure 2b ) seen between model variants were much smaller than for the models conditioned on the first infection (Figure 2a ), the interaction model allowing for time varying cattle infectivity gave the most adequate fit (measured by both mean deviance and DIC, see Table 2 ). We cannot statistically compare the two sets of models in Table 1 and Table 2 , as the data used are different for the two cases. However, the parameter estimates from the best-fitting models of each table are largely consistent. Each post-23 rd February estimated value from the best-fit model in Table 1 is included in the corresponding pre-31 st March 95% credible interval of the best fit model in Table 2 (and vice-versa). The most important message from the second set of models is that all models with cattle time varying infectivity (best fit) indicated higher values of infectivity after 31 st March than before (m = 0.73 (0.63, 0.83)) ( Table 2 ). This may seem paradoxical but reflects the fact that while culling (the effect of which is explicitly included in the input data) dramatically reduced case incidence in April, from May to September 2001, case incidence maintained itself at a low level -but almost entirely within cattle farms. This increase in cattle infectivity may therefore really reflect the impact of reduced biosecurity and/or increased non-compliance with movement controls. It is informative to examine what our parameter estimates imply in terms of geographic variation in transmission potential. Given the parameter estimates for each model, we can define the relative risk of transmission an infectious farm j would pose to all susceptible farms in the country r j : r j~Ic n c j zI s n s j P i=j Sc Ss n c i zn s This quantity multiplied by the average duration of infectiousness of a farm (time from end of latency to culling) gives the reproduction number R 0j of the farm j. We divided the UK into 5 km squares and then calculated the average transmission risk of all farms in each square (local R 0 ). Figure 4 shows how geographic risk changed before and after 23 rd February for our best fit model conditioned on the first infection. The kernel shape has a major influence on the average risk distribution throughout the country. Figure 5 shows the corresponding risk maps for the estimates inferred from our best fit model conditioned on 23 rd February. A slightly higher risk is predicted after 31 st March by the model conditioned on 23 rd February due to the increase in the cattle infectivity after this date. The risk estimates after 23 rd February from the first set of models appear consistent with those obtained from the models conditioned on 23 rd February, though a rigorous statistical comparison is not appropriate. We have made the strong assumption for this study that the only infected farms during the 2001 epidemic were the reported IPs, and hence that any farms which were infected but culled before clinical diagnosis were not responsible for causing any infections. It is therefore interesting to calculate how many of the proactively culled farms our model predicts might have been infected (but, by definition, not diagnosed). To calculate the probability p i that a particular proactively culled farm i was infected, we need to adjust the infection hazard by the probability that the farm would have not been reported as a clinical case before its culling date T i c . From the outbreak data, we calculate the probability density of the time from infection to report for reported IPs and hence the cumulative probability distribution of the time from infection to report, denoted by F. Then, with l i (t) being the force of infection on a proactively culled farm i at time t (from the best fit model conditioned on 23 rd February), the probability that that farm gets infected and escapes reporting between its potential infection time and culling time T i c is We calculate the expected number of infections in different classes (e.g. DCs, CPs) of proactively culled farms culled within a particular time interval (T i c [ T 0 ,T 1 ½ ). For instance, the expected number of CPs culled at the time T i c [ T 0 ,T 1 ½ which are predicted to have been infected can be formally written as This is a simplification, as in reality the delay from infection to report almost certainly depends on the size and species mix on a farm, but the result is nevertheless indicative of the expected level of infection in proactive culling. Also, at this stage, the calculations are made as if culling was a non-informative censoring process. This is a reasonable assumption for all proactively culled farms except for DCs (which by definition had been identified by veterinarian as having had a high risk of exposure) but our method may underestimate the infection rate. In calculating the infection to report delay distributions, we divided the epidemic after 23 rd February into 3 time periods: 23 rd February-31 st March, 31 st March-1 st May and 1 st May-5 th October. In these intervals a total of 1332, 4498 and 1627 farms were slaughtered, respectively. Our best fit model conditioned on 23 rd February predicts different infectivity regimes before and after 31 st March (see Parameter Estimates and Table 2 ) but we split further the second period of time due to different delays in reporting to culling. The infection to report delay is 8.6 and 8.8 days for the last two periods of time respectively but the infection to cull delay drops from 9.4 and 8.8 days respectively. Applying this approach to the interaction model with time varying cattle infectivity which conditioned on the 23 rd of February, we calculated the expected proportion of proactively culled farms which were infected. We estimate that approximately 1.3% (1%, 1.6%) of 7457 culled non-IP farms may have been infected -97 in total (Figure 6a ). Of the 1332 farms culled between 23 rd February and 31 st March, 1.7% (1%, 2.4%) may have been infected (23 farms). Of the 4498 farms culled between 31 st March and 1 st May, we estimate 0.7% (0.5%, 1%) were infected (34 farms). In the period 1 st May to 5 th October, we estimate that 1.6% (1%, 2.3%) of 1627 farms culled were infected (27 farms). The proportion of CPs estimated to have been infected is 2% (1.5%, 2.5%), equating to 62 farms (Figure 6b ). Over the whole epidemic, we estimated 1.5% (0.8%, 2.1%) of farms designated as DCs were infected (19 farms). This estimate (Figure 6c ) does not allow for higher risk of infection implied by the veterinary judgement that led to those DCs being identified, which may mean that a higher proportion were in fact infected. If we assume that DCs were 3 times more likely to be infected due to their status than the model would predict, then the incidence of infection in DCs goes up accordingly, i.e. to 4.6% or 59 farms. Farms culled neither as DCs or CPs (typically those culled under the 3 km and local sheep cull policies in the Cumbria, Dumfries and Galloway areas) had the lowest estimated rate of infectiona mere 0.5 % (0.2%, 0.8%) or 16 out of 3067 farms. In this section we examine the sensitivity of our results to a number of factors: leaving pigs out of the analysis, possible errors in the estimated IP infection dates, and the assumed latent period. To justify the simplification of the analysis by discarding the number of pigs in a farm, we present some more detailed statistics regarding this variable. We also fit the simplest model conditioned the last two farms are exclusively pig farms. We denote by n p i ,S p ,I p the number of pigs in farm i, pigs susceptibility and pigs infectivity respectively. The simplest model similar to (1.2) conditioned on the first infection has been fitted, reducing the number of parameters in the same manner. In addition we estimated pig:sheep susceptibility ratio and pig infectivity, assuming all parameters constant through time. We found that cattle:sheep susceptibility ratio is 6. Table 1 shows parameter estimates for cattle and sheep are largely unaffected by ignoring the pig population, with none of the estimates from the two analyses being significantly different. We conclude that including pigs would not change the conclusions presented in Table 1 regarding cattle and sheep (given the very small number of IPs which had pigs) but it would decrease the power of the analysis and increase model complexity. To understand to what extent our estimates are affected by the assumption that the infection dates have been accurately observed, we randomized the estimated infection dates by adding a Gaussian noise with zero mean and a standard deviation of 2 days. This is motivated by the substantial proportion in the observed standard deviation (73.5% less or equal than 2 days) of the distribution time from the estimated infection date to the report date of IPs. We then fitted the simplest model (conditioned on both first infection and 23 rd February) to 10 such randomised datasets. The average estimates across them are given in Table 3 . They lie well within the confidence intervals we predicted in Table 1 . The average cattle:sheep infectivity ratio is also very close to the values estimated using the original data. The average estimates across 10 randomized datasets using the most appropriate model conditioned on 23 rd February (i.e. cattle infectivity and interaction model) are also in Table 3 . The values are within the 95%CI presented in Table 2 . We assessed a sensitivity analysis for the estimated proportion of infections in proactively culled farms (see the previous section) with respect to infection times. Using the predicted parameters for each dataset, we calculated the average proportions across all of them, for each category of proactively culled farms. The average proportion of infections between DC farms is 1.37% (2%, 0.78% and 0.72% for each period of time, respectively). For CP farms, the same quantities evaluate to 1.9% with 1.8%, 1.3% and 1.98%, respectively. Overall proactively culled farms, we obtained an average percentage of 1.25% with 1.64%, 0.81% and 1.6% for each considered period of time. All the values are well within the 95%CIs predicted by the original data (see the previous section and Figure 6 ). All the results presented above assume a fixed latent period of 1 day. We tested the sensitivity of parameter estimates to this assumption by examining latent periods of 2 and 3 days. Overall, we would expect infectiousness parameters to increase to compensate for the shorter infectious period, and thus slightly increased generation time (namely the mean time from infection of one case and the time of infection of the cases that case generates). Interestingly, however, it is the kernel parameter estimates which are altered as the latent period is varied with the kernel becoming slightly less local with increasing latent period. For two and three days latent period, pre 23 rd February, the values of c dropped from 1.69 (Table 1) This paper has presented a statistical analysis of the spatiotemporal evolution of the 2001 foot and mouth outbreak in GB. Qualitatively, the results agree with those obtained by Keeling et al. [5] in identifying cattle as being the key species in the 2001 epidemic. Using the interaction model conditioned on 23 rd February with time varying cattle infectivity, we estimated that 88% of IPs between 23 rd Feb-31 st March were infected by cattle and only 12% by sheep. Sheep-to-sheep transmission only accounts for 3.1% of IPs in that period. After 31 st March (when we estimated that cattle infectivity increased slightly, see Table 2 ) Allowing for non-random mixing between species indicates contacts between farms are assortative on the basis of species composition of the farm; i.e. like species mix with like. This agrees with intuition about the nature of farming practices (e.g. sharing of personnel and equipment is likely to be more common if 2 farms have the same livestock species). The implications of the moderate degree of assortativity we found for control measures remains to be explored. We did not use data collected during the epidemic on traced contacts between farms to fix the spatial kernel function in our analysis, since in the final version of the FMD epidemic data warehouse [http://www.defra.gov.uk/footandmouth/cases/index.htm] very few of the contacts apparently identified early in the epidemic remain confirmed. Also we shared the concern of earlier work that the distribution of contact distances in traced contacts may well be biased [3] . We therefore estimated the kernel function, using an offset power-law functional form. The higher value of the kernel power parameter we estimated after 23 rd February (2.67 vs. 1.70 before - Figure 3a) is consistent with the expected dramatic shortening in the typical contact distance following the national movement ban. This localized spread together with the higher estimated level of infectivity in cattle after 31 st March explains the long tail of the epidemic seen in 2001. In estimating the transmission risk between farms, we assumed a dependence on the Euclidian distance between them. In reality, other metrics (e.g. the time required to travel between two farms) might be more reasonable, and should be examined in future work. We also did not include information on landscape (e.g. height above sea-level, location of rivers, trees etc). The estimated risk maps (Figure 4 and Figure 5 ) match the areas of the country where highest case incidence rates were seen -with the notable exception of Wales. The discrepancy between the high predicted risk in Wales and the small number of cases observed may reflect inaccuracies in the input data set -Keeling et al. [5] reduced farm-level sheep population numbers by 30% in Wales and obtained a better geographic match to the data (Matt Keeling, personal communication). However, the discrepancy may also reflect model inadequacy. We have not here allowed for other farm-level risk factors, such as the farm fragmentation index considered by Ferguson et al. [3] . We have not explored more complex non-linear models of the dependence of susceptibility and infectiousness on the number of animals on a farm or relaxed our implicit assumption that contact rates between farms scale linearly with the local density of farms. All these assumptions are being relaxed in ongoing work. The most important issue to be revised in future work is to allow for proactively culled farms which were not diagnosed as IPs to be potentially infected and infectious to other farms. This requires modification of the inference model used to allow for an arbitrary number of unobserved infections. The very low numbers of proactively culled farms we estimated as infected suggested that the effect of this model refinement may be limited. It should be noted though that these infection prevalence estimates are in part a result of the relatively non-local kernel estimated simultaneously. If kernel estimates change in a refined analysis -and if DCs were attributed a much higher risk of infection than estimated here due to their status -then it is possible that estimated infection rates in DCs and other proactively culled farms may increase somewhat. However, even if these factors increased our estimated infection prevalence among proactively culled farms 5 fold (which seems unlikely from ongoing work), it would still mean that only a small proportion (,10%) of DCs and CPs culled were infected. This does not imply that proactive culling had no effect on the epidemic -as the largest expected effect of such culling is via the targeted depletion of susceptible animals. In this regard, proactive culling has the same epidemiological impact as vaccination. Future work will revisit past estimates of exactly how important such culling was for the control of the 2001 FMD epidemic.
88
Efficient replication of pneumonia virus of mice (PVM) in a mouse macrophage cell line
Pneumonia virus of mice (PVM; family Paramyxoviridae, subfamily Pneumovirinae) is a natural respiratory pathogen of rodent species and an important new model for the study of severe viral bronchiolitis and pneumonia. However, despite high virus titers typically detected in infected mouse lung tissue in vivo, cell lines used routinely for virus propagation in vitro are not highly susceptible to PVM infection. We have evaluated several rodent and primate cell lines for susceptibility to PVM infection, and detected highest virus titers from infection of the mouse monocyte-macrophage RAW 264.7 cell line. Additionally, virus replication in RAW 264.7 cells induces the synthesis and secretion of proinflammatory cytokines relevant to respiratory virus disease, including tumor necrosis factor-α (TNF-α), interferon-β (IFN-β), macrophage inflammatory proteins 1α and 1β (MIP-1α and MIP-1β) and the functional homolog of human IL-8, mouse macrophage inflammatory peptide-2 (MIP-2). Identification and characterization of a rodent cell line that supports the replication of PVM and induces the synthesis of disease-related proinflammatory mediators will facilitate studies of molecular mechanisms of viral pathogenesis that will complement and expand on findings from mouse model systems.
Pneumonia virus of mice (PVM) infection in mice was originally described by Horsfall and colleagues [1, 2] , but until relatively recently, the sole interest in this virus was as a pathogen of laboratory rodent colonies [3] [4] [5] . Over the past several years, we and others have built on Horsfall's early studies, and have developed and characterized an in vivo model of severe respiratory virus infection using PVM [reviewed in [6, 7] ]. Among our findings, we have shown that a minimal, physiologically relevant inoculum of PVM (typically <100 pfu) results in robust virus replication in lung tissue, accompanied by influx of granulocytes in response to local production of specific proinflammatory chemokines [8] . The pathophysiology of PVM bronchiolitis leading to pneumonia and acute respiratory distress syndrome (ARDS) is similar to that observed in response to severe respiratory syncytial virus (hRSV) infection in human infants [9] . While PVM clearly replicates efficiently in mouse lung tissue, the in vitro propagation of this pathogen is significantly less straightforward. The primate BS-C-1 epithelial cell line supports minimal rates of PVM replication in vitro [10] . The BS-C-1 cell line has been used for traditional plaque assays, but PVM-induced plaques develop slowly, have relatively indistinct borders, and are difficult to eval-uate quantitatively [see Figure 1A ]. Furthermore, from an evolutionary perspective, one would prefer to perform molecular studies of virus pathogenesis in cells from a relevant species, i.e...a rodent cell type or cell line. We have demonstrated that PVM replicates in the mouse LA4 respiratory epithelial cell line [11] , but virus growth is similarly slow, even at temperatures permissive for virus propagation in vitro. In this work, we explore PVM replication in several independent cell lines and identify the mouse macrophage The rodent L2, LA4, RAW 267.4, J774A.1, RLE and 3T3 and primate A549, BS-C-1, and HEp-2 cell lines obtained from American Type Culture Collection (Manassas, VA) were maintained in Iscove's Modified Dulbecco's medium with 10% heat-inactivated fetal calf serum, 2 mM glutamine and penicillin-streptomycin at 5% CO 2 and 32°C (permissive for virus growth in culture) unless otherwise indicated. Mouse-passaged PVM prepared as described was stored in liquid nitrogen at ~10 6 pfu/ml [12] . Virus replication in RAW 264.7 cells was determined by both Q-RT-PCR detection of the virus SH gene [see reference [13] ] for complete method] and by western blot [14] probed with a 1:200 dilution of polyclonal anti-PVM N peptide antibody prepared against sequence SQQLN-IVDDTPDDDI encoding amino acids 379 -393 of the PVM N protein. Proinflammatory cytokines in culture were evaluated by ELISA (R&D Systems, Minneapolis, MN). Q-RT-PCR detection of interferon-β was via standard methods using primer -probe set Mm00439546_s1 (ABI, Columbia, MD) normalized as described [13] on RNA prepared from infected and control uninfected cells in culture (RNazol B, Friendship, TX). The cell lines evaluated for the ability to support virus replication included rat epithelial L2 and RLE, mouse epithelial LA4, mouse macrophage RAW 267.4 and J774A.1, and primate epithelial A549, BS-C-1, and HEp-2. All were inoculated with PVM on day 0 (MOI = 0.02, 10 4 pfu per 5 × 10 6 cells). On day 7, virus titer in the culture supernatants was determined by standard plaque assay [12] . Although pneumoviruses maintain strict host-pathogen specificity in vivo, we determined that PVM replicated to a limited extent in vitro (< 10 3 pfu/ml supernatant) in each of the aforementioned cell lines. The mouse monocyte/ macrophage RAW 264.7 cell line (established from a tumor induced by Abelson murine leukemia virus) generated the highest virus titers (10 4 pfu/ml) under culture conditions described. Cells of the RAW 264.7 line also support replication of other unrelated virus pathogens, including murine hepatitis virus and Japanese encephalitis virus [15] [16] [17] . To evaluate the kinetics of virus replication and production of proinflammatory mediators in the RAW 264.7 cell line, cells at 50% confluence were inoculated with PVM (MOI 0.1) on day 0 and harvested on days 2 -5 thereafter. RAW 264.7 is a semi-adherent cell line, and is not well-suited for plaque assays. Here, virus replication was examined qualitatively on western blot of cellular homogenates probed with PVM-specific antisera [ Figure 1B ]. Virus was first detected in infected cultures on day 3 post-inoculation, and then in increasing amounts through day 5. No immunoreactive PVM N protein was detected in uninfected control cultures. Virus replication was also examined quantitatively by Q-RT-PCR using the virus SH gene as a target sequence [13] , [ Figure 1C ]. PVM replication was readily detected in inoculated RAW 264.7 cells, reaching ~4 × 10 5 copies per microgram total RNA on day 5 of infection. No copies of the virus SH gene were detected in uninfected cells. RAW 264.7 cells respond to infection with PVM by producing a variety of proinflammatory mediators. Transcription of interferon-β in response to virus infection was detected by Q-RT-PCR [ Figure 1D ]. Cytokines MIP-2, TNF-α, MIP-1α, and MIP-1β were detected in culture supernatants by ELISA [ Figure 1E ]. Interestingly, MIP-1α and MIP-2 are among the most prominent mediators detected in BAL fluid of infected mice; MIP-1α levels correlate directly with the severity of pneumovirus disease in both PVM and hRSV infection [18, 19] . In parallel to our findings, hRSV replicates in the human monocytic THP-1 cell line [20] , and several groups have provided evidence consistent with hRSV and bovine RSV (bRSV) replication in alveolar macrophages, although this point remains controversial [21] [22] [23] [24] [25] . Furthermore, hRSV infection of the human monocytic U937 cell line was associated with production of the proinflammatory mediator, platelet-activating factor (PAF) [26] . In summary, PVM has recently emerged as a useful novel model for the study respiratory disease in mice [7, [27] [28] [29] [30] ; this has provided significant incentive toward identifying tissue culture systems for virus propagation. The mouse RAW 264.7 cell line supports efficient replication of PVM in vitro and responds to infection by augmenting production of cytokines implicated in the pathogenesis of respiratory disease. Use of this ex vivo model of PVM infection will permit further study of biological responses associated with virus infection and the cellular and molecular level.
89
Designing and conducting tabletop exercises to assess public health preparedness for manmade and naturally occurring biological threats
BACKGROUND: Since 2001, state and local health departments in the United States (US) have accelerated efforts to prepare for high-impact public health emergencies. One component of these activities has been the development and conduct of exercise programs to assess capabilities, train staff and build relationships. This paper summarizes lessons learned from tabletop exercises about public health emergency preparedness and about the process of developing, conducting, and evaluating them. METHODS: We developed, conducted, and evaluated 31 tabletop exercises in partnership with state and local health departments throughout the US from 2003 to 2006. Participant self evaluations, after action reports, and tabletop exercise evaluation forms were used to identify aspects of the exercises themselves, as well as public health emergency responses that participants found more or less challenging, and to highlight lessons learned about tabletop exercise design. RESULTS: Designing the exercises involved substantial collaboration with representatives from participating health departments to assure that the scenarios were credible, focused attention on local preparedness needs and priorities, and were logistically feasible to implement. During execution of the exercises, nearly all health departments struggled with a common set of challenges relating to disease surveillance, epidemiologic investigations, communications, command and control, and health care surge capacity. In contrast, performance strengths were more varied across participating sites, reflecting specific attributes of individual health departments or communities, experience with actual public health emergencies, or the emphasis of prior preparedness efforts. CONCLUSION: The design, conduct, and evaluation of the tabletop exercises described in this report benefited from collaborative planning that involved stakeholders from participating health departments and exercise developers and facilitators from outside the participating agencies. While these exercises identified both strengths and vulnerabilities in emergency preparedness, additional work is needed to develop reliable metrics to gauge exercise performance, inform follow-up action steps, and to develop re-evaluation exercise designs that assess the impact of post-exercise interventions.
Since 2001, state and local health departments in the US have accelerated efforts to prepare for bioterrorism and other high-impact public health emergencies. These activities have been spurred by federal funding and guidance from the US Centers for Disease Control and Prevention (CDC) and the Health Resources and Services Administration (HRSA) [1] [2] [3] . Over time, the emphasis of this guidance has expanded from bioterrorism to include "terrorism and non-terrorism events, including infectious disease, environmental and occupational related emergencies" [4] as well as pandemic influenza [5] . For any locality, the rarity of major public health emergencies necessitates the use of practice-based exercises to simulate real life experiences in order to develop and improve skills and to assess response capabilities over time. The US Federal Emergency Management Agency (FEMA) describes six levels of exercises, increasing in complexity from informational seminars that minimally exercise response capacities to simulations that mimic reality and exercise participants' capacity to implement emergency response functions [6] . Intermediate in this progression is the tabletop exercise, which FEMA describes as a "facilitated group analysis of an emergency situation." As practiced in public health, there is considerable variability in how tabletop exercises are designed and conducted. Tabletop exercises may be structured discussions of evolving events or unstructured reactions to short scenarios; participants may be limited to public health staff or involve representatives from partner agencies or organizations; scenarios may range from simple to complex; and facilitation may range from being minimally directive, allowing participants to assume responsibility for managing the discussion through "role play," to highly directive, enabling the facilitator to assure that specific questions are addressed. Recognizing the need to exercise public health emergency response, and enabled by funding and directives from CDC and HRSA, health departments throughout the US have implemented exercise programs. These exercise programs have had varying goals, including building relationships among stakeholders [7, 8] , training staff [9] [10] [11] , and evaluating preparedness levels [12, 13] , and they have been used for a variety of purposes, including to identify gaps in preparedness [14] , make recommendations for improving preparedness [15] , and identifying variations in preparedness across health departments [16] . These exercises have involved diverse groups of stakeholders involved in public health preparedness, such as representatives from public health [17] , health care [18] , agriculture [19] , and emergency medical services [20] . Despite the commonality of preparedness and response compo-nents across a variety of biological threats, most exercises have been designed for single use and focus on single disease, such as smallpox [21] , pandemic influenza [22] , or a novel virus [23] . These exercises have focused attention on the interaction between preparedness goals and exercise strategies, and have illuminated strengths and vulnerabilities in public health emergency decision making and response capacities. The increased utilization of tabletop exercises by health departments has not been accompanied by a parallel increase in knowledge sharing about lessons learned from them, either with regard to identifying common challenges that confront health departments or strategies for effective exercise design and management. Further, the literature dealing with tabletop exercises to date consists almost entirely of case studies and descriptions of a single exercise or a single disease. This paper describes lessons learned by public health researchers at RAND, and their collaborators, about the process of developing and conducting tabletop exercises in collaboration with state and local health departments in the US and their implications for public health emergency preparedness. Data for this paper come from four related projects conducted from 2003-2006. Taken together, these projects involved developing, conducting, and evaluating 31 tabletop exercises with state and local health departments of different sizes and structures in 13 different states across the northeast, south, mid-west, and west regions of the country (Table 1) . Participating health departments did not incur any expenses through their involvement in these exercises other than the staff time required to participate. Two of these projects, one in California and the other in Georgia, involved the conduct of exercises in multiple jurisdictions in the same state. In California, the Little Hoover Commission, a bipartisan, independent state body, asked RAND to assess California's public health infrastructure. A key component of the project, described in greater detail elsewhere [16] , was the development of a tabletop exercise that simulated a smallpox outbreak. This exercise was conducted in seven local health departments across California. In Georgia, RAND collaborated with the Georgia Division of Public Health and the Rollins School of Public Health at Emory University to develop, conduct, and evaluate a series of tabletop exercises focusing on different biologic agents in seven local health departments across Georgia, as well as one exercise focused at the state level. The two remaining projects were funded by US Department of Health and Human Services (HHS) and involved the participation of multiple local health departments. The first project involved developing ten different tabletop exercise templates and formats focusing on the local public health response to bioterrorist agents. These were tested in 13 local health departments in 12 different states. The second project involved developing a tabletop exercise to examine the interface between local health departments and health care systems in a hypothetical influenza pandemic. This exercise was tested in three local health departments in different states. Greater detail on the structure for these tabletop exercises as well as the tabletop exercise templates themselves can be found elsewhere [7, 17] . All exercises focused on at least one of three related objectives: training, relationship-building, and evaluation. The structure and design of the tabletop exercises varied from project to project because their objectives were somewhat different. The key domains covered are outlined in Table 2 . The level of facilitator involvement varied with the exercise objectives. At one extreme, the facilitator's role was limited to introducing the exercise scenario and periodically interjecting updates. During these exercises, the participants were encouraged to lead the discussion themselves, based on their respective roles in their agency or organization. At the other extreme, the facilitator took a very active role by leading the discussion and interjecting questions or prompts. In between were exercises in which the facilitator turned the discussion over to participants but occasionally joined the discussion to request clarifications from the participants or assure that issues critical to the exercise objectives were discussed. Despite these differences, all of the exercises shared common elements, including: evolving hypothetical scenarios, facilitated group discussions, and some level of collective decision making by participants emphasizing *< 100,000 = small, 100,000-1,000,000 = medium, > 1,000,000 = large; **Mild involvement-most of exercises was role played by participants, with very little intervention or direction from facilitators; Moderate involvement-most of exercises was role played or issue discussion, with the facilitator inserting additional probes and ensuring the discussion stayed on track; Active-most of the exercise was more discussion based, with facilitator asking questions or identifying issues that were subsequently discussed. the role of local health departments in recognizing and initiating a response to an emergency. The scenarios typically began with a single case report or series of case reports that heralded a nascent disease outbreak and required a public health assessment. These situations exercised the internal communication and coordination across disciplines within health departments as well as the communication and coordination with partner agencies and organizations such as health care facilities and emergency medical service agencies. Several exercises extended beyond this initial response and included scenarios that progressed days or weeks into an outbreak, requiring greater interactions between local-and state-level authorities and attention to health care surge capacity. Every exercise concluded with a "hot wash" in which participants discussed their collective performance, identified strengths and weaknesses, and when relevant, related their performance to experience with actual outbreaks or crises. In the latter exercises, participants were prompted to develop an initial 'action plan' that addressed key vulnerabilities identified in the exercise. The facilitators subsequently generated a written "After Action Report" (AAR) that summarized the exercise experience and highlighted the observed strengths and areas for improvement. In addition, participants completed exercise evaluation forms. These consisted of a series of structured and semistructured questions that asked participants to discuss what they learned during the exercise and to evaluate aspects of the exercise structure and conduct. For example, participants were asked to identify key gaps in preparedness that occurred during the exercise and to identify the most useful thing they learned during the tabletop exercise. The observations reported here are based on reviewing the after action reports, participant evaluations, as well as internal team discussions and consensus following the exercise debriefings. The performance of health departments that participated in our tabletop exercises varied from agency to agency. However, there were consistent themes that emerged across the agencies, regardless of the structure or the biologic agent/disease discussed; nearly all agencies struggled with a common set of challenges. These challenges, summarized in Table 3 and described below, represent critical dimensions of an outbreak response. Many local health departments did not have a structured process for notifying or soliciting case reports from health care providers in the community other than those in hospitals, largely because they did not have reliable contact information for private providers or a sure means to reach them rapidly. In most instances local health departments had good relations with staff in local hospitals (e.g., emergency department staff and infection control practitioners) but did not appear to have similar working relationships with non-hospital based practitioners. Local public health officials were sometimes unsure about their direct role in following up with suspected ill patients and collecting and shipping clinical samples for laboratory testing. For example, there was frequently confusion around whether it was the responsibility of the local health department, the state health department, or the medical personnel at the hospital to collect laboratory samples. Once the samples were collected there was often confusion around whose responsibility it was to transport the samples, and in a few sites, local law enforcement were surprised to find out that they were the responsible party. A related issue was the ability of health departments to realistically generate enough surge capacity in their public health workforce to investigate or respond to a large event, especially one that encompassed multiple jurisdictions in the same state, thereby limiting the state health department's ability to shift manpower and resources from one jurisdiction to the next. Few of the health departments in which we conducted exercises were proactive in their contacts with the media, and most waited until they were contacted by the media to begin communicating with the public. One consequence of this passive approach was that public health officials often responded defensively to early and sometimes unexpected media requests and in turn, had trouble quickly formulating an initial message to the public that was clear, informative, and alleviated anxiety. Health departments consistently expressed uncertainty about how to effectively communicate with vulnerable or underrepresented population groups in their jurisdictions, and few had well established relationships with community leaders or organizations that could serve as messengers or communications channels to these groups. In several sites, law enforcement and EMS personnel present in exercises had greater familiarity with these groups and could help identify trusted community messengers. Further, in some communities, health departments had limited language capacities or were not sufficiently familiar with community leaders to communicate effectively with these groups. Communicating fully and effectively with response partners (e.g., law enforcement and EMS) about their occupational health risks and personal protection was also a challenge for local public health. In particular, while public health officials were usually quick to notify response partners soon after determining an event to be significant, response partners in many cases felt that public health officials were slow to provide them with critical information about the disease in question, what their risks might be, or what actions they should take to protect themselves. As a result, response partners frequently reported feeling either left out of the process or expressed concerns about continuing to work unless the risks to them were clarified and more was done to ensure their safety on the job. The use of the National Incident Management System (NIMS) and its associated Incident Command Structure (ICS) structure is relatively new to public health. This was evident in the exercises, in that nearly all health departments had difficulties deciding if and when to implement the ICS process and in identifying the party who would serve as incident commander. Similar challenges were seen in the decision and processes related to opening an Emergency Operations Center (EOC). As a result, in many exercises, local public health officials delayed taking these steps and preferred maintaining a more informal management process. This approach was preferred even as the outbreak became progressively larger, thereby stressing these informal networks. As outbreaks evolved, there was often a lack of clarity about whether and when local health departments should hand off control to the state health department, how responsibilities should be jointly shared between local and state authorities, and whether or when federal agencies, such as CDC, should become involved. In many of the exercises, state health departments were surprised by the level of assistance requested by their local health departments especially in the early stages of the outbreak; in other more rare examples, state health departments surprised local health departments by assuming roles and responsibilities local health departments regarded as their own. Regardless, the general consensus among local public health participants in most exercises was that CDC staff would be on the ground to help them fairly quickly, particularly in situations where bioterrorism was considered likely. Most local health departments articulated some type of plan for increasing medical surge capacity by developing alternative care sites. In most instances however, these plans were unable to hold up to even a modest amount of scrutiny during the exercise because they were superficial and lacked sufficient detail necessary for rapid implementation. Related to this issue, local health departments frequently reported that there were not enough local health care workers to manage these sites even if they could create them. For example one participant noted, "We have pop-up tents and beds to increase capacity, we just don't have pop-up people to staff them." Even obtaining a census of available staff members turned out to be challenging as many health care participants noted that some staff would likely be double-counted, particularly nurses and security officers who might work in several institutions. Increasing staff capacity through the use of community volunteers, including retired medical personnel, while often recognized as one potential solution to staffing shortages, proved to be extremely difficult to actually implement. Public health participants universally recognized the importance of volunteers, but learned that their plans to recruit, train, and mobilize large numbers of volunteers were too vague and lacked concrete actionable steps for realistic application during a real emergency. In nearly all exercises, we also identified a number of strengths within participating health departments. However, there was far less commonality in these strengths than we observed with the areas for improvement. Universally, we observed public health leaders and staff who were committed and struggling to 'do the right thing.' The most commonly observed strengths were strong relationships between epidemiologists and hospital infection control practitioners and between public health workers and other emergency coordinators. In some instances, prior experience with emergency planning or response, such as involvement of health departments and emergency service agencies in coastal areas in preparing for or responding to hurricanes, was associated with stronger and more facile interactions between health department officials and partner agencies. Our exercises were conducted over a period of several years. While we did not conduct exercises with any health department more than once and did not employ an experimental design to assess changes over time, we were struck by how the performance of health departments overall improved over time. First, compared to earlier exercises, local health departments appeared far more sophisticated about their early internal processes related to notification, enhanced surveillance and large outbreak investigations. In addition, by the end of the exercise period, health departments had considered plans for surge capacity, and participating hospitals had explicit plans for cancelling emergency surgery and discharging less severely ill patients. They also appeared more acutely aware of the challenges in assuring adequate numbers of staff to provide care. The large number of tabletop exercises we conducted allowed us to test and compare different strategies for designing and conducting tabletop exercises. These comparisons enabled us to modify our exercises over time to build upon lessons learned from previous exercises. Below we briefly highlight five lessons we learned from this experience. Exercises should be designed to achieve a specific objective When first developing the tabletop exercises, our assumption was that a single exercise could achieve multiple objectives, such as training, relationship building, and evaluation. While these objectives are interrelated and opportunities often exist to achieve them concomitantly in the same exercise, it is critical to define the priority objective for the exercise because different objectives have different implications for exercise design. For example, if exercise participants outlined a response that was flawed or problematic, in an exercise primarily focused around the objective of training it would be appropriate for the facilitator to pause and help the participants re-think their approach. On the other hand, if the objective of the exercise is evaluation, this type of facilitator involvement can lead the participants to choose a different course of action and therefore bias the overall outcome being evaluated. Taken further, in an exercise designed to build relationships and links across disciplines or agencies, a facilitator intervention implying that a participant had made a mistake could be embarrassing or diminish that person's credibility, depending on the level of trust among participants. Exercises should be as realistic as possible while remaining logistically feasible Taken together, the optimal mix of design elements represents a balance between exercise objectives and logistic feasibility. The ideal balance is one that assures sufficient realism to provide a meaningful experience while minimizing distractions associated with the necessary artifice of exercise scenarios. Some departures from reality may be inadvertent if scenarios are developed with insufficient attention to local routines, forcing participants to sidestep usual procedures. Even seemingly minor design errors, such as using an outdated name for a hospital, or a time course for a disease that is inconsistent with its known epidemiology, can undermine the credibility of the exercise and can distract participants enough to take them out of their roles, thus disrupting the flow of the exercise. The desire for a realistic exercise scenario can lead to the development of tabletop exercises around scenarios rather than issue areas based on local preparedness needs and priorities. This does not ensure that the participants will address the important issue areas. A broad mix of challenges related to a given scenario must be addressed, often simultaneously, ranging from conducting epidemiologic or environmental investigations, implementing and modifying interventions as information becomes available, communicating within and across agencies, and communicating with political leaders and the public. Introducing this full set of tasks into an exercise scenario in a way that meaningfully exercises relevant capacities is unlikely to align with the exercise's objective. Moreover, different stakeholders may want to address different issue areas and may become frustrated if their expectations are not met. It is therefore important for stakeholders to agree on a limited number of priority issue areas for the exercise and then to focus the design of the scenario around these areas. For example, one set of exercises we designed focused on pandemic influenza preparedness in local health departments. Because it was infeasible to exercise the entire pandemic plan around a single scenario and in a single exercise, we developed the scenario and then the exercise by first meeting with local stakeholders to decide on the issue areas that would be covered in the exercise. These issue areas included disease surveillance, medical surge capacity, non-pharmacological disease control, and the use of antiviral medications. The scenario was then customized to unfold to deal with each of these issue areas. Key decisions for each discussion point were then developed, as well as facilitator probes and instructions based on the specific objective of the exercise. If not designed or facilitated properly, an exercise can lack focus and resolution, leaving participants to wonder what exactly was accomplished during the exercise. Therefore, it is paramount that exercises are designed to focus on issue areas that require concrete decisions over a limited period of time. For example, an exercise dealing with a simulated smallpox outbreak that unfolds over time might involve a discussion period dealing with movement restrictions in which participants are asked one or more questions such as, "Should schools be closed at this point?" Participants should then be given a limited amount of time to discuss this issue and make a decision. It is the facilitator's job to keep the discussion focused on the issue area and the specific question(s) at hand and to ensure that at the end of discussion, participants have collectively made the decision(s) they were tasked to make. Exercises can be designed to have multiple such issue area discussions as the scenario unfolds. Depending on the goals and objectives of an exercise, an exercise can involve a narrow or wide range of potential participants. While broader inclusion would likely be more realistic, such inclusiveness needs to be weighed against the logistics of effectively managing a larger number of participants and the potential adverse effects of inclusion. For example, participants may be less comfortable discussing ideas, taking risks, or making mistakes, depending on who is in the room. Such constraints may impede the exercise process and undermine achievement of exercise objectives. One solution to this problem would be to sequentially stage the involvement of different participants or to physically separate different groups in a way that more closely mimics actual situations. For example, some conversations that involve airing uncertainties or weighing difficult alternatives may normally involve a limited group of people, and members of that group may be more comfortable exercising such a conversation apart from colleagues from other agencies or organizations. The disadvantage of this approach is that it is substantially more difficult logistically and it diminishes the opportunity for people from different groups to gain an understanding of one another's role and approach to problems. In those exercises where certain participants, notably law enforcement, joined the scenarios at different stages, the feedback was generally critical, and participants felt that staging participation diminished learning and teambuilding opportunities. Another solution is to split exercise participants into two or more groups that allow everyone to participate, often placing people at similar levels of responsibility in the same group, and to conclude the exercise with a session that brings everyone together to share what they learned. The exercises described in this report represented collaborations between people familiar with local circumstances and people from outside the participating jurisdictions who had expertise in exercise design and facilitation. Because we did not test an alternative approach that exclusively involved local personnel, it is difficult to generalize from this experience about the value of engaging people from outside the participating agencies. Nonetheless, it was our impression that at certain points in the development, facilitation, and feedback steps, there was value in involving people who were not personally invested in local relationships or situations and who could offer seemingly independent advice or perspectives. Tabletop exercises can provide useful insights into both strengths and vulnerabilities in public health preparedness. It is important to recognize, however, that exercise outcomes are influenced by the way they are designed and conducted. The exercises described in this report emphasized varying dimensions of public health preparedness, reflecting differences in state or local priorities for prioritizing exercise objectives. For example, some emphasized the early response to initial reports of suspect illness while others emphasized management of surges in demand for health care services that are likely to occur later emergency scenarios. Given the intellectual and emotional demands of participation in an exercise, participants (and facilitators) may be less energetic during later rather than earlier stages of an exercise scenario, affecting perceived capacity to execute different elements of a response. Potential gaps between observed and actual preparedness should be considered in interpreting after-action reports and evaluating exercises themselves. The utility of tabletop exercises as tools to identify areas for improvement and make improvements on these problems is still evolving. Our ability to evaluate exercise performance is hampered by the lack of an evidence base about what constitutes optimal performance and by the lack of standards for assessing public health preparedness. There is a need to move beyond qualitative performance measures to ones that are quantifiable and can be measured over time. These quantifiable measures can range from simple checklists to Likert rating scales to scorecards. For example, in a series of our exercises we used checklists to assess the performance of health departments related to surveillance, risk communication and other functions. One fairly consistent observation was that health departments identified gaps that had been identified in prior exercises or actual experience, but had not yet been addressed. Reasons for this included lack of time, and lack of knowledge about how to make change. We now conclude exercises by having health departments prioritize the challenges observed during the exercise and then have them develop initial action plans related to up to three priority items. There are important limitations to our work and its interpretation that must be recognized. First, the nature of our exercises changed over time on a number of important dimensions, including the scenario, priority objectives, facilitation, exercise designers and facilitators, and attention to beginning an action plan after the hot wash. As a result of this variation, we are unable to provide a numerical tabulation of the numbers of health departments that struggled with each gap or displayed given strengths. Second, because we did not employ a methodology that could conclusively assess change over time, we cannot be certain that the improvements we identified were truly reflective of improvement, and not due to the inclusion of more sophisticated health departments in the latter part of our exercise period. We doubt this is the case, however, given the national emphasis on preparedness and planning and the ways in which health departments participating in later years qualitatively described their improvement. Furthermore, similar observations regarding improvements in public health preparedness during the same time period have recently been reported by others [24, 25] . We also cannot asses the potential influences that external events (e.g., hurricanes, outbreaks) may have had on health departments during the time period of our work, but it is noteworthy that all exercises were concluded before Hurricane Katrina struck. In addition, our exercises were not conducted in a random sample of health departments, and the findings may not be generalizable to all health departments. Finally, as discussed above, the evidence base for determining best practices in the design and conduct of exercises is extremely thin. We share our experience in the hope that it will help others, but do not propose that our recommendations constitute best, proven practices. Developing, conducting, and evaluating tabletop exercises requires considerable planning and the perspectives of a variety of stakeholders. While these tabletop exercises identified both strengths and vulnerabilities in emergency preparedness, additional work is needed to develop reliable metrics to gauge exercise performance, inform followup action steps, and to develop re-evaluation exercise designs that assess the impact of post-exercise interventions.
90
Transcript-level annotation of Affymetrix probesets improves the interpretation of gene expression data
BACKGROUND: The wide use of Affymetrix microarray in broadened fields of biological research has made the probeset annotation an important issue. Standard Affymetrix probeset annotation is at gene level, i.e. a probeset is precisely linked to a gene, and probeset intensity is interpreted as gene expression. The increased knowledge that one gene may have multiple transcript variants clearly brings up the necessity of updating this gene-level annotation to a refined transcript-level. RESULTS: Through performing rigorous alignments of the Affymetrix probe sequences against a comprehensive pool of currently available transcript sequences, and further linking the probesets to the International Protein Index, we generated transcript-level or protein-level annotation tables for two popular Affymetrix expression arrays, Mouse Genome 430A 2.0 Array and Human Genome U133A Array. Application of our new annotations in re-examining existing expression data sets shows increased expression consistency among synonymous probesets and strengthened expression correlation between interacting proteins. CONCLUSION: By refining the standard Affymetrix annotation of microarray probesets from the gene level to the transcript level and protein level, one can achieve a more reliable interpretation of their experimental data, which may lead to discovery of more profound regulatory mechanism.
Microarray technology was invented to rapidly profile the quantities of mRNA transcripts in a particular cellular context [1] [2] [3] . Its application has become universal in biomedical researches. Although it is mRNA that is actually detected by microarray experiments, and it is mRNA that has the direct relationship with protein, the methodology and algorithms for data analysis are commonly gene based. As evident in the probe annotation file provided by Affymetrix, gene-level annotation is widely accepted even though it fails to discriminate multiple mRNAs transcribed from the same gene. As a result, the analysis results are usually summarized at the gene level, such as differentially expressed genes [4] [5] [6] or gene-sets [7, 8] . Even in the recent works that integrate protein-protein interaction data and microarray data [9] [10] [11] , the analysis unit is reduced to the gene instead of mRNA. This practice could be attributed to the fact that most of the functional knowledge is at gene level instead of transcript level, and the functional differences between mRNA variants transcribed from the same gene are seldom clear. Although the gene-level analysis ignores the difference among mRNAs variants, this strategy is still biologically meaningful considering that the diversity of genes is much higher than that of transcripts encoded by the same gene. It has been well established that alternative splicing increases mRNA diversity, and over 60% of human genes are involved in this mechanism [12] . In addition, other RNA processing events, such as RNA editing, also account for the increased diversity at the mRNA level [13] . Since these events enable one gene to encode multiple proteins which might be functionally heterogeneous, we feel it necessary to separate transcript-level synonymous probesets from gene-level synonymous ones. The probesets that hybridize to more than one transcript variant of the same gene are referred to as gene-level synonymous probesets; while the ones that hybridize to a single variant are named as transcript-level synonymous probesets. It has been noticed that transcript-level synonymous probesets tend to have similar expression profiles, while gene-level synonymous probesets may have distinct expression profiles [14, 15] . In fact, individual reports demonstrated that the expression of transcript variants could be totally different [16, 17] . These phenomena indicate that the gene-level strategy of microarray data analysis is imprecise enough that one may overlook the expressional inconsistencies among gene-level synonymous probesets. As a matter of fact, Affymetrix suffixes their probeset ID according to the probeset's specificity. For example, probesets that recognize unique transcript variants are suffixed with _at, and probesets that recognize multiple alternative transcripts from a single gene are suffixed with _a_at or _s_at, and so on. This suffix system gives a hint on the varied specificities of the probesets, and could be considered as an endeavor trying to do away with customers' worry about the gene-level data analysis strategy. However, the correctness of the suffix system has been in doubt [14, 18] . Therefore it is not reliable to perform transcriptlevel analysis on the basis of this imperfect suffix system. As the standard annotation files and most of the analysis algorithms are gene-oriented, analysts often average out the expression heterogeneity of the same gene when dealing with probeset level data [19] . In this paper, we linked the probesets of two widely used Affymetrix arrays with the International Protein Indexes (IPIs) [20, 21] through proper association and rigorous alignment procedures, and demonstrated the statistically significant advantage of interpreting microarray data at the transcript-level or protein-level. Our results can be viewed as a more precise annotation of Affymetrix array's probesets, with which one may achieve a more reliable interpretation of their experimental data. Moreover, the application of this new annotation substantially increased the expression correlation between interacting proteins. Two Affymetrix arrays, MOE430A_2 and HG-U133A, were chosen for this study, in which 21,097 and 16,213 putative protein-coding probesets for the two arrays were subject to the alignment investigation (Table 1) . Candidate probeset-mRNA relationships were compiled based on the probeset-gene mapping information specified in the standard Affymetrix annotation files and gene-mRNA mapping relationships provided by separate mRNA transcript sources. Rigorous blast procedure similar to that described in [22] were performed to filter these candidate probeset-mRNA relationships. Finally, the mRNA targets passing the filtering criteria were linked to protein IDs in the IPI database. Through the rigorous association and alignment, we obtained precise annotations for 18,894 and 15,288 probesets in Affymetrix arrays MOE430A_2 and HG-U133A respectively (see Additional file 1). These annotations discriminate alternative mRNA variants transcribed from a same gene, thus are at transcript level as opposed to the standard gene-level annotation files provided by Affymetrix. It is worth noting that since the transcript data we used were quite redundant, a conceptual transcript variant may be represented by multiple redundant transcript accessions in the transcript database. In our transcriptlevel annotation file, each conceptual transcript is identified with one IPI ID, as we only investigated the probesets associated with protein-coding transcripts. Statistics on the non-control, investigated and annotated probesets, together with the number of involved genes and proteins, are shown in Table 1 . It is evident that the proportion of genes covered by our annotated probesets to those covered by all non-control probesets ('gene retaining percentage' in Table 1 ), 95.2% for MOE430A_2 and 85.4% for HG-U133A, are higher than the corresponding probeset retaining percentages, 83.5% and 68.8%, indicating that the gene coverage has only been slightly reduced by our filtering procedures. This observation is in support of our primary goal, that is to refine gene-level probeset annotations to transcript-level, but not to simply remove the poor-quality gene-level annotations. Furthermore, we classified the probesets based on the way they linked to proteins. By checking the alignment results, the probesets in our annotation tables were divided into two groups, namely one-to-one and one-to-many. In our annotation table, a one-to-one probeset was linked to only a single protein, while a one-to-many probeset was linked to multiple proteins due to alternative splicing of one gene. The rest of the investigated probesets that were not linked to any protein were categorized into the third group of "one-to-null". The statistics of these three groups are shown in Table 2 . As we know, Affymetrix suffixes their probeset ID according to the probeset's specificity. The over ten types of probeset suffixes can be categorized into three groups: transcript-specific, with '_at', gene-specific, with '_a_at' or '_s_at', and other suboptimal probesets that may cross-hybridize with unrelated sequences. The first two groups are comparable to our one-to-one and one-to-many probesets respectively. However, we found that, out of a total of 21,097 and 16,213 investigated probesets in MOE430A_2 and HG-U133A respectively (Table 1) , our one-to-one type accounts for only 53.7% and 43.3%, which were significantly lower than that of _at probesets, 74.0% and 72.6%, in the respective arrays. This suggests that some of _at probesets are not really specific for a transcript. To further clarify the issue of probeset specificity, we grouped the investigated probesets with the varied Affymetrix suffixes as well as our own categorizing system (Table 3 ). It is evident that overall one-to-one and one-tomany take up the majority of the _at group and the _a_at/ _s_at group respectively, but there are some disagreements between the two classifications. For example, the Affymetrix _at group has 30.1% one-to-many probesets and 11.7% one-to-null probesets, suggesting that these so-called 'transcript-specific' probesets do not have the expected high specificity, and that they should be treated with caution in data analysis. On the other hand, the Affymetrix _a_at/_s_at group contains 42.0% one-to-one probesets. We attributed this mainly to the trimming of Table 2 : One-to-one and one-to-many probesets in our annotation tables. Annotated probesets One-to-one probesets 11327 (53.7% of investigated) 7014 (43.3% of investigated) One-to-many probesets 7567 8274 Sum 18894 15288 One-to-null probesets 2203 925 Total (Investigated probesets) 21097 16213 the originally false transcripts during the update of sequence information, as they may lead to the absence of the transcript targets of some probesets. Similarly, there are a large number of one-to-one probesets in the socalled Affy others group, 1,206 out of 2,619 for MOE430A_2 and 476 out of 1,555 for HG-U133A. These probesets, however, might not be really one-to-one mapping to the identified transcript targets, as our strategy was based on the premise that most probesets were specific for certain genes, and our blast was limited to the candidate transcripts associated with the probesets, but not the entire transcript pool. These subgroups of the investigated probesets with different Affymetrix suffixes indicate the imperfection of the Affymetrix suffix system, thus affirming the necessity of our transcript-level or protein-level probeset annotations. In fact, several other research groups have addressed the misleading nature of the Affymetrix suffix system and the imperfection of its standard annotation file, including some re-annotation works for array HG-U133 [14, 18, 19, 23] . We will discuss these related works in detail in next section. Since the probesets were linked to transcripts and proteins through rigorous association and alignment procedures, the expression profiles of transcript-level synonymous probesets were supposed to be more consistent than those of gene-level synonymous probesets [14, 15] . This was taken as the basis for the evaluation of our annotations. We downloaded 30 expression datasets assayed with MOE430A_2 from the Gene Expression Omnibus database (GEO) [24, 25] , and calculated the Pearson correlation coefficients (PCCs) of expression profiles of synonymous probesets at gene level and transcript level. Since transcript-level synonymous probesets in our work recognized a single protein-coding transcript variant identified with a unique IPI ID, transcript-level synonymous probesets were also named as protein-level synonymous ones. Practically, we grouped gene-level synonymous probesets according to gene ID, and protein-level ones according to IPI ID. The probeset-protein mapping tables provided at NetAffx, the official protein-level annotation of Affymetrix probesets [26] , was used as comparison. Probesets corresponding to a single protein according to NetAffx were designated as 'Affy-protein' level synonymous probesets, which was a third level of synonymy. Within each synonymous group, all pair-wise PCCs were calculated and then summarized to one value indicating the expression correlation of this group. The expression correlations of all synonymous groups at one level were then averaged into an overall value for a dataset (see Method section for details). The correlations of synonymous groups for 30 datasets at three different levels were depicted in parallel in Figure 1A . It is noticeable that the synonymous probesets at Affy-protein level show higher correlations than those at gene level, but the synonymous ones based on our protein level annotation show even higher correlations consistently over 30 datasets (p < 0.05 for 30 datasets under student's t-test, see Additional file 2 for details). The comparison proves the rationality and necessity of our protein-level probeset annotation. These results support the argument that it is more reliable to interpret microarray data at transcript level than at gene level. The inferiority of the Affy-protein level annotation to our annotation could be attributed to the technical details in their alignment and association procedures [26] . First, they used the representative mRNA sequence ('consensus' or 'exemplar' sequence of each probeset), instead of the probe sequences themselves, as the query sequence in the alignment. Second, they aligned against the GenBank non-redundant protein database, rather than a comprehensive pool of mRNA sequences. In microarray experiment hybridization takes place between the probe sequences immobilized on the array and the cDNA sequences from the sample, so one can deduce that the alignment between the representative mRNA sequences and the protein sequences cannot precisely simulate the hybridization between probes and mRNAs. Finally, NetAffx filtered the blast results according to a cut- off of E-value, which indicates the likelihood of the observed alignment by chance [27] . Although frequently adopted for sequence homology analysis in closely related species, E-value is not sensitive enough to grade the many well aligned targets from the same species. In our practice, we took the probe sequences as the query and the mRNA sequences as the alignment targets, and adopted the matching nucleotide proportion as the filtering criterion (see Methods). A similar comparison was conducted for HG-U133A array, involving another transcript-level annotation by Harbig et al. [14] . Across the whole 28 datasets, the annotation by Harbig et al. showed advantage over Affy-protein-level annotation, while our transcript-level annotation performed best (p < 0.05 for 18 datasets under Student's t-test, see Figure 1B and Additional file 2). As our work was done two years later than Harbig et al.'s, the updated mRNA sequences in the probe-mRNA alignment is one of factors contributing to the increased performance. The other contributing factor is different approach we used to identify the mapping of probes to mRNA targets. Harbig et al. performed a two-phase blast: first, blast probeset target sequences against mRNA sequence pool, and then blast probe sequences against the retrieved mRNA sequences. Efficient as it was, this two-phase-blast strategy reduced the alignment precision as compared to our direct probe-against-mRNA blast strategy. Moreover, Harbig et al. accepted the mRNA with the highest average probe matches as the target of a probeset, even if the highest value could be suboptimal. According to our data, the puzzling fact that probesets for one gene may show variable expression profiles can be clarified when viewing the data at the protein level. Such genes are likely to involve alternative transcript variants which are constitutively different in expression levels. We looked at two genes as examples. Shown in Figure 2A , three probesets in GDS1277 dataset, 1448556_at, 1421382_at and 1451844_at, were mapped to the mouse Prlr gene (GeneID: 19116), with the first probeset correlating to protein IPI00321091, or PRL-R3, and the other two correlating to protein IPI00408593, or PRL-R2. The Prlr gene was reported to encode at least seven isoforms of prolactin receptor precursor in mouse [16] . The two probesets corresponding to the same protein IPI00408593, showed similar expression profiles as expected, with a PCC value of 0.6738 (P = 4.57e-6); while the probeset corresponding to another isoform, IPI00321091, showed a significantly different expression profile, with PCCs of 0.2530 and -0.0331 separately for 1421382_at and 1451844_at ( Figure 2A ). These data were consistent with the previous report that these two isoforms were predominantly expressed in liver and kidney, where PRL-R3 was highly expressed and PRL-R2 was weakly expressed [16] . We noticed that the shape of the expression curve of PRL-R3 seemed to be somewhat comparable to those of PRL-R2, although PRL-R3 probeset presented an overall higher expression level. These phenomena may suggest that PRL-R3 and PRL-R2 are partially co-regulated so that they show a comparable expression pattern, but they are expressed at different levels probably due to diverse roles of the regulatory elements. Shown in Figure 2B , three probesets in GDS1076 dataset, 1419114_at, 1419115_at and 1419116_at, were mapped to the mouse gene Alg14 (GeneID: 66789), with the former two corresponding to IPI00132168 and the latter one corresponding to IPI00405947. Both proteins were indicated in IPI as homologs of yeast asparagine-linked glycosylation 14 without any further information. We notice two interesting phenomena in this case. First, the two probesets correlating to a same protein do not show similar expression profiles. Instead, they behave like the probesets correlating to different variants characterized by a PCC value of 0.4080 (P = 0.0415). This issue might be due to some factors causing microarray hybridization efficiency shift. Secondly, according to our calculation, the expressions of these two variants are negatively correlated with PCC values of -0.7740 (P = 5.04e-5) for 1419116_at and 1419115_at, and -0.4402 (P = 0.0296) for 1419116_at and 1419114_at. So far there are no reports on expression regulation of these two transcript variants of gene Alg14, and no function reports of the corresponding protein isoforms. This negative correlation suggested that these two proteins might perform different roles thus should be distinctly annotated. In recent years, considerable efforts have been devoted to identifying and characterizing protein-protein interaction (PPI). Besides investigations on the molecular events involved in PPI, functional annotation of an unclassified protein according to its interacting partners is also an important topic [28] . Since it is too bold to infer protein functions according to the "majority rule" that utilizes only the PPI network structure [29, 30] , many studies integrate other data sources into the functional characterization of PPI, among which the gene expression data is the favorite [9, 31, 32] . All these works assumed that interacting protein pairs were characterized with higher expression correlation than random ones. However, previous investigations indicated that the relationship between expression correlation and PPI was weak on a genomic scale [33] [34] [35] although a recent work strengthened the association by integrating cross-species conservation information [10] . We noticed that in these genome-scale studies PPI information was always first converted to gene pairs, after which the Pearson correlations of the probeset pairs corresponding to the gene pairs were evaluated. That is, the analysis targets were expanded from real interacting protein pairs to all possible cross-gene protein pairs for which interaction may not always exist. As illustrated in Figure 3 , suppose we have gene a (abbreviated to Ga) and gene b (Gb), with Ga encoding protein a1 (Pa1) and protein a2 (Pa2), Gb encoding protein b1 (Pb1) and protein b2 (Pb2). Among these protein variants, only proteins Pa2 and Pb1 interact with each other, while the other three possible cross-gene interactions, including Pa1-Pb1, Pa1-Pb2 and Pa2-Pb2, do not really happen. The four probesets, Pst_a1, Pst_a2, Pst_b1, and Pst_b2, recognize transcript variants Ta1, Ta2, Tb1 and Tb2 respectively, producing proteins Pa1, Pa2, Pb1 and Pb2. In the conventional genome-scale studies mentioned above, besides the probeset pair (Pst_a2, Pst_b1) corresponding to the real interacting protein pair, the other three cross-gene pairs, (Pst_a1, Pst_b1), (Pst_a1, Pst_b2) and (Pst_a2, Pst_b2), were also included, which would blunt the expression cor-Expression profiles of synonymous probesets Figure 2 Expression profiles of synonymous probesets. A) Three probesets on the array MOE430A_2 are associated with the mouse gene Prlr. Probeset 1448556_at correlates to protein IPI00321091, or PRL-R3; probesets 1421382_at and 1451844_at both point to protein IPI00408593, or PRL-R2. GEO dataset GDS1277 was analyzed. B) Three probesets on the array MOE430A_2 are associated with the mouse gene Alg14. Probesets 149114_at and 149115_at both point to protein IPI00132168; probeset 1419116_at correlates to protein IPI00405947. GEO dataset GDS1076 was analyzed. relation between the real interacting entities according to our preceding observations. We propose that this might partly explain the weak coherency between PPI and expression correlation. In order to validate our proposition, we investigated the relationship between protein interactions and expression correlation at both gene-level and protein-level perspectives, using 1,037 interacting protein pairs from the HPRD [36] and 28 microarray datasets assayed with HG-U133A from the GEO. As depicted in Figure 3 , the probeset pairs corresponding to all possible cross-gene protein pairs are termed as GGI pairs, out of which only the probeset pair corresponding to the real PPI, such as Pst_a2-Pst_b1, is a PPI pair. We calculated PCCs of both PPI pairs and GGI pairs for all available 1,037 interactions, evaluated the statistical significances of these PCC values under one-tailed t-test, and estimated the corresponding false discovery rates (FDR) using the SPLOSH FDR estimation method [37] . For all datasets, we observed strengthened expression correlations between interacting proteins under the PPI schema relative to the GGI schema. Taking the GDS987 dataset including 41 arrays as an example ( Figure 4A 153, respectively, and the former was significantly smaller than the latter (Kolmogorov-Smirnov test p value is 9.9E-12). Summarizing the comparisons at the positive side and the negative side, we conclude that the expression correlation between interacting proteins is strengthened when the non-interacting protein pairs are excluded from the interacting gene pairs, that is, with the PPI PCC calculation in place of the conventional GGI PCC calculation. In Figure 4A , we notice that the negative correlation is also strengthened by the PPI PCC calculation as well as the positive correlation. That is to say, PPI pairs seem to be either positively correlated or negatively correlated, but not exclusively 'co-expressed' as previous publications reported [10] . This phenomenon is more evident when we examine the PCC values for each coupled PPI pair and GGI pair. Figure 4B shows a scatter plot of the PPI PCC values versus the corresponding GGI PCC values. It is evident that most points fall into the 1st and the 3rd quadrants, indicating that each pair of PPI PCC value and GGI PCC value tends to have the same signs. The scatter plot suggests a linear relationship between the PPI PCCs and GGI PCCs, and indeed we get a linear regression formula, y = 0.5612x + 0.0046, at high confidence (p < 2e-16). Since the estimated coefficient, 0.5612, is far less than 1, we may conclude that the absolute PCC values of PPI pairs are often larger than those of the corresponding GGI pairs. So the PPI PCC calculation preserves the original positive or negative correlation tendency revealed by the conventional GGI PCC calculation, and strengthens it with larger absolute correlation values. Such correlation tendencies between interacting proteins, especially those negative ones, would very likely be submerged under the Figure 3 Illustration of Entities and terminologies involved in protein-protein interaction. A) Relationship among genes, transcripts, proteins and probesets. B) Two types of pairs mentioned in the text. PPI pair indicates the probeset pair correlating to a real interacting protein pair; while GGI pairs are those probeset pairs correlating to all possible cross-gene pairs. PPI may not exist in all cross-gene pairs. background correlations of random pairs if the non-interacting protein pairs are included in the analysis. Similar observations were made over all datasets (see Additional file 4 and Additional file 5). Given the results from all 28 datasets, we were also able to compare the PPI PCC calculation and GGI PCC calculation at a higher level. Under a FDR threshold of 0.1, the list of significant PPI pairs or GGI pairs from each dataset was determined. Based on these lists, we counted the significantly correlated PPIs or GGIs (termed 'significant PPIs' or 'significant GGIs') over each dataset, and the datasets on which a PPI or GGI demonstrates significant correlation (termed 'PPIsignificant datasets' or 'GGI-significant datasets'). The former is a 'twenty-eight by two' table, shown in Table 4 ; the latter is a '1,037 by two' table, shown fully in Additional file 4 and partly in Table 5 . In both summary tables, we find that the statistics of PPI PCC calculation are mostly larger than the counterparts of GGI PCC calculation, indicating that PPI PCC calculation can detect more correlated PPI pairs in a certain experiment setting ( Table 4 ), and that it can detect the correlation tendency across more experiment settings ( Table 5 ). The same experiments were also implemented with 274 PPI pairs extracted from the IntAct database [38] , and sim- Protein pairs with the false discovery rate of PCC value less than 0.1 were deemed as significantly correlated pairs. 'Significant PPIs' are pairs detected by the PPI calculation; 'Significant GGIs' are those detected by the GGI calculation. ilar conclusions were obtained. More details can be found in Additional file 6 and Additional file 7. In this work, we re-annotated the probesets of two widely used Affymetrix arrays, MOE430A_2 and HG-U133A, via proper association and rigorous alignment procedures in a transcript perspective, and demonstrated the necessity and advantage of exploring microarray data at the transcript or protein level, instead of the conventional gene level. Although Affymetrix utilized the most complete information available at the time of array design, tremendous progress in genome sequencing and annotation in recent years renders existing probeset designs and target identifications suboptimal. In recent years, there have been continuous reports on systematic false expression signals of Affymetrix probesets [39] , spurious expression correlation caused by cross hybridization [18] , and expressional inconsistency among different microarray platforms or even different generations of one platform [40] [41] [42] [43] . A few research groups performed probe-against-mRNA blast similar to ours [22, 42, 44] , but mostly they centered around UniGene [45] and therefore improved the accuracy of annotation only at gene level. A major trend among these efforts was to redefine probesets so that probes matching the same molecular target were placed into custom probesets, as proposed by [19, 23, 39, 42] , but as the authors of [19] pointed out, 'these transcript-targeted probesets are not transcript-specific, as probesets targeting transcripts from the same gene may share many or even all probes'. Thus the probe re-organization strategy may be used to make distinction at the level of genes, but not at the level of transcripts or splice variants [18] . Besides, this strategy takes the probe-level intensity file (the CEL file) as a prerequisite, however only around half of the expression datasets deposited in public databases like GEO were found with CEL files. In order to make distinction precisely at the transcript level, we preserved the classical Affymetrix probesets, but distinguished them among their alternatively spliced transcript targets according to the consistent alignments of probes against up-to-date mRNA sequences. Our annotation table clearly divides the Affymetrix probesets into three groups with increased transcript-level specificity (reliability): one-to-null probesets that do not recognize any transcript, one-to-many probesets that hybridize to multiple alternative transcript variants of the intended gene, and one-to-one probesets that hybridize to unique alternative transcript variants of the intended gene. We IPI1 IPI2 Gene1 Gene2 PPI-significant datasets GGI-significant datasets IPI00003326 IPI00015161 ARL2 PDE6D 15 7 IPI00003894 IPI00019930 RNF11 UBE2D1 17 1 IPI00007411 IPI00021831 AKAP11 PRKAR1A 15 2 IPI00008529 IPI00008527 RPLP2 RPLP1 25 5 IPI00011118 IPI00026689 RRM2 CDC2 17 16 IPI00013871 IPI00011118 RRM1 RRM2 14 10 IPI00015952 IPI00029012 EIF4G2 EIF3S10 14 3 IPI00016910 IPI00290460 EIF3S8 EIF3S4 16 10 IPI00018350 IPI00184330 MCM5 MCM2 16 15 IPI00021700 IPI00026689 PCNA CDC2 14 13 IPI00022865 IPI00026689 CCNA2 CDC2 16 15 IPI00026689 IPI00015105 CDC2 CKS2 15 13 IPI00027462 IPI00007047 S100A9 S100A8 23 17 IPI00028266 IPI00026689 CCNB2 CDC2 17 15 IPI00165506 IPI00025616 POLDIP2 POLD2 14 2 IPI00246058 IPI00025277 PDCD6IP PDCD6 16 1 IPI00290461 IPI00029012 EIF3S1 EIF3S10 15 4 IPI00291006 IPI00018206 MDH2 GOT2 16 11 IPI00294696 IPI00026689 CCNB1 CDC2 17 14 IPI00306708 IPI00026689 PBK CDC2 16 14 IPI00328118 IPI00026689 SPAG5 CDC2 15 10 IPI00604664 IPI00291006 NDUFS1 MDH2 18 10 IPI00647217 IPI00552920 SKIV2L2 EXOSC8 18 2 A PPI-significant dataset is a dataset where the false discovery rate of the PCC value for the PPI pair is less than 0.1, while a GGI-significant dataset is a dataset where the false discovery rate of the PCC value for the GGI pair is less than 0.1. Only PPI pairs with more than 14 'PPI-significant datasets' datasets are shown. The full table is provided in Additional file 4. discriminate the intended alternative transcript variants of Affymetrix probesets based on the NetAffx's gene-level annotation for the first time. Given the fact that existing solutions are accompanied with imperfections and no consensus has been reached on an overwhelming strategy, our alternative solution to the problematic standard annotation points out a new way to improve the interpretation and exploitation of Affymetrix microarray data. Although the transcript collections were not identical and the reannotation strategies differed more or less, we made out some similar discoveries to previous reports. For example, Harbig et al. found that a number of probesets did not detect any transcript and attributed this phenomenon to the elimination of the target sequence in the process of sequence update [14] . In our study, altogether 16.5% and 31.2% of non-control probesets in the MOE430A_2 and HG-U133A arrays were not found with any transcript targets in the pool of GenBank, RefSeq and Ensembl. Using newer and larger collection of transcript sequences, we even obtained a quite similar statistics of the percentage of 'multiple-targeting' probesets to that estimated in a foregoing study [18] , specifically 54.6% for MOE430A_2 and 54.1% HG-U133A (see Table 2 ). The significant mutual agreement among the related researches justifies the necessity to set up an improved annotation mechanism of the Affymetrix probes in the face of the continual growth of genomic and transcriptomic knowledge, ideally at transcript or protein level. Over the past few years, the analysis of alternative splicing has emerged as an important new field in bioinformatics, and several recent large-scale studies have shown that alternative splicing can be analyzed in a high-throughput manner using DNA-microarray methods [46, 47] . Most of these studies used arrays particularly manufactured for analyzing alternative splicing, such as genomic tiling array and exon-exon junction array. Constructed without any priori knowledge of the possible exon content of a genomic sequence, the genomic tiling array [48, 49] is in principle capable of detecting novel alternative splicing events of diverse types, but it is in doubt whether their data will be readily interpretable as successful experiences remain insufficient [46] . On the other hand, although designed particularly to address the alternative splicing issue, exon and exon-exon junction arrays [49] were reported to be plagued by problematic probe specificity and unsatisfying hybridization efficiency because of the necessity of probe coverage across the full length of the gene (including 5' end) [5] . Many questions about the reproducibility of the amplification protocol, the quantitative accuracy, and the data analysis need to be addressed as a prerequisite to reliable quantitative analysis using these splicing arrays [50] . Given the current imperfection of splicing array techniques and inconvenience in deci-phering their generated data, it is an economic way to do large-scale investigations of alternative transcribing events with standard gene expression arrays, provided that the recognizing targets of the probes can be rigorously defined at the transcript level. Hu et al. proposed a primitive analysis method to explore alternative splicing with Affymetrix 3' gene expression arrays, though they regretted that only alternative splicing biased toward the 3'end of the gene can be detected in their way [51] . In the present paper, we conducted a large-scale alignment of the probe sequences in traditional gene expression arrays against the currently most comprehensive collection of transcript sequences, highlighting the probesets mapping to unique alternative transcripts unambiguously. For each of the two Affymetrix expression arrays tested in this study, we found over 40% of all probesets could be mapped to transcripts in a oneto-one manner, so our work strongly validate that it is feasible to analyze alternative splicing using traditional gene expression arrays. While the foregoing work contributed by Hu et al. remains as a qualitative analysis method aiming at detecting novel alternative splicing events, our work gives explicitly the relationship between the probesets and the currently known alternative transcript variants, which can be immediately exploited to facilitate quantitative analysis of alternative variants. As our mapping relationships are defined for the standard probesets of the traditional gene expression arrays, they can be conveniently exploited as the standard NetAffx annotation information, without any ad hoc influence on the widely applied experiment protocols or the routine data processing algorithms. In the demonstrative implementations of the novel annotation tables, we actually observed several examples of negatively correlated alternative variants (see Figure 2B for one of them), which will shed light on further studies of expression regulation of alternative transcript variants. To sum up, we re-annotated two popular Affymetrix gene expression arrays, MOE430_2 and HG-U133A, in a transcript-level perspective, aiming at identifying probesets' detecting targets precisely at the transcript level. Although previous works addressed similar issues [14, 15, 18, 19, 22, 23] , we are the first to rigorously link existing Affymetrix probesets to their specific transcript targets and their corresponding proteins. Armed with this new annotation, we re-examined a number of previous studies, 30 datasets for MOE430_2 and 28 datasets for HG-U133A from GEO, and revealed increased expression consistency among synonymous probesets and closer expression correlation among interacting proteins. This transcript-level annotation of Affymetrix probesets allows for a more reliable gene expression data analysis and a more accessible protein-level correlation study. The Affymetrix 3' eukaryotic gene expression analysis arrays MOE430A_2 and HG-U133A were selected for this study. Probe sequence files and corresponding annotation files, 'Mouse430A_2_annot_csv.zip' (annotated on 2005-12-19) and 'HG-U133A_annot_csv.zip' (annotated on 2006-04-11), were downloaded from Affymetrix website [52] . Also downloaded there were the NetAffx probesetprotein mapping files for MOE430A_2 (file 'Mouse430A_2_blast_csv', updated on 2005-12-18) and HG-U133A ('HG-U133A.na21.blast.csv.zip', updated 2006-04-11), which were the blast results of the representative mRNA sequence of probes against protein sequence databases [26] . Ensembl transcripts: 37,854 mouse transcript sequences were obtained from the Ensembl database (release 38) [56] . Mapping tables between Ensembl Gene ID, Ensembl Transcript ID, and Ensembl Peptide ID were obtained from Ensembl martview [57] . IPI entries and their mappings to external protein accession numbers were acquired from the International Protein Index (IPI) database [21] (release 3.17). Also obtained there were the mapping relations between IPI numbers and transcript IDs (GenBank, RefSeq, and Ensembl). The counterpart file for human was downloaded there too (release 3.16). Microarray datasets were downloaded from the Gene Expression Ominibus on April 15, 2006 . Array MOE430A_2, indexed as GPL339, was associated with 2,276 samples in GEO, ranking the second among all registered Affymetrix mouse arrays. All 31 GDS datasets profiled with MOE430A_2 were used in the analyses except for GDS1057, which contains only two samples. Array HG-U133A, indexed as GPL96, was associated with 8,698 samples in GEO, ranking the first among all registered Affymetrix human arrays. For our analysis, we downloaded 31 GDS datasets with largest sample sizes, and used 28 of them in our analyses, excluding GDS534, GDS1329, and GDS1324 as they are in a data format inconsistent with the others. Details about the used datasets can be found in our Additional file 8. Out of the total 22,690 and 22,283 probesets in arrays MOE430A_2 and HG-U133A, respectively, 64 and 68 control probesets were firstly removed. The remaining probesets were associated with genes according to the probeset-gene mapping information provided in Affymetrix's standard annotation file. The probeset-transcript mapping relationships were obtained based on the gene-mRNA mapping tables from GeneBank, RefSeq and Ensembl. In the process, we only included probesets that were identified with one Entrez Gene ID or one Ensembl gene ID. We ignored the probesets that were associated with multiple entities or no entity in the two gene-centric databases, since their gene-level specificity appears doubtful in the standard annotation file. This filtered out 3.2% and 5.2% of non-control probesets in MOE430A_2 and HG-U133A, respectively. For the rest of the probesets, we linked the candidate transcript targets to their corresponding protein entries in IPI database. IPI is currently the least redundant yet most complete protein database for featured species, with one protein sequence matching each transcript variant. Those probesets of which transcript tar-gets do not have any protein counterparts were also excluded from the following blast validation in order to focus our attention to the transcripts with well-characterized functions at protein level. The remaining probesets, 21,097 for MOE430A_2 and 16,213 for HG-U133A, were selected for the BLAST procedure. We then filtered the candidate probeset-mRNA mapping relationships by aligning probe sequences in these probesets against their corresponding transcripts. Probes were blasted against their candidate mRNA targets using the bl2seq program [61] , and the probe to transcript matches were accepted if no more than one mismatch was found. At the level of probesets, the probeset to transcript matches were accepted only if more than 90% of all probes within a probeset (over 10 probes for the typical 11-probeset) were mapped to the transcript in the same orientation. The probeset-transcript-protein links related to the above probesets passing BLAST filter were retrieved. After reducing the redundancy information of multiple transcripts corresponding to the same IPI, we finally obtained rigorous probeset annotation files for Affymetrix arrays MOE430A_2 and HG-U133A. There are two types of probesets in the new annotation file: one-to-one probesets, where one probeset maps to only one IPI ID; and one-to-many probesets, where one probeset maps to two or more IPI IDs. Only the one-to-one probesets were used in the subsequent analyses. We grouped gene-level synonymous probesets according to gene ID (gene-level), and protein-level ones according to IPI ID (protein-level). Additionally, probesets corresponding to a single protein according to NetAffx probeset-protein mapping tables (see Materials) are grouped as 'Affy-protein' level synonymous probesets. In the case of the HG-U133A array, we included a fourth level, the 'Harbig-protein level', for comparison. The Harbig-protein level reflects the probeset-protein association transformed from a recent alternative annotation of the Affymetrix U133 plus 2.0 array [14] , also proposed in a transcript-level perspective. The downloaded annotation file mapped 33,579 probesets to 287,791 GenBank mRNA accessions, among which 21,669 were found on HG-U133A array, mapped to 186,085 GenBank mRNA accessions. The probeset-mRNA associations related to HG-U133A array were further linked to IPI IDs, finally giving rise to 26,960 probeset-IPI mapping relationships among which 12,146 were one-to-one. The following treatments were the same as those implemented to the standard gene-level annotation, the NetAffx Protein annotations, and our novel annotations. 30 and 28 expression datasets were selected from GEO respectively for MOE430A_2 and HG-U133A, and the original intensity data within each GEO dataset (GDS) were transformed to log 2 base and normalized to a constant median across all samples. For a synonymous group, if the expression values of all probesets in all samples were no larger than the constant median value, the probesets in this group were regarded not moderately expressed, and their expression profiles not informative enough. Therefore, we only kept the synonymous groups with at least one expression value above the constant median value, similar to the filtering procedure used by Tian et al. [62] . For the remaining synonymous groups, Pearson correlation coefficients (PCCs) were calculated for the expression profiles of each probeset pair. The minimum value of these PCCs was taken as a measurement of the expression coherence of this group. We used the minimum aggregation because the gene level synonymous probesets gave rise to within-protein PCCs (which are theoretically higher) and across-protein PCCs (which are theoretically lower), and the former was identical to the result of the protein-level synonymous group. In such a setting, the maximum did not result in any difference, and the average aggregation was not as sensitive as the minimum in terms of differentiating the two groups, so we adopted the minimum aggregation. Finally, the mean of the expression coherences of all synonymous groups over a dataset was calculated. In this way we obtained an evaluation of expression consistencies within synonymous probesets for a microarray dataset, and may compare the expression consistencies at the three levels over different microarray datasets. Given protein-protein interaction data from HPRD or IntAct, we first transformed the binary relations of protein accessions to IPI-IPI pairs, and also got the corresponding Gene-Gene pairs. For each PPI, we assembled the PPI probeset pairs and the GGI probeset pairs as illustrated in Figure 3 , where PPI pairs are those associated with the interacting IPI IDs while GGI pairs are those associated with the corresponding Gene IDs. For all probeset pairs associated with the IPI-IPI pair (PPI pairs) and those associated with the corresponding gene-gene pair (GGI pairs), the PCCs were calculated and averaged into a PPI PCC and GGI PCC, respectively. These PCCs of interacting pairs were further calculated to obtain the accompanying false discovery rates using the SPLOSH FDR estimation method. The distributions of the PPI PCCs and the GGI PCCs were plotted in a same figure to show the contrast ( Figure 4A ). In addition, a background distribution of the PCCs of ran-dom probeset pairs was overlaid on the same figure. We let the number of random pairs equal to the number of PPI or GGI pairs, but repeated the process of calculating random PCC distribution 20 times and averaged over the 20 separate random distributions in order to cut down on random fluctuation. Within each run of calculating random PCC distribution, we randomly compiled 2 × n (n = 1037 or 274, for HPRD or IntAct, respectively) pairs of probesets, where each two probeset pairs formed a group. The two PCCs of each group were firstly averaged into a group-level PCC, and the distribution was calculated over the n group-level PCCs. The group-level averaging was devised to mimic the counterpart operation in PPI PCC or GGI PCC calculation.
91
Selective redox regulation of cytokine receptor signaling by extracellular thioredoxin-1
The thiol-disulfide oxidoreductase thioredoxin-1 (Trx1) is known to be secreted by leukocytes and to exhibit cytokine-like properties. Extracellular effects of Trx1 require a functional active site, suggesting a redox-based mechanism of action. However, specific cell surface proteins and pathways coupling extracellular Trx1 redox activity to cellular responses have not been identified so far. Using a mechanism-based kinetic trapping technique to identify disulfide exchange interactions on the intact surface of living lymphocytes, we found that Trx1 catalytically interacts with a single principal target protein. This target protein was identified as the tumor necrosis factor receptor superfamily member 8 (TNFRSF8/CD30). We demonstrate that the redox interaction is highly specific for both Trx1 and CD30 and that the redox state of CD30 determines its ability to engage the cognate ligand and transduce signals. Furthermore, we confirm that Trx1 affects CD30-dependent changes in lymphocyte effector function. Thus, we conclude that receptor–ligand signaling interactions can be selectively regulated by an extracellular redox catalyst.
Disulfide bonds have long been recognized as structural elements stabilizing proteins in harsh extracellular environments. More recently, an additional concept has emerged: some disulfide bonds operate as dynamic scaffolds capable of regulated rearrangement into a variety of functional forms (Jordan and Gibbins, 2006) . Consistent with this notion, various cell surface processes have long been known to depend on catalyzed thiol-disulfide exchange including cell adhesion (Essex, 2004) , uptake of bacterial toxins (de Paiva et al, 1993) and viral fusion with the host membrane (Sanders, 2000) . Moreover, a variety of cell surface signaling receptors appear to exist in more than one thiol-disulfide configuration, for example CD28 (Greene et al, 1996) . However, in most cases, neither the catalyst driving thioldisulfide exchange nor the functional differences between the redox forms have been elucidated. A number of thiol-disulfide oxidoreductases are known to be secreted and to act on the cell surface. One of these redox catalysts is protein disulfide isomerase (PDI), a member of the thioredoxin (Trx) superfamily. Cell surface-PDI has been found to act on transmembrane and surface-associated proteins, including the envelope protein of HIV-1, to cause its fusogenic conformation (Markovic et al, 2004) and integrins, to mediate platelet adhesion (Lahav et al, 2003) . Another thiol-disulfide oxidoreductase associated with extracellular functions is Trx1. Best known for its intracellular roles, Trx1 reduces transiently formed disulfide bonds of cytosolic and nuclear target proteins and thereby participates in a multitude of fundamental processes, ranging from oxidant scavenging and DNA synthesis to regulation of apoptosis and cell proliferation (Powis and Montfort, 2001) . In addition, Trx1 is released to the extracellular environment by a variety of normal and neoplastic cells (Rubartelli et al, 1992) . Human Trx1 was first purified as a cytokine-like factor from supernatants of virally transformed lymphocytes and initially named adult T-cell leukemia-derived factor (Tagaya et al, 1988) , Tac-inducing factor (Tagaya et al, 1989) , B-cell stimulatory factor or 'B cell IL-1' (Wakasugi et al, 1990) . Extracellular Trx1 is present in the circulation of healthy subjects and its levels increase under inflammatory conditions, including viral infection (Nakamura et al, 2001a) . Circulatory Trx1 acts as a chemoattractant for monocytes, neutrophils and lymphocytes (Bertini et al, 1999) , and inhibits neutrophil migration into inflammatory sites both in vitro and in vivo (Nakamura et al, 2001b) . More recently, Trx1 was found to be secreted by dendritic cells upon cognate T-cell recognition and to contribute to subsequent T-cell activation (Angelini et al, 2002) . At present, the mechanism(s) and pathway(s) by which extracellular Trx1 influences cellular behavior remain unknown. As many of its reported extracellular activities depend on a functional active site, it appears likely that Trx1 catalyzes thiol-disulfide exchange in one or more cell surface target proteins through its enzymatic activity. However, thiol-disulfide exchange reactions, even if highly specific, are too transient to be detected by conventional techniques. To date, only a single cell surface receptor, CD4, a member of the immunoglobulin superfamily, has been shown to be susceptible to Trx1 redox activity (Matthias et al, 2002) . Other cell surface proteins targeted by the enzymatic activity of Trx1 await identification. In this study, we address the question as to which cell surface receptors expressed on lymphocytes specifically interact with extracellular Trx1 by way of disulfide bond exchange. Using a kinetic trapping technique that enables the detection and isolation of otherwise short-lived reaction intermediates on the surface of intact cells, we identify and validate the tumor necrosis factor receptor superfamily member CD30 (TNFRSF8) as the principal target molecule for Trx1 on infected and transformed lymphocytes. The cell surface activity of Trx1 is highly selective, discriminating between different members of the TNFR superfamily. Trx1-mediated thiol-disulfide exchange leads to a structural change in the CD30 ectodomain that can be detected with conformationsensitive antibodies. We demonstrate that disulfide exchange between Trx1 and CD30 interferes with binding of the CD30 ligand (CD30L) to its cognate receptor and that Trx1 affects CD30-dependent changes in lymphocyte effector function. As CD30 is implicated in both stimulatory and apoptotic signaling, our findings suggest that Trx1 interacts with CD30 to modulate lymphocyte behavior and survival under conditions of infection and inflammation. To identify candidate Trx1 target proteins on the cell surface of lymphoid cells, we applied a trapping technique based on the reaction mechanism. This approach makes use of the finding that mutant thiol-dependent oxidoreductases lacking the C-terminal cysteine of the CXXC active site motif form long-lived mixed disulfide intermediates with target proteins. Thus, target proteins remain covalently linked to the mutant oxidoreductase and become amenable to isolation and analysis (principle shown in Figure 1A ). Kinetic trapping has been applied previously to identify interaction partners of Trx family members in plants (Motohashi et al, 2001) and in the secretory pathway of human lymphocytes . In these studies, the CXXC-based trapping technique identified both established and novel target proteins, subsequently confirmed by independent techniques, demonstrating the competence of this technique to identify bona fide interaction partners. To determine whether kinetic trapping can be applied to human Trx1, we created recombinant wild-type and mutant Trx proteins, each equipped with a C-terminal dual affinity tag composed of a streptavidin-binding peptide (SBP) and a hexahistidine tag. To create a trapping mutant, the second cysteine of the 32 CXXC 35 motif was replaced by serine (C35S). Trx1 harbors three additional cysteine residues distal to the active site (cysteines 62, 69 and 73). As these residues are dispensable for catalytic activity but may cause oxidative inactivation by either intra-or intermolecular disulfide bond formation (Casagrande et al, 2002; Watson et al, 2003) , we also created mutants containing amino-acid substitutions for those additional cysteines ( Figure 1B ; CCCCC, CCAAA, CSCCC, CSAAA and SSAAA annotate the identity of residues 32, 35, 62, 69 and 73). To test whether Trx1(C35S)-based trapping is capable of identifying known Trx1 target proteins, Trx1(CSAAA) was allowed to react with cytosolic proteins released from digitoninpermeabilized cells. Incubation led to the formation of a reproducible pattern of distinct mixed disulfide conjugates as visualized by silver staining of a SDS-PAGE gel under nonreducing conditions ( Figure 1C , lane 7). In accordance with the trapping mechanism, conjugation strictly depended on the availability of the N-terminal thiol (Cys-32) and the concurrent absence of the C-terminal thiol (Cys-35), as wild-type or cysteine-free Trx1 did not capture any proteins ( Figure 1C, lanes 3, 5 and 9 ). The pattern of trapped proteins was not significantly influenced by the presence or absence of the non-catalytic cysteines (data not shown). The main cytosolic interaction partner of Trx1(CSAAA) was identified as peroxiredoxin-1 (Prx1) by liquid chromatography tandem mass spectrometry (LC-MS/MS). The Trx1-Prx1 association was further confirmed by immunoblotting (data not shown). The Trx1-Prx1 disulfide-linked conjugate is maintained under non-reducing conditions ( Figure 1C , lane 7) and cleaved into its monomer constituents under reducing conditions ( Figure 1C , lane 8). To test whether Trx1(CSAAA) would also undergo authentic interactions under conditions more typical of an extracellular environment, we allowed Trx1(CSAAA) to react with human plasma proteins. To avoid nonspecific absorbance to high-abundance serum proteins, we applied Trx1(CSAAA) to a o30 kDa plasma ultrafiltrate, leading to the capture of a small number of proteins, as visualized by colloidal Coomassie staining ( Figure 1D ). Using LC-MS/MS, the principal interaction partner from the plasma ultrafiltrate was identified as peroxiredoxin-2 (Prx2), a well-established target protein of Trx1, recently found to be present in human plasma (Chen et al, 2004) . These experiments provided proof-of-principle evidence that kinetic trapping is capable of capturing and identifying proven target proteins of human Trx1 from both intra-and extracellular environments. To further investigate whether the capture of proteins by Trx1(CSAAA) is Trx-specific, we directly compared Trx1(CSAAA) with the corresponding trapping mutant of another member of the Trx superfamily, glutaredoxin-1 Grx1(CSAAA) (CSAAA annotates the identity of residues 22, 25, 7, 78 and 82). Grx1, like Trx1, uses its active site thiol to act as a disulfide reductase in the cytosolic environment. However, in contrast to Trx1, Grx1 is specialized in the recognition and reduction of protein-glutathione mixed disulfide bonds and forms mixed disulfide intermediates with glutathione rather than with proteins (Yang et al, 1998; Peltoniemi et al, 2006) . As expected, we did not detect trapping of peroxiredoxins (or other Trx1-interacting proteins) by Grx1(CSAAA), neither on silver gels ( Figure 1E , lanes 3 and 4) nor by immunoblotting (data not shown). The activity and thiol reactivity of the Grx1 trapping mutant was confirmed in independent experiments (data not shown), thus demonstrating that the mere availability of an active site thiol does not explain the profile of proteins captured by the Trx1 trapping mutant. Instead, our results support the notion that Trx-mediated reducing activity is steered toward distinct target disulfide bonds by specific protein-protein interactions. Having established the Trx1 kinetic trapping approach for soluble target proteins, we asked whether the kinetic trapping technique can also be applied to the surface of intact cells in culture. Given previous indications of disulfide bond exchange between CD4 and wild-type Trx1 (Matthias et al, 2002) , we asked whether kinetic trapping would enable us to detect this interaction on the surface of the CD4 positive promyelocytic cell line U937. In brief, we allowed mutant Trx1 to interact with the surface of live cells, removed unreacted oxidoreductase by washing and collected disulfide-linked Trx1 complexes from cellular lysates by and non-trapping (SSAAA) mutants of Trx1 were incubated with a o30 kDa fraction prepared from fresh human serum. Disulfide-linked complexes were analyzed by colloidal Coomassie staining under non-reducing and reducing conditions. The Trx1-Prx2 conjugate and the Trx1 dimer as well as monomeric Prx2 and Trx1 are indicated. (E) Kinetic trapping is mediated by specific protein-protein interactions. Cytosolic proteins from digitonin-permeabilized Jurkat cells were incubated with SAv sepharose, Grx1(CSAAA) or Trx1(CSAAA). Disulfide-linked complexes were analyzed by silver staining under nonreducing and reducing conditions. Trx1 conjugates formed with Prx1 and Prx2 are indicated. Other bands correspond to additional cytosolic proteins interacting with Trx1. streptavidin (SAv) affinity purification. We found that cell surface CD4 forms a mixed disulfide with exogenously added Trx1(CSAAA) (Supplementary Figure S1 , left panel), which could be dissociated by DTT treatment (Supplementary Figure S1 , right panel). This result confirmed that kinetic trapping can indeed be used to identify specific Trx1-reactive cell surface proteins and should therefore allow de novo identification of previously unknown cell surface target proteins. Human Trx1 was first identified as an autocrine factor secreted by and acting on virally transformed lymphoid cell lines. We therefore applied cell surface trapping to a human EBV-transformed lymphoblastoid B-cell line (LCL-721.220) derived from the same parental cell line (LCL-721) as the 3B6 cell line, originally used in the description of the costimulatory factor '3B6-IL1', later identified as Trx1 (Wakasugi et al, 1990) . Mixed disulfide complexes, which formed on the surface of LCL-721.220 cells were isolated and analyzed by Trx1-specific immunoblotting to visualize overall mixed disulfide conjugates. Interestingly, we found that Trx1 predominantly engages a single protein on the lymphoblastoid surface, suggesting a highly selective interaction ( Figure 2A , lane 3). As expected, the trapping product, a mixed disulfide conjugate of about 160 kDa, was susceptible to reduction ( Figure 2A , lane 4). The 160 kDa conjugate did not form on the surface of the EBV-negative Burkitts lymphoma cell line BL-41 ( Figure 2B ). In contrast, a conjugate of the same size was found to be formed on the surface of CCRF-CEM T cells ( Figure 2C ) and YT large granular lymphoma cells (data not shown). Pretreatment of the cell surface with the alkylating agent iodoacetamide (IAA) did not interfere with the formation of the 160 kDa mixed disulfide conjugate, thus confirming that the conjugate was formed by the expected disulfide exchange mechanism (rather than by de novo disulfide bond formation between two thiol groups). To identify the unknown Trx1 target protein, we performed cell surface trapping on a larger scale (5 Â10 9 LCL-721.220 cells), purified the Trx1-interacting surface protein by SAv affinity purification and visualized the protein by colloidal Coomassie staining ( Figure 3A , left panel). The 160 kDa band was absent in the control precipitation with Trx1(CCAAA). Corresponding bands from non-reducing and reducing lanes were subjected to tryptic digestion and LC-MS/MS analysis. From both samples the unknown protein was identified as TNFRSF8 (CD30), a member of the TNFR superfamily. To validate the direct covalent interaction between Trx1(CSAAA) and CD30, an aliquot of trapped complexes from the same experiment was separated under non-reducing and reducing conditions and subjected to immunoblotting analysis with anti-Trx1 ( Figure 3A , middle panel) and anti-CD30 antibodies ( Figure 3A , right panel), respectively. The observed mobility difference between non-reducing (NR) and reducing (R) lanes demonstrated the formation of a mixed disulfide conjugate ( Figure 3A , right panel). Additional immunoblotting experiments demonstrated that low nanomolar concentrations of Trx1(CSAAA) are sufficient to detect the interaction with CD30 ( Figure 3B ) and also confirmed that trapping of CD30 depends on the N-terminal cysteine of the CXXC motif ( Figure 3C, lanes 1-8) . Application of the Grx1 trapping mutant under the same conditions did not lead to its conjugation to CD30 ( Figure Ectodomains of TNFR superfamily proteins typically are composed of one to four cysteine-rich domains (CRDs), each normally harboring three disulfide bonds (Bodmer et al, 2002) . To exclude the possibility that Trx1 interacts with CRDs uniformly, we tested whether Trx1 discriminates between distinct members of the TNFR superfamily. As shown in Figure 4A , CD95 (TNFRSF6) did not form a mixed disulfide conjugate with Trx1(CSAAA) on the same cells under identical conditions. The same result was obtained for the epidermal growth factor receptor (EGFR), which contains a total of 25 ectodomain disulfide bonds, and is expressed at substantial levels on A431 cells (Gill and Lazar, 1981 ) ( Figure 4B ). These findings indicate that Trx1 reactivity of cell surface receptors is selective and is not determined by the presence of CRDs or the number of ectodomain disulfide bonds. To further exclude that Trx1 reactivity of receptors is determined or limited by surface expression levels, we ectopically overexpressed CD30 or other TNFR superfamily members under control of the same promoter in HeLa cells and analyzed Trx1 cell surface trapping by indirect immunofluorescence. While mock-transfected HeLa cells did not capture Trx1(CSAAA) on their surface ( Figure 4C , lower row), expression of CD30 led to a strong Trx1 surface association and colocalization of both proteins ( Figure 4C , upper row). In contrast, expression of CD95 ( Figure 4D ), TNFR1 or NGFR (data not shown) did not promote Trx1 interactions with the cell surface, further strengthening the notion that Trx1 reactivity is a specific property of CD30. The domain structure of human CD30 differs from other members of the TNFR superfamily and from its murine homologue by the presence of two additional CRDs, arising from the internal duplication of two exons (Burgess et al, 2004) . We asked whether this unusual feature might confer Trx1 reactivity to human CD30 and tested whether the shorter murine CD30 could also interact with Trx1. As shown in Figure 4E , murine CD30 expressed on the Rauscher murine leukemia virus-induced T-cell lymphoma line RMA is efficiently targeted by Trx1(CSAAA), thus demonstrating that the additional CRDs in human CD30 are not required for Trx1 reactivity. The result also suggests that the enzymatic affinity of Trx1 for CD30 has been conserved during mammalian evolution. To demonstrate by an independent method that Trx1 attacks and breaks a disulfide bond in CD30, we used thiol-specific cell surface biotinylation to verify that Trx1 activity generates free thiols within the CD30 ectodomain (Supplementary Figure S3 ). Subsequent analysis of CD30 cell surface expression by flow cytometry and fluorescence microscopy revealed that a brief treatment of CD30 þ cells with wild-type Trx1 led to the complete loss of CD30 recognition by the Ki-1 antibody ( Figure 5A , upper panel and Figure 5B , second column). Similar results were obtained when another monoclonal antibody against CD30, MAB229 (R&D Systems; Clone 81337), was used to examine CD30 expression ( Figure 5A , middle panel). Under the same conditions, recognition by the Ber-H2 antibody was only slightly affected ( Figure 5A , lower panel and Figure 5B , third column), indicating that Trx1-mediated disulfide bond reduction (and possibly rearrangement) induces a structural alteration in the CD30 ectodomain, which disrupts or conceals the Ki-1 epitope. Pursuant to the observation that recognition of CD30 by antibodies Ki-1 and MAB229 is affected by CD30 redox state, we used flow cytometry to analyze the response of CD30 to with Trx1(CSAAA) and complexes were analyzed by anti-EGFR immunoblotting. Cellular lysate was included as control. (C) HeLa cells were transiently transfected with an expression construct for human CD30 or empty vector incubated with Trx1(CSAAA) and analyzed by immunofluorescence microscopy using CD30-and Trx1-specific antibodies (scale bar, 20 mm). (D) HeLa cells were transiently transfected with expression constructs for human CD30 or CD95, treated as described in (C) and analyzed by immunofluorescence microscopy using CD30-, Trx1-and CD95-specific antibodies (scale bar, 20 mm). (E) RMA mouse lymphoma cells were incubated with Trx1(CSAAA) or Trx1(CCAAA). Disulfide-linked Trx1 complexes were analyzed by anti-mouse CD30 (mCD30) immunoblotting under non-reducing and reducing conditions. The disulfide-linked Trx1-mCD30 complex and monomeric mCD30 are indicated. The additional band of higher molecular weight has not been further characterized but might represent a conjugate between Trx1 and a dimer of mCD30. reduction in greater detail. Addition of oxidized Trx1 did not influence antibody reactivity of CD30 (data not shown). When reduced Trx1 was applied in the absence of a regenerating system, a conformational change in CD30 could be observed, but remained incomplete (data not shown). Complete and sustained loss of antibody reactivity required a source of reducing equivalents for the oxidoreductase. Both DTT and Trx reductase (TrxR)/NADPH were found to be effective as regenerating systems. Importantly, neither DTT nor TrxR/NADPH had an effect when applied in the absence of Trx1 ( Figure 5A , middle panel and data not shown). Wild-type Grx1 ( Figure 5C , left panel) and the redox-inactive mutant of Trx1 ( Figure 5C , middle panel) did not alter the redox-sensitive CD30 epitopes. Other cell surface receptors, for example CD28, analyzed in parallel on the same cells were unaffected by Trx1 treatment ( Figure 5C, right panel) . Loss of CD30 antibody recognition typically occurred within minutes ( Figure 5D , upper panel). Testing the influence of Trx1 concentration under the same conditions, consistent effects on CD30 conformation became evident at concentrations around 100 nM ( Figure 5D , lower panel). The observation that antibody binding to CD30 is influenced by Trx1 suggested a redox-dependent conformational change within the CD30 ectodomain. To test whether Trx1-mediated reduction of CD30 influences binding of CD30 to its ligand (CD30L), we analyzed the interaction between CD30 and recombinant soluble CD30L (sCD30L) on the cell surface by flow cytometry. A brief incubation of CD30 þ Hodgkin's lymphoma HDLM-2 cells with wild-type Trx1 led to a substantial loss in CD30L binding to the cell surface ( Figure 6A) . A similar result was obtained for the large granular lymphoma Figure S4) . The same effect was evident in the absence of an exogenously added reducing system (Supplementary Figure S5) , thus demonstrating that under the given conditions Trx1-substrate interactions are not limited by oxidative inactivation. As shown by fluorescence microscopy, sCD30L binds to the surface of CD30-transfected HeLa cells and colocalizes with its receptor ( Figure 6B , upper row). Treatment of HeLa cells with Trx1(CCAAA) ( Figure 6B , middle row) but not a redox-inactive mutant ( Figure 6B , lower row) strongly interferes with CD30L binding and colocalization. Given its influence on ligand binding, we reasoned that Trx1 might interfere with CD30-mediated signaling. To test whether Trx1-mediated conformational alteration of CD30 affects CD30-dependent cellular responses, we made use of the YT large granular lymphoma line that has previously been used to study signals emanating from CD30 and to define the genes regulated by such signals (Muta et al, 2000) . Consistent with previous results (Bowen et al, 1993) , we observed that stimulation of CD30 with either agonistic antibodies or sCD30L led to upregulation of the IL-2Ra chain (CD25) within 24 h ( Figure 7A , lower panel, compare columns 3, 4 and 7). Treatment of YT cells with Trx1(CCAAA) (but not with redox-inactive Trx1) prevented CD25 upregulation ( Figure 7A , lower panel, columns 5 and 8), concomitant with the redox change in CD30 ( Figure 7A , upper panel, column 4), thus demonstrating that the redox interaction between Trx1 and CD30 has a pronounced influence on CD30-mediated gene expression. YT cells respond to CD30 signals by downregulating the expression of cytotoxic effector molecules, including FasL, thus decreasing their cytotoxicity against Fas-expressing target cells (Bowen et al, 1993; Muta et al, 2000) . To test if Trx1 influences CD30-mediated suppression of cytotoxicity, we treated YT cells with Trx1(CCAAA) or Trx1(SSAAA) before stimulation with agonistic anti-CD30 antibody and quantified cytotoxicity against Cr-labeled Raji cells. Upon CD30 stimulation, cytotoxicity was markedly reduced ( Figure 7B , compare columns 1 and 4). The decrease in cytotoxicity was completely reversed by Trx1(CCAAA) but not the catalytically inactive mutant (SSAAA) ( Figure 7B , columns 5 and 6). As judged by RT-PCR, changes in cytotoxicity correlated with changes in FasL mRNA expression ( Figure 7B, lower panel) . These results confirm that the catalytic activity of Trx1 modulates CD30-dependent changes in cellular behavior and function. Accumulating evidence indicates that the reduction and rearrangement of disulfide bonds constitutes a mechanism controlling protein function on the cell surface (Hogg, 2003) . The idea that disulfide bonds can act as dynamical redox switches, specifically operated by secreted redox catalysts, represents a novel concept in signal transduction (Jordan and Gibbins, 2006) . However, technical difficulties in detecting and analyzing individual disulfide rearrangements on the cell surface have made progress slow. Trx1 is recognized as one of the most important regulators of cellular and organismal redox homeostasis (Gromer et al, 2004) . In particular, intracellular Trx1 counteracts oxidative stress, promotes cell growth and inhibits apoptosis. Under conditions of oxidative stress, Trx1 is released by cells and accumulates at sites of inflammation (Nordberg and Arner, 2001) . Numerous studies have reported that secretory Trx1 influences effector functions and proliferation of lymphocytes (Nakamura et al, 1997) . However, proteins and pathways coupling extracellular Trx1 redox activity to defined cellular responses have remained unknown. In this study, we addressed the question regarding which lymphocyte surface receptors are targeted and regulated by the redox activity of extracellular Trx1. For this purpose, we made use of a mechanism-based kinetic trapping approach to capture mixed disulfide intermediates formed between exogenous Trx1 and its target proteins on the cell surface of living cells. Activity-based techniques offer the opportunity to identify interactions too short-lived to be detectable by conventional methods. To our knowledge, this is the first reported application of kinetic trapping to identify novel target proteins of mammalian Trx1 and the first application of this technique to the surface of intact cells. We demonstrate that Trx1 interacts with intraand extracellular target proteins in a highly selective manner, guided by specific protein-protein recognition rather than random encounters with disulfide bonds. Applying the approach to the surface of cell lines representative of the lymphoid lineage, we observed that Trx1 basically targets a single cell surface protein, subsequently identified as TNF receptor superfamily member 8, also known as CD30. The pronounced preference of Trx1 for one particular target protein might seem surprising, but could be due to the fact that we assessed Trx1 reactivity of proteins as they are embedded in their natural microenvironment, namely the intact surface of the active plasma membrane of living cells. It is conceivable that protein disulfide exchange interactions are limited and controlled by their native context and location. To scrutinize the specificity of the observed interaction, we asked if the preference of Trx1 for CD30 might be caused by an unusual density of disulfide bonds within CD30 and/or exceptional cell surface expression levels. Although the CD30 ectodomain harbors a significant number of predicted disulfide bridges within CRDs, it does not appear to be unusual in terms of disulfide bond composition/density when compared to other members of the superfamily. When tested experimentally, Trx1 failed to interact with other CRD-containing proteins, including the EGFR featuring a total of 25 ectodomain disulfide bonds. The preference for CD30 could not be explained by exceptional surface expression levels either. While Hodgkin's disease cell lines typically express high levels of CD30, other cell lines including LCL-721.220 or CCRF-CEM show at least 20-to 100-fold lower expression as determined by flow cytometry, yet the same selective targeting was observed. Conversely, ectopic overexpression of several related TNFR superfamily members in HeLa cells did not lead to their interaction with Trx1, yet CD30 strongly interacted on the same cells under the same conditions. Consistent with these findings, recent experiments demonstrate that Trx1 targets a particular site within the CD30 ectodomain (Y Balmer and TP Dick, unpublished data) . To facilitate identification of low-abundance cell surface proteins, in vitro trapping experiments were typically performed using Trx1 concentrations of 1-3 mM. However, when disulfide exchange was subsequently tested at lower concentrations, Trx1(CSAAA) concentrations in the low nanomolar range (4-40 nM) were found to give rise to the formation of proportional amounts of Trx1-CD30 mixed disulfide intermediates ( Figure 3B ), thus demonstrating that the observed interaction is compatible with the expected physiological concentration range of secretory Trx1 (see below). Wild-type Trx1 is known to act as a multiple-turnover catalyst if a suitable reducing system and electron source is provided for its regeneration. In agreement with these considerations, we observed that sustained reduction of CD30 in cell culture requires a Trx1 regenerating system. Using flow cytometry to monitor conformational changes in the CD30 ectodomain, CD30 was found to respond to Trx1 concentrations in the nanomolar range, starting at around 100 nM ( Figure 5E , lower panel). However, the minimal Trx1 concentration required for sustained CD30 reduction might be substantially lower in specific environments, which are efficient in delivering reducing equivalents and preventing oxidative inactivation of Trx1. At present, it is not clear how extracellular Trx1 is regenerated in vivo. Despite the overall oxidizing character of the extracellular compartment, reductive processes are known to take place on the cell surface. On the one hand, there is longstanding evidence for the existence of transplasma membrane redox systems delivering electrons to the cell surface (Crane et al, 1985) . On the other hand, Trx1 may be regenerated by co-secreted reductants, as Trx1 secretion in DC-T co-culture is accompanied by the release of reduced cysteine and the creation of a reducing microenvironment between interacting cells (Angelini et al, 2002) . In addition, TrxR was found to be secreted by activated monocytes and might be part of an extracellular Trx1 reducing system (Soderberg et al, 2000) . The concentration of Trx1 in human plasma is in the low nanomolar range (1-5 nM), and is found to be elevated several-fold under inflammatory conditions (Yoshida et al, 1999) . However, plasma Trx1 is oxidized and appears to represent systemic dilution of Trx1 previously released within tissues. Accordingly, local tissue concentrations of secretory Trx1, for example, within activated lymph nodes, are expected to be markedly higher than in plasma. Overall Trx1 concentrations in mammalian tissues can be as high as 20 mM (Gromer et al, 2004) . Certain Trx1-secreting cell types, including macrophages and dendritic cells, distinctly upregulate expression of Trx1 upon activation (Angelini et al, 2002) . In vitro studies of Trx1 secretion suggest that a substantial fraction of intracellular Trx1 can be released within a few hours (Rubartelli et al, 1992) . Although direct measurements of extracellular Trx1 within tissues are not available, physiologically relevant extracellular Trx1 concentrations may well reach into the upper nanomolar, if not lower micromolar range. We found that Trx1-mediated disulfide reduction changes the conformation and functional properties of the CD30 ectodomain. In the reduced state, CD30 lost its ability to interact with its cognate ligand CD30L or agonistic antibodies. The presence of catalytically active Trx1 impeded CD30-dependent signaling in the YT lymphoma cell line, as demonstrated by its effect on CD25 and FasL expression, as well as its influence on cytotoxicity against Fas-expressing target cells. The physiological role of the CD30-CD30L system has remained unclear. In vitro studies focusing on CD30 þ lymphoid malignancies showed that triggering of CD30 signaling can induce either proliferation, activation, growth arrest or apoptosis, depending on cell type and stimulatory conditions (Schneider and Hubinger, 2002) . In vivo, cell surface expression of CD30 appears to be tightly regulated and restricted to B and T lymphocytes undergoing activation in lymphoid tissues. It has been proposed that CD30 provides proliferation and/or survival signals during lymphocyte responses (Croft, 2003) . Under inflammatory conditions, CD30 expression is markedly induced. In vivo activation of CD30 can be monitored by the release of sCD30, shed from the plasma membrane upon CD30L binding (Hansen et al, 2000) . Similar to serum Trx1, serum sCD30 is increased in infection, autoimmunity and allergy, for example systemic lupus erythematosus, rheumatoid arthritis and atopic dermatitis (Horie and Watanabe, 1998) . Both Trx1 and CD30 appear to play a role in the regulation of the antiviral inflammatory response. Both Trx1 secretion and CD30 expression have been associated with virally transformed lymphocytes. Elevated levels of sCD30 occur during viral infection. Likewise, viral infection leads to elevated Trx1 plasma levels and several studies indicate that secreted Trx1 modulates the antiviral inflammatory process (Nakamura et al, 2001b (Nakamura et al, , 2002 . In this study, we have identified an enzyme-substrate relationship between Trx1 and CD30, a receptor of activated lymphocytes involved in the regulation of inflammation. As lymphocytes migrate between different microenvironments, it is conceivable that Trx1 catalyzes disulfide exchange dynamically, activating or inactivating the CD30 pathway in response to the redox environment. The interaction between Trx1 and CD30 might represent a regulatory link between oxidative stress and lymphocyte function. Understanding of this relationship in vivo awaits the generation of suitable experimental tools. Cell culture BL-41, CCRF-CEM, HDLM-2, Jurkat, RMA and U937 cells were cultured in RPMI 1640 (Gibco) supplemented with 10% heatinactivated fetal bovine serum, 2 mM L-glutamine, 100 U/ml penicillin and 100 mg/ml streptomycin (Gibco). LCL-721.220 and YT cells were cultured in IMDM (Gibco), HeLa and A431 cells in DMEM (Gibco) with the same supplements. Depending on the type of experiment, recombinant trapping mutant was applied to cytosolic preparations, human serum ultrafiltrate or intact cells. A detailed description of the different substrate trapping protocols is provided as Supplementary information. For reduction of cell surface CD30, 2.5 Â10 5 cells were incubated with 5 mM Trx1 together with 200 mM DTT or 100 nM human Trx reductase (TrxR)/500 mM NADPH for 30 min at 41C. To monitor reduction of cell surface CD30, cells were stained with anti-human CD30 monoclonal antibody MAB229 (R&D Systems), anti-human CD30 monoclonal antibody Ki-1 (Santa Cruz) or anti-human CD30 monoclonal antibody Ber-H2 (DakoCytomation) followed by incubation with R-PE-conjugated goat F(ab') 2 anti-mouse Ig's (Biosource). For control, cells were stained with R-PE-conjugated anti-human CD28 monoclonal antibody (BD Pharmingen). Cells were analyzed using a FACSCalibur (Becton Dickinson) and CellQuest software. A total of 2.5 Â10 5 cells were incubated with 5 mM Trx1 (SBP-CCCCC) and 200 mM DTT for 30 min at 371C, washed three times and incubated with 250 ng/ml recombinant CD30L-His 10 (R&D Systems) for 10 min at RT. After washing, cells were stained with anti-polyHis monoclonal antibody (Sigma) followed by incubation with R-PE-conjugated goat F(ab') 2 anti-mouse Ig's (Biosource). HeLa cells were seeded on coverslips and transfected with expression constructs using CaCl 2 precipitation. After 2 days, transfected cells were fixed with 3% formaldehyde and 2% sucrose in PBS for 7 min at RT. Fixed cells were washed three times with PBS and incubated with different Trx1 constructs or recombinant CD30L (R&D Systems). Proteins were visualized using appropriate primary antibodies (Anti-CD30L polyclonal antibody (R&D Systems), anti-CD30 monoclonal antibodies Ki-1 (Santa Cruz) or Ber-H2 (DakoCytomation), anti-CD95 monoclonal antibody (a kind gift from Dr P Krammer), anti-Trx1 polyclonal antibody (M Preuss and TP Dick, unpublished) followed by FITC-conjugated anti-goat IgG, FITC-conjugated anti-rabbit IgG or TRITC-conjugated anti-mouse IgG and analyzed with a Nikon C1Si confocal microscope. Supplementary data are available at The EMBO Journal Online (http://www.embojournal.org).
92
Screen for ISG15-crossreactive Deubiquitinases
BACKGROUND: The family of ubiquitin-like molecules (UbLs) comprises several members, each of which has sequence, structural, or functional similarity to ubiquitin. ISG15 is a homolog of ubiquitin in vertebrates and is strongly upregulated following induction by type I interferon. ISG15 can be covalently attached to proteins, analogous to ubiquitination and with actual support of ubiquitin conjugating factors. Specific proteases are able to reverse modification with ubiquitin or UbLs by hydrolyzing the covalent bond between their C-termini and substrate proteins. The tail regions of ubiquitin and ISG15 are identical and we therefore hypothesized that promiscuous deubiquitinating proteases (DUBs) might exist, capable of recognizing both ubiquitin and ISG15. RESULTS: We have cloned and expressed 22 human DUBs, representing the major clades of the USP protease family. Utilizing suicide inhibitors based on ubiquitin and ISG15, we have identified USP2, USP5 (IsoT1), USP13 (IsoT3), and USP14 as ISG15-reactive proteases, in addition to the bona fide ISG15-specific protease USP18 (UBP43). USP14 is a proteasome-associated DUB, and its ISG15 isopeptidase activity increases when complexed with the proteasome. CONCLUSIONS: By evolutionary standards, ISG15 is a newcomer among the UbLs and it apparently not only utilizes the conjugating but also the deconjugating machinery of its more established relative ubiquitin. Functional overlap between these two posttranslational modifiers might therefore be more extensive than previously appreciated and explain the rather innocuous phenotype of ISG15 null mice.
Posttranslational modification by ubiquitin regulates processes such as proteasomal degradation, intracellular trafficking, and transcription. Ubiquitin is attached to substrates in covalent isopeptide linkage or as an N-terminal fusion [1] [2] [3] . Ubiquitination, however, is reversible: the ubiquitin moiety can be released from substrates through the action of deubiquitinating proteases, which may rescue ubiquitinated substrates from their degradative fate [4] . In contrast, proteasome-associated DUBs enhance the rate of proteasomal degradation by removing bulky poly-ubiquitin chains from substrate proteins prior to proteolysis. Such DUBs enhance the processivity of the proteasome toward target proteins, and also recycle ubiquitin, a modifier that itself turns over slowly [5, 6] . DUBs are furthermore required to hydrolyze the ubiquitin precursor and generate the active ubiquitin monomer. Inspection of mammalian genomes shows the presence of more than 100 genes that encode putative DUBs, consistent with their specific and diverse regulatory functions. Ubiquitin-specific proteases (USPs) are the dominant family among DUBs [7] . Ubiquitin-like molecules show sequence and structural similarity to ubiquitin. Unlike ubiquitination, modification by UbLs usually does not target proteins for destruction by the proteasome. A notable exception may be FAT10, a modifier that serves as a ubiquitin-independent signal for proteasomal degradation [8] . The conjugation of UbLs to target proteins follows reaction pathways similar to those involved in ubiquitination [9] . The enzymes that attach or cleave UbLs are generally distinct from the ligases or proteases of the ubiquitin pathway. A closely related homolog of ubiquitin in vertebrates is the UbL polypeptide ISG15, an interferon-inducible gene product that is strongly upregulated following viral or bacterial infection [10] . However, the molecular and regulatory consequences of ISGylation remain unknown [11] . ISG15 consists of two ubiquitin domains in a tandem arrangement, similar to FAT10. Unlike other members of the UbL family, ISG15 co-opts at least one of ubiquitin's conjugating enzymes, Ubc8 [12, 13] and the ubiquitin ligase Herc5 [14] [15] [16] [17] . USP18 constitutes the only presently appreciated isopeptidase specific for ISG15, and its absence has profound effects on innate immunity, leading to increased resistance to certain viral infections [18, 19] . Notably, these effects appear not to be contingent upon proteolytic activity of USP18 [20, 21] . Apart from USP18, additional proteases for ISG15 must exist, since the ISG15 precursor protein is cleaved properly in USP18 knockout mice [19] . The C-terminal six amino acids of ubiquitin and ISG15 are identical. This tail region is required for specific recognition of ubiquitin by conjugating enzymes, and also for recognition of ubiquitin adducts by isopeptidases [22, 23] . The overlap in conjugation between ubiquitin and ISG15, as well as their Cterminal similarity, imply the existence of promiscuous DUBs, capable of removing both ubiquitin and ISG15 from substrate proteins. Here, we report on the identification of new ISG15specific proteases measured by reactivity toward active-site directed probes and isopeptide-linked substrates [24] [25] [26] . Activity-based profiling of DUBs Figure 1 shows a consensus phylogram based on the alignment of catalytic core sequences of DUBs, including the majority of known human USP homologs. In this tree, the ISG15-protease USP18 clusters close to USP5 (IsoT1) and its isoform USP13 (IsoT3). Previous work had identified USP5 as a protease with affinity for both ubiquitin [27] and ISG15, as shown by its reaction with an electrophilic ISG15 derivative, ISG15-vinyl sulfone (ISG15VS) [28] . To probe for additional ISG15-reactive proteases, we have cloned and expressed a total of 22 human DUB homologs from different clades of this phylogram (indicated with arrows), 17 of which reacted with a ubiquitin-based probe and/or an ISG15based probe (see below). The screen was based on in vitro transcription and translation (IVT) of cloned cDNAs, which affords a rapid method to generate radiochemically pure proteins. This technique allows the generation of DUBs that cannot be readily expressed in bacterial systems, or that are sequestered in subcellular compartments or in multimolecular complexes when expressed in cell lines. To profile for DUB specificity, we used recombinantly expressed ubiquitin, SUMO1 and ISG15, and installed an electrophilic trap at their C-terminus to obtain the active-site probes ubiquitin-vinylmethyl ester (UbVME), SU-MO1VME and ISG15VS, respectively [28] . DUBs generated by IVT were incubated with each of these three probes (Figure 2 ), followed by direct analysis of the reaction mixture by SDS-PAGE. We have determined by X-ray crystallography that probes of this type form a covalent adduct with the catalytic cysteine residue of active DUBs to yield a thioether-linked adduct between enzyme and probe [26] . When unmodified IVT products are run adjacent to samples incubated with these activity-based probes, the adduct is readily detected through a shift in apparent molecular mass. USP2, USP5, USP13 and USP14 reacted with ISG15VS ( Figure 2B , C, D, E). The following observations confirm the validity of our assay. First, the bona fide ISG15-isopeptidase USP18 displayed reactivity only towards ISG15VS (Figure 2A ), whereas most of the DUBs reacted only with UbVME. An example is shown with CGI-77 ( Figure 1 , 2F), a previously uncharacterized Otubain-homolog. Second, as a negative control, the SUMO protease SENP2 formed an adduct exclusively with SUMO1VME ( Figure 2G ). We found no evidence for any of the DUBs evaluated here to display reactivity toward the SUMO1 probe (not shown). The presence of the reactive group alone is clearly not sufficient for binding to the active-site cysteine of a protease and specificity of a DUB thus depends on the peptide moiety of the probe, containing either ubiquitin, SUMO1 or ISG15. Lastly, all covalent modifications of DUBs by active-site directed probes were blocked by pretreatment of the translated polypeptides with the sulfhydryl alkylating agent N-ethylmaleimide (NEM) (Figure 2 ), confirming the cysteinedependency of adduct formation. In accordance with previous results [28] , we observed a non-linear decrease in electrophoretic mobility of the ISG15VS adducts. The ISG15 probe has a mass of 17.4 kDa, whereas the size increase observed for each of the DUBs investigated here is in the order of 25-110 kDa, when bound to ISG15VS. The correlation between the abnormal shift and the initial mass of the unmodified DUB suggests that the decrease in gel mobility is based on steric properties of the branched adduct, and is not caused by covalent modification of a single DUB by multiple ISG15VS molecules. In fact, in our sample set, the observed size increase of the ISG15VS-DUB adducts based on SDS-PAGE very closely matched a logarithmic equation ( Figure 3A and see Methods). To further verify that DUBs are modified by only a single ISG15VS probe per molecule, we replaced the catalytic cysteine at position 114 in USP14 with a serine residue. As expected, this mutation abolished all labeling ( Figure 3B ). Our assay was conducted in IVT lysate and the size increase of the ISG15VS adduct could potentially reflect modification of USP14 by additional factors. However, even USP14 that was recombinantly expressed in bacteria and .95% pure showed the same abnormal electrophoretic mobility for its ISG15VS adduct ( Figure 3C ). Mass spectrometry confirmed the monovalent modification of purified USP14 by ISG15VS and excluded covalent binding of additional factors to the complex ( Figure 3D ). Collectively, these experiments establish that the observed shift in apparent molecular mass of the ISG15VS-DUB adducts is solely a consequence of its unusual electrophoretic behavior. It also underscores the uncertainties in estimating the degree of modification of a target protein with UbLs by SDS-PAGE alone. USP14 reacts more efficiently with ISG15VS in its proteasome-associated form USP14 and its yeast counterpart Ubp6 show significantly higher activity when bound to the 26S proteasome [29] . This activity may in fact be strictly dependent on association with the proteasome, as shown for Ubp6 [30] . As further evidence for a physiological role of the interaction of USP14 with ISG15, we investigated whether the allosteric activation of USP14 also influences its reactivity toward ISG15VS. We examined labeling of USP14 with the ubiquitin-and the ISG15-based probes as a function of the concentration of added purified proteasomes. As a negative control, we evaluated USP5, a DUB that is not a known interaction partner of the 26S complex in vivo. As anticipated, the inclusion of purified proteasomes had no effect on the ISG15VS-or UbVME-reactivity of USP5 ( Figure 4B ). In contrast, we observed a dose-dependent increase in ISG15VS adduct formation of USP14 with increasing proteasome concentration, indicating enhanced activity of this DUB. The effect was similar in magnitude to that seen for the ubiquitin probe ( Figure 4A , C). While recombinant USP14 in its purified form bound to electrophilic probes ( Figure 3C ), we did not detect robust hydrolytic activity against ubiquitin-AMC or against ubiquitinor ISG15-linked isopeptide fusion proteins (data not shown). However, using sequential ultracentrifugation to obtain a cytosolic fraction that is enriched in 26S proteasomes [29] , we could show that proteases in this fraction efficiently and specifically cleave an ISG15-isopeptide linked substrate ( Figure 5A , B). The absence of proteolytic intermediates suggests specific cleavage of the isopeptide bond. In addition, the same bait peptide linked to SUMO1 was stable and not hydrolyzed, even upon prolonged incubation for over 24 hours with the proteasome fraction. Proteolysis of the ISG15-linked peptide substrate was inhibited by inclusion of NEM, indicating cleavage by cysteine proteases. Analysis by reaction with ISG15VS supports that USP14 is the only active Red arrows depict DUBs that bound to neither probe (UbVME, ISG15VS, or SUMO1VS), whereas black arrows indicate DUBs that formed covalent adducts with the indicated probes. Our screen represents the first biochemical proof for protease activity of USP13 and the Otubainhomolog CGI-77 (DUB homologs without publication record regarding biochemical function are marked with an asterisk). Otubain1 (OTU1) is an exception in that it binds to alkylhalide-or aldehyde-based probes, but not to the Michael acceptors employed in this study (data not shown). doi:10.1371/journal.pone.0000679.g001 ISG15-specific protease in the proteasome-enriched fraction ( Figure 5C ). While we cannot formally exclude the possibility of a yet undefined enzyme binding to ISG15VS, we consider this unlikely: such a protease would have to display a mass highly similar to that of USP14 and, furthermore, it would have to sediment after centrifugation for 5 hours at 100,000 g. However, only few deubiquitinating enzymes are sedimentable, none at a level comparable to USP14 [29, 31] . The wealth of USPs found in the human proteome likely reflects substrate specificity, but potentially also complementation in terms of expression profiles and subcellular distribution. We therefore sought to analyze the intracellular distribution pattern of a subset of our crossreactive DUBs, using confocal microscopy. The analysis of a genome-wide set of C-terminal GFP fusion proteins for yeast had shown remarkably few with altered function or subcellular distribution (,5%), validating the choice of such Cterminal modifications [32] . Using anti-G/YFP antibodies, we also utilized this tag to assay for activity of DUBs in cell lysate. We cloned and transiently expressed five USP-EYFP constructs in 293T cells: USP5, USP13, USP14, USP3, and USP36 ( Figure 6A ). Lysate of USP14 EYFP transfected cells was incubated with the ubiquitin and the ISG15 probe, and assayed by anti-YFP immunoblot analysis ( Figure 6B ). Whereas the USP14 EYFP construct reacted with both probes, the respective C114S mutant Mutation of the catalytic cysteine residue to serine (C114S) in USP14 abolishes its reactivity toward UbVME and ISG15VS. When stored for longer periods at room temperature, the probes polymerize covalently, presumably by formation of secondary amine bonds between internal lysine residues and the reactive Michael acceptor at the C-terminus, thus resembling isopeptide-linked polyubiquitin. Such polymeric probes of UbVME likely caused the additional high-molecular mass adducts observed for USP14. Note that the smallest version of these adducts (a UbVME dimer) has a maximum electrophoretic mobility similar to that of the diubiquitin-like ISG15VS when complexed to USP14. The absence of any adducts in the C114S mutant of USP14 excludes the possibility of multiple binding sites for the probes. (C) Purified recombinant USP14 labels with UbVME and ISG15VS and results in the same abnormal mobility shift for ISG15VS-USP14 as seen in the IVT labeling experiments. (D) MALDI-TOF mass spectroscopic analysis of USP14 after incubation with ISG15VS. As described above for UbVME, ISG15VS also engages in internal polymerization. Molecular masses consistent with tri-and tetrameric ISG15VS are marked in this spectrogram by the numbers in superscript. Monovalently modified USP14 results in an adduct of predicted size, indicated with a red arrow. This complex is unique to the mixture containing both USP14 and ISG15VS, and is absent in the mass spectra of either component alone (data not shown). doi:10.1371/journal.pone.0000679.g003 did not, in agreement with the results of our IVT screen. With respect to subcellular distribution, we were particularly interested in the expression pattern of USP5 and USP13, given that these two isoforms displayed different specificity in UbVME and ISG15VS labeling experiments. USP5 EYFP was found throughout the cell ( Figure 6C, upper left panel) , similar to USP18 [33] . In contrast, its close relative USP13 EYFP was expressed mainly in the nucleus in a speckled pattern ( Figure 6C, upper right panel) . Consistent with the in vivo interaction between USP14 and the proteasome [5] , we observed USP14 EYFP predominantly in the cytoplasm, though we also noticed fluorescence in the nucleus ( Figure 6C, lower left panel) . To demonstrate that the presence of the C-terminal YFP fusion does not interfere with the endogenous distribution of these enzymes, we analyzed USP3 EYFP ( Figure 6C , lower right panel) and USP36 EYFP ( Figure 6C , lower right panel). USP3 is predicted to be a nuclear protein [34] and USP36 was . Reactivity of USP14 toward ISG15VS is augmented by proteasomal association. USP14 and USP5 were generated by IVT. Their activity toward ISG15VS and UbVME was analyzed in the presence of increasing concentrations of purified human 26S proteasomes. (A) Activity of USP14 toward UbVME and ISG15VS increases as a function of the concentration of added purified 26S proteasomes (in mg/ml). (B) The activity of USP5 remains unaffected. (C) Quantification of the radioactive signal of covalently modified USP5 and USP14. Binding affinity is depicted on the y-axis as percent in labeling intensity, determined by the ratio of labeled versus unlabeled USP5 or USP14. The ratio in the absence of exogenous proteasomes is defined as 100%. doi:10.1371/journal.pone.0000679.g004 Figure 5 . Proteasome-associated USP14 has ISG15-specific isopeptidase activity. (A) Scheme depicting the UbL-peptide conjugate used to assay isopeptidase activity. The biotinylated peptide heptamer is attached to either ISG15 or SUMO1 in isopeptide-linkage. Upon hydrolysis of the isopeptide bond by a specific DUB, the heptamer is released and the biotin signal lost. (B) Incubation of proteasomeenriched fraction (''5 hr pellet'') with UbL-peptide conjugate. After overnight incubation, the ISG15-peptide conjugate is completely cleaved, resulting in loss of the biotin-signal (significant proteolysis occurs already after one hour, data not shown). This activity is sensitive to NEM. Hydrolysis is not observed for the SUMO1-peptide conjugate. (C) Anti-HA immunoblot of HA-ISG15VS treated subcellular fractions. Based on previous identification [28] and on electrophoretic mobility, USP5 is the dominant ISG15-reactive DUB in the five-hour supernatant, which is enriched for uncomplexed proteins of light and moderate size (red asterisk). The five-hour pellet represents heavy cytosolic complexes, in particular the 26S proteasome, and contains USP14 as the only ISG15reactive DUB (blue asterisk). doi:10.1371/journal.pone.0000679.g005 identified as a nucleolar protein [35] . Both proteins were detected in the expected subcellular compartment. Combined, these results indicate that ISG15-specific proteases are expressed throughout the cell -a feature also proposed for DUBs. This observation supports the notion that unlike SUMOylation, which is believed to mostly occur in the nucleus [36] , ISG15-modification affects many cellular compartments. Our data show the existence of multiple ISG15-reactive DUBs, a finding that further strengthens the similarities between ubiquitin and ISG15. USP2 is a highly active protease [37] and it represents one of only few mammalian DUBs with a known target. USP2 exhibits oncogenic potential in prostate cancer by stabilizing its substrate Fatty Acid Synthase (FAS) [38] , and FAS has indeed been identified as a target of ISG15 modification [39] . Furthermore, USP2 has been implicated in the regulation of the p53 pathway [40] . The recently solved structures of USP2 and USP14 [41, 42] show that both proteases accommodate the ubiquitin molecule in a shallow pan-like protrusion. Based on the orientation of the ubiquitin protein in both structures, ISG15 easily fits into the catalytic domain of USP2, USP14, and USP5 (data not shown) without apparent steric clashes (Figure 7 ) [23, 43] . Stimulation by interferons alters the composition of the proteolytic proteasome core [44] , tailoring its activity toward generation of peptide-MHC complexes for inspection by the immune system. Interferon treatment also results in enhanced modification of proteins by ISG15 -a factor that evolved in the vertebrate lineage, and whose origin thus coincides with that of the adaptive immune system. Interestingly, inhibition of the proteasome leads to rampant accumulation of ISG15-modified substrates [45] . While nothing is known about the molecular functions of ISG15, its structural relative FAT10 is a modifier that destines proteins for degradation by the proteasome [8] . We now have demonstrated that USP14 exhibits proteasome-associated isopeptidase activity toward ISG15. Could therefore ISG15 be a (co-)modifier of proteins destined for proteasomal degradation? We have found no evidence that USP14 markedly changes the amount of ISG15-modified substrates in cells (data not shown). However, USP14 is not a vital protease [30, 46] and its low catalytic turnover does not affect overall ubiquitin conjugation either [47] . A recent study suggests that USP14 might have a more complex role, by inhibiting the proteasome in addition to acting as a deubiquitinase [48] . The close sequence relationship between USP5 and USP13 (61.4% identity, 26.9% similarity) is contrasted by the functional differences and localization of these two proteases. These enzymes provide a unique opportunity to investigate the structural features that may contribute to ubiquitin versus ubiquitin-like specificity. A characteristic of USP5 and USP13 is the tandem occurrence of a UBA domain, which has been implicated in the binding of ubiquitin [49] . Our results raise the possibility that UBA domains in general interact not only with ubiquitin, but also with ISG15. Alternatively, ISG15 with its multiple lysines could act as a ubiquitination anchor, and USP5 may be a protease responsible for depolymerization of such chains [50, 51] . Moreover, the Cterminal hydrolase that processes the ISG15 precursor has not been identified yet, and any of the novel ISG15-specific proteases described here are potential candidates. The ubiquitin gene is prone to duplications and insertions, leading to the formation of new fusion proteins [52, 53] . ISG15 likely emerged approximately 400-600 million years ago, when a nucleotide stretch from the polyubiquitin precursor gene, encoding a ubiquitin-dimer, was accidentally inserted in an area of the genome that was or that came under control of an interferon promoter. From an agnostic point of view, one could argue that ISG15 simply has no relevant function. The moderate or absent phenotype of the ISG15 knockout in mice [11] , the fact that ISG15 has not (yet) established its very own family of conjugating and deconjugating enzymes, and ISG15's relatively low degree of conservation between species would all support this view. Yet, ISG15's massive expression upon interferon challenge [54] likely reflects a role in anti-microbial or anti-viral defense [55] [56] [57] [58] [59] . And adaptation to the specific needs of host immunity often demands polymorphism. As a result, some genes most critical to the immune response are paradoxically least conserved. For example, cytokines and cytokine receptors substantially differ between species [60] and it is interesting to note that ISG15 and ubiquitin were initially reported to be cytokines [61, 62] . Similar observations were made for the ubiquitin-like modifier FUBI (also known as Fau or MNSFb) [63, 64] . If true, how do these factors gain access to the secretory pathway? It may pay to approach ISG15 from a less conventional perspective and from this vantage point, we might uncover new functions of ubiquitin as well. USP2 and USP18 cDNAs were cloned from a human kidney cDNA library (BioChain Institute, Inc.). The cDNAs encoding the other human DUBs were obtained from ATCC. All cDNAs encoding fulllength DUBs were subcloned into pcDNA3. The protein sequences of human DUBs were obtained from the National Library of Medicine and the core domains were aligned with the CLUSTALW algorithm (EBI server) [65] and manually edited with Genedoc (http://www.psc.edu/biomed/genedoc/) by K.B. Nicholas & HB Nicholas Jr., using the putative active-site cysteine as an alignment anchor. The phylogram represents a consensus tree based on 100 bootstrap iterations, calculated by the Minimum Evolution method under default parameters [66] . IVT was performed using the ''TNT-T7 Quick Reticulocyte Lysate System'' kit (Promega) for 30 to 45 min (0.25-1 mg DNA per reaction). Then, aliquots of the reaction mix were treated with RNase B (1 mg/ml, Sigma) for 10 min and incubated with the probes as described below. SDS-PAGE followed by fluorography was performed as described [67] . The synthesis of human ubiquitin and UbL probes has been described [24, 28] . IVT products were incubated with saturating amounts of the individual probes (0.2-0.4 mg/10 ml IVT lysate). Preincubation with NEM was performed for 10 min at room temperature at a final concentration of 10 mM, after RNase B treatment. Autoradiograms were subjected to quantification of the optical density with NIH Image software (version 1.32j) as ratio of labeled versus unlabeled IVT products. Purified human 26S proteasomes for the experiments in Figure 4 were purchased from Biomol International and inhibited with MG132 (50 mM). E. coliexpressed human recombinant USP14 for the experiments shown in Figure 3C To 50 mL of conjugation buffer (100 mM Tris, pH 7.4, 5 mM MgCl 2 , 20 mM DTT, 40 mM ATP) was added ISG15 (Boston Biochem, 9.4 mM final) or GST-SUMO1 (Boston Biochem, 10.4 mM final) and biotinylated peptide 7-mer (biotin-VKAKIQD-OH, 250 mM final). The solution was mixed thoroughly and to this was added ISG15 activating enzyme (Boston Biochem, ISG15 E1, 50 nM final) and UbcH8 (Boston Biochem, 250 nM final), or SUMO activating enzyme (SAE1/SAE2 heterodimer, 50 nM final) and UbcH9 (250 nM final). The solution was mixed and incubated for 15 hours at 37uC. The reaction mixture was transferred to a microcentrifuge membrane filter (Vivascience, 5000 Da MWCO), diluted to 600 mL total volume with 50 mM Tris, pH 7.4 and concentrated at 4uC to 50 mL. This dilution/concentration procedure was repeated six times. The products were transferred to a clean tube and diluted with 50 mM Tris, pH 7.4 to a final volume of 100 mL. The SUMO1-and ISG15-linked biotinylated isopeptide was detected after transfer to a PVDF membrane (Perkin Elmer) with Streptavidin-HRP (Amersham). EL4 cells were lysed with glassbeads and the proteasome-enriched fraction was retrieved by consecutive ultracentrifugation steps as previously described [29] . Proteasome activity was inhibited with MG132 (50 mM). 10 mg of fraction protein were incubated with 0.2 mg of N-terminally HA-tagged ISG15VS in a total volume of 10 ml for two hours at room temperature to detect ISG15VSreactive proteases by anti-HA immunoblotting. The cleavage assay for ISG15-or SUMO1-branched peptides was performed at 37uC using 20 mg of total protein from the proteasome-enriched fraction and 5 ml of branched peptides in a total volume of 10 ml. HA-ISG15VS treated subcellular fractions were analyzed with a monoclonal anti-HA antibody (12CA5). Immunoblotting was performed as published [67] . 293T cells were maintained in DME medium as described [67] . Various constructs were expressed by transient transfection, using a liposome-mediated transfection protocol (5-10 mg of DNA/ 20 ml of Lipofectamine-2000 per 10 cm dish; Invitrogen) as described [67] . Cells were analyzed between 24 and 48 h after transfection. C-terminal EYFP fusion proteins of DUBs were generated by subcloning from pcDNA3.1 into pEGFP-N1 (Clontech). NP40 lysates of USP14 EYFP and USP14 C114S-EYFP transfected 293T cells were prepared and incubated with activesite probes as described [24] . Due to the high similarity with GFP, EYFP fusion proteins can be detected with a polyclonal anti-GFP rabbit serum [68] . Immunoblotting was performed as published [67] . Immunofluorescence experiments were performed as described [67] with minor modifications. Cells were allowed to attach to slides overnight. After fixation with 3.7% paraformaldehyde for 20 min at room temperature, subcellular localization of EYFP fusion proteins was analyzed with a Perkin Elmer spinning disk confocal microscope Ultraview RS system. The microscope used was a Nikon TE2000-U inverted unit with a Nikon 1006 1.4NA DIC lens. The imaging medium was Nikon type A immersion oil.
93
s-RT-MELT for rapid mutation scanning using enzymatic selection and real time DNA-melting: new potential for multiplex genetic analysis
The rapidly growing understanding of human genetic pathways, including those that mediate cancer biology and drug response, leads to an increasing need for extensive and reliable mutation screening on a population or on a single patient basis. Here we describe s-RT-MELT, a novel technology that enables highly expanded enzymatic mutation scanning in human samples for germline or low-level somatic mutations, or for SNP discovery. GC-clamp-containing PCR products from interrogated and wild-type samples are hybridized to generate mismatches at the positions of mutations over one or multiple sequences in-parallel. Mismatches are converted to double-strand breaks using a DNA endonuclease (Surveyor™) and oligonucleotide tails are enzymatically attached at the position of mutations. A novel application of PCR enables selective amplification of mutation-containing DNA fragments. Subsequently, melting curve analysis, on conventional or nano-technology real-time PCR platforms, detects the samples that contain mutations in a high-throughput and closed-tube manner. We apply s-RT-MELT in the screening of p53 and EGFR mutations in cell lines and clinical samples and demonstrate its advantages for rapid, multiplexed mutation scanning in cancer and for genetic variation screening in biology and medicine.
Screening for genetic changes to unveil molecular attributes of human specimens is important for a variety of medical applications, including genotyping for inherited disorders, prediction of the pathologic behavior of malignancies, identification of cancer biomarkers and can affect treatment decisions for individual patients (1) (2) (3) . For example, mutations in genes like EGFR can profoundly influence chemotherapeutic response in lung cancer (2) (3) (4) (5) and the response is modulated by mutations in other genes of the same signaling pathway [e.g. K-ras, HER2, ErbB-3 (1, 6) ]. Therefore there is a need for efficient and high-throughput mutation screening of multiple genes along identified signal transduction pathways in tumor samples. Because a large portion of cancer-causing genetic changes remains unknown and can occur in numerous positions along tumor suppressor genes (e.g. p53, ATM, PTEN) mutation scanning rather than detection of specific mutations is frequently required for molecular cancer profiling. Sequencing is often considered the gold standard for comprehensive mutation analysis. Multi-capillary electrophoresis, re-sequencing arrays or pyrosequencing provide platforms for highly parallel genetic analysis (7) (8) (9) (10) (11) (12) (13) . However, the expense associated with these techniques is currently high both for instrumentation and for runningcosts. Since somatic mutations for most genes are relatively rare events it can be inefficient to scan for mutations using expensive approaches that in several cases provide unnecessary data (14, 15) . Another issue with direct sequencing or re-sequencing arrays is the difficulty *To whom correspondence should be addressed. Tel: +1-617-525-7122; Fax: +1-617-587-6037; Email: mmakrigiorgos@partners.org ß 2007 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. in detecting a small fraction of mutated alleles in the presence of a high excess normal alleles, which is frequently the case with clinical cancer samples (16) . As a less expensive alternative, rapid pre-screening methods such as SSCP, DGGE, dHPLC, CCM, CDCE or HR-melting are widely utilized to identify DNA fragments that contain mutations prior to performing full sequencing (14, (16) (17) (18) (19) (20) . Enzymatic mutation detection based on mismatch scanning enzymes like MutY, TDG or T4 endonuclease VII for mutation pre-screening has also been employed (21) (22) (23) (24) (25) , albeit with modest success since these enzymes cannot detect all possible mutations and deletions (22) and some of them have substantial activity on homoduplex DNA (16) . Recently an enzymatic mutation scanning method based on the Surveyor TM (CELI/II) nuclease (26, 27) combined with dHPLC or gel electrophoresis detection was introduced that shows satisfactory selectivity and reliability (1% mutant to wild-type alleles is detectable) while it also identifies all base substitutions and small deletions that are important to cancer (17, 28) or to biotechnology and plant genetic applications [TILLING method (29) (30) (31) (32) (33) (34) ]. While reliable, the use of dHPLC for examining Surveyor TM -generated DNA fragments is a slow endpoint detection method restricted to examining a single DNA fragment at a time and the resulting DNA fragments cannot be sequenced. This limits analysis of cancer specimens when numerous samples or genetic regions need to be screened. We introduce a new approach that enables Surveyor TM to scan for mutations over one or several PCR products simultaneously and selectively amplify and isolate the mutation-containing DNA fragment(s) via linkermediated PCR. By selectively amplifying mutationcontaining DNA from wild-type fragments, the present approach de-couples enzymatic mutation scanning from the endpoint detection step. As a result, following enzymatic action on mismatches any chosen DNA detection method (real-time PCR, gel/capillary electrophoresis, microarray-based detection) can potentially be used to identify the mutated DNA fragments in a simplex or multiplex fashion. Here we utilize real-time PCR coupled with melting curve analysis (Surveyor TMmediated Real Time Melting, s-RT-MELT) to validate the new technology. We demonstrate that this approach increases the mutation scanning throughput by 1-2 orders of magnitude when several (4100) samples are to be pre-scanned for mutations, enables mutation scanning over several PCR fragments simultaneously and mutationpositive samples can be directly sequenced when somatic mutations are at a low-level ($1-10% mutantto-wild-type ratio) in surgical cancer specimens. Genomic DNA from cell lines with defined mutations in p53 exons, DU145 (exon 6), SW480 (exons 8 and 9), DLD1 (exon 7) and BT483 (exon 7) was extracted from cell lines purchased from the American Type Culture Collection (ATCC), or purchased as purified DNA when available. Surgical colon and lung cancer tumor samples were obtained from the Massachusetts General Hospital Tumor Bank following Internal Review Board approval. DNA from the EGFR mutation-positive cell lines A549, HCC827, H1975 and LU011 and from formalin-fixedparaffin-embedded lung cancer samples were obtained from the Lowe Center for Thoracic Oncology, Dana Farber Cancer Institute following Internal Review Board approval. We isolated genomic DNA using DNeasy TM Tissue Kit (Qiagen). PCR with primers containing 5 0 -GC-clamp and 5 0 -M13 Sequences for the 5 0 M13 and GC-clamp portion of the primers, as well as the gene-specific portion of the primers used in this investigation are listed in Supplementary Table 1 . The M13f and GC-clamp sequence was added to the 5 0 end of forward and reverse gene-specific primers respectively, or vice versa. Twenty microliter PCR reactions were performed from genomic DNA with final concentrations of reagents as follows: 1X JumpStart TM buffer (Sigma), 0.2 mM each dNTP, 0.2 mM forward and reverse primer, 1X JumpStart TM Taq polymerase (Sigma). PCR cycling was done on a Perkin Elmer 9600 PCR machine. The cycling conditions were: 948C, 90 s; (948C, 20 s/658C, 20 s/688C, 1 min) Â 10 cycles, with annealing temperature decreasing 18C/cycle, touch-down PCR; (948C, 20 s/558C, 20 s/688C, 1 min) Â 30 cycles; 688C, 5 min. This PCR program was linked to a program for denaturation and re-annealing of the PCR product over 10 min. Five-microliter PCR product (300-500 ng) was mixed with 0.5 ml Enhancer TM and 0.5 ml Surveyor TM (Transgenomic) and incubated at 428C for 20 min followed by adding 0.5 ml Stop-solution, as per manufacturer's protocol. The inactivated Surveyor TM -digested product was purified with PCR QiaQuick TM purification kit (Qiagen) and eluted in 35 ml water. In some experiments, the PCR product was mixed with an approximately equal amount of PCR product from wild-type DNA prior to forming cross-hybridized sequences, to facilitate detection of homozygous mutations. Addition of polyA-tail on the 3 0 -end Following purification of the Surveyor TM -treated sample, Poly-adenine 'tail' was added to the 3 0 -ends of DNA fragments. For each reaction, we added 5 ml purified surveyor-digested PCR product to a final volume of 20 ml with final concentration of 1X reaction buffer-4, 1X CoCL 2 , 0.2 mM dATP, 4 U Terminal Transferase (New England Biolabs). The reaction was incubated at 378C for 10 min and inactivated by heating at 758C for 10 min. The real-time PCR amplification was performed using Titanium-Taq TM polymerase (BD-Biosciences -Clontech) in a Smart Cycler (Cepheid) real-time PCR machine. For each reaction, we added 0.5 ml polyA-tailed DNA to a final volume of 20 ml with final concentration of 1X Titanium buffer, 0.2 mM each dNTP, 0.1 Â LCGreen (Idaho Technologies), 0.2 mM m13f primer, 0.2 mM oligodT-anchor mix GACCACGCGTATCGATGTCG ACTTTTTTTTTTTTTTTTV [V represents A, C and G each oligodT-anchor concentration is 0.067 mM, as per RACE protocol (35)], 1 Â Titanium TM polymerase (Clontech-BD Biosciences). The thermocycling program was as following: 1 cycle of 948C for 2 min, 25 cycles of 948C for 15 s, 558C for 20 s and 688C for 30 s for reading fluorescence. Temperature titration was performed using different denaturation temperatures, 94-828C to experimentally determine conditions that selectively enable mutation-containing fragments to amplify. The real-time PCR step was immediately followed by real-time differential melting curve analysis using the SmartCycler TM machine. DNA melting was performed immediately following PCR on the Smart Cycler I machine. Samples were heated from 708C to 958C at 0.18C/s. Differential fluorescent intensity curves (ÀdF/dT) were smoothed and used for identification of melting peak (s). Altenatively, real-time PCR products were examined via dHPLC chromatography on a WAVE TM system (Transgenomic). Mutation-positive PCR products were purified via PCR purification kit (Qiagen) and sequenced using the M13f primer. All experiments were repeated at least three times in independent runs from genomic DNA. The OpenArray TM high-throughput, massively parallel real-time PCR platform (36) (BioTrove) was tested for compatibility with s-RT-MELT. Forty-eight samples of p53 exon 8 PCR products were generated from 48 different lung adenocarcinoma samples and mutation-containing cell lines and processed via the hybridization and enzymatic steps of s-RT-MELT. Real-time PCR in the OpenArray TM platform was performed with the LightCycler FastStart TM DNA Master SYBR Green TM I (Roche) using 0.2 mM M13f and 0.2 mM oligodT-anchor-mix as primers pre-positioned on the array through-holes (36) and polyA-tailed DNA as template. The cycling conditions were as follows: 1 cycle at 948C for 2 min, 25 cycles of 908C for 15 s, 558C for 20 s and 688C for 30 s for reading fluorescence using a high sensitivity imaging camera (36) . The real-time PCR step was immediately followed by real-time differential melting curve analysis. Raw data were exported in Excel software for further analysis. The OpenArray TM experiment was repeated twice at the company's headquarters. To estimate T m,min, the PCR denaturation temperature below which PCR is not efficient it was assumed, as an initial approximation, that495% hypochromicity must be present for PCR to work (i.e. any given sequence must be completely denatured, otherwise it re-forms immediately when temperature is lowered in the reaction and inhibits primer binding). The percent melting (hypochromicity)versus-temperature relations for GC-clamp-containing PCR products and Surveyor TM activity-generated products were estimated using the POLAND algorithm (37) , and the thermodynamic parameters determined by Blake and Delcourt for 75 mM NaCl in the solution (38) were used. In order to force agreement at a single point, predicted and observed values for a p53 exon 8 sequence containing a short GC-clamp were normalized at 888C. This shift accounts for the influence on T m,min of NaCl and Mg++ content in the reaction, the presence of the SYBR-GREEN/LC-GREEN dyes and the proprietary composition of PCR buffers. The T m,min of all other PCR products was then estimated using these semi-empirically determined parameters. The 'enriched PCR' method by Behn et al. (39) was used to sequence codon 273 mutation of p53 exon 8 from sample CT20 and wild-type samples. In addition, a second method [Amplification via Primer-Ligation At The Mutation (40, 41) ] was used to distinguish mutant and wild-type samples by virtue of the de novo Nla-III site generated in the mutant sample by the p53 codon 273 G4A mutation. The s-RT-MELT assay converts PCR fragments generated at positions of mutations by the Surveyor TM enzyme to fully amplifiable sequences that enable selective PCR amplification in a subsequent quantitative PCR detection method. Following denaturation and re-annealing of PCR products that leads to formation of cross-hybridized sequences at the positions of mutations ( Figure 1A ) the sample is exposed to Surveyor TM endonuclease that recognizes base pair mismatches or small loops with high specificity (28) and generates a break on both DNA strands 3 0 to the mismatch. The resulting DNA fragments participate in a terminal transferase (TdT) reaction that leads to polynucleotide 'tailing' (sequential addition of adenine, poly-A-tail) at the 3 0 -ends. A real-time PCR reaction is subsequently performed using adjusted conditions that enable selective amplification of the mutantonly fragments, followed by real-time melting curve analysis for identification of mutations in the presence of SYBR-GREEN TM or LC-GREEN TM DNA dye. To enable selective amplification of the mutationcontaining fragments in the real-time PCR step, modified primers are employed for the original amplification from genomic DNA ( Figure 1B ). The forward primer contains a region specific to the target gene and a high melting domain (GC-clamp), while the reverse primer contains a region specific to the target gene and an M13 tail (or vice versa). Following the TdT tailing reaction, the M13 primer is used for real-time PCR in conjunction with a primer that binds to the poly-A tail. The denaturation temperature of the real-time PCR reaction is lowered to enable PCR amplification only for fragments that do not contain GC-clamps. Because the PCR products that escape digestion by Surveyor TM contain GC-clamps ( Figure 1B ), these fragments do not amplify efficiently during PCR, thereby enabling selective amplification of Surveyor TM -selected fragments, i.e. an effective 'purification' of mutation-containing fragments. The subsequent closed-tube melting curve analysis enables clear separation of true mutant sequences from PCR dimers or other artifacts. Because s-RT-MELT does not require size-separation for identification of enzymatically generated fragments, more than one sequence can be scanned in parallel for unknown mutations in a single-tube reaction of Surveyor TM . This simple procedure enables the specificity of the Surveyor TM enzyme to be combined with the throughput and convenience of real-time PCR for rapid mutation scanning. Finally, because the amplified mutated sequences contain defined primers at their ends, direct sequencing of enzymatically selected PCR products is readily possible following the real-time melting step, enabling sequencing of low-level mutations identified by Surveyor TM . To provide initial proof of principle for unknown mutation scanning using s-RT-MELT we utilized cell lines and tumor samples containing sequencing-identified mutations at several positions of p53 exon 8. Figure 2A depicts dHPLC chromatograms of the products obtained using a sample containing a p53 exon 8 G4A mutation or a wild-type sample. The standard Surveyor TM -dHPLC approach (28) was first employed to identify the mutation following PCR amplification of exon 8 from genomic DNA. The resulting dHPLC traces contain a single product for the wild-type and two products for the mutation-containing sequences (Figure 2A , curves 1 and 2, respectively). Next, s-RT-MELT was used to screen the same p53 exon 8 sequence. Following PCR amplification Simplex or Multiplex PCR amplification of one or more exons PCR PRODUCT(s) G C -C L A M P M 1 3 Self-hybridize or cross-hybridize with wild type DNA: generate mismatches at positions of mutations in one or more PCR fragments Scan for mismatches all fragments simultaneously using CEL I /Surveyor TM enzyme. CEL I /Surveyor TM enzyme. Use TdT enzyme to add oligonucleotide tail (e.g. oligo-dA) to 3′OH ends, to serve as primer anchor Un-digested fragments Digested fragments TdT tailing of 3′ DNA ends Amplify only mutated fragment(s) coupled w. real time melting analysis (see B) Detect mutations via closed tube, high-throughput melting curve analysis. IF positive, sequence the amplified mutated DNA fragment with GC/M13-modified primers we cross-hybridized PCR products and exposed them to Surveyor TM and TdT tailing. The subsequent real-time PCR was run at different denaturation temperatures and the products were examined either via dHPLC or via real-time melting-curve analysis. At the standard denaturation temperature of 948C the mutation-containing sample contains two peaks, corresponding to the anticipated amplification of both Surveyor TM -digested and un-digested fragments (Figure 2A, curve 3) . However, when the PCR denaturation temperature is lowered (e.g. 86-888C) a single PCR product is generated for the mutant sample, while the wild-type sample demonstrates no product (Figure 2A , curves 4-7). In Figure 2B , real-time differential melting curves for the PCR reaction run at 888C are depicted. A peak corresponding to the PCR product from the mutant sample is again clearly evident, which is absent in the wildtype sample. Finally, Figure 2C depicts sequencing of the s-RT-MELT-generated PCR fragment, as well as the direct sequencing from genomic DNA. The G4A mutation is evident in both samples. In the s-RT-MELT product the anticipated addition of the poly-A tail at the 3 0 -position next to the mutation is illustrated. To examine the selectivity of s-RT-MELT, dilutions of mutant to wild-type DNA were performed using DNA from SW-480 cells that harbor a p53 exon 8 14487G4A homozygous mutation. The real-time PCR reaction was again performed at 888C and mutant-to wild-type ratios of $1-10% were distinguished from the wild-type using either dHPLC ( Figure 2D ) or melting curve analysis ( Figure 2E ). In these samples, direct di-deoxy-sequencing could not identify a mutation if the ratio of mutant-towild-type was 5$30-40% (data not shown). On the other hand, sequencing of s-RT-MELT products was possible including the lower dilutions ( Figure 2F ). sRT-MELT sequencing generated traces with poly-A tails depicting the presence and the position of the mutation, although the exact nucleotide change was less clear than the one in exon 5 (i.e. the position AE1 base from the mutation might also be confused to be a mutation). The reason for this AE1 base ambiguity of the exact position of the mutation can be probably understood. The PCR performed following poly-A tail addition contains an equimolar mixture of three reverse primers (3 0 ending in V = G, A or C). Depending on the exact nucleotide at the mutation, the correct primer should in theory be preferred, while the other two primers should not allow efficient polymerase extension due to the mismatched 3 0 -end. However, in practice this 'allele-specific PCR' step occasionally allows 3 0 -mismatched primer extension, enabling more than one version of the primer to amplify over the position of the mutation, or alternatively the incorporation of the poly-A tail may occur AE1 base from the exact position of the mutation. We conclude that in certain cases sRT-MELT indicates the position of the mutation to within 1 base, while in others (e.g. p53 exon 5) it indicates the position 'and' the actual nucleotide change. Next, p53 exon 8 was amplified using DNA from a group of 48 surgical lung adenocarcinoma samples and s-RT-MELT was used for the screening of unknown mutations via melting curve analysis. Mutations at different positions along exon 8 were present in several of these clinical samples, as indicated by the shift in melting profiles obtained ( Figure 2H ) and subsequently verified via sequencing. In this set of samples, sRT-MELT-sequencing detected a low-level mutation on a colon cancer specimen (CT20) that direct sequencing failed to identify ( Figure 2I ). As with Figure 2F , sequencing of sample CT20 indicated the position of poly-A tail addition to within one base, but the actual nucleotide change was difficult to identify. To exclude the possibility for a false positive, two independent RFLP-based methods were used to verify the presence of the mutation. Thus, since the position of poly-A tail addition was known ( Figure 2I , codon 273 of p53 exon 8) the mismatched primer approach by Behn et al. (42) was used to introduce an MluI restriction site for the wild-type p53 sample but not for the codon 273 mutants. Subsequent restriction with MluI enzyme followed by PCR generated a product with a 14487G4A mutation for the CT20 sample but not for the wild-type sample (Supplementary Figure 1 , Frame A). As an additional verification for the low-level CT20 mutation, we observed that G4A mutation generates a de-novo Nla-III site at the position of the mutation. Accordingly, we applied 'Amplification via Primer-Ligation At The Mutation', a method that we described previously (40, 41) to ligate a primer at the Nla-III-digested site, and preferentially amplified the mutant fragment in a second PCR. The sequenced PCR product identified again the 14487G4A mutation (Supplementary Figure 1, Frame B) . In conclusion, sRT-MELT identified correctly a p53 codon 273 low-level mutation on CT20 that was missed by regular sequencing. This is very significant as p53 exon 8 mutations at codon 273 have been associated with bad prognosis in cancer (43, 44) . Table 2 of Supplementary Data depicts a good agreement between standard Surveyor screening, s-RT-MELT screening and di-deoxy-sequencing, except for the low-level mutation discovered on sample CT20 via s-RT-MELT. s-RT-MELT-sequencing traces for two samples with p53 exons 6 and 7 mutations are also depicted. The data in Figures 2A-D and H indicate a lack of substantial PCR amplification at denaturation temperatures 4888C for fragments containing the GC-clamp and a selective amplification of the mutation-containing fragments for several different mutation positions on p53 exon 8. To estimate the influence of the GC-clamp length on PCR efficiency versus temperature and the PCR amplification of fragments generated for mutations lying at different positions along the sequence, a calculation based on the POLAND algorithm (37) was performed. The predicted minimum temperatures for substantial PCR amplification were then plotted versus the experimentally observed values. Three possibilities were simulated, no GC-clamp, 26 nucleotides (nt) GC-clamp and 117-nt GC-clamp. DNA fragments corresponding to mutations at several positions along exon 8 were also simulated and compared to the experimentally observed minimum temperatures for generating a PCR product for three samples that contained mutations at different positions along p53 exon 8 (SW480, CT5 and TL50). The results ( Figure 2G ) indicate agreement to within $1.08C between theoretical prediction and experimental observation. For denaturation temperatures in the region 85-888C in combination with a 26-nt GC-clamp all the available mutations on p53 exon 8 are predicted to result in selective amplification of the mutation-containing fragment and inhibition of the GC-clamp-containing fragment. This prediction is consistent with the experimental results obtained from PCR temperature-titration experiments ( Figure 2G ). The developed calculation algorithm can thus be used to predict the appropriate PCR denaturation temperature for additional PCR fragment/GC-clamp combinations. As a further validation for s-RT-MELT, we utilized the method to identify mutations in additional p53 exons. Figure 3A depicts the chromatographs obtained when a 1:1 mixture of DNA from SW-480 cells (homozygous mutation at p53 14686 C4T exon 9) and from wild-type cells was screened. The real-time PCR reaction was performed at different denaturation temperatures and the products were examined both via dHPLC and via melting curve analysis for comparison. As also observed for p53 exon 8, at 948C denaturation temperature both the Surveyor TM -digested and the undigested PCR products are amplified during real-time PCR ( Figure 3A , curves 1 and 2, mutant and wild-type, respectively). By lowering denaturation temperature to 858C or 848C, a single PCR product is obtained from the mutant while no product, other than primer dimer, is obtained by the wild-type sample ( Figure 3A , curves 3-6). Figure 3B depicts the melting curves obtained following real-time PCR at 858C denaturation temperature for the mutant and wild-type samples. s-RT-MELT was subsequently applied in the same manner to screen for p53 mutations in exons 5-7 from cell lines and surgical colon samples harboring sequencing-identified mutations including a single-base frameshift mutation in exon 7 (listed in Supplementary Table 2 ). The melting curves from mutant and wild-type samples in p53 exons 5-7 are depicted in Figure 3C -E. The data indicate that results similar to those obtained for p53 exon 8 are also obtained for p53 exons 5, 6, 7 and 9. Detection of mutations in EGFR exons 18-21 is of particular clinical interest as these alterations can modulate response to EGFR inhibitors in lung adenocarcinoma patients (2,3) . Figures 3F, G and H depict the application of s-RT-MELT for screening DNA from lung cancer cell lines that harbor dHPLC-identified alterations in EGFR exons 19-21, including a two-codon deletion (del L747-E749, exon 19). The ability of s-RT-MELT for detecting low-level EGFR mutations was evaluated by performing DNA dilutions of a heterozygous EGFR exon 20 into a homozygous sample. A 1-10% mutant-to-wildtype ratio was detectable in this dilution experiment ( Figure 3F) . Finally, the application of s-RT-MELT in detecting mutations in DNA from formalin-fixed paraffinembedded (FFPE) samples was examined by screening four clinical FFPE lung adenocarcinoma specimens. Two of these samples were known to harbor EGFR exon 21 mutations (L858R), while the other two were negative for mutations when independently evaluated via dHPLC (28) . Figure 3I demonstrates the identification of the mutational status of these samples via s-PCR-MELT. Multiplex s-RT-MELT or OpenArray TM -based s-RT-MELT increases the throughput of mutation scanning A significant potential advantage of enzymatic mutation scanning is the ability to screen several sequences simultaneously for mutations. To demonstrate that s-RT-MELT can be used for parallel scanning of mutations in several PCR products, we mixed equimolar amounts of PCR products from p53 exons 5-9 containing mutations either in exon 8 or in exon 9. We then formed 'cross-hybridized sequences' and screened the mixture for mutations in p53 exons 5-9 in a single tube using s-RT-MELT, as depicted in Figure 1A . Following real-time PCR and melting curve analysis, the exon 8 or exon 9 mutants were clearly distinguished from the wild-type sample ( Figure 4A, curves 1-3) . Next, the mutant exon 8 DNA sample was first diluted 10-fold into wild-type exon 8 and the equimolar mixture of p53 exons 5-9 was prepared and screened again in a single tube via s-RT-MELT. The exon 8 mutation was again distinguished from the wild-type mixture of exons ( Figure 4B , curves 1-3). Since 480% of p53 mutations in human tumors are encountered in exons 5-9 (45), the multiplex single-tube s-RT-MELT reaction could be used to identify most p53 mutations encountered in clinical tumor samples. Combined with multiplex PCR directly from genomic DNA, this approach could result to a convenient, high-throughput method for mutation scanning. By adopting a real-time PCR platform as endpoint detection for s-RT-MELT, the throughput for mutation scanning increases drastically over other mutation prescreening approaches that utilize dHPLC, or capillary and gel electrophoresis. To demonstrate better this point, a highly parallel nano-technology platform was adopted for the real-time PCR step of s-RT-MELT that enables an array of 3072 nl volume real-time PCR reactions (OpenArray TM system) to be carried-out simultaneously followed by differential melting curve analysis (36) . As a proof of principle of the compatibility of s-RT-MELT with OpenArray TM , p53 exon 8 PCR products were generated from 48 different lung adenocarcinoma samples and mutation-containing cell lines and processed via the hybridization and enzymatic steps of s-RT-MELT. The 48 samples were each dispensed in 10 replicate nano-liter volume reactions on OpenArray TM plates pre-fabricated to contain the appropriate primers and amplified in 3072 real-time PCR reactions using a denaturation temperature of 908C in the presence of SYBR-GREEN I dye. Melting curves were subsequently obtained using the OpenArray TM melting curve analysis mode. The PCR growth curves and smoothed differential melting curves obtained distinguish clearly the mutation-containing samples from wild-type samples ( Figure 4C and D, representative results from 3072 reactions). Furthermore, identification of mutation-containing samples is in good agreement between the conventional and the nano-technology platforms ( Figure 4D versus Figure 2H ). These data indicate that s-RT-MELT is compatible with high-throughput nano-technology detection formats and reiterates the advantage of de-coupling enzymatic selection from the detection step. Comparison of the throughput using conventional pre-screening method (dHPLC or dHPLC/Surveyor TM ) to s-RT-MELT (Table 1) indicates that s-RT-MELT is 1-2 orders of magnitude faster when a large number of samples (4100) are screened for mutations. If the multiplex s-RT-MELT format is adopted, the throughput can increase further. The intrinsic potential of enzymatic mutation scanning for parallel identification of mutations can, in principle, be very high since the enzyme operates on numerous distinct mismatch-containing sequences on a molecule-tomolecule basis thus providing highly parallel mutation scanning. However, in the past the selectivity of the enzymes used and the endpoint detection method has limited the realization of this potential. Here we enabled Surveyor TM , an endonuclease that recognizes selectively mismatches formed by mutations and small deletions following 'cross-hybridized sequence' formation, to generate mutation-specific DNA fragments that are amplified and screened via differential melting curve analysis. The replacement of size-separation methods (capillary/gel electrophoresis, dHPLC) by real-time PCR technology as the endpoint detection platforms and the ability to scan more than one sequences in parallel result in a highly increased throughput for s-RT-MELT while retaining the ability to detect diverse mutations at low-levels. Cel I/II endonucleases have also been known to have exonuclease activity on 5 0 DNA-ends (26, 27) . For this reason, s-RT-MELT was designed to attach an oligonucleotide linker to the 3 0 -DNA ends via terminal transferase (TdT) instead of using the 5 0 -DNA ends. The exonuclease activity also tends to degrade the attached 5 0 -GC-clamps in s-RT-MELT, thereby eliminating their influence in reducing amplification of un-digested fragments. We found that if exposure of DNA 'cross-hybridized sequences' to Surveyor TM is limited to 15-20 min, the substantial degradation of 5 0 -GC-clamps is avoided. For multiplexing mutation detection using several PCR products simultaneously, the size of the GC-clamp on each PCR amplicon may need to be individually adjusted to ensure that mutations along all sequence positions of the PCR products included in the mixture can be screened at a single real-time PCR temperature and that undigested fragments do not amplify. The calculational tools developed in this work can be used to guide the individual design of GC-clamps. s-RT-MELT detects heterozygous SNPs as well as mutations. As with other mutation prescreening techniques, the presence of a SNP concurrently with a mutation might be difficult to identify without performing sequencing. Because SNPs occur at fixed positions, melting peaks originating from SNPs have a reproducible pattern and melting temperatures (46, 47) thus in many cases they should be distinguishable from mutations. Finally, it is noteworthy that s-RT-MELT is a general methodology that may also be applied to isolate mutations using mismatch-cutting enzymes other than Surveyor TM when enzymes with satisfactory properties for mutation detection become available. Detection platforms other than real-time PCR/melting (e.g. DNA microarraybased) may also be envisioned following enzymatic mutation selection. In summary, we developed a new method for rapid mutation scanning, s-RT-MELT that utilizes the Cel I/II (Surveyor TM ) and terminal deoxy-nucleotide transferase (TdT) enzymes to isolate and amplify mutation-containing DNA fragments without the requirement of DNA sizedependent techniques. Besides enabling highly increased throughput, multiplexed mutation screening and direct sequencing of the identified mutant DNA fragments, s-RT-MELT also retains the advantages of the Surveyor endonuclease over alternative pre-screening methods, such as reliability and identification of genetic alterations present at low (1-10%) fractions in the sample. s-RT-MELT provides a significant advancement in unknown mutation scanning in cancer research and diagnostics as well as for general medical, biological and biotechnology applications.
94
Rapid Identification of Malaria Vaccine Candidates Based on α-Helical Coiled Coil Protein Motif
To identify malaria antigens for vaccine development, we selected α-helical coiled coil domains of proteins predicted to be present in the parasite erythrocytic stage. The corresponding synthetic peptides are expected to mimic structurally “native” epitopes. Indeed the 95 chemically synthesized peptides were all specifically recognized by human immune sera, though at various prevalence. Peptide specific antibodies were obtained both by affinity-purification from malaria immune sera and by immunization of mice. These antibodies did not show significant cross reactions, i.e., they were specific for the original peptide, reacted with native parasite proteins in infected erythrocytes and several were active in inhibiting in vitro parasite growth. Circular dichroism studies indicated that the selected peptides assumed partial or high α-helical content. Thus, we demonstrate that the bioinformatics/chemical synthesis approach described here can lead to the rapid identification of molecules which target biologically active antibodies, thus identifying suitable vaccine candidates. This strategy can be, in principle, extended to vaccine discovery in a wide range of other pathogens.
Human Plasmodium falciparum (Pf) infection is a dramatic public health problem. Today approximately forty percent of the world's population is at risk of malaria. Malaria causes more than 300 million acute clinical cases, and at least one million deaths annually. Ninety percent of malaria deaths occur in sub-Saharan African countries mostly among young children and pregnant women (http://rbm.who.int). Thus, there is an urgent need of a malaria vaccine. However, vaccine discovery, in general and particularly in malaria, is still a very empirical process. In fact, protective antigens do not bear any structural, physico-chemical or sequence-related characteristics that would allow their identification. For protective humoral responses, the only recognized characteristics of antigens are their antigenicity/immunogenicity and accessibility. In addition, the lack of surrogate markers of protection renders the vaccine discovery process difficult and time consuming. Hence, in spite of a constant and impressive progress in molecular biology techniques and antigen identification and expression [1, 2] , vaccine discovery is still labor-intensive, making the approach fastidious, costly and poorly adapted to highthroughput screening. Proper protein folding and solubility remain a limitation in numerous cases. Thus, overcoming the bottlenecks of manufacturing and identification of fragments/proteins, as possible targets of a protective immune response still constitute a scientific and technical challenge. We addressed this challenge by combining bioinformatics, chemical peptide synthesis and functional protection assays. We focused on the search for a-helical coiled coil motifs that, in general, do not exhibit a folding problem, and are a target of effective antibodies for several current malaria vaccine candidates (eg. LSA-1 [3] , LSA-3 [4] , MSP-3 [5] , and MSP-6 [6] and other pathogens [7] . MSP-1, another leading malaria vaccine [8] , also contains predicted a-helical coiled coil regions (unpublished results). Our choice was based on the following considerations. First, the a-helical coiled coil motif bears a characteristic seven amino acid residue repeat (abcdefg) n with hydrophobic residues located in a and d positions and hydrophilic residues generally elsewhere. This motif can be easily identified by bioinformatic analysis. Secondly, an important known characteristic of the ahelical coiled coil domains is that, taken separately from the whole protein, they frequently and readily fold into the same stable oligomeric structure [9] . Thirdly, for this reason, the a-helical coiled coil fragments are frequently recognized by conformational dependent antibodies, and can similarly elicit antibodies reactive with structurally ''native'' epitopes. In addition, these domains are short (about 40 residues) and can be rapidly produced by chemical synthesis. Furthermore, when the antibodies have an anti-parasite biological activity, this designates the corresponding proteins/ fragments as potential, novel vaccine candidates to be further developed and assessed. Here, we focus on the Pf parasite erythrocytic stage, a target of protective antibodies and describe a straightforward, rapid procedure based on bioinformatic analysis of a-helical coiled-coil motifs and peptide synthesis. The screening of the Pf genome [10] using generalized sequence profiles [11] identified several hundred proteins containing putative a-helical coiled coil motifs. Through proteome and transcriptome data [12] [13] [14] we assessed which of these molecules are expressed in the Pf parasite erythrocytic stage. The combined analysis/assessment identified over 100 segments associated with this stage and displaying the putative a-helical coiled coil motifs with high probability score (Table S1) . Out of these a-helical coiled coil fragments, in general 30-40 amino acids long, present either in the same protein or in different ones, 95 were chemically synthesized and HPLC purified. Among them, longer peptides (up to 70 amino acids), which contained one or more a-helical coiled coil domains, were also synthesized (antigenS 1, 12 and 83; Table S1 ). The selected antigens were then tested in ELISA assays for reactivity with three panels of sera obtained from adult donors from Burkina Faso, Tanzania and Colombia, respectively. To our surprise, all of the a-helical coiled coil fragments were antigenic, though the prevalence of responders varied greatly (Tables 1 and S1 ). In this manner, 71 proteins were identified whose lengths varied from 200 to 10,000 amino acids. Twenty-one peptides with the highest prevalence of responders and ELISA mean OD value were selected for further studies. Variation in recognition among the three panels of sera may be due to differences in the genetic background of the hosts, of the parasites and, most likely, to distinct malaria transmission conditions in the three regions. The high level of recognition of the a-helical coiled coil motifs may be explained by the fact that taken separately from the whole protein these fragments readily fold into the same stable structure in aqueous solution. Indeed, circular dichroism (CD) studies of selected peptides associated with biological activities (Tables 1 and 2) indicate that they predominantly assume an a-helical conformation in water. Peptides 14, 27 and 45 ( Figure S1A ) exhibit a CD pattern characteristic of a high a-helical content, whereas the remaining peptides show CD profiles similar to that shown for peptide 12 ( Figure S1B ) or intermediate between those shown in Figures S1A and S1B characteristic of a partial a-helical organization. When analyzed by size exclusion chromatography on FPLC columns, peptides presented elution profiles between those exhibited by chymotrypsin and ribonuclease (MW 24 and 13kDa, respectively). The CD and size exclusion chromatography results suggest that peptides adopt an a-helical coiled-coil structure, which need to be unambiguously ascertained by NMR and ultra-centrifugation studies. To test the biological activity of peptide-specific antibodies, the latter were purified by affinity chromatography using three serum pools obtained from Papua New Guinean adults. The 3 serum pools were first tested in ELISA assays against 21 peptides that were the most antigenic (Table 1) ; from these, 18 peptide-specific antibodies were purified from the most positive serum pool and tested again in ELISA. These 18 antibodies all reacted with parasite native proteins in infected red blood cells as shown by IFAT ( Figure 1A ; Table 2 ). Reactivity was restricted to blood stages, since the antibodies did not react with sporozoites stages (data not shown), and this reactivity was also peptide-specific as shown by IFAT competition assays with the corresponding peptide ( Figure 1A ). The specificity of the antibodies obtained was investigated in detail, particularly since several peptides contain glutamic acid (Glu)-rich sequences which are known to generate cross reactivity among several malarial Glu-rich proteins [15] . Cross-reactions were systematically investigated using each of the 18 affinitypurified antibodies on each of the 18 peptides. Results show thatwith few exceptions-each antibody preferentially recognizes the peptide against which antibodies were affinity-purified, i.e. they are specific for the corresponding peptide (Table S2) . To determine if non-specific antibody binding to solid phase-adsorbed antigens could be responsible for the rare cross-reactivities detected, ELISA competition assays were performed. To this end, binding of antibodies to the solid phase-adsorbed antigen was competed against increasing concentrations of the homologous or cross-reacting peptides. Only homologous peptides competed best whereas peptides having sequence similarity did not (Figures 2A, 2B , and S2B), or at a much higher concentration ( Figure S2A ) including the shorter Glu-rich peptides derived from the Cterminus of peptide 27 and 45 (Figures 2A and 2B) . Finally, the pattern of recognition of peptides by the various sera tested, which differ markedly from one to the other (data not shown), confirms the above results i.e., specificity of antibodies to the corresponding peptide. Antibodies corresponding to the 18 selected peptides were tested for direct and cell-mediated anti-parasite activity. Clinical experiments have shown that the Antibody-Dependent Cellmediated Inhibition (ADCI) of P. falciparum malaria represents one of the mechanisms controlling parasitemia and thereby clinical manifestations in humans [16] . Twelve peptide-specific antibodies proved able to induce a strong (more than 40%) and intermediate (lower than 40%) monocyte-dependent parasite killing (Table 1) , whereas, in the absence of monocytes, no direct effect of antibodies on parasite growth was observed. The effects were in the range observed with antibodies from African adults who have the highest natural protection known against malaria. Therefore peptidespecific, human affinity-purified antibodies were functionally effective as shown by their ability to react with parasite proteins and to inhibit parasite growth. Thus, in vitro functional assays show that peptide-specific antibodies elicited by natural exposure to the parasite can induce protective mechanisms effective against malaria. Sixteen peptides -twelve targeted by ADCI positive antibodies and four controls-were used to immunize CB6F1 mice ( Table 2) . Eleven of them elicited an intermediate or high antibody response, four of which also recognized the parasite protein in infected erythrocytes as determined by IFAT ( Figure 1B ; Table 2 ). As seen before for human antibodies, recognition was restricted to blood stages since sporozoites were negative in IFAT assays (data not shown) and by IFAT competition assays with the corresponding peptide ( Figure 1B ). Anti-peptide 27 mouse antibodies, which are positive in IFAT, are also specific for the homologous peptide 27 but not for the sequence related peptides 9, 12 and 45 (Table S2) . Thus, peptides, which were chosen for their propensity to form ahelical coiled coil, can induce the production of antibodies that recognize epitopes present in the native protein. Improvement of the immunogenicity and structural specificity of the remaining peptides might be achieved in the future by a) a short elongation at the N-and C-terminal ends, b) stabilizing the a-helix as suggested by Cooper et al. and Lu and Hodges [17, 18] and/or c) use of other adjuvants. Genetic polymorphism in current vaccine candidates is a major limitation to vaccine development. However available genotyping studies of our peptide sequences in parasite isolates of worldwide origin indicate very limited polymorphism (PlasmoDB 5.2, and unpublished sequencing data). A few peptide DNA sequences show deletion of one entire heptad repeat so that the shorter region still preserves its potential for the a-helical coiled coil formation. With regard to the structural features and cellular location prediction of the proteins corresponding to the peptides selected for ADCI assays ( Table 1) , 15 of the proteins contain a pentapeptide conforming to the PEXEL consensus [19, 20; 21, 22] , but that none of these have a position within the amino acid sequence that conforms to the location of known active PEXEL motifs (see Materials and Methods and membrane segments, and none of them has a GPI anchor. Only one protein contains a signal sequence. Fourteen proteins are predicted to be in the cytoplasm, one in the nucleus, one in the mitochondria, and one in the peroxysomes (Table S3 ). The prediction of the sub-cellular localization of these proteins should be taken with caution because gene annotation is being constantly updated and/or protein trafficking of the parasite is complex and not fully elucidated [23] . Further investigations will be required to determine the actual localization of the corresponding antigens. The predicted localization is a priori surprising for molecules able to trigger an ADCI activity. However, recent studies have shown that in addition to merozoite surface proteins, soluble proteins released at the time of schizont rupture were equally effective at triggering ADCI provided they defined at least two epitopes [24] , which is the case for a-helical coiled coil heptad repeats. Therefore, molecules expressing a trans-membrane domain that can be exported to the parasite or host cell membrane, as well as molecules present in the cytoplasm of maturing schizonts and released by bursting schizonts can trigger antibodies to cross-link Fc-c receptors on monocytes to achieve Pf parasite killing. In conclusion, an approach combining a genome-wide search by bioinformatics of a-helical coiled coil protein motifs and chemical synthesis can lead to the rapid identification and development of new malaria vaccine candidates. In fact, this approach is straightforward and easy to scale up; vaccine formulations may comprise mixtures of peptides or single constructs made up of several epitopes. In principle, this strategy can be extended to the discovery of proteins and vaccine candidates in other complex pathogens. The Pf 3D7 genome [10] was used for the bioinformatics analysis. The generalized sequence profile method and the pftools package [11] were used to search for the short a-helical coiled coil domains. The coiled coil profiles were constructed using an alignment of several amino acid sequences corresponding to the known coiled coil domain. Two profiles containing four and five heptad repeats were used for the analysis. The cut-off levels of the profiles were chosen by tests performed against sequence database of proteins with the known 3D structures. Subsequently, the coiled coil Figure 1 . Immunofluorescence microscopy analysis of Pf 3D7 parasites with peptide specific antibodies. Acetone/methanol-fixed schizonts and merozoites were reacted with A: human peptide specific, affinity purified antibodies obtained with peptides 12 and 14 (Table 1) and B: sera from mice immunized with peptide 27 (Table 1) . Grey: bright field images; blue staining: indicates DAPI nuclear staining of schizont stage parasites; red staining shows labeling of peptide specific antibodies by Cy3-conjugated anti-human or anti-mouse IgG specific antibody. Merge picture is an overlay of the blue and red fluorescence channel. doi:10.1371/journal.pone.0000645.g001 regions selected by this approach were tested manually for the presence of the characteristic heptad repeats. These proteins were also analyzed by the COILS program [25] . The selected a-helical coiled coil containing proteins were further tested on their possible surface location and GPI anchoring by using the following programs: identification of potential signal peptides, SecretomeP and SignalP (http://www. cbs.dtu.dk/services/) [26] ; transmembrane spanning regions (TMPRED http://www.ch.embnet.org/software/TMPRED_ form.html and TMHMM http://www.cbs.dtu.dk/services/ TMHMM; [27, 28] ), and GPI-anchored proteins (http://mendel. imp.univie.ac.at/sat/gpi/gpi_server.html; [29] ) and prediction of sub-cellular localization (pTARGET http://bioinformatics. albany.edu/,ptarget; [30] ). To identify the PEXEL-like motifs in sequences of the selected proteins we used the following pattern [ [DEQ] that represents a combination of the PEXEL patterns indicated in recent papers [21, 22] . The presence of the identified proteins in the asexual erythrocytic stages was also checked using the published data on the transcriptome and proteome of this stage of development of P. falciparum (www.PlasmoDB.org; [31] ). Peptides were synthesized on the Advanced ChemTech (Hatley St George, UK) AC T348 Omega multi channel synthesizer and the Applied Biosystem synthesizer 431A and 433A (Foster City, CA) using solid-phase Fmoc chemistry. Crude peptides were purified by RP-HPLC (C18 preparative column) and analyzed by mass spectrometry (MALDI-TOF; Applied Biosystem, Foster City, CA). Chemicals and solvents used for peptide synthesis were purchased from Fluka (Buchs, Switzerland) and Novabiochem (Laufelfinger, Switzerland). Circular dichroism (CD) spectra of peptides were recorded on a JASCO J-810 spectrometer (JASCO corporation, Tokyo, Japan) equipped with a temperature controller and a 0.1 cm path length cuvette. The measurements were made in water at pH 7.3 and 22uC and at a peptide concentration of 0.2 mg/ml. The sera from Burkina Faso were collected in the village of Goundry located in the central Mossi Plateau, between 15 and 50 km north of the capital Ouagadougou, in the province of Oubritenga. The climate is characteristic of areas of Sudanese savannah, with a dry season from November to May and a rainy season from June to October. Malaria transmission is very high during the rainy season and markedly seasonal. Ethical clearance was obtained from the Ministry of Health, Burkina Faso. After obtaining informed consent from parents and caretakers, heparinized venous blood samples were collected during a crosssectional survey during the malaria low transmission season 1998. The Tanzanian sera came from a large-scale community based study undertaken in Kikwalila village, Kilombero District, Morogoro Region from 1982 to 1984. Blood samples from adults (.15 years) were taken by finger prick and the serum was kept at -70uC until use. Research and ethical clearance for the study was obtained by the Tanzanian Commission for Science & Technology. The Colombian sera were collected in Buenaventura the main port on the Colombian Pacific Coast after human informed consent, during a cross sectional survey carried out from February to May 2002 within the framework of a project supported by the Colombian Research Council, COLCIENCIAS. The area has unstable transmission of both P. falciparum and P. vivax malaria. Ethical clearance to draw blood from human volunteers was Human purified antibodies were used at 5 mg/mL; IFAT was not performed on ring stages. Valle. Blood was taken by venipuncture into tubes containing EDTA and sera fractionated and stored frozen until use. The sera from adults from Papua New Guinea (PNG) pooled for affinity purification were collected in the Maprik district of the East Sepik Province, during a cross sectional survey in July 1992 within the framework of the Malaria vaccine Epidemiology and Evaluation Project (MVEEP) supported by the United States Agency for International Development [32] . The area is highly endemic for malaria. Ethical clearance for MVEEP was obtained from the PNG Medical Research Advisory Committee. Blood was taken by venipuncture into tubes containing EDTA. The pool of immune African globulins (PIAG) used for ADCI was prepared from immune individuals living in endemic areas and negative control IgG (N-IgG) was obtained from a pool of more than 1000 French adult donors with no history of malaria. Briefly, the IgG fractions from both positive and negative controls were purified using a size exclusion TrisacrylH GF05M (Pall BioSepraH; Pall life Sciences, NY) column followed by an ionic exchange DEAE Ceramic HyperDH F column (Pall BioSepraH). Purified IgG were then extensively dialyzed against RPMI and kept at 4uC until use. CB6F1 mice were injected 3 times with 20 mg of the indicated peptide in Montanide ISA 720 at the base of the tail on day 1, 22 and 78. Bleeding was performed 10 days after the second and third immunization. ELISA was performed according to Lopez et al. [33] and anti human IgG-or anti mouse IgG conjugated to alkaline phosphatase was used (Sigma, St Louis, MO) as second antibody. Individual human sera from 37, 42 and 39 adults donors from Burkina Faso, Tanzania and Colombia respectively were used at 1:200 dilution. Serum was considered positive if the optical density (OD) reading was higher than the mean OD value+3 standard deviation (SD) of the negative controls (individual serum samples from 8 to 11 naïve Swiss donors) or if the OD ratio of the mean of duplicate experimental values to the mean OD of the negative control was higher than 2. For mouse sera, the end point value was determined as the last dilution of the mean OD value+3 standard deviation (SD) of the negative control (non immune sera). ELISA competition assays were performed by incubating either each of the 18 selected human affinity-purified antibodies, or antibodies elicited in mice, together with each of the 18 antigens over the indicated range of concentrations for 30 minutes at room temperature prior to addition to the ELISA peptide-coated plate wells. Antigen-Sepharose conjugate preparation: 5 mg of antigen was dissolved in 1 mL of coupling buffer (0.1 M NaHCO 3 containing 0.5 M NaCl, pH 8.0). The CNBr-sepharose 4B (Amersham Bioscience AB, Uppsala, Sweden) was activated by swelling in 1 mM HCl and then washed with coupling buffer. The antigen solution was added to the gel and the mixture was stirred for 1h at RT. After the coupling reaction, excess antigen was washed away with coupling buffer. The unreacted activated groups were blocked by treatment with ethanolamine (0.25 M; pH 8.0) for 30 min at RT. The gel was then washed with sodium acetate buffer (0.1 M; pH 4.0), followed by coupling buffer. The antigensepharose beads were either used or stored at 4uC in PBS (1x) containing 1 mM azide. Isolation of specific antibody: Pooled human serum was diluted five times with PBS (1x) containing 0.5 M sodium chloride and mixed with antigen-sepharose conjugate. This mixture was then stirred gently on a wheel O/N at 4uC. After centrifugation the supernatant was collected and stored at 220uC for further use. The antigen-sepharose beads were then washed with 5 mL of trizma base TRIS (20 mM containing 0.5 M NaCl, pH 8.0) then with 5 mL of TRIS (20 mM, pH 8.0). The elution of bound antibody was achieved with glycine (0.1 M, pH 2.5). The fractions obtained were instantly neutralized with TRIS (1 M, pH 8.0), dialyzed against phosphate buffer (0.1M, pH 7.0) and the antibody concentration was determined by the absorbance of the solution at 280 nm. Slides coated with Pf sporozoites were dried at RT for 30 minutes, fixed with 100% acetone at 4uC for 10 minutes, washed 2 times in PBS-0.05% Tween 20, dried carefully and blocked with 20 mL/ well of PBS-3% bovine serum albumin (BSA) for 30 minutes at RT. Slides coated with Pf merozoites were fixed with 100% acetone at 220uC for 15 minutes and dried O/N at RT. The appropriate antibody or serum dilutions prepared in PBS-3% BSA were distributed (10 mL/well) and incubated for 1h at RT in a humid chamber. After washing with PBS-0.05% Tween-20, goat anti-human or goat anti-mouse polyvalent immunoglobulins conjugated to Cy3 (Molecular Probes) diluted 1/500 in PBS-3% BSA or anti-human IgG (Fc specific) FITC conjugate (Sigma) diluted 1/50 in Evans blue solution (1/50000) was added (400 mL/slide) and incubated for 1h at RT in a humid chamber in the dark. Slides were washed as above, covered with 50 % glycerol, sealed and read using a fluorescence microscope (Leica DMIRB DC200). The Uganda Palo Alto strain (FUP/C) was cultured in RPMI-1640 supplemented with 0.5% albumax I (GibcoBRL-Invitrogen, San Diego, CA). For ADCI assays, blood stage parasite cultures were synchronized by at least two successive sorbitol treatments followed, after maturation over 24 h, by floatation on 1% porcine skin gelatin type A (Sigma). Blood monocytes (MN) were prepared from cytapheresis samples obtained from healthy blood donors with no previous history of malaria (Lecourbe Blood Bank, Paris, France). Peripheral blood mononuclear cells (PBMC) were separated on Ficoll density gradients J PREP (TechGen, Les Ulis, France) and washed in Ca 2+ and Mg 2+ free HBSS buffered with 10 mM HEPES (both from GibcoBRL-Invitrogen). Cells were then distributed on polystyrene 96-well flat-bottomed culture plates (TPP, Trasadingen, Switzerland) and adherent MN were selected by incubation for 2 h at 37uC, in a humidified 5% CO 2 atmosphere. More than 90% of the adherent cells obtained in this manner were MN as estimated by the non-specific esterase test (a-naphtyl acetate esterase; Sigma). MN from each donor were tested prior to ADCI assays and only those without direct inhibitory effect were used in assays. To wells containing 2610 5 MN purified as described above, 50 ml of an asynchronous parasite culture at 0.5 % parasitemia and 4 % hematocrit were added. Wells were then supplemented with test or control antibodies (Ab) and the total volume adjusted to 100 ml with culture medium. After 48 h and 72 h, 50 ml of culture medium were added to each well and after 96 h the ADCI assay was stopped and the final parasitemia was determined by light microscopy on Giemsa-stained smears by counting $50,000 red blood cells. For each Ab tested, duplicate wells included the following controls 1) non-specific monocytic inhibition, both MN+parasite, and MN+N-IgG+parasites and 2) direct inhibition by control or test IgG, both N-IgG+parasites, and test Abs+parasites. PIAG and N-IgG were used at a final concentration of 1 mg/ml as positive and negative controls respectively. Immunopurified tests Abs were used at 15 mg/ml. The Specific Growth inhibitory Index (SGI) which considers the parasite growth inhibition due to the effect of test Abs cooperating with MN was calculated as follows: SGI = 1006[12(% parasitemia with MN and test Abs/% parasitemia test Abs)/(% parasitemia with MN and N-IgG/% parasitemia N-IgG)]. Figure S1 CD spectra of the peptides 45 (S1A) and 12 (S1B) Found at: doi:10.1371/journal.pone.0000645.s001 (12.32 MB TIF) Figure S2 ELISA inhibition assay using anti-human peptide specific antibodies. Binding of peptide specific antibodies to peptides 76 (S2A) and 9 (S2B) absorbed on ELISA plates was inhibited by incubating specific antibodies (1-2 mg/ml) with peptides 14, 76 and 81 (S2A) and peptides 8 and 9 (S2B), respectively (see Material and Methods). Peptides 14, 76 and 79 share NNM or MNN as sequence similarity while peptides 8 and 9 do not exhibit any apparent sequence similarity. Table S3 Structural feature and cellular location prediction of the proteins containing the peptides whose specific antibodies were tested in ADCI (Table 1) . Found at: doi:10.1371/journal.pone.0000645.s005 (0.05 MB DOC)
95
FluGenome: a web tool for genotyping influenza A virus
Influenza A viruses are hosted by numerous avian and mammalian species, which have shaped their evolution into distinct lineages worldwide. The viral genome consists of eight RNA segments that are frequently exchanged between different viruses via a process known as genetic reassortment. A complete genotype nomenclature is essential to describe gene segment reassortment. Specialized bioinformatic tools to analyze reassortment are not available, which hampers progress in understanding its role in host range, virulence and transmissibility of influenza viruses. To meet this need, we have developed a nomenclature to name influenza A genotypes and implemented a web server, FluGenome (http://www.flugenome.org/), for the assignment of lineages and genotypes. FluGenome provides functions for the user to interrogate the database in different modalities and get detailed reports on lineages and genotypes. These features make FluGenome unique in its ability to automatically detect genotype differences attributable to reassortment events in influenza A virus evolution.
Infections with influenza A viruses continue to be a public health problem, causing seasonal epidemics and sporadic but devastating pandemics. Each year in the US, influenza epidemics cause more than 200 000 hospitalizations and result in over 30 000 influenza-related deaths (1) . Influenza pandemics are infrequent but they can result in high mortality. It is estimated that $20-100 million people were killed worldwide by the 1918-1919 influenza pandemic (2) (3) (4) . The current level of pandemic alert is at the highest level, phase 3, since the most recent pandemic of 1968 (5) . Influenza viruses belong to the family Orthomyxoviridae and are classified into three types, A, B and C based on the identity of major internal protein antigens (6) . Influenza A and C viruses can infect multiple mammalian species, while influenza B virus is almost exclusively a human pathogen (7) . Influenza A viruses cause the greatest morbidity and mortality in humans. Interestingly, the largest pool of influenza A viruses is maintained by horizontal spread in wild aquatic birds, in which the virus does not normally cause any disease (6, 8) . Food and companion animal populations such as poultry, swine, horses and dogs support sustained replication of certain lineages of influenza A, with minimal to lethal disease depending on the virulence of the strain (6) . Influenza viruses have evolved in association with their various hosts in different continents for extended periods of time (9) . This co-evolution has resulted in extensive genetic divergence among the extant viruses currently available for analysis. Influenza A viruses are classified into subtypes on the basis of antigenic analysis of hemagglutinin (HA) and neuraminidase (NA) glycoproteins. So far, 16 HA subtypes and 9 NA subtypes have been found (10) . In recent years, gene sequences have become available for a large number of viral strains creating a diverse pool of influenza A viruses from historical and current isolates collected in multiple geographic regions. Comparison of the deduced amino acid sequences of the HA and NA revealed an excellent agreement between the results of clustering viruses by the antigenic reactivity and sequence similarity. However, molecular genetic analysis allows a comprehensive analysis of the entire viral genome and is gaining popularity because it is more practical for most laboratories as a method for classification (11) . Most importantly, study of the influenza genomic structure, namely genotyping, could reveal mechanisms of virus evolution, spread and disease pathogenesis. The influenza A genome consists of eight negativestranded RNA segments that encode at least 10 viral proteins (12) . The viral genome evolves through accumulation of mutation by the viral RNA-dependent RNA polymerase which lacks proofreading ability and through reassortment of entire gene segments (13) . Forces selecting viral variants such as the neutralizing antibody response of vertebrate hosts as well as species-related structural variation can also promote rapid evolution (14) . Each of the segments can evolve at a different rate if they are subject to differential selective pressures and functional constraints (15) (16) (17) (18) (19) . The segmented nature of the viral genome allows for segment exchange (termed reassortment) when two distinct viruses co-infect a cell and generate progeny with a mixed genome (20, 21) . Reassortment may theoretically yield 254 (2 8 -2) different combinations of gene segments from two parent viruses. A comprehensive influenza genotype database that can be searched using a web tool for the genotyping viruses is not available. Unlike HIV and HCV, the influenza A virus has a segmented genome, so eight separate phylogenies must be analyzed to establish a genotype. We approached the problem of genotyping influenza A viruses by analyzing each gene segment independently, segregating gene segments into subtypes and subsequently into lineages. The genotype of an influenza A viral strain is the sequential aggregate of the eight assigned gene segment lineages. A nomenclature for influenza A viral genotypes will allow researchers to unequivocally describe influenza A viral genotypes to analyze, compare and communicate the molecular epidemiology of the virus. In this report, we define a nomenclature for influenza A viral genotypes and describe a web tool developed for genotyping influenza A viruses from genome sequences. Our tool facilitates identification of reassortment events between divergent lineages. Two nomenclature conventions are used routinely in influenza research: (i) the eight segments in the influenza A genome are numbered from 1 to 8 for PB2, PB1, PA, HA, NP, NA, M and NS, respectively; (ii) There are currently 16 alleles of the HA gene termed subtypes. Likewise, there are nine alleles for NA, and two alleles for non-structural (NS) proteins. Since influenza A viruses have an unusual genomic structure, we approached the genotyping problem by first analyzing each gene segment separately. According to the above conventions and considering that the evolutionary rate varies from segment to segment, we defined a genotype as a sequential combination of the lineages for each of the eight segments in a genome. A letter was assigned to each lineage of PB2, PB1, PA, NP and M, and a number followed by a letter was assigned to each lineage of HA, NA and NS with the number representing the subtype or allele. For example, [A,D,B,3A,A,2A,B,1A] is the genotype of a human seasonal subtype H3N2 virus with PB2 lineage A, PB1 lineage D, PA lineage B, HA subtype 3, lineage A and so on, following the convention for numbering of influenza genome segments. With this nomenclature, identifying genotypes and reassortment becomes an easy task accomplished by comparing the predicted genotype against all genomes that have been classified previously. Genomic sequences of all influenza A viruses with 475% of the full segment length were downloaded from NCBI Influenza Virus Resource (http://www.ncbi.nlm.nih.gov/ genomes/FLU/FLU.html). Alignments were performed for each individual gene segment using the ClustalW program (22) . The MEGA software was used to construct the phylogenetic trees with the neighbor-joining method and the HKY-85 model selected (23) . The goal of our genotype method is to determine when a reassortment event with a gene segment from a non-traditional host or location has occurred. The lineages of each viral gene were carefully determined as detailed subsequently: (i) using the phylogenetic trees constructed, significant clusters (which were segregated by $10% nucleotide difference by p-distance) were assigned lineages; (ii) bootstrap analysis was used on a smaller set of sequences with values 490% considered significant; (iii) the initial lineages were evaluated for nucleotide differences within and between other lineages and for strength of bootstrap support; (iv) approximately 10 sequences from each lineage were randomly selected for the maximum likelihood (ML) analysis for each gene segment, serotype (for HA, NA) or allele (for NS) on the MultiPhyl server (24) . The lineage assignment of each influenza gene available in the public databases was uploaded into the Segment Table in the database as described subsequently. The FluGenome database contains three tables: Segment, Genome and Genotype. The Segment table contains information-related to sequences, including assigned lineage, strain name, segment, serotype, host, country, year, GenBank accession number, nucleotide sequence and sequence length. The Genome table contains the information for complete genomes, including assigned genotype and accession numbers of each gene segment. When more than one sequence was available for a gene segment, the longer of the two sequences was kept for the genome accession. Unique genotypes are stored in the Genotype table along with the total number of genomes that have that genotype. The Genotype table was created by querying the Genome table for distinct genotypes. Host categories were created to separate the genomes of each genotype, which include Human (Hu), Avian (Av), Swine (Sw), Equine (Eq), Canine (Ca) and Others (ONHM). The FluGenome database is updated automatically every night. New sequences are downloaded from the NCBI Influenza Virus Resource (ftp://ftp.ncbi.nih.gov/ genomes/INFLUENZA/) and added into the FluGenome database. The lineage information predicted for new sequences is used to update Segment, Genome and Genotype tables if necessary. For sequences already in the database, the script checks to see what information needs to be updated, and the sequences entries are flagged for further validation. The web interface and databases were implemented with the LAMP strategy. The server used Linux (L) for the operating system, along with Apache (A) as the web server. The genotyping database was built with the MySQL database management system (M). PHP and PERL (P) were used to code the two parts of the web tool: the back end program and the front end interface. JavaScript and HTML were used sparingly in the front end interface. A domain name, http://www.flugenome.org, was acquired to provide access to the database and the web tool. The BLAST algorithm is used for sequence comparison, because of its advantages such as fast computation and accurate results in detecting local highly similar sequence regions. To overcome its inherent disadvantage (i.e. not a global alignment algorithm), we used a parameter called 'coverage' to detect gene-wide sequence similarity (25) . The default thresholds for identifying lineages were set to be 95% identity and 95% coverage. The user can reset the thresholds to any allowable value. The top BLAST results for a user-submitted query sequence are sorted by identity and coverage, and the best result is used to assign a lineage to the query sequence. If a result from BLAST falls below the thresholds, the lineage will be flagged with an asterisk ( Ã ). To determine the genotype of a complete or partial influenza virus genome, a script is executed that first establishes the lineage of each viral gene segment. The genotype will be created by the sequential incorporation of the lineages for each of the eight segments, arranged per convention as shown in Table 1 . If a lineage does not meet the thresholds specified (95% default for both identity and coverage), the lineage will be assigned an asterisk ( Ã ) indicating the query sequence does not meet criteria and may be from a new lineage. If no BLAST results are found a blank lineage will be displayed. If all segments belong to known genotypes, the genotype of the query genomic sequence will be provided as output. The resulting genotype can be compared to previously identified genotypes in the Genotype database. This analysis can reveal reassortment events and host switching. If the genotype determined by FluGenome is not found in the Genotype database, the genome will be flagged as a virus with a potentially new genotype. Viral genotypes reported as new by FluGenome can simply result from identification of a gene from a novel phylogenetically defined lineage or the presence of genes from known lineages in novel combinations. The online tool presents two query options to the user; entering gene segment sequence(s) or genotype sequence(s) (Figure 1) . The segment query 'Determine Individual Gene Segment Lineage' is used to identify the lineage of a viral gene segment of interest, for example PB2. In this case, the input FASTA file can contain one or many sequences, but all must correspond to the same gene segment. To analyze data sets from more than one gene simultaneously; e.g. both the PB1 and PB2, the user must first enter the number of different gene segments and then provide each sequence data set in a separate FASTA file. The genotype query 'Determine Genotype' analyzes incomplete or complete genomes. Sequences from each genome must be in a separate FASTA file. Alternatively, the user can cut and paste sequences of one genome at a time. Multiple genomes can be analyzed simultaneously. Nearly 30 000 sequences were collected from public databases and used for the lineage analyses, resulting in 184 lineages. The viral gene segments showed a wide range of diversity; HA was partitioned into 78 lineages whereas MP only into seven (Table 1) . Mining the aforementioned sequences resulted in $2300 complete genomes, which consists of 156 unique genotypes with 50 serotypes (http://www.flugenome.org/show_genotypes.php). Serotypes may comprise as many as 15, different genotypes; Step 1. Enter number of different gene segments to analyze Step 2. Select which gene segment(s) to analyze Step 3. Enter sequence(s) in FASTA format Step 4. Results page with lineage(s) Step 5. Show viruses with the same gene lineage(s) Step 1. Enter number of influenza A genomes to compare Step 2. Input genome sequence(s) in FASTA format Step 3. Results page with genotype(s) We propose a nomenclature system for naming influenza A viral genotypes. This nomenclature was exploited to analyze $2000 complete viral genomes (nearly full-length or full-length segment sequences), revealing 156 unique genotypes. The FluGenome web server implementation also includes facilities for analysis and sorting of lineages and genotypes which allow the user to explore the evolutionary history of the viral strains. In particular, the FluGenome web server can provide genotype information that greatly facilitates the inference of genetic reassortment among influenza viruses.
96
Influenza pandemic intervention planning using InfluSim: pharmaceutical and non- pharmaceutical interventions
BACKGROUND: Influenza pandemic preparedness plans are currently developed and refined on national and international levels. Much attention has been given to the administration of antiviral drugs, but contact reduction can also be an effective part of mitigation strategies and has the advantage to be not limited per se. The effectiveness of these interventions depends on various factors which must be explored by sensitivity analyses, based on mathematical models. METHODS: We use the freely available planning tool InfluSim to investigate how pharmaceutical and non-pharmaceutical interventions can mitigate an influenza pandemic. In particular, we examine how intervention schedules, restricted stockpiles and contact reduction (social distancing measures and isolation of cases) determine the course of a pandemic wave and the success of interventions. RESULTS: A timely application of antiviral drugs combined with a quick implementation of contact reduction measures is required to substantially protract the peak of the epidemic and reduce its height. Delays in the initiation of antiviral treatment (e.g. because of parsimonious use of a limited stockpile) result in much more pessimistic outcomes and can even lead to the paradoxical effect that the stockpile is depleted earlier compared to early distribution of antiviral drugs. CONCLUSION: Pharmaceutical and non-pharmaceutical measures should not be used exclusively. The protraction of the pandemic wave is essential to win time while waiting for vaccine development and production. However, it is the height of the peak of an epidemic which can easily overtax general practitioners, hospitals or even whole public health systems, causing bottlenecks in basic and emergency medical care.
The recent spread of highly pathogenic avian influenza from Asia to Europe and the transmission to humans has intensified concerns over the emergence of a novel strain of influenza with pandemic potential. While still being in an inter-pandemic stage, nations plan for pandemic contingency following recommendations of the WHO [1, 2] . National influenza preparedness plans are constantly being refined, aiming to mitigate the effects of pandemic influenza on a national, regional and local level. Even in the absence of a pandemic strain, seasonal influenza causes substantial morbidity and mortality [3] . Seasonal outbreaks put pressure on general practitioners and strain hospital resources, leading to bottlenecks in outpatient treatment and hospital admission capacities. Various intervention strategies reduce the impact of influenza on individuals and public health systems. In interpandemic phases, vaccination is the most important tool to reduce morbidity and mortality, but a potent vaccine will probably not be generally available in the initial phase of a pandemic [4] . Other control strategies like pharmaceutical (antiviral) [5, 6] and non-pharmaceutical interventions (reduction of contact rates) [7, 8] will have to be implemented. The use of antiviral drugs during a pandemic seems to be the treatment of choice at present [9] [10] [11] [12] , but not all countries can afford stockpiling enough drugs. Furthermore, concerns about the over-reliance of a "pharmaceutical solution" have been expressed [13] . An epidemic can also be mitigated by reducing contact rates in the general population and by decreasing the infectivity of cases [9] . Such reductions can be achieved by measures like quarantine and case isolation [14] , closing day care centres and schools, cancelling mass gathering events, voluntary self isolation and general behavioural changes in public and increasing social distance [8] . The effectiveness of such interventions depends on various factors which must be prospectively explored by sensitivity analyses, based on mathematical models. Here, we use the freely available Java applet InfluSim [15] to investigate how effectively pharmaceutical and non-pharmaceutical interventions contribute to mitigate an influenza pandemic while vaccines are not available. In particular, we examine how intervention delays determine the course of a pandemic and constrict the success of interventions. InfluSim is a deterministic compartment model based on a system of over thousand differential equations which extend the classic SEIR model by clinical and demographic parameters relevant for pandemic preparedness planning. Details of the simulation and a discussion of the standard parameter values have been described previously [15] ; a summarizing description of the model is provided in the Appendix. The program and its source code are publicly available [16] to offer transparency and reproducibility. The simulation produces time courses and cumulative numbers of influenza cases, outpatient visits, applied antiviral treatment doses (neuraminidase inhibitors), hospitalizations, deaths and work days lost due to sickness, all of which may be associated with financial loss. The analyses presented here are based on InfluSim 2.0, using demographic and public health parameters which represent the situation in Germany in 2006. Interventions include antiviral treatment, isolation of patients, social distancing measures and the closing of day care centres and schools as well as cancelling mass gathering events. Using the standard set of InfluSim parameters (freely accessible from [15] ), about one third of all infected individuals is expected to become severely ill and to seek medical help. Patients seeking medical help will be referred to as "outpatients" throughout this paper. An exponential distribution is used to model the delay between onset of symptoms and seeking medical help; on average, patients visit a doctor after 24 hours. If a patient seeks medical help within 48 hours after onset of symptoms, he or she is given antiviral treatment unless the stockpile of antivirals is exhausted. Antiviral treatment reduces the duration and degree of infectivity of the case and the number of hospitalizations (Table 1 ) [17] . For more detailed descriptions see [15] or the Appendix. Non-pharmaceutical interventions examined in this paper are contact reduction measures and the isolation of cases. The latter effectively leads to reduced contact rates between individuals, too. In the scenarios presented below, we assume that everybody in the population avoids a given percentage of contacts (e.g. by improved hygiene, wearing masks, or behavioural changes) and that sick patients are isolated which reduces the contact rates of moderately sick, severely sick (but non-hospitalized) and hospitalized cases by 10%, 20% and 30%, respectively. Further interventions which comprise the closing of day care centres and schools, and the cancelling of mass gathering events will be examined in detail in a separate paper. Assuming a basic reproduction number of R 0 = 2.5 and using the standard parameter set of InfluSim [15] , an epidemic in a population of 100,000 individuals reaches the peak about 40 days after introduction of the infection and is practically over three weeks thereafter if no interventions are performed (Figure 1 ). During the whole epidemic, 87% of the population become infected, 29% seek medical help, 0.7% are hospitalized and 0.2% die. Figure 1 shows how pharmaceutical and non-pharmaceutical interventions can mitigate this scenario. Contact reduction by isolation of cases alone (see Appendix), protracts the peak of the epidemic by about one week. Distribution of antivirals or additional contact reduction measures delay the epidemic by approximately 10 days and are hardly sufficient to provide a substantial delay. A combination of antiviral treatment, isolation of cases and social distancing in the general population seems to be necessary to delay the epidemic in the order of weeks. This example furthermore shows that an efficient mitigation of the epidemic is not necessarily associated with a significant reduction in the number of infections. For information on the proportions of infected people and outpatients see the legends to the Figures. The mitigating effect of antivirals strongly depends on the onset of their distribution ( Figure 2 ). Antivirals can delay the epidemic if distributed very early while few cases exist in the population. Late distribution of antivirals (e.g. starting on day 30) leads to the paradoxical effect that the stockpile is exhausted even quicker compared to early distribution (shaded areas und the curves in Figure 2 ). Additionally, the mitigating effect of the intervention drastically diminishes and benefits are restricted to lowering the peak of the epidemic. Unrestricted availability of drugs (grey curves in Figure 2 ) still leads to an epidemic because (i) asymptomatic and moderately sick cases are not eligible for treatment, (ii) patients visit a doctor on average 24 hours after onset of symptoms while already being highly infectious and (iii) antivirals cannot fully prevent infectivity. Figure 3A ). In contrast, the mitigating effect becomes negligible, if antivirals are distributed with delay ( Figure 3B ). Independent of the delay in the distribution of antivirals, their quantitative availability affects only the height of the peak of the epidemic, but hardly the mitigation of the epidemic ( Figure 3A , B). For considerations into the final size of the epidemic see below. In summary, delaying the epidemic depends on early action, whereby lowering the peak depends on the quantitative availability of antivirals. Contact reduction measures, comprising social distancing and the isolation of cases, can be an effective part of mitigation strategies; they have the advantage over antiviral treatment to be not limited per se, i.e. they can be continued for a sufficiently long period of time. Figure 4 examines the effect of isolation of cases and social distancing measures (see figure caption for details) in the absence of antiviral treatment. The peak of the epidemic is protracted by about 1 day for every percent of contact reduction if this intervention starts immediately after the introduction of the infection. Thus, a peak shift is not only possible by early action, but also by the degree of contact reduction. If contact reduction is initiated later, the peak shift diminishes, but the proportionality remains. For example, if the intervention starts three weeks after the introduction of infection, the peak of the epidemic is only mitigated by about half a day per 1% contact reduction ( Figure 4B ). Premature cessation of contact reduction measures restores the infection rates to the pre-intervention values which fuels the epidemic. It can lead to a delayed course and a higher total number of infections, involving a plateau or even a second peak of the epidemic ( Figure 4C ). The preceding examples with interventions based on antivirals or contact reduction alone yielded peak delays only in the order of weeks, whereas months may be required for vaccine development and production, demanding for a combined intervention scheme ( Figure 5 ). We examine an optimistic scenario where antivirals are distributed immediately after the infection is introduced (dark bars in Figure 5 ), while varying the onset of social distancing measures. The antiviral stockpile lasts longer if social distancing measures are initiated earlier (pale bars in Figure 5 ). Immediate initiation of contact reduction can protract the epidemic by months, whereas a delayed initiation leads to a plateau in the epidemic curve at a time when antivirals are used up. Without interventions, N i = 87% of the population become infected during the course of the epidemic and the cumulative number of outpatients reaches N o = 29%, reflecting the assumption that approximately one third of infected individuals becomes sufficiently sick to seek medical help. These outcomes remain surprisingly stable even for interventions assuming optimistic resources (cf. footnotes to Figures 1, 2 , 3, 4, 5). For instance, immediate and unlimited availability of antivirals reduces these fractions only to N i = 72% and N o = 24% ( Figure 2 ). This minor effect has three reasons: only about one third of cases seeks medical help and will receive antiviral treatment, many infections are passed on before cases seek medical help and antiviral treatment does not fully prevent further transmission. These disadvantages do not apply to contact reduction measures. For instance, a reduction of 20% of contacts reduces these fractions to N i = 68% and N o = 22% ( Figures 4A, B) . A combination of antiviral treatment and contact reduction can further reduce these values to N i = 53% and N o = 18% ( Figure 5 ). In the preceding analyses it was assumed that parameter values are precisely known; in a real world scenario, however, uncertainty arises from biological variability, stochastic influences, heterogeneities, etc. We illustrate with a concluding example to which extent simulated epidemics are affected by uncertainty in the parameter values. As shown in Figure 6 , epidemics can be highly variable, although only four parameters have been varied within moderate ranges. Varying more parameters would further increase this variability. For the interventions and parameter variations considered, the cumulative number of outpatients ranges from a few thousand to over twenty thousand (see inset in Figure 6 ). Among the four parameters, R 0 is the strongest predictor of the number of outpatients (analysis not shown) as it strongly determines how quickly antivirals become exhausted. In two out of 1,000 simulations the randomly chosen parameter combinations involved values for R 0 around 1.8 which led to very minor outbreaks given the intervention scheme. The cumulative number of outpatients escalates when antiviral stockpiles become exhausted while the proportion of susceptibles is still large enough to allow for further propagation of infectives. In this case, the epidemic curve proceeds with a second wave or a plateau. With pandemic influenza, we have to "expect the unexpected" [18] . Historical reports frequently mention the surprising speed at which a pandemic wave travels through the population [19] [20] [21] . Predicting the course of a future pandemic which will be caused by a virus with unknown characteristics is based on substantial uncertainties and we must rely on sensitivity analyses, performed with mathematical models like InfluSim. Because of the short serial interval of influenza, timely action is essential. Different control measures must be regarded as complementary and not as competing. Neither antiviral treatment nor non-pharmaceutical measures should be used exclusively to mitigate a pandemic influenza wave. Infectious disease models have suggested that an upcoming influenza epidemic with a low basic reproduction number might be contained at the source through targeted use of antiviral drugs [9, 12] . The published scenarios concern WHO phases 4 and 5 (inter-pandemic alert period) and assume that an outbreak starts in a rural area with low population density. It can be expected that the pandemic virus will be introduced into Europe and the US after a local epidemic (i.e. in WHO phase 6). Communitybased prophylaxis, however, is of limited use for several reasons. Under a high prevalence of infection in phase 6, a wide distribution requires an enormous number of antiviral courses; with available stockpiles, it will be virtually impossible to locally contain the pandemic with targeted antiviral prophylaxis. Development of resistance, limited production capacities and extremely high costs are further limitations of this strategy, so that population-wide prophylaxis has not been recommended by the WHO for the final phase of the pandemic [1]. The discussion of pandemic influenza preparedness planning has frequently focussed on the amounts of drugs to be stockpiled and to whom and when they should be supplied [22] . Even if the currently stockpiled antiviral drugs will be fully effective against the pandemic strain, their use may not be able to sufficiently prevent the spread of influenza because (i) transmission of the infection may occur before the onset of clinical symptoms (as assumed in the InfluSim model) [23] , (ii) asymptomatic and moderately sick cases [6] are usually not treated despite contributing to transmission, and (iii) the occurrence of cases with influenza-like illness caused by other pathogens may lead to an accelerated depletion of the antiviral stockpile. Likewise, moderately sick cases or even healthy people may seek medical help and succeed in receiving antiviral treatment which would further deplete the stockpile. These factors reduce the efficacy of pharmaceutical control measures [24] , indicating the demand of extending this strategy by non-pharmaceutical intervention measures. Especially if antivirals are limited, they should be supplied as early as possible. If their distribution is delayed, cases become so abundant that resources will quickly be exhausted without having much impact on the spread of the disease (Figures 2 and 3 ). This confirms that the amount of antivirals needed strongly depends on the number of infections that are present when the intervention is initiated [25] . If antiviral drugs are extremely limited, they should be used to preferably treat severe cases Onset and sustainability of antiviral intervention Intervention with limited amounts of antivirals Figure 3 Intervention with limited amounts of antivirals. Number of outpatients expected during a pandemic wave, varied by the availability of antivirals. Parameter values are based on the InfluSim standard configuration [15] with R 0 = 2.5, except those listed at the end of this legend and indicated by superscripts 1 . Antiviral availability ranges from 0% (no antivirals available, dashed curves 2 ) to 10% (antivirals available for 10% of the population 3 Figure 4 Effects of contact reduction measures. Number of outpatients expected during a pandemic wave if contact reduction measures are implemented additionally to the isolation of cases. Parameter values are based on the InfluSim standard configuration [15] with R 0 = 2.5, except those listed at the end of this legend and indicated by superscripts 1 . The dashed curve shows the epidemic without intervention. Contact reduction involves social distancing 2 and isolation of cases 3 . The curves show the effects caused by social distancing, where contacts are reduced by 0% 4 (grey curve) up to 30% 5 in steps of 2% 6 (black curves, from left to right). Bars at the bottom of each graph illustrate the periods of contact reduction, which are in A: full, from day 0 to end, in B: delayed, from day 20 to the end, and in C: temporarily, from day 20 to day 50. 1 :Parameter modifications are given in the following and terms in italics refer to terms in the InfluSim user interface. InfluSim output: N i = cumulative proportion of the population infected, and N o = cumulative proportion of outpatients in the population. 2 that need hospitalization. Although this has practically no effect on the pandemic wave per se, it helps to reduce the death toll in the population (results not shown). Rather than relying on a pharmaceutical solution, pandemic preparedness should also involve non-pharmaceu-tical measures (see above). Early self-isolation and social distancing measures can be highly effective, as shown for the SARS epidemic [26] : after the WHO's global alert and the implementation of massive infection control measures, the effective reproduction numbers in Hong Kong, Vietnam, Singapore and Canada fell below unity. Rigorous social distancing measures in the entire population, Parameter values are based on the InfluSim standard configuration [15] with R 0 = 2.5, except those listed at the end of this legend and indicated by superscripts 1 . The sensitivity analysis extends the scenario shown in Figure 5 , where antivirals are available for 10% of the population and are distributed from day zero 2 , and where contact reduction measures 3 , including the isolation of cases 4 , are initiated three weeks after the introduction of infection (scenario "day 21"). Right panel: parameter values for each realization are sampled independently from normal distributions as shown (means given in bold, 99% of the values lie within the range specified by dotted lines, except b A which is truncated). R 0 : basic reproduction number, x 50 : cumulative infectivity during the first half of the symptomatic period, b A : relative infectivity of asymptomatic cases, f c : antiviral treatment reduces infectivity by a factor of 1-f c . For each parameter, an increase of the value aggravates the epidemic. Large plot: from a hundred random realizations, we selected the two most extreme epidemics, and eight epidemics homogeneously placed between them. The epidemic with N 0 = 20800 is caused by parameter values drawn from the left tail of the corresponding distributions, and the epidemic with N 0 = 5000 is caused by parameter values drawn from the right tail of the corresponding distributions (see right panel). The epidemic curves show a plateau or a second wave when antiviral stockpiles are exhausted while the proportion of susceptibles is still large enough to allow for further propagation of infectives (thin curves in black); for optimistic parameter combinations (e.g. small R 0 ), the available stockpiles last over the whole period of the intervention and the epidemic curve proceeds without a plateau (bold curves in grey). Inset: distribution of cumulative number of outpatients obtained from 1,000 random realizations. 1 :Parameter modifications are given in the following and terms in italics refer to terms in the InfluSim user interface. 2 however, will tax the social and economic structure and the population may not be willing or able to reduce contacts during the whole course of a pandemic wave. For Figure 5 , we assumed that contact reduction measures (e.g. improved hygiene, wearing masks, or behavioural changes) could add up to reduce contacts by 20%. Studies on the SARS outbreak suggest some preventative effect of wearing masks [27] [28] [29] , but compliance, availability of masks and their effectiveness against influenza infection remain unknown factors. Stockpiling surgical masks for the population results in exorbitant high numbers and may not be feasible [30] and individual stockpiling may be impossible due to economic limitations, especially in crisis situations. Since the specific effects of such behavio-ral changes remain uncertain, we modeled their contribution as a general reduction in contact rates. In contrast to SARS, we will not be able to rely on isolating hospitalized cases when a new influenza pandemic emerges. Using the standard parameter settings of InfluSim, we expect only a total of 0.7% of the population to be hospitalized. Even for the worst case scenario of the US Pandemic Preparedness Plan, where this value may be up to ten times larger [31] , the wide majority of infected individuals is never hospitalized. With influenza, we have to rely on self-isolation of moderately sick cases and of bed-ridden patients who stay at home. As these cases form the majority of infections and exert the highest force of infection, even a moderate reduction of contacts between them and the general population can substantially change the pandemic wave. Time is of the essence when controlling infectious diseases that spread at high speed and thus, interventions are most effective in the beginning when only few people are infected. Only a timely application of antiviral drugs (even with limited supplies) and a quick implementation of contact reduction measures will notably protract the peak of the epidemic and substantially reduce its height in a pandemic influenza wave. Whereby the protraction of the pandemic wave is essential to win time while waiting for vaccine development and production, it is the height of the peak of a pandemic wave which can easily overtax general practitioners as well as hospitals and whole public health systems, and can lead to dangerous bottlenecks in basic and emergency medical care. Vaccinating a small fraction of the population with a pre-pandemic vaccine would have a similar effect on the course of the epidemic as reducing the basic reproduction number by the percentage of immunized individuals (e.g. by 10%). The sensitivity analyses at the end of the Results section shows that the planning of intervention strategies must not only be based on single parameter values, but must also address their variability. More detailed analyses into this will be presented in a subsequent publication. Mathematical models like InfluSim should not only be used to predict a specific outcome, but also to explore best and worst case scenarios. The author(s) declare that they have no competing interests. InfluSim is a deterministic compartment model based on a system of over 1,000 differential equations which extend the classic SEIR model by clinical and demographic parameters relevant for pandemic preparedness planning. It allows for producing time courses and cumulative numbers of influenza cases, outpatient visits, applied antiviral treatment doses, hospitalizations, deaths and work days lost due to sickness, all of which may be associated with economic aspects. The software is programmed in Java and open access [16] , it operates platform independent and can be executed on regular desktop computers. The model structure of InfluSim is represented by Figure 7 , with descriptions given below. Susceptible individuals (S) are infected at a rate which depends on their age and on the interventions applied at the current time. Infected individuals (E) incubate the infection for a mean duration of 1.9 days. To obtain a realistic distribution of this duration, the incubation period is modelled in 7 stages yielding a gamma distributed incubation period with a coefficient of variation of 37.8%. The last 2 incubation stages are regarded as early infectious period during which patients may already spread the infection. This accounts for an average time of about half a day for the standard set of parameters. After passing through the last incubation stage, infected individuals become fully infective and a fraction of them develops clinical symptoms ( Figure 8A ). The course of disease depends on their age and risk group: one third remains asymptomatic (A), one third shows a moderate course of disease (M, "moderately sick") and the remaining third a severe course of disease (V, "very sick"); a small fraction of the latter third shows an extremely severe course of disease (X, "extremely sick") and needs hospitalization. The rationale for distinguishing extremely sick cases is that only these can die from the disease and need to be hospitalized; in all other aspects, both groups of severe cases are identical. The period of infectivity is gamma distributed and depends on the course of the disease and on the age of the case. To allow for an infectivity which changes over the course of disease, we apply weighting factors which depend on the stage of infectivity. Our standard value results in an infectivity which is highest immediately after onset of symptoms and which declines in a geometric progression over time ( Figure 8B ). Severe cases seek medical help on average one day after onset of symptoms, whereby the waiting time until visiting a doctor is exponentially distributed. Very sick and extremely sick patients who visit a doctor may be offered antiviral treatment. Very sick patients are advised to withdraw to their home (W) until the disease is over whereas extremely sick cases are immediately hospitalized (H). Death rates of extremely sick and hospitalized cases are age-dependent. Whereas asymptomatic and moderately sick patients who have passed their duration of infectivity are considered healthy immunes, very sick and extremely sick patients first become convalescent before they resume their ordinary life (gamma distributed with a mean of 5 days and coefficient of variation of 33.3%). Fully recovered patients who have passed their period of convalescence join the group of healthy immunes; working adults will return to work, and children again visit day care centres or schools. Antiviral treatment: Severe and extremely severe cases who visit the doctor within at most two days after onset of symptoms are offered antiviral treatment, given that its supply has not yet been exhausted. Antiviral treatment reduces the patients' infectivity by 80 percent, the duration of being diseased by 25%, and the risk of hospitalization by 50 percent. Extremely sick patients, whose hospitalization is prevented by treatment, are sent home and join the group of treated very sick patients. Contact rates in the general population can be reduced by increasing "social distance", by closing schools and day care centres, by cancelling mass gathering events, or by behavioural changes. Isolation of cases reduces their contact rates. Contacts are not necessarily reduced by 100%, but between 0 and 100%, as specified by the user. Our standard scenario considers reductions of 10%, 20% and 30% for moderately sick cases, very sick cases (at home) and extremely sick cases (hospitalized), respectively. For the mixing of the age classes, we employ a "whoacquires-infection-from-whom matrix" (WAIFW matrix) which gives the relative frequency of contacts of infective individuals by age. InfluSim assumes bi-directional contacts (e.g. children have the same total number of contacts with adults as adults with children). In order to match the user-specified basic reproduction number R 0 , the diseasespecific infectivity and the durations of infectivity in this matrix must be incorporated, resulting in the next generation matrix. This matrix is multiplied with a scaling factor chosen such its largest eigenvalue is equal to the chosen value of R 0 . The force of infection is given as the product of the number of infective individuals and the corresponding age-dependent contact rates. At the start of the simulation, one infection is introduced into the fully susceptible population. To avoid bias between simulations, the initial infection is distributed over all age and risk classes.
97
DNA Vaccines against Protozoan Parasites: Advances and Challenges
Over the past 15 years, DNA vaccines have gone from a scientific curiosity to one of the most dynamic research field and may offer new alternatives for the control of parasitic diseases such as leishmaniasis and Chagas disease. We review here some of the advances and challenges for the development of DNA vaccines against these diseases. Many studies have validated the concept of using DNA vaccines for both protection and therapy against these protozoan parasites in a variety of mouse models. The challenge now is to translate what has been achieved in these models into veterinary or human vaccines of comparable efficacy. Also, genome-mining and new antigen discovery strategies may provide new tools for a more rational search of novel vaccine candidates.
In spite of the success of vaccines in public health, there are still numerous pathogens, and in particular protozoan parasites such as Plasmodium falciparum, Trypanosoma sp., or Leishmania sp. against which there are still no effective vaccine. However, the discovery that the direct injection of plasmid DNA encoding foreign proteins could lead to endogenous protein biosynthesis and a specific immune response against it opened new perspectives in vaccine development. Over 15 years later, DNA vaccines have gone from a scientific curiosity to one of the most dynamic fields of research and may offer new alternatives for the control of infectious diseases [1] . Indeed, the first two DNA vaccines have been licenced, in recent years, to protect horses from west nile virus and salmons from infectious hematopoietic necrosis virus, confirming the usefulness of this biotechnology. We review here some of the advances and challenges for the development of DNA vaccines against two well-studied protozoan parasites, Leishmania sp. and Trypanosoma cruzi. Both belong to the trypanosomatidae family and are ranked among the three major protozoan parasites affecting humans. Leishmaniasis is a complex disease caused by at least 18 species of parasites from the Leishmania genus and transmitted to humans by hematophagous sandflies. With an estimated 12 million cases, it has a major public health impact in several regions, and in particular in India, Sudan, and Brazil [2] . Clinical manifestations range from self-healing cutaneous lesion to fatal visceral form, and this variety can be attributed in part to the respective parasite species, and each presents specific relationships with the host and diverse mechanisms of pathogenesis [3, 4] , which represents an additional difficulty for the development of treatments and vaccines. On the other hand, T. cruzi is the agent of Chagas disease, which is present from southern Argentina to the southern USA. An estimated 16-18 million persons are infected in the Americas and close to 100 million people are at risk of infection. After a short benign acute phase (a few weeks) and a very long (several years) asymtomatic phase, about 30-40% of infected patients develop chronic chagasic cardiomyopathy and eventualy die of heart failure. Current chemotherapy relies on nitrofurans (Nifurtimox), or nitroimidazoles (Benznidazole). However, the usefulness of these drugs is limited by their reduced efficacy (mostly during the early stages of the infection), serious side effects, and the emergence of drugresistant strains of parasites, and new treatments are slow to develop [5] . DNA vaccines induce a complete immune response against the encoded antigen. The exact mechanisms involved in this proccess are still poorly understood, and particularly the type of CD4 + and CD8 + effector and memory cells activated, and some of these aspects have been reviewed in detail elsewhere [1] . Apart from their immunogenicity and efficacy that will be discussed below, there are several features of DNA vaccines that make them very advantageous against tropical diseases. First, they are extremely safe as they do not contain any pathogenic organism that may revert in virulence. The major concern of genomic integration of the plasmid DNA has also extensively been studied in safety studies and found to be rather unlikely [6] . Additional safety issues such as anti-DNA antibodies or autoimmunity have also been addressed in a growing number of preclinical and clinical studies [7] , which confirmed the high safety of these vaccines. With respect to manufacturing, storage, and distribution, they also present major benefits in that the production process is the same for any DNA vaccine, which is not the case for other types of biologicals and vaccines, for which a specific protocol has to be developed for each. This makes production easy and costs will likely go down as this type of vaccines become mainstream and future technological improvements are implemented. Also, plasmid DNA is a very stable molecule, specially compared to recombinant or live attenuated vaccines, which would greatly facilitate storage and distribution of DNA vaccine in tropical settings with limited health infrastructure as the huge costs associated with the cold chain may be offset. Administration is also easy as simple IM or ID injections can be sufficient, and multiple plasmids can be combined for the elaboration of multivalent vaccines [1] . Overall, DNA vaccines may thus represent an ideally affordable alternative for disease control, which explains in part the growing interest in their development for the control of tropical parasitic diseases such as malaria, leishmaniasis, or Chagas disease. As mentioned above, leishmaniasis is caused by at least 18 species of parasites with diverse relationships with the host and mechanisms of pathogenesis [3, 4] . Early studies of cross-protection between Leishmania species clearly showed that it is a complex problem, with infection by one species protecting or not from subsequent infection by another species, depending on the species and the order of infections. Most vaccine studies have thus been focusing on homologous protection, although a single vaccine able to protect against all pathogenic species would be ideal. The correlates for protection have been extensively studied in the case of L. major, and contributed considerably to the development of the Th1/Th2 paradigm [8] . Thus, there is a general agreement that a Th1-type immune response, characterized by a high IFNγ and low IL-4 and IL-10 production, leads to control of L. major infection, while a Th2type immune response does not [8] . Antibodies may have an exacerbatory role [9] , but may also contribute to T cell responses [10, 11] . Both IFNγ producing CD4 + and CD8 + T cells seem to contribute to protective immunity, and induction of NO production by macrophages is central to parasite elimiation [12, 13] . While it was assumed for a long time that this Th1/Th2 paradigm was applied to all Leishmania species, it has become clear in recent years that each species has a distinct relationship with the host, different mechanisms of pathogenesis, and possibly different correlates for protection [3, 4] . Nonetheless, IFNγ production seems to be a general requirement, although not necessarily sufficient, for protection against most if not all Leishmania species. The earliest DNA vaccine experiments against Leishmania used L. major GP63 antigen, which has been extensively used as a recombinant or peptide vaccine. Immunization with a plasmid encoding GP63 was able to induce a Th1-type cytokine profile and a significant reduction of lesion size after challenge of the immunized mice with L. major [14] [15] [16] [17] . Subsequent studies investigated DNA vaccines encoding a large variety of Leishmania proteins ( Table 1 ) and showed that many different DNA vaccines were able to induce a Th1 immune response, and confer variable degrees of protection as assessed by reduction in skin lesion size and/or parasite burdens in mouse models. However, given the large variety in experimental models and designs, it is difficult to compare the effectiveness of the different vaccines to induce a protective immune response. Nonetheless, it is clear from studies comparing different DNA vaccines that the nature of the antigen encoded by the vaccine is a key parameter for efficacy. Also, a few studies provided interesting comparisons of the same antigens administered as recombinant protein or DNA vaccines and showed that the latter were overall more effective than their recombinant protein counterparts. Indeed, DNA vaccines were able to induce a stronger Th1 bias in the immune response, a longer-lasting immunity, and/or a better protection against disease progression [19, 31, 32, 35, 40, 44] . While most of these studies have used a rather artificial infectious challenge based on the injection via nonnatural routes of high parasite doses, an experimental system criticized by some authors, the superior efficacy of DNA vaccines was also observed using a low-dose intradermal challenge in the ear, which was proposed to more closely mimick natural infection [45] . In these studies, both DNA and protein vaccination were able to induce very similar level of short-term (2 weeks postvaccination) protection against infection with L. major, but only DNA vaccine was able to induce long-term (12 weeks postvaccination) protection [45] . These results thus confirmed the strong potential of DNA vaccines against Leishmania, but also indicated that in most cases only partial protection was achieved. Prime-boost immunization protocols have been tested with various antigens to increase vaccine potency (Table 1) . They are based on priming the immune response with a DNA vaccine and boosting with the corresponding recombinant vaccine based on recombinant virus or protein (Table 1) . In some studies, such immunization protocol resulted in increased immunogenicity of the vaccines and better protection levels [26, 27] , but in others, DNA only remained the best formulation for optimum efficacy [32] . Nonetheless, a major drawback of such vaccine formulation remains its complexity, which may limit their practical use. An alternative way to broaden vaccine immunogenicity and increase its efficacy has been to use combination of plasmids encoding various antigens. For examples, cysteine proteinase (CP) a and b DNA vaccines are not protective when used individually, but immunization with a combination of both plasmids induces long-term protective immunity [34] . Alternatively, gene fusion has also been successively used to achieve expression of an antigenic fusion protein from a single plasmid construct [33] . Overall, expression of several antigens mostly resulted in increased efficacy, but this also depended on the antigen combination [13, 22, 23, 41, 45] . Most authors thus argue that a successful Leishmania vaccine is likely to be based on multiple antigens. Immunization with large number of plasmids is also the basis for expression library immunization, a powerful but laborintensive strategy for vaccine discovery [46] , which has been used with Leishmania. Immunization of mice with L. major genomic expression library fractions was able to induce significant protection, but these authors did not pursue library fractionation further [47] . In another study, the identification of protective library subsets from an L. donovani amastigote cDNA library and their successive fractionation into smaller protective libraries lead to the identification of novel protective antigens [48] . Interestingly, most of the antigens identified would not have been predicted to be good vaccine candidates. Indeed, they were not surface or secreted proteins, neither stage-specific, but were intracellular and some very conserved such as histones, or ribosomal proteins [48] . Vaccine discovery is also the next logical step following the recent completion of the L. major genome sequencing [49] . In one approach, the random screening of 100 genes upregulated in amastigotes tested as DNA vaccine allowed the identification of 14 novel protective and 7 exacerbating antigens [50, 51] . Again, function and cellular localization would have been poor predictors of the protective efficacy of these antigens, as most were not predicted to be localized on the surface, but shared similarity with ribosomal proteins, cytoskeleton, or metabolic enzymes [51] . It is thus becoming increasingly clear that there is little rationale to limit Leishmania vaccine discovery searches to surface or secreted antigens. Rather, new criteria need to be considered for the rational identification of vaccine candidates as strategies based on such random screening cannot be applied to large genomes such as that of Leishmania, with over 8000 annotated genes. An additional advantage of DNA vaccines is their potential as therapeutic vaccines, aimed at reinforcing or redirecting the immune response of an infected host to control disease progression [58] . The major advantage of this strategy in addition to its efficacy is that it relies on short treatment regimens, and it is thus an attractive alternative to chemotherapy, particularly in the case of Leishmania with so few chemotherapeutic options. Thus, administration of as little as two doses of a DNA vaccine encoding PSA-2 can control an ongoing infection with L. major in mice [59] . The therapeutic effect is due to a shift of the immune response towards a Th1 immune response [59] . Similarly, a DNA vaccine encoding L. donovani nucleoside hydrolase NH36 has therapeutic activity against murine visceral leishmaniasis caused by L. chagasi [60] . The simplicity of such treatment makes them very advantageous compared to chemotherapy. In addition, the fact that the same DNA vaccine can be effective for both the prophylaxis [40] and the therapy of Leishmania infection is thus very promising as this would provide a versatile tool for the control of this parasite. As mentioned above, an added challenge to Leishmania vaccine development is the large number of species, as well as the variability within species. Indeed, studies on the polymorphism of leading antigens such as GP63 quickly revealed that it was a very polymorphic [61, 62] . Such polymorphism has important implication for vaccine development as it may limit their efficacy against variant strains of parasites or novel escape mutants, and thus restrict vaccine protection to a single species [63, 64] . Antigen polymorphism between multiple strains and species is thus becoming a major issue in many vaccine development studies [65, 66] . In the case of Leishmania, few DNA vaccines have been tested against mul-tiple species. LACK antigen, initially identified in L. major, and found to be very conserved between Leishmania species, can protect mice against L. major [20] and L. amazonensis [67] , but not against L. mexicana [23] , L. donovani [25] , or L. chagasi [24] . On the other hand, L. amazonensis nuclease protein P4 can protect against both L. amazonensis and L. major, but cross-protection requires a different formulation (IL-12 or HSP70 as adjuvant, resp.) [38] . In other studies, antigens from one species were used to induce protection against another species [31] , but the extent of crossprotection against various species was not investigated. More recently, a single formulation of L. donovani NH36 DNA vaccine was found to induce a very good protection against both L. chagasi and L. mexicana, suggesting that this DNA vaccine may be able to provide broad protection against various Leishmania species [40] . Importantly, no DNA vaccine has yet been tested against L. braziliensis, in spite of this species being responsible of most cases of cutaneous leishmaniasis in South America. While all the above DNA vaccines were based on Leishmania antigens, an alternative approach has used antigens derived from sand-fly saliva. Indeed, it has been shown that sand-fly saliva can exacerbate Leishmania infection [68, 69] , and preexposure of mice to saliva components may be sufficient to induce protection against infection [70] . Thus, a number of salivary antigens have been tested as vaccines against Leishmania. Maxadilan is a potent vasodilator from sand-fly saliva and was found to be responsible of most of the exacerbatory effects of whole saliva on Leishmania infection [71] . Immunization with this antigen (as a recombinant vaccine) protected mice against L. major infection [71] . Other salivary components, such as Phlebotomus papatasi SP15, have been tested as DNA vaccines and found to protect mice against L. major and while the vaccine induced both humoral and DTH responses, protection seemed to be mostly accounted for by the latter, as B-cell deficient mice remain protected [72] . Thus, characterization of sand-fly salivary proteins may lead to the identification of new vaccine candidates [73, 74] . However, as for Leishmania antigens, salivary protein polymorphism remains an important issue and may limit the usefulness of such antigens as vaccine candidates [75, 76] . Based on the success of many of these DNA vaccine studies in mice, a few vaccine candidates have been tested in additional animal models, possibly more relevants for the development of a veterinary or human vaccine (Table 2) . PFR-2 and KMP11 antigens were tested as DNA vaccines in hamsters, a highly susceptible animal model. PFR-2 was tested as protein, DNA, or DNA-protein immunization, and protection levels against L. mexicana varied greatly depending on vaccine formulation, route of immunization, and sex of the animals [52] . Also, contrary to mouse studies, protein vaccination seemed more protective than DNA only vaccination. However, as in mouse studies, heterologous prime-boost vaccination with DNA and protein seemed better than DNA only [52] . Another DNA vaccine encoding PapLe22 was found to be immunogenic in hamsters and decreased parasitemia after infection with L. infantum, but further assessment of disease was not performed [54] . Immunization with KPM11 DNA induced a mixed Th1/Th2 response, but was able to protect hamsters against visceral leishmaniasis caused by L. donovani [53] . In dogs, while several protein vaccines have been tested and a purified protein vaccine has now been licenced for veterinary use [77] , very few DNA vaccine studies have been performed. A heterologous prime-boost strategy using CPa and CPb DNA and protein was reported as immunogenic and protective [57] , but the study was of limited power given the reduced number of animals. In another study, dogs were immunized with a mixture of DNA vaccines encoding 10 different antigens previously tested in mouse models, and this immunization induced a very good immune response, with a high production of IFNγ [56] . However, evaluation of protection was limited to an acute in vitro assay [56] and further studies will be required to assess the potential of this vaccine in dogs. In spite of their limitations, these studies clearly showed that several DNA vaccines can induce a potent immune response in nonmurine animal models, and it is likely that a good level of protection can be achieved in these as well, provided the correct antigens and vaccine formulation are used. Vaccine development against Chagas disease has been dramatically limited because of extensive debate on the mechanisms involved in this pathology [78, 79] . Indeed, some studies suggested that tissue damage was associated with the presence and replication of intracellular amastigotes, while others proposed that autoimmunity induced by parasite antigens mimicking host proteins was responsible for it. It was thus unclear if the immune response needed to be inhibited, to reduce autoimmunity, or stimulated, to eliminate the parasite. It is now accepted that the presence of parasites in cardiac tissue is necessary to initiate and maintain the inflammatory response, and that therapeutic treatments or vaccines aimed at eliminating T. cruzi would limit or prevent the progression towards chronic chagasic cardiomyopathy [80, 81] . There is a growing consensus that protection against T. cruzi relies on a Th1 immune response and the activation of cytotoxic CD8 + T cells [82] [83] [84] [85] . The first DNA vaccines to be tested against T. cruzi encoded an antigen from the well characterized trans-sialidase family of proteins. There are over 1400 members in this family, making it one of the largest protein families of the parasite, and they are very abundant surface proteins. Several studies have used different members of this family, such as TS or TSA-1 (Table 3 ) [84, [86] [87] [88] . Immunization with TS was found to induce significant antibody titers able to inhibit trans-sialidase enzyme activity, a strong DTH, and lymphoproliferative response [86] . This immune response was protective as determined by an increase in survival and a decrease in parasitemia. Immunization with TSA-1 DNA was found to induce a specific CTL response which also lead to a lower parasitemia and increased survival in both BALB/c and C57BL/6 mice [88] . As in Leishmania vaccine studies, a few authors addressed the question of comparing protein and DNA vaccines encoding the same antigen [90, 98] . In A/Sn mice, immunization with recombinant TS induced a higher antibody titer than TS DNA, but a comparable decrease in parasitemia. However, the DNA vaccine was unable to increase survival, which the author attributed to the strain of the mice used, since this DNA vaccine was protective in BALB/c mice [90] . On the other hand, immunization with recombinant CRP or CRP DNA induced a comparable Th1 immune response, but only the DNA vaccine was protective against infection [98] . A number of other studies showed that DNA vaccines encoding various antigens could induce significant protection against T. cruzi infection, as evidenced by decreased parasitemia and improved survival of vaccinated mice (Table 3 ). In addition, a few studies also presented evidence of a reduction in cardiac tissue damage and inflammation at the histopathologic level [87, 97] . Furthermore, T cell analysis confirmed that protection relied on CD8 + T cells [84, 91] and recent studies showed that these cells were very rapidly activated following infection of mice immunized with DNA vaccines [101] . DNA vaccines based on defined T cell epitopes from TS antigen have also been tested and it was found that 6 Journal of Biomedicine and Biotechnology IM: intramuscular; CTL: cytolytic activity; −: no protection; +: little protection; ++: fair protection; +++: very good protection. both CD4 + and CD8 + T cell epitopes were necessary and sufficient to induce a protective immune response [102] . Taken together, these data clearly demonstrated that vaccination did not result in increased pathology, as initially feared, but allowed at least partial control of disease progression, thus confirming the central role of parasite persistence for Chagas disease pathogenesis and opening the way to further assessment of DNA vaccines against T. cruzi. However, it has to be noted that many of the antigens tested belonged to the trans-sialidase family of protein, so that there is still little diversity in terms of the antigens tested as vaccines against T. cruzi (Table 3) . Because protection induced by single antigen DNA vaccine remained partial, a number of studies have evaluated strategies to increase vaccine efficacy. These include the use of cytokine/chemokine encoding plasmids to potentiate the immune response induced by the vaccine, and two of the most studied molecules have been IL-12 and GM-CSF, which both were generally able to potentiate protection ( Table 2) . Alternatively, mixtures of plasmids encoding distinct antigens were used for immunization, and as mentioned above for Leishmania vaccines. For example, immunization of mice with plasmids encoding TS and ASP-2 proteins resulted in a specific immune response against both antigens and an increased protection against infection [97] . On the other hand, an immunization with a mixture of DNA vaccines encoding up to 6 proteins from the mucin family resulted poorly protective, while a mixture of up to 7 proteins from the TS family was protective, but not as much as a single antigen vaccine encoding the TS-like antigen ASP-2 [96] . Similarly, a mixture of DNA vaccines encoding ASP-1, ASP-2, and TSA-1 had a similar protective activity as TSA-1 alone [87] . The lack of efficacy of these multivalent vaccines may be attributed to the presence of shared or immunodominant epitopes since they have significant sequence similarity that may not have resulted in a broader immune reponse. Heterologous prime-boost approach has also been evaluated and immunization with some combinations of DNA and recombinant TS was found to enhance Th1 immune response, but protection was not significantly different from that obtained with DNA alone [103] . Taken together, these studies suggest that additional strategies need to be investigated to potentiate DNA vaccine efficacy against T. cruzi. Therapeutic administration of DNA vaccines to control an ongoing infection with T. cruzi may represent an additional Eric Dumonteil 7 alternative for Chagas disease control. The concept was demonstrated in mice acutely or chronically infected, and in both cases the administration of only two doses of DNA vaccine encoding TSA-1 or Tc24 antigens was sufficient to limit disease progression, as treated mice presented increased survival and reduced cardiac tissue damage, as assessed by histopathologic analysis [104] . A comparative study of different DNA vaccines identified Tc52 antigen as another therapeutic vaccine candidate, while DNA vaccines encoding antigens from the TS family previously found to be protective had no signifiacnt therapeutic effect [105] . It was found that therapeutic vaccination rapidly induced spleen cell proliferation, including IFNγ-producing CD4 + and CD8 + T cells, while the effects on cardiac tissue inflammation and parasite burden take longer to be detectable [106] . Importantly, in all these studies, therapeutic vaccination of T. cruzi infected mice did not result in an increased inflammatory reaction in the heart, confirming that it is safe to stimulate the immune response of T. cruzi infected mice and that attacking the parasite can lead to a reduction of pathology. These studies thus open very attractive perspectives for the control of T. cruzi infection, and further studies on the efficacy of DNA vaccines encoding other antigens and on the immune mechanisms underlying their therapeutic effect should provide clues for the optimization of this strategy. As for any vaccine, the nature of the antigen used remains a key factor for vaccine efficacy, and there is still little variety in terms of antigens evaluated as DNA vaccine candidates against T. cruzi. Thus, a number of studies have aimed at identifying novel antigens through various strategies. The most classical approach has been the screening of cDNA libraries using antibodies and screening an amastigote library allowed the identification of a novel antigen Tcβ3, and two previously characterized ones, LYT1 and FcaBP/Tc24 [96] . DNA vaccines encoding these antigens induced variable levels of protection, the best one being LYT [96] . Alternatively, expression-library immunization, described above for Leishmania, was also tested with T. cruzi, and found to be immunogenic, but there was no attempt at fractionating the library or identifying protective antigens [107] . A likely reason is that such strategy may be too labor-intensive for large genomes/libraries, and its usefulness may be limited to pathogens with small genomes. The availability of T. cruzi genome sequence also opens new possibilities for antigen discovery. In one of the first studies using such resource, a combination of bioinformatics analysis were used to identify GPI-anchored or secreted proteins, and most of the identified clones were immunogenic as DNA vaccines [108] . Further studies may confirm the usefulness of these new vaccines to protect against T. cruzi infection. Nonetheless, as discussed above for Leishmania, the rationale for limiting antigen searches to surface proteins may not be totally relevant, and additional strategies should also be used to include unbiaised genome-wide surveys for antigen discovery. As detailed in this review, there have been considerable advances in DNA vaccines against Leishmania and T. cruzi in recent years. Taken together, these studies clearly validated the concept of using DNA vaccines for both protection and therapy against these protozoan parasites in a variety of mouse models. While sterile immunity seems to be an irrealistic goal for either Leishmania or T. cruzi, a reduction in disease severity and in the development of the pathology seems clearly within the reach of DNA vaccines. Nonetheless, the relevance of such mouse models for the development of veterinary or human vaccines against these parasites has been challenged by some authors. The few DNA vaccine studies in nonmurine models of leishmaniasis suggest that some extrapolation may be feasible, but certainly not completely. Additional advanced preclinical studies of DNA vaccine candidates in nonmurine animal models such as rats, hamsters, dogs, or monkeys are thus warranted in the next few years, to further explore the immunology and efficacy of DNA vaccines against these parasites. As already observed in such studies for other pathogens, this will lead to the challenge of achieving in these species an immunogenicity of comparable level and protective efficacy as that obtained in murine models. However, advances in adjuvants, DNA vaccine formulation, and delivery systems are likely to contribute to such results [1, 109] . Another major issue is that of antigen discovery, and while a number of DNA vacines tested so far against Leishmania or T. cruzi have shown promise, we are still unsure if these are the best possible antigens, particularly since these parasites have relatively large genomes, and only a limited variety of antigens have been tested. The availability of the genome sequences of these parasites will without doubt be a key resource for genome-wide screenings for new protective antigens. A key lesson from the initial studies reviewed here [48, 51, 108] , together with other similar antigen discovery studies, seems to be that cellular localization and protein function are poor predictors of the antigenicity and protective efficacy of a protein. Alternative criteria should thus be used so that potent vaccine candidates are not missed, and the important development of genome-mining and bioinformatic tools is providing new tools for a more rational search of vaccine candidates [110] . To conclude, those DNA vaccines represent a promising approach for the control of Leishmania sp. and T. cruzi, and such vaccines would have a major impact in developing endemic countries. Thus the question does not seem to be if DNA vaccines can control these parasites, since many studies have clearly showed that this is the case, but how to translate what has been achieved in mouse models into veterinary or human vaccines of comparable efficacy. This work was funded by Grant SEP-2004-C01-47122 from the Consejo Nacional de Ciencia y Tecnología (CONACYT).
98
Estimating Individual and Household Reproduction Numbers in an Emerging Epidemic
Reproduction numbers, defined as averages of the number of people infected by a typical case, play a central role in tracking infectious disease outbreaks. The aim of this paper is to develop methods for estimating reproduction numbers which are simple enough that they could be applied with limited data or in real time during an outbreak. I present a new estimator for the individual reproduction number, which describes the state of the epidemic at a point in time rather than tracking individuals over time, and discuss some potential benefits. Then, to capture more of the detail that micro-simulations have shown is important in outbreak dynamics, I analyse a model of transmission within and between households, and develop a method to estimate the household reproduction number, defined as the number of households infected by each infected household. This method is validated by numerical simulations of the spread of influenza and measles using historical data, and estimates are obtained for would-be emerging epidemics of these viruses. I argue that the household reproduction number is useful in assessing the impact of measures that target the household for isolation, quarantine, vaccination or prophylactic treatment, and measures such as social distancing and school or workplace closures which limit between-household transmission, all of which play a key role in current thinking on future infectious disease mitigation.
The household is a fundamental unit of transmission for many directly transmitted infections. In addition, the household provides a ''laboratory'' within which key measures of transmission such as infectiousness, generation time and the effect of immunity or vaccination can be studied [1] . In recent years considerable effort has gone into understanding the dynamics of transmission within populations organised into households using mathematical models [2, 3, 4, 5, 6] . Most effort has gone into analysing the asymptotic behaviour of these models, elucidating the threshold levels of transmission required for infection to be self-sustaining, calculating final epidemic sizes, or predicting the impact of generalised or targeted interventions designed to reduce or eliminate transmission. In parallel, methods have been derived to estimate the parameters which govern transmission within the household from detailed case reports [7, 8, 9, 10] . However, scant effort appears to have been paid to how to apply household structured models to the analysis of epidemics, either retrospectively or in real time. Concurrently, mathematical models have played an ever greater role in interpreting and responding to emerging pathogens. These models have typically been either of the ''simple but tractable'' variety which ignore or average over demographic structure and social mixing patterns [11, 12] or the ''complex computer simulation'' variety that capture many details of demographic structure and dynamics, but of whom the behaviour can only be determined by intensive numerical analysis [13, 14, 15] . The aim of this study is to develop methods of a perhaps ''slightly less simple but still tractable'' variety that capture some of the detail that micro-simulations have shown is important, but which can be rapidly applied (say on a daily basis) in an emerging outbreak situation, to inform policy. More specifically, the aim is to arrive at a method to estimate the key transmission and control parameters for a model of transmission within and between households from as few detailed observations as are likely to be gathered in the heat of a major outbreak. The resulting analysis will still be based on major simplifications in respect to all the spatial and other social constructs that govern disease transmission, but less so than those based on the very simplest assumption of free, homogeneous mixing. In this context, it should be stated that even in the best, most robustly parameterised microsimulations, gross approximations are made in describing the fabulously complex web of human behaviour, and even they are only attempts to characterise the statistical properties of the system as a whole. Extensive effort is, and should continue to be, spent on identifying the conditions where different types of simplification (household models, static network models, spatial metapopulation models…) can and can't be justified, and in developing analytical approximations to describe disease transmission within such simplified structures. Individual based simulations of influenza and smallpox pandemic spread and control, incorporating detailed information on population density, age structure, commuting patterns, workplace sizes and long-distance travel have highlighted the particular importance of the household as a fundamental unit of transmission [13, 14, 16, 17, 18] (and reviewed in [19] ). Pure household models have been used fruitfully to explore detailed policy options in a city-wide response to an influenza pandemic [20] . It thus seems a priori that household models are a natural starting point in terms of extending theory previously developed for the simplest assumption of homogeneous mixing. The analysis presented here will focus on deriving new estimators for individual and household reproduction numbers, denoted R (t ) and R * (t ) respectively. The individual reproduction number R (t ) is defined roughly as the average number of people someone infected at time t can infect over their entire infectious lifespan; as I will show below, there are several ways of defining this more precisely. The household reproduction number R * (t ) is defined here as the average number of households a household infected at time t can infect [3, 6] . The individual reproduction number R (t ) rightly plays a privileged role in epidemiology, as it is a meaningful measure within any contact network. However, of the possible summary measures of epidemic progress, it is not necessarily the most useful. For example, for an emerging directly transmitted pathogen, such as pandemic influenza virus, public health interventions may target the household rather than the individual, enforcing household quarantine as well as offering antivirals to the household to limit transmission within the household. In such a situation, the household reproduction number R * (t ) is more directly related to the parameters which characterize the intervention, and is thus a better measure of the effect of these interventions. These quantities (R (t ) and R * (t )) share the two essential properties of reproduction numbers, namely that they increase when infectiousness increases and decrease when infectiousness decreases (monotonicity), and that they mark a threshold that separates exponentially growing epidemics (when R (t ).1 or equivalently R * (t ).1) from exponentially declining epidemics (when R (t ),1 or equivalently R * (t ),1) [3, 6] . The structure of the paper focuses first on deriving estimators for individual reproduction numbers, then on household reproduction numbers and finally on examples of pandemic influenza dynamics and measles. Though less well known than their compartmental counterparts (SIR, SIS, etc…), time-since-infection models offer a more intuitive starting point for modelling infectious disease transmission, and importantly for this application, they provide two other major advantages. First, it is typically easier to identify their key parameters, and second they more readily adapt to describe multi-level transmission (by multi-level, I mean here withinhousehold and between household). A disadvantage is that it can be harder to include heterogeneities. Nomenclature is confusing, since both types of model have their origin in the same classic paper of Kermack and McKendrick [21] , and both the SIR model and the simplest time-since-infection model are known as ''the Kermack-McKendrick model''. The model, in the formalism chosen here, predicts the changing incidence rate I (t) as a function of calendar time t in terms of the transmissibility, denoted b (t, t ), an arbitrary function of calendar time t and time since infection t. b (t, t) typically reflects pathogen load, or perhaps more precisely pathogen shedding. It is commonly a single peaked function reflecting pathogen growth followed by immune suppression, or host death, but can be more exotic such as the double peaked profile associated with early and late transmission of HIV [22] , or the repeated peaks of malaria [23] . b (t, t) also reflects the effective contact rate between infectious and susceptible individuals, which can change during the course of a single infection, increasing for example if a person coughs or sneezes due to respiratory disease, or decreasing if a person takes to bed with illness, and during the course of the epidemic as public health measures are implemented. More discussions of the components (infectiousness and contact) of b (t, t) can be found in [24] . Because I am interested in outbreaks of emerging infections, I will not describe explicitly reductions in the susceptible population caused by the epidemic. Formally this corresponds to working in the infinite population limit. This assumption is not essential for this section however, since b (t, t) could also be thought of as incorporating the proportion of cases that are susceptible; the assumption becomes more important in the later sections on household models. Mathematically, transmission is defined by a Poisson infection process such that the probability that, between time t and t+d, someone infected a time t ago successfully infects someone else is b (t, t)d, where d is a very small time interval. This assumption then results in a prediction that the mean incidence I (t ) at time t follows the so-called renewal equation This equation states that the number of newly infected individuals is proportional to the number of prevalent cases multiplied by their infectiousness. It may often be convenient (and realistic) to truncate the function b (t, t) at a time t m such that b (t, t) = 0 for all t.t m . The asymptotic behaviour of incidence I (t ) is determined by reproduction numbers [21, 25] . Two intuitively defined reproduction numbers are the case reproduction number, which I denote R c (t ), and the instantaneous reproduction number, which I denote R (t ). The case reproduction number R c (t ) is a property of individuals infected at time t, and is the average number of people someone infected at time t can expect to infect. For a person infected at time t it is the total infection hazard from time t onwards, i.e. While the case reproduction number has been widely used, it may also be worth considering a quantity which I call the instantaneous reproduction number R (t ), a property of the epidemic at time t. It is the average number of people someone infected at time t could expect to infect should conditions remain unchanged. It is given by To illustrate the distinction between R c (t ) and R (t ), consider a situation where the transmission rate is abruptly reduced at a time t = t I . The instantaneous reproduction number R (t ), which estimates how many people one case would infect if circumstances were to remain fixed, would abruptly switch from a high to a low value at time t I . The case reproduction number R c (t ), on the other hand, estimates how many people each case actually infects. It will thus account for the fact that someone infected at time t,t I may spend part of their infectious period before and after the reduction in transmission which occurs at time t I and thus R c (t ) will smoothly transition from higher to lower values. To derive simple estimating equations for R (t ), I consider the case where this function is separable, which corresponds to saying that the relative progression of infectiousness as a function of time since infection is independent of calendar time. In this case b (t, t) can be written as the product of two functions w 1 (t ) and w 2 (t), i.e. b t,t ð Þ~w 1 t ð Þw 2 t ð Þ ð4Þ A counter-example might be when reactive patient isolation is introduced and acts to reduce infectiousness in late stage infection, in which case b (t, t) can't be decomposed in this way. For this type of situation, it may be reasonable to assume the b (t, t) can be decomposed separately in different stages of the epidemic, pre-and post-implementation of isolation measures, for example. Since b (t, t) is a product, I can arbitrary normalise one or other of the functions w 1 (t) and w 2 (t), so without loss of generality, I choose w 2 (t) to have total integral 1, i.e. Ð ? 0 w 2 t ð Þdt:1. The function w 1 (t) is equal to the instantaneous reproduction number R (t). The function w 2 (t) is then the distribution of how these infection events are distributed as a function of time since infection t. This is an idealised definition of the generation time distribution, which I denote w (t). Thus, infectiousness can be decomposed as the product of the instantaneous reproduction number and the generation time distribution, i.e. The relationship between the idealised generation time distribution w (t) and the distribution of observed generation times can be rather complex for a number of reasons. First, infections are rarely observed, and thus must be either backcalculated or the generation times must be based on a surrogate such as the appearance of symptoms [1, 12] . Second, right censoring can cause the observed generation times to be shorter or longer than expected for a growing or declining epidemic, respectively [26] . Third, as apparent here, if the reproduction number R (t) changes due to depletion of susceptibles, changes in contact rates or public health measures, then this will also change the observed generation times for infectious individuals during that period of change. Thus the distribution w (t) is really intended as a measure of infectiousness which will correspond to generation times for an index case in an ideal large closed setting where contact rates are constant. It can be inferred from data on the timing of cases, as in [10, 13] . Inserting (6) into (1) yields a novel estimator for instantaneous reproduction number By substituting the decomposition (6) into equation (2) , a relation between the instantaneous and case reproduction number is obtained: i.e. the case reproduction number is a smoothed function of the instantaneous reproduction number. Usually, incidence is reported as a discrete time series of the form I i incident cases reported between time t i and time t i +1, in which case the generation time distribution should be appropriately discretised into a form w i such that P n i~0 w i~1 . The estimators for the reproduction numbers become and Equation (10) was proposed by [12, 27] as a real time estimator of the reproduction number, while equation (9) was first used for analysing polio transmission in India [28] (based on the work presented in this manuscript). While the case reproduction number is an intuitively appealing quantity, the instantaneous reproduction number estimated by equation (9) should also be considered for practical applications as it may suffer fewer problems of right censoring in an incompletely observed epidemic. Right censoring is a real problem in using the case reproduction number to track an epidemic in real time, since the estimator for R c (t) at time t is seen in equation (10) to rely on knowing the incidence at future time-points. An algorithm to deal with this issue was proposed by [29] , but switching instead to the instantaneous reproduction number estimated by equation (9) may be a simpler solution. Right censoring is not however the only complication associated with estimating reproduction numbers in practice, and is not completely absent from (7) due to the delay in detecting infections. Left censoring may also arise due to not knowing the baseline number infected if an epidemic has been unfolding for some time before observations are recorded. Finally, estimating the generation time distribution may not be straightforward. Several strategies are possible to deal with the fact that one never observes infections, but rather as a time series of cases of the form C i , where case definitions could be based on symptoms, hospitalisation or seroconversion. One strategy, used in [12] , is simply to ignore this and use cases as surrogates of infection for estimation of both the generation time and the reproduction numbers. Often though, it may be possible to characterise a distribution of the time from infection to becoming a case, say j i where P n i~0 j i~1 . If a case is defined by symptoms then this would be the incubation period distribution. One can then backcalculate incidence as follows A drawback of this approach is that the estimated incidence time series Iˆi will tend to be over-smoothed relative to the original time series I i . It also makes clear that there is still a problem of right censoring in an incompletely observed epidemic in the estimator of equation (9), though less than in equation (10) . Statistical properties of these estimators are straightforward [12, 28] . One previously noted point [12, 28] is that because these estimators are essentially ratios of incidences, they can be used in cases where only a fraction of cases are observed, such as for polio where only a tiny fraction of infections lead to disease (of the order of 1 in 200), though the confidence intervals will change. A special case applicable to many cases where surveillance is poor is when only the epidemic growth/decline rate r is known. In this case the incidence takes the form I (t) = I (0)exp (rt) and both estimators (7) and (8) for the reproduction number become where the reproduction numbers are now expressed as a function of the exponential rate of change r. This is likely a useful formula, presented and studied in detail in [30] , where the links to earlier ecological and demographic modelling were also highlighted. Much of the subsequent analysis will concern itself with deriving an equation equivalent to (12) for the household reproduction number R * (r). The model defined above assumes that the function b (t, t) describes the ''natural history'' of infection in each infected individual. Before specialising to the model of household transmission, it is first worth considering the case where different individuals experience different ''natural histories'', defined here by the susceptibility to infection, and infectiousness after infection. I denote a vector of random variables X = {X1, X2, …} to describe factors which influence susceptibility or infectiousness. For example for the standard SEIR model of infection the random variables would be the durations of the latent period (L) and the infectious period (D), i.e. X = {L,D}. Let f (X ) denote the probability distribution of these random variables amongst new infections (taking into account differences in susceptibility), defined such that where the integral is taken over the domain of the random variables. In other words, f (X ) is the proportion of new infections that have state X. Let b (X, t, t) denote the infectiousness profile of an individual with state X. Assuming that all individuals mix homogeneously, then the transmission model defined earlier by equation (1) is generalised to where I (X , t) is the incidence of infections with state X. I define the function K(t) to denote the integral which clearly depends only on time t and not state X. The total incidence at time t is defined by the integral By substituting equation (14), which can be rewritten as I (X, t ) = f (X ) K (t ), into equation (16), I obtain that K (t ) = I tot (t ) and thus that I can now substitute (17) into (14) to obtain Dividing both sides of this equation by f(X)yields an equation for the total incidence If I define the average infectiousness as follows then equation (19) can now be seen to be the standard Kermack-McKendrick model of equation (1), i.e. In other words, in this model of an emerging infectious disease epidemic with heterogeneities in susceptibility and infectiousness, the dynamics of mean total incidence of infection is exactly equivalent to the basic model where the infectiousness is appropriately averaged using equation (20) . Once an expression is derived for the average infectiousness b (t, t), the results such as equations (9) or (12) can be used without further consideration of the heterogeneities in infectiousness or susceptibility. Heterogeneities which are transmitted or preserved from one infection to the next, for example due to non-random mixing between different risk groups, a situation not considered here, lead to a more complex result. Some public health interventions such as isolation and contact tracing can induce such heritability even if it is not a basic property of the transmission process [31, 32] . A useful exercise in applying this formalism (not elaborated here) is the derivation of standard formulae for the basic reproduction number as a function of the exponential growth rate r for the SEIR model [30] . One approach to estimating household reproduction numbers is simply to switch perspective from individual to household, directly estimate the generation time distribution (times taken for one household to infect another) and incidence of infection of households, and apply the results of equations (9) or (12) to estimate reproduction number as a function of time, R * (t), or exponential growth rate, R * (r). Because, as I have shown, the linearised Kermack-McKendrick model is applicable even when susceptibility and infectiousness are heterogeneous, this method is acceptable despite the fact that households may be quite heterogeneous in size and in the number of people infected. One analogous situation where this approach has been used is in estimating farm-to-farm reproduction numbers in the 2001 UK foot-and-mouth virus epidemic [27] . However, unless specifically tailored to this task, it is unlikely the data will be collected in the requisite form for this approach to be used in the human household situation. Thus, in this section I explore the alternative approach of explicitly modelling transmission within and between households. Homogeneous transmission models can be interpreted as twolevel hierarchical models, where the processes which guide the natural history of infection within the host are considered separate from those which drive transmission between hosts. The link between the two can be thought of as the function b (t, t) which translates the impact of changing processes within the host into changing infectiousness as a function of time since infection. The approach taken here to modelling household transmission is to study a three-level hierarchical model of transmission. The three levels are within-host, within-household, and between households. The natural history of infection is described by the individual infectiousness function b (t, t). I assume in this section that individuals are homogenous in infectiousness and susceptibility. I then use this to predict the course of epidemics within households, and derive a function b * (t, t * ) which describes the average infectiousness of a household towards other households as a function of the time since the household was infected, t * (from here-on, I use the starred symbols to denote properties of households, and un-starred symbols to denote properties of individuals). The basic idea behind this analysis is illustrated in Fig 1. To simplify the notation, and because the main aim of this section is to study the case of an epidemic growing exponentially, I consider the situation where infectiousness is independent of calendar time t. This could be relaxed, though only if variation in time is somewhat slower than the typical duration of infection within a household. More specifically, the model assumptions are that: N individuals are distributed into households, and mix randomly and homogeneously outside of their household; N within a small time interval d, an individual who has been infected a time t ago infects a person at random in the population with probability b G (t )d; N within this same time interval he or she infects each susceptible individual in his or her household with probability b L (t, n)d (this is allowed to depend on the household size n, since empirical evidence suggests such variation may occur [10] ); N the population is large, and the disease has low prevalence, so that the probability of a household being repeatedly infected is negligible; N the functions b G (t ) and b L (t, n ) are proportional to each other as functions of the time since infection t. As a result of the last assumption and of the discussion around equation (6), the infectiousness functions can be decomposed as b G (t) = R G w (t) and b n (n, t) = r n w (t), where R G is the average number of people each infected individual infects through random (non-household) contacts, w (t) is the generation time distribution for between household transmission, and r n is a parameter describing infection within the household whose interpretation will be clarified below. I start by analysing the process of transmission within a single infected household of size n in terms of the functions r n and w (t). Consider first a household of size 2, where one individual is infected at time t * = 0. Given the Poisson process described by the assumptions listed above, the probability that the second individual The probability that the second person is never infected is The distribution of times of infection of the second individual, conditional on infection, is then where 2L q 2 (t * )/Lt * is the rate of change of the cumulative probability of not being infected, i.e. the probability density of being infected at time t * , and the normalising factor 1-Q 2 is the total probability of being infected. The difference between w 2 (t * ) and the standard generation time distribution w (t) is a saturation effect, so that the second case tends to get infected earlier as the infectiousness of the index case (r 2 ) is increased. The infectiousness of the second individual towards other nonhousehold members of the population, conditional on his or her infection, and described as a function of the time t * since the infection of the household is thus the convolution of w 2 (t * ) and b G (t), so that the total infectiousness of the household is Generalising this exact result to larger households involves some complications. Consider for example a household of size 3, where one individual is infected at time t * = 0. The probability that neither of the other two individuals is infected by the first individual at time t * is q 3 t à ð Þ~exp {r 3 Ð tà 0 w s ð Þds À Á directly analogous to the situation for households of size 2. However this is somewhat greater than the actual probability that they are not infected at all, since once one of these two is infected, they can also infect the other, and thus the probability that they each escape infection is somewhat less than Q 3 :q 3 ? ð Þ~exp {r 3 ð Þ. To progress further with analysing this system, I propose to approximate the process by assuming that infections within a household can be approximately described by a discrete generation Reed-Frost model, i.e. where the probability of not being infected in each generation is (Q n ) m where m individuals are infected in the previous generation and Q n u exp (2r n ). Q n is the escape probability of each infectious-susceptible pair of individuals considered in isolation. In the formalism proposed by Ludwig, this corresponds to using infectious rank as a surrogate for infectious generation [33] . Dynamics are recovered by assuming the times between generations are described by the standard generation time distribution w (t). The ordering of infection events has no influence on the final number of individuals infected [33] , and therefore this approximation will produce exact results for the final number of people infected in each household. Because of the possibility of ''later'' generations preceding ''earlier'' ones, as noted in the case of households of size 3 above, and because of ignoring the saturation effect present in equation (22) in terms of the actual generation times within households, this approximation will overestimate the time taken for individuals to become infected in the household. Because of the general form of the relation between generation time and reproduction number seen in equation (12) , this will result in over-estimates of the household reproduction number R * (r). To provide a counter-balancing under-estimate of R * (r), I also consider an alternative approximation obtained by assuming the same total number of cases as predicted by this Reed-Frost model, but where all cases are assumed to be infected by the first index case. This is not a formal lower bound, since in the limit of infinite infectiousness within the household, all members of the household will be infected simultaneously upon introduction of the infection into the household. I find however that even for the example of highly infectious measles virus (below), the under-approximation is sufficient to provide a practical lower bound. The probability of different chains of infection within households can easily be computed from the assumed Reed-Frost model [2] . I denote pr( {m 1 , m 2 , …, m n }|n) the probability of a chain of infection occurring in a household of size n where m 1 index cases infects m 2 , who in turn infect m 3 tertiary cases and so on, up to a maximum of n generations of infection. It is an assumption of the where pr m iz1 j m 1 , . . . ,m i f g ,n ð ÞBinomial The second approximation is that the time taken for one infected to infect the next is distributed according to the standard generation time distribution w (t). The time at which someone in the (i+1) th generation of infection is infected is as a result drawn from the i th auto-convolution of this distribution, denoted here w [i] (t * ) and defined by the recursive convolution equation which satisfies . Consider now an individual in the i th generation of infection in the household, and consider this household at a time t * after the first index case was infected. This individual must have been infected at some earlier time s ( t * distributed according to the distribution w [i-1] (s). His or her infectiousness to others outside of the household will be given by b G (t * -s). Thus, by averaging over all possible values of s, the average infectiousness of such an individual in the i th generation is Thus having averaged over all possible times of infection in the chain of transmission events in the household, infectious households are stratified by their size and by the number of cases in each generation. Using the notation defined earlier, I define the state vector X = {n, m 1 , …, m n } of variables which define the infectiousness and susceptibility of the infected household, where n is the household size and m i is the number of infected individuals in the i th generation of infection in the household. The infectiousness of a household with this state X towards other households, mediated by random mixing of individuals between households, is the sum of the infectiousness of all the individuals each given by equation (27) Given that this infection process involves random mixing of individuals outside their household, the distribution of sizes of households which get infected is the so-called size-biased household distribution. This is the distribution of sizes one obtains by sampling individuals at random in the population and recording the size of their household, as opposed to the more commonly recorded household size distribution which is obtained by sampling households at random. If k n denotes the household size distribution, then is the size-biased household size distribution. Given a household of size n gets infected, the probability of a chain of infections is given by the Reed-Frost probabilities pr ({m 1 , …, m n }|n). The distribution of the random variables X = {n, m 1 , …, m n } at infection is thus f X ð Þ~k n : pr m 1 , . . . ,m n f g j n ð Þ ð 30Þ The mean infectiousness of a household is Let m = S i m i be the average total number of cases in an infected household. The household reproduction number takes an intuitive and well known form derived in [3, 6] , expressed in terms of the parameter R G as follows: i.e. the household reproduction number is the product of the expected number of infections in a household multiplied by the number of people each individual infects out of their household. The mean household generation time distribution (time for one household infecting the next) is The mean generation time for households, T g * , can be expressed in terms of the individual generation time T g as The generation time distribution w * (t * ) can be used for the previously defined estimators of reproduction numbers (7)-(12) using household incidence data or just exponential growth rates. The exponential growth rate r for an exponentially growing epidemic is the same whether measured for individual or household incidence. For an exponentially growing or declining epidemic, one obtains the estimator Now consider the integration where the first equality uses the definition of the auto-convolution, the second is a re-ordering of integrals, the third involves changing variables to u = t * -s, the fourth is a factorisation and the fifth arises by induction. The sixth uses the definition of the individual reproduction number R (r ) one obtains ignoring household structure from equation (12) . The household reproduction number can be expressed in terms of the individual reproduction number R (r ) as Examination of equation (33) immediately reveals that the estimate for the number of people each person infects out of the household is I have thus derived a simple analytic relation between the individual and household reproduction numbers. Both are approximations, ignoring the effects of local saturation on the generation time, which will tend to produce overestimates of the reproduction number. An alternative approximating to the household reproduction number, which provides an underestimate, is found when all secondary household cases are assumed to arise in the second generation, i.e. using equation (38) There are two reasons for considering household structure in analysing the pandemic influenza situation. First, influenza transmission is known to be concentrated within the household, and thus parameter estimates which ignore this heterogeneity are likely to be frail. Second, many public health policies for future pandemics are likely to be organised around the household. The net effect of social distance measures such as school and workplace closures and cancellation of social gatherings is effectively to reduce transmission out of households (and perhaps inadvertently to increase transmission within them). Furthermore, antiviral treatment and prophylaxis and quarantine measures are likely to be targeted at whole households rather than individuals (though restricting families with one suspect case to stay together without any other support is possibly undesirable) [16, 17, 20] . A number of studies have identified the parameters needed to estimate the household reproduction number for influenza [8, 10, 11, 17] . It is important to bear in mind that these parameters could be quite different in future pandemics, and thus that robust methodology may be more useful in responding to new outbreaks than numerical estimates obtained for past outbreaks. While it would be straightforward to use demographic data and exponential growth rates from earlier pandemics combined with interpandemic data on the transmissibility of influenza within households to obtain estimates of R * for historical pandemics, it has not been shown that the within household transmission parameters for inter-pandemic influenza adequately describe the pandemic situation, so I focus instead on providing illustrative examples using current demographic data (on the household size distribution from the UK) [34], and recent data on the transmissibility of influenza in modern households [10] . The household size data from 2001is truncated to size 6, and I assume that all households of size 6 or greater have size exactly 6. The data are k 1 = 29% (i.e. 29% of households are single person households), k 2 = 35%, k 3 = 16%, k 4 = 14%, k 5 = 5% and k 6 = 2%. The size of the mean household is thus 2.38 (average size of households where households are sampled at random), while the household of the mean individual has size 3.06 (average size of household to which individuals belong, where individuals are sampled at random). From the French influenza study [10] , I obtain maximum likelihood estimates of the within household transmission parameter of r n = 1.35/n 1.0 (which is consistent with the best fit to the Tecumseh data [8] of r n = 1.27/n 0.97 ). The former study followed seronegative households for a two week winter outbreak of seasonal influenza. The corresponding escape probabilities are Q 2 = 50.9% (i.e. the probability of not being infected by the other household member in a household of size two is 50.9%), Q 3 = 63.8%, Q 4 = 71.4%, Q 5 = 76.4% and Q 6 = 79.9%. On the scale of other infections, this places influenza as being approximately as infectious as mumps, but a lot less infectious than either varicella-zoster or measles [1] . By applying the Reed-Frost model to these data with this distribution of households, I obtain estimates of the average number of infections in each generation of infection of m 1 u 1, m 2 = 0.64 (i.e. the first index case directly infects an average of 0.64 people in his or her household), m 3 = 0.19, m 4 = 0.036, m 5 = 0.0037 and m 6 = 0.00021, and thus the estimate for the total expected number of cases in an infected household is m = S 6 i = 1 m i = 1.87, to be compared to the mean size of 3.06. These calculations are performed in Microsoft Excel 2007 using equation (25) . There is not yet a consensus on the generation time of influenza [13, 14, 16, 30, 35] , with estimates ranging from 2.6 days in [13] to 5.3 days in [14] . I use a Gamma distribution with mean T g = 2.85 days and standard deviation 0.93 days, as reported in [30] . Based on these data, I compare the predicted and simulated infectiousness of households in Fig. 2 , which shows the average over all households sizes and compares this to the final analytical approximation given by equation (31) for b * (t * ), and also the alternate approximation which considers all secondary infections to arise in the second generation of infection ; the simulations and the first approximation are clearly in good agreement. Individual based stochastic simulations were programmed using Berkeley Madonna, and are described in Appendix S1. For the case of an exponentially growing epidemic, the estimates of the individual and household reproduction numbers, R and R * respectively, are shown in Figure 3 , along with the estimate of the number of people one person infects outside their household, R G . For R * , both the under-and over-estimating approximations are shown, along with estimates obtained from the simulated generation time distribution. As expected for this low-infectiousness scenario, the simulated values are closer to the overestimating approximation. The range between these approximations which bracket the true value is rather narrow, indicating that the method is predictive. For the 1918 ''Spanish Flu'' H1N1 pandemic, the median growth rate in large US cities was r = 0.20 per day [30, 35] , with comparable estimates in the UK [17] . This value also serves as an upper estimate for the spread of the H2N2 pandemic virus in 1957 [17] . Based on this growth rate, the estimated individual reproduction number is R = 1.74, while the estimated household reproduction number is R * = 2.26, and thus the out-of-household reproduction number is R G = 1.21. Of course, households were bigger in 1918 than now, so that the actual value of R * was likely higher than this. These estimates would imply that a proportion 121/R * = 56% of between household transmission would need to be blocked to prevent epidemic spread. Figure 3 could provide a rough guide to the likely values of R * and R G for a new influenza pandemic where the rate of exponential growth can reliably be determined. Consider someone who the index case in their household; they would be expected to infect R G = 1.21 people out of their household and m 2 = 0.64 within their household. This validates assumed proportions of transmission within and between households from earlier simulation studies [17, 20] . The sum of these is greater than R since the reproduction number R is an average over different generations of infection within the household. For this value, the estimate of R which takes into account local saturation effects was determined numerically to be R = 1.79. Fig 3 shows that for all values of r, numerically estimated values for R (r ) are close to the curve estimated from application of equation (12) which ignores local saturation effects. As a final check of the method, epidemics within a community of 2,000 households were simulated using an individual based stochastic model (see Appendix S1). I choose R G = 1.21 as inferred from an epidemic growth rate of r = 0.20 per day, and the other parameters as described above. The exponential rate of growth was then re-estimated directly from the simulated incidence timeseries to be r = 0.19 (Figure 4) , close to the predicted value of r = 0.20. This provides further support for the validity of this method, especially since no restrictions were placed on re-infection of households within this small simulated community. As noted above, influenza is relatively uninfectious compared to other common viruses. For a contrasting application of the method, I now focus on measles which was the most infectious of the pathogens studied in [1] . Measles also has a more peaked generation time distribution, so that generations of infection are more distinct, and to make the contrast with the influenza estimates yet greater, I also use demographic data on household size chosen from the national census in 1961, when household sizes were greater than they are now. This analysis is perhaps a little artificial when applied to measles, since a large proportion of the population will have immunity either due to past infection or vaccination with the live MMR vaccine. The principal motivation is to further test and illustrate the methods in a case where good data on the transmission dynamics within households are available. Stratification by household of the recent outbreaks of measles caused by decreasing uptake of the MMR vaccine could reveal whether household heterogeneities should have be accounted for in estimating the changing reproduction number of measles [36] . The household size data from 1961 is truncated to size 6, and I assume that all households of size 6 or greater have size exactly 6. The data are k 1 = 14% (i.e. 14% of households are single person households), k 2 = 30%, k 3 = 23%, k 4 = 18%, k 5 = 9% and k 6 = 7%. The size of the mean household is thus 2.99 (average size of households where households are sampled at random), while the household of the mean individual has size 3.66 (average size of household to which individuals belong, where individuals are sampled at random). Hope-Simpson reported susceptible-infectious escape probabilities of Q = 69.9% for mumps, Q = 39% for varicella, and Q = 24.4% for measles in under 15s [1] . The results were reported independent of household size, and were regarded as unreliable in over-15s. Based on applying the Reed-Frost model to the measles (t) ) is shown, as is the infectiousness of the typical infected household (denoted b * (t * )). This latter curve is obtained by simulating over 10,000 epidemics of transmission within households starting from one infected case. The two analytical approximations described in the text are also shown. ''Approx 1'' is the main approximation described, while ''Approx 2'' is the one obtained by assuming that all infections occur in the second generation of infection within the household. Parameters are as described in the main text, and the curves are arbitrarily scaled such that each individual infects on average one person outside of the household (i.e. R G = 1). doi:10.1371/journal.pone.0000758.g002 estimate with this distribution of households, I obtain estimates of the average number of infections in each generation of infection of m 1 u 1, m 2 = 2.01 (i.e. the first index case directly infects an average of 2.01 people in his or her household), m 3 = 0.50, m 4 = 0.020, m 5 = 0.00036 and m 6 = 0.0000031, and thus the estimate for the total expected number of cases in an infected household is m = S 6 i = 1 m i = 3.54, to be compared to the mean size of 3.66. Hope-Simpson also reported the intervals between linked cases in households using different case definitions [1] ; the intervals for what he regarded as the most reliable case definition, ''maximum rash''. These data is well described by a Gamma distribution (not shown). The maximum likelihood estimate of the generation time is T g = 10.5 days with standard deviation 2.4 days. Based on these data, I repeat the simulations of the previous section on influenza but with parameters for measles in Figs 2, 3 and 4. Figure 2B shows that, as expected, the average infectiousness of a household is less well approximated by either approximation than for the much less infectious case of influenza. In this case, multiple peaks of infectiousness corresponding to generations of infection within the household can be clearly distinguished, and there are more cases in the second generation of infection than in the first. In terms of the predicting of the household reproduction number R * , the method is still found to be strongly predictive (as evidence by the small gap between upper and lower estimate) and reliable (compared to numerical estimates). While in influenza, the simulations were close to the upper approximation, here they are closer to the lower approximation, as expected for the more infectious situation of measles transmission. Simulations of transmission within a community of households were again found in Figure 4B to validate the approach. The difference in the shape of the epidemic curve with influenza reflects the different shape of the generation time distribution, though the exponential growth rate is the same. New methods were presented to estimate both the individual and household reproduction number during an epidemic. The new method presented for estimating the individual reproduction number relates closely to earlier work [12, 27, 30] , but provides an alternative and possibly simpler solution to the problem of incomplete observations during an unfolding epidemic [29] . It also provides an alternative and perhaps more satisfying solution than the incidence-to-prevalence ratio method [37, 38] to the problem of long generation time distribution infections such as HIV, where epidemiological circumstances can change substantially within the course of a single infection, and thus the case reproduction number represents too much of an average to convey secular changes in behaviour and transmission. Nothing in this study challenges the central role of the individual reproduction number as an epidemiological measure; because the empirical measures of reproduction number proposed here and in [12, 27, 29] use incident observed cases as the base, all of the complication in defining the 'typical' or 'eigen' case for structured models discussed most clearly in [24] are neatly sidestepped. What this study does highlight is that much complexity is hidden in effectively defining and estimating the generation time distribution for a structured population. In the case studied here, generation times between individuals are shorter for within household transmission than between household transmission, particularly for more infectious pathogens, and this resulted in systematic biases associated with estimating the reproduction number while ignoring this effect, which were quite substantial in the case of highly infectious measles virus. The methods presented for the estimation of household reproduction numbers were not affected by this problem in the same way. Analytical approximation were derived which bracketed estimates between a lower and upper bound, and numerical simulations showed the range within these brackets to be narrow. These approximations were shown to be robust, but it is worth noting that assumptions are made about the population mixing randomly out of their households and results are only valid in the scenario of an emerging pathogen where overall prevalence is low. The usefulness of these methods is likely to be found in predicting and understanding the impact of household targeted infection control measures in an emerging epidemic. This actually covers a wide class of interventions since the household is a central living and administrative unit in most populations. Decisions regarding isolation, quarantine, vaccination and prophylaxis may often be made for entire households. Similarly school and workplace closures as well as restrictions on leisure activity can be thought of as trying to reduce between household transmission. Analytical approaches are also invaluable in calibrating and providing independent checks on more detailed individual based micro-simulations, such as [13, 17, 20] . Some control interventions require more subtle analyses; for example it has been shown that vaccinating whole households is not the most effective strategy for a given vaccine coverage rate, and that alternative strategies such as preferentially vaccinating larger households could be considered [39] . Further avenues of research include studying the statistical properties of these estimators for different situations. The assumption made here, that individuals mix nearly homogeneously out of their household may be an appropriate approximation for describing transmission within a neighbourhood or even a city [20] , but ultimately one should also consider developing the estimators for more complex demographic situations such as a hierarchy of organisations (household, to village, to region, to country, etc…) or a more complex overlap of households, workplaces and regular social spaces. Also of interest is the study of intervention measures, particularly those that respond to the presence of a symptomatic cases; the measures of pre-symptomatic transmission presented in [25] clearly generalise to a household, but analytical results on the efficacy of isolation and quarantine are not evidently obtainable. The estimators of the household reproduction number have been shown here to be robust on their own terms, but I have not addressed the issue of model misspecification, for example to inaccurate determination of the generation time distribution or to individual heterogeneity in infectiousness or susceptibility within households. Further scenarios could be explored both to test the method with different infections and to address the issue of model misspecification. There are many cases where it may be desirable to quantify household transmission, but where a degree of natural or vaccineinduced immunity may be present in the population, a problem not addressed here. In considering these more complex situations, while it may not be possible to obtain analytic forms for the infectiousness of a household, numerical forms can usually be obtained quickly and still offer benefits over full individual based micro-simulations in easily exploring a wide range of parameters. Finally, the likely practical benefits of estimating household transmission parameters in an emerging epidemic need to be clearly established and communicated, and the most effective ways to enhance data collection protocols to allow their rapid estimation need to be identified. Appendix S1 Description of the simulations Found at: doi:10.1371/journal.pone.0000758.s001 (0.08 MB PDF)
99
Natural Killer Cells Promote Early CD8 T Cell Responses against Cytomegalovirus
Understanding the mechanisms that help promote protective immune responses to pathogens is a major challenge in biomedical research and an important goal for the design of innovative therapeutic or vaccination strategies. While natural killer (NK) cells can directly contribute to the control of viral replication, whether, and how, they may help orchestrate global antiviral defense is largely unknown. To address this question, we took advantage of the well-defined molecular interactions involved in the recognition of mouse cytomegalovirus (MCMV) by NK cells. By using congenic or mutant mice and wild-type versus genetically engineered viruses, we examined the consequences on antiviral CD8 T cell responses of specific defects in the ability of the NK cells to control MCMV. This system allowed us to demonstrate, to our knowledge for the first time, that NK cells accelerate CD8 T cell responses against a viral infection in vivo. Moreover, we identify the underlying mechanism as the ability of NK cells to limit IFN-α/β production to levels not immunosuppressive to the host. This is achieved through the early control of cytomegalovirus, which dramatically reduces the activation of plasmacytoid dendritic cells (pDCs) for cytokine production, preserves the conventional dendritic cell (cDC) compartment, and accelerates antiviral CD8 T cell responses. Conversely, exogenous IFN-α administration in resistant animals ablates cDCs and delays CD8 T cell activation in the face of NK cell control of viral replication. Collectively, our data demonstrate that the ability of NK cells to respond very early to cytomegalovirus infection critically contributes to balance the intensity of other innate immune responses, which dampens early immunopathology and promotes optimal initiation of antiviral CD8 T cell responses. Thus, the extent to which NK cell responses benefit the host goes beyond their direct antiviral effects and extends to the prevention of innate cytokine shock and to the promotion of adaptive immunity.
The development of antiviral immune responses involves the orchestration of a complex network of innate and adaptive immune cells to promote health over disease. Natural killer (NK) cells, plasmacytoid dendritic cells (pDCs), CD11b and CD8a conventional dendritic cells (cDCs), B cells, and CD8 T cells have all been demonstrated to be important for the generation of protective immunity to various viral infections [1] [2] [3] [4] [5] [6] . However, how the antiviral defense as a whole is coordinated, and in particular how the functions of different types of immune cells impact the shaping of the global immune response to viruses in vivo, is not thoroughly understood. The importance of an efficient NK cell response for the promotion of a favorable outcome to viral infections has been demonstrated in both mice and humans [2, 7] , and the rapid activation of NK cells after infection is a hallmark of their potency as innate immune system effectors [2] . Increasing evidence supports the idea that optimal coordination of immune responses involves an intricate relationship between NK cells and other innate leukocytes. For example, several reports have documented the importance of NK/DC ''crosstalk'' for the reciprocal activation of these cell types and in the promotion of antitumor immunity as recently reviewed [3, 8] . Others have shown the involvement of macrophages as an intermediary in the activation of NK cells via Va14i NK T cells [9] . NK cells have also been shown to have the capacity to interact with neighboring NK cells as well as T cells to stimulate cellular proliferation [10, 11] . pDCs were initially identified in humans and mice based on the unique ability of these cells to secrete enormous amounts of IFN-a/b early in response to viral challenge [12] . pDCs respond in this way as a consequence of their ability to recognize molecular signatures of viruses in a manner that is independent of pDC infection [12, 13] . Early during the course of murine cytomegalovirus (MCMV) infection, pDCs are the major producers of interferons a and b (IFN-a/b) [13, 14] , which are critical for host survival [15, 16] . IFN-a/b have direct antiviral effects because they can inhibit viral replication in infected cells as well as convey to noninfected cells defense mechanisms that protect them from viral infection. IFN-a/b also mediate a variety of immunoregulatory functions, either activating or inhibitory [17, 18] . They contribute to the proliferation of NK cells and induce them to express functional cytotoxic granules, promote cDC phenotypic and functional maturation, and can help to activate antiviral CD8 T cells. In contrast, depending on the context, IFN-a/b can also compromise the immune response of the host by inhibiting DC differentiation [19] or by directly leading to the attrition of CD8 T cells [20, 21] . Indeed, different viruses have recently been suggested to induce a profound immunosuppression in the host by inducing overwhelming levels of IFN-a/b [19] . The importance of the host genotypes in the efficiency of IFN-a/b antiviral functions in the context of the infection with specific viruses are well documented [22, 23] . In contrast, it is not known whether naturally occurring differences in the interactions between a virus and its hosts exist that may shape IFN-a/b responses by specifically dampening the immunosuppressive functions of these cytokines that are detrimental to the host and beneficial to the pathogen. The identification of such conditions will help to better understand the physiopathology of viral infection and may lead to the development of innovative treatments to fight these infections. In the early phase of MCMV infection there is a tight temporal relationship between the activation and exertion of effector functions between NK cells and pDCs because the production of IFN-a/b by pDCs promotes NK cell proliferation and cytotoxicity [13, 24] . However, the impact of NK cell functions on the pDC response to MCMV remains unexplored. The goal of this study was therefore to evaluate the effect of efficient NK cell responses on pDC IFN-a/b production and on the development of the adaptive immune responses of the host. C57BL/6 mice are able to mount NK cell responses that can control MCMV replication very efficiently, due to their expression of the Ly49H activating receptor, which endows NK cells with the ability to recognize and kill viral-infected cells expressing the viral ligand m157 (reviewed in [1] ). Although the NK cells from BALB/c and other strains of mice lacking the Ly49H gene are strongly activated for the acquisition of the cytotoxic machinery and for the production of IFN-c early after MCMV infection, they fail to recognize and kill infected cells and can therefore be considered inefficient. Inbred mouse strains that are resistant or sensitive to MCMV infection in an NK cell-dependent manner differ in many other immune parameters, including major histocompatibility (MHC) haplotype as well as DC subset frequency and function [25] . Therefore, rigorous evaluation of the impact of NK cell responses on pDC and adaptive responses requires comparison between mice of the same genetic background differing only in the ability of their NK cells to control the infection. For this, we took advantage of the fact that the Ly49H/m157 activation axis is sufficient for NK cell control of MCMV when introduced into the BALB/c genome [26] . The response of pDCs and CD8 T cells to MCMV were therefore compared between BALB/c mice (Ly49H À ) and animals on the BALB/c background that are congenic for the C57BL/6 natural killer complex and in particular for Ly49H (C.B6-Klra8 Cmv1-r /UwaJ mice, hereafter named Klra8 [27] ). This system thus provides a unique model to study antiviral immune responses in vivo in the context of both efficient (Ly49H þ ) and inefficient (Ly49H À ) NK cell activation, in the absence of any broad defect in NK cell development or activity that may affect the homeostasis or functions of other immune cells. This system allowed us to demonstrate, to our knowledge for the first time, that NK cells accelerate CD8 T cell responses against a viral infection in vivo. Moreover, we identified the underlying mechanismthe ability of NK cells to limit pDC IFN-a/b production to levels not immunosuppressive to the host-which itself results from the early control of viral replication. Ly49H triggering by m157 is required for the proliferation and accumulation of NK cells between days 3 and 7 after MCMV infection [28] . Therefore, we first used NK cell expansion as an indicator to show that Klra8 and BALB/c mice generate an NK cell response of different quality to MCMV infection ( Figure S1A ). We next compared the ability of Klra8 and BALB/c mice to control early viral load. Consistent with reported differences in viral load seen between mice lacking or possessing the Ly49H gene [26, 29, 30] , there was a substantially higher viral burden within the spleens of BALB/c mice when compared to those of congenic Klra8 animals ( Figure S1B ). In addition, NK cell depletion in Klra8 mice increased viral replication to levels similar to those observed in BALB/c animals ( Figure S1C ). Thus, these data confirm previous reports demonstrating that Klra8 mice mount a strong NK cell response that is able to efficiently control MCMV replication, in contrast to BALB/c animals. To evaluate the impact of efficient antiviral NK cell activity on the induction of IFN-a/b, we first examined the levels of the cytokines present in the serum of Klra8 and BALB/c mice. The serum levels of IFN-a/b were dramatically lower (100fold) at the peak of the response (day 1.5), and decreased at all To fight viral infections, vertebrates have developed a battery of innate and adaptive immune responses aimed at inhibiting viral replication or at killing infected cells. These responses include the early production of innate antiviral cytokines, especially interferons a and b (IFN-a/b), and the activation of cytotoxic lymphocytes such as the innate natural killer (NK) cells and the adaptive CD8 T cells. While critical for antiviral defense, cytokine or CD8 T cell responses can be detrimental or even fatal to the host when deregulated. Therefore, we need to better understand how the different arms of antiviral immunity are regulated. In particular, NK cells are proposed to play a protective role in a variety of viral infection in humans, but the underlying mechanisms remain poorly understood. Here, in a mouse model of cytomegalovirus infection, we demonstrate that NK cells prevent an excessive production of IFN-a/b and promote more efficient antiviral CD8 T cell responses. We thus show that NK cells can help promote health over disease during viral infections by regulating both innate and adaptive immune responses. It will be important to examine in humans whether NK cells control innate cytokine production to prevent immunopathology and to promote adaptive immunity against herpesviruses, HIV-1, influenza, or SARS. other times points examined, in Klra8 mice as compared to BALB/c mice ( Figure 1A) . The serum levels of other innate cytokines were also significantly lower in Klra8 mice, including IL-12p70 ( Figure 1A ) and TNF-a (unpublished). Similar results were observed when comparing serum cytokine levels between control-treated and NK cell-depleted Klra8 mice ( Figure S2 and unpublished data). Thus, while the systemic production of IFN-a/b and of other innate cytokines in response to MCMV infection is very high in susceptible animals, it is greatly reduced in the presence of an NK cell response that controls viral replication early and efficiently. In MCMV-infected 129 or C57BL/6 immunocompetent animals, most of the systemic production of IFN-a/b and a significant proportion of that of IL-12p70 comes from pDCs [13, 14, 31] , which are not productively infected [13, 32] , and does not come from infected cells ( [13] and unpublished data). To our knowledge, the contribution of pDCs to IFN-a/b or other innate cytokine production has not been assessed in animals on a BALB/c genetic background, which are the most susceptible to MCMV infection. cDCs infected with MCMV in vitro have been reported to produce high levels of IFN-a/b [33, 34] , and more generally, any virus-infected cell could theoretically produce these cytokines. Therefore, the increase in the systemic levels of IFN-a/b observed in BALB/c animals could result either from a high activation of pDCs or from a more significant contribution to this function from other cell types or from the high numbers of infected cells in these animals [33, 34] . Therefore, we next compared the activation of pDCs for IFN-a/b or IL-12p70 production between BALB/c and Klra8 animals by intracellular staining for these cytokines on ex vivo isolated splenocytes ( Figure 1B ). Less than 1% of the pDCs were found to produce IFN-a/b or IL-12 in Klra8 as compared to 10%-12% in BALB/c mice. Consistent with previous reports on 129 or C57BL/6 animals [14, 35] , pDCs were the sole producers of IFN-a/b at all the time points examined in BALB/c and Klra8 mice (unpublished data). Both pDCs and cDCs contributed to the production of IL-12 in Klra8 mice. However, the high IL-12p70 serum levels observed in BALB/c mice could solely be attributed to pDCs ( Figure 1B and unpublished data). Therefore, a very high activation of pDCs for innate cytokine production was observed in BALB/c mice, which was drastically reduced in congenic animals endowed with efficient antiviral NK cell functions. To further link the early viral load and the level of IFN-a/b production by pDCs, we performed infections of BALB/c and Klra8 mice with 5-fold serial dilutions of viral inoculums, leading to challenge doses from 10 3 to 1.25 3 10 5 pfu/mouse. Systemic IFN-a titers in these animals were measured by (B) Splenic leukocytes were isolated from Klra8 and BALB/c mice and analyzed for the frequency of IFN-a/b þ pDCs and IL-12 þ pDCs directly ex vivo. Numbers in dot plots represent the percent of IFN-a/b þ and percent of IL-12 þ cells within the total pDC population. One representative animal from groups of three mice for day 1.5 post-MCMV infection is shown. Graphs represent the total numbers of IFN-a/b þ pDCs and IL-12 þ pDCs present in the spleens of Klra8 and BALB/c mice on days 0, 1.5, 2, and 3 post-MCMV infection. Results are expressed as mean 6 SD of three mice per group and in ten thousands of cells. One experiment representative of three for (A, B) is shown. *p 0.05; **p 0.01; § ¼ not detected. doi:10.1371/journal.ppat.0030123.g001 ELISA on serum samples, and the numbers of pDCs producing IFN-a/b by flow cytometry ( Figure S2 ). The results clearly showed that the amount of IFN-a/b secreted by pDCs increases dramatically at high viral dose inoculums in Klra8 mice to levels similar to those observed in BALB/c animals. A 5-fold difference in the virus inoculums from 5 3 10 3 to 2.5 3 10 4 pfu/mouse is sufficient to switch on IFN-a/b production by pDCs in Klra8 mice from very low levels to nearly maximal levels similar to those observed in BALB/c animals. At low dose viral infection (10 3 pfu/mouse), BALB/c mice are still activated for nearly maximal pDC cytokine production. Thus, BALB/c and Klra8 animals are confirmed to exhibit dramatically different levels of pDC IFN-a/b production over more than a 10-fold range of viral inoculums. These data also show that the pDCs of Klra8 mice are able to produce high levels of IFN-a/b under conditions of stimulation with high viral inoculums. Moreover, NK cell depletion in Klra8 mice infected with low virus inoculums was also shown to lead to a dramatic increase in pDC IFN-a/b production, almost to the levels observed in BALB/c animals ( Figure S2 ). Altogether, these data thus show that the ability of the host to control MCMV replication early through NK cell responses limits the activation of pDCs and therefore prevents the induction of very high systemic levels of IFN-a/b and IL-12. These data also demonstrate that the pDCs of Klra8 mice are not intrinsically deficient for IFN-a/b production in response to MCMV infection. High systemic levels of IFN-a/b production have been shown to lead to the ablation of cDCs [19] and to the attrition of both antigen-specific and bystander CD8 T cell populations during infection with lymphocytic choriomeningitis virus (LCMV) [20, 21] . We therefore hypothesized that the high systemic levels of IFN-a/b induced during MCMV infection in the context of inefficient NK cell activity could lead to the ablation of cDCs and may impinge on the development of antiviral CD8 T cell responses. Indeed, it has been previously reported that MCMV infection leads to a severe ablation of CD8a cDCs in BALB/c mice that is prevented by efficient antiviral NK cell functions in C57BL/ 6 animals [36] . However, the cause for this ablation of cDCs in BALB/c animals has not been identified, and the effect of this ablation on the development of antiviral CD8 T cells has not been evaluated. Thus, we next compared the numbers of DC subsets and the activation of antiviral CD8 T cells at different time points after infection between Klra8 and BALB/c mice. We first confirmed that pDC numbers remained at homeostatic levels in BALB/c mice throughout the early phase of MCMV infection, while a severe ablation of cDCs occurred for both the CD11b and CD8a subsets ( Figure 2 ). In contrast, in the presence of an efficient NK cell response, in Klra8 mice, both the cDC and pDC compartments were preserved, consistent with the previous observations reported in C57BL/6 animals [36] . We next compared the kinetics of CD8 T cell expansion between Klra8 and BALB/c mice ( Figure 3A ). We observed a significant increase in total CD8 T cells in the spleen at day 4 post-infection in Klra8 mice but only later in BALB/c animals. We investigated whether this accelerated expansion of CD8 T cells was a result of a faster activation of antiviral effector CD8 T cells. We first analyzed total splenic CD8 T cells for the expression of an O-glycosylated, activation-associated isoform of CD43. The 1B11 monoclonal antibody (mAb) specifically recognizes this isoform of CD43 on CD8 T lymphocytes [37] and identifies the subset of effector cells capable of IFN-c production and potent cytolytic activity ex vivo during the acute phase of a variety of viral infections [38] [39] [40] [41] [42] . We observed the appearance of CD43 hi effector CD8 T cells starting at day 4 post-infection in Klra8 mice but only later in BALB/c animals ( Figure 3B ). We then used MHC class I tetramers loaded with a peptide derived from the IE-1 protein of MCMV in order to quantify the subset of antiviral CD8 T cells specific to this epitope ( Figure 3C ). We observed a faster expansion of the anti-IE-1 CD8 T cells in Klra8 mice, similar to what was observed for total or effector CD8 T lymphocytes. Similar results were obtained when analyzing the CD8 T cell response against another viral peptide, derived from protein m164 (unpublished data). Conversely, and in Splenic leukocytes were isolated from Klra8 and BALB/c mice and analyzed for the frequency of pDCs (120G8 þ CD11c int ), CD11b cDCs (120G8 À CD11c hi CD8a À ), and CD8a cDCs (120G8 À CD11c hi CD8a þ ) within the DX5 À and TCRb À population. Numbers in dot plots represent percent pDCs, percent CD11b cDCs, and percent CD8a cDCs within the total splenocyte population for one representative animal from groups of three mice for days 0 and 2 post-MCMV infection. Graphs represent the total numbers (in millions) of pDCs, CD11b cDCs, and CD8a cDCs present in the spleens of Klra8 and BALB/c mice on days 0, 1.5, 2, and 3 post-MCMV infection. Results are expressed as mean 6 SD of three mice per group. One experiment representative of three is shown. *p 0.05; **p 0.01. doi:10.1371/journal.ppat.0030123.g002 accordance with what has been previously reported in C57BL/ 6 mice depleted of NK cells by antibody treatment [43] , we observed accumulation of effector CD8 T cells in BALB/c mice at later time points after challenge. In order to investigate the effector potential of the MCMVspecific CD8 T cells observed at day 4 post-infection in Klra8 mice, we measured different parameters associated with the protective functions of antiviral CD8 T cells, namely their proliferation, by monitoring the expression of the marker Ki-67 ( Figure 4A ), their capacity to produce IFN-c in response to antigen-specific restimulation in vitro using intracellular staining ( Figure 4B ) and ELISPOT ( Figure S3 ), and their cytotoxic potential by measuring antigen-specific target cell killing in vivo ( Figure 4C ). For each of these parameters, we observed an accelerated acquisition of effector functions by CD8 T cells from Klra8 mice when compared to BALB/c animals. As Klra8 and BALB/c mice differ for the whole NK cell locus, which includes genes expressed in several leukocyte subsets, we sought to ensure that the differences observed in the activation kinetics of CD8 T cells between these two mouse strains i) were not due to intrinsic differences in the reactivity of CD8 T cells and ii) were dependent on Ly49Hmediated NK cell activation. This was achieved by showing that no accelerated expansion of CD8 T cells occurs in Klra8 animals i) when the antibacterial responses of Klra8 and BALB/c mice are compared during Listeria monocytogenes infection ( Figure S4 ), or ii) in response to MCMV infection when Klra8 mice are depleted of NK cells ( Figure 5A ), or when the CD8 T cell responses of Klra8 and BALB/c mice are compared during infection with a Dm157 virus as opposed to infection with wild-type or revertant viruses ( Figure 5B ). Moreover, the impact of efficient NK cell activity on the development of antiviral CD8 T cell responses was confirmed in mice of another genetic background, as a delay in antiviral CD8 T cell activation was observed in B10.D2 animals deficient for Ly49H-mediated NK cell activation due to the inactivation of the associated adaptor molecule DAP12 (B10.D2.DAP128), as compared to wild-type controls ( Figure 5C ). Altogether, these data show that efficient NK cell responses to a viral infection accelerate the development of effector antiviral CD8 T cell responses. To determine whether the dramatic reduction of pDC IFN-a/b production by NK cell functions in Klra8 mice could in part account for the ability of these mice to preserve an intact cDC compartment and to mount early CD8 T cell responses, we examined whether exogenous administration of the cytokines in these animals could directly impact the initiation of cDC and CD8 T cell responses despite efficient NK cell-mediated control of viral load. MCMV-infected Klra8 mice were treated with recombinant mouse IFN-a as described in Materials and Methods, and their DC or CD8 T cell compartments were compared to those of controltreated Klra8 and BALB/c mice ( Figure 6A ). The treatment of Klra8 mice with IFN-a indeed led to a drastic reduction in both CD11b and CD8a splenic cDCs, but not in pDCs, when compared to their control-treated Klra8 counterparts. Likewise, IFN-a treatment induced a delay in the expansion of total, CD43 hi , and IE-1 Tet þ CD8 T cell numbers ( Figure 6B ). Of note, the IFN-a treatment administered to the Klra8 mice did not significantly change the level of viral replication in these animals. Very low but detectable levels of infectious viral particles were observed in both control-treated and IFN-a-injected Klra8 mice (1.9460.35 versus 1.9260.11 log pfu/spleen), which contrasted sharply with the high viral replication observed in BALB/c animals (5.260.04 log pfu/ spleen). These results demonstrate that exogenous injection of IFN-a in Klra8 mice is sufficient to decrease the numbers of cDCs and to ablate early antiviral CD8 T cell responses, and, therefore, that excessive levels of IFN-a can have a direct negative impact on antiviral immune cell responses in a manner that is independent of the level of viral replication in the host. The possibility remains that other innate cytokines, such as IL-12 and TNF-a, which are produced at much higher levels in BALB/c as compared to Klra8 mice, may also bear some contribution to this function, in a synergistic or redundant manner with IFN-a/b. In any case, our data strongly suggest that the ability of Klra8 mice to preserve an intact cDC compartment and to mount early CD8 T cell responses is in part due to their ability to control viral replication very early without the need for the host to produce high systemic levels of IFN-a/b. Altogether, our data thus identify how naturally occurring differences in the interactions between a virus and its host can tilt the balance between the various functions of IFN-a/b, and eventually other innate cytokines, towards conditions promoting the induction of early adaptive immunity, rather than the development of a state of transient immunosuppression. We next aimed at further understanding the link between NK cell functions, IFN-a/b production, and downstream effects on DCs and CD8 T cells. As BALB/c mice have a much higher viral burden after infection with MCMV, it seemed plausible that the difference in viral load between Klra8 and BALB/c mice could have an impact on the intensity of the DC and CD8 T cell responses. In order to test the impact of the extent of viral replication in vivo on pDC cytokine production and on antiviral CD8 T cell activation, independently of the function of NK cells, we sought a strategy to control viral replication efficiently in BALB/c mice with a tightly controlled timing and magnitude. To achieve this, we utilized an MCMV (DN-SCP-MCMV) that is genetically engineered so that its in vivo replication can be effectively arrested by the administration of doxycycline [44] . Using this system, we were able to design a doxycycline treatment protocol which, in BALB/c mice, closely mimics the extent and the kinetics of efficient NK cell-mediated control of viral replication observed in Klra8. Indeed, under the experimental conditions selected, the viral burdens in Klra8 mice and in DN-SCP-MCMV-infected, doxycycline-treated BALB/c mice were comparable at the different time points examined, and much lower than in untreated DN-SCP-MCMV-infected BALB/c animals ( Figure 7A ). The reduction in viral burden in doxycycline-treated DN-SCP-MCMVinfected BALB/c animals led to a significant, strong decrease in the serum levels of the pDC-derived cytokines IFN-a and IL-12p70 ( Figure 7B ) and allowed the maintenance of cDCs ( Figure 7C ). Moreover, we observed an accelerated activation Figure 7D ). The treatment of DN-SCP-MCMV-infected Klra8 mice with doxycycline had no effect on any of the parameters tested (unpublished data). Thus, our data indicate that, in Klra8 mice, the ability of NK cells to dampen innate cytokine production by pDCs and thus to promote optimal conditions for the initiation of antiviral cDC and CD8 T cell responses result from their exquisite capacity to control viral replication early and efficiently after infection. The results from this study demonstrate that efficient NK cell responses promote the accelerated generation of effector antiviral CD8 T cells during infection in vivo, in part by preventing the generation of very high, immunosuppressive levels of antiviral cytokines. Thus, our study demonstrates that NK cells can serve not only as effector lymphocytes but also as a regulator of immune system function for defenses against viral infections. Therefore, our data bring important and original advances to the understanding of the contributions NK cells make to immunity against infectious disease, by demonstrating that i) they control the functions of another critical player in the innate antiviral responses, the pDC, and modulate the cytokine milieu induced early after challenge, and ii) they accelerate the generation of effector antiviral CD8 T cells. Other mechanisms through which NK cell activities can modulate immune responses to pathogens have been summarized in several recent reviews [8, 45, 46] and include in vivo i) the regulation of the homeostasis and of the maturation of DCs and ii) the prevention of a detrimental persistence of the activation of the CD8 T cells at later time points during the immune response. Very recently, it was also shown that perforin-mediated NK cell killing down-modulates the activation of macrophages and prevents the development of a hemophagocytic lymphohistiocytosis-like syndrome during MCMV infection [47] . Our results demonstrate that NK cells can dramatically decrease the intensity and duration of pDC activation by controlling viral burden, which prevents the production of very high systemic levels of IFN-a/b, and eventually other innate cytokines, that can have detrimental effects for the host. We show that this mechanism also protects against the MCMV-mediated loss of splenic cDCs [36] . It has been reported that that both measles virus and LCMV can exploit the host's IFN-a/b response to inhibit cDC development and drive cDC loss in vivo [19] . Our results suggest that MCMV also can induce high production of IFN-a/b to promote its own survival by ablating cDCs and delaying the activation of antiviral effector CD8 T cells. In light of this observation, it is interesting to note that MCMV has developed strategies to actively counteract the antiviral responses to IFN-a/b or IFNc within infected cells [48] , as opposed to the mechanisms employed by negative-strand RNA viruses, which act to shut down the production of these cytokines by infected cells [49] or pDCs [50] . Indeed, even though complete deficiency in IFN-a/b responses is associated with a dramatic increase in the susceptibility of mice to MCMV infection [16] , it clearly appears that the benefit of high level IFN-a/b production for the host is less than that brought by an efficient NK cell response (since Klra8 mice show viral titers that are 1,000fold lower than those seen in BALB/c mice, even though Klra8 mice produce 100-fold less IFN-a/b). Thus, it is tempting to speculate that the efficient NK cell activity driven by the Ly49H activating receptor and its ability to dampen pDC IFN-a/b production and to promote adaptive immunity is a direct host countermeasure to the subversion of the IFN-a/b response by MCMV. Altogether, our results suggest that the NK cell response governs the balance between the positive and negative effects of IFN-a/b, and eventually other innate cytokines, for the optimal orchestration of the immune response to MCMV. Our data indicate that efficient NK cell activity contributes to the adaptive arm of the immune response to MCMV by promoting the accelerated expansion of antigen-specific CD8 T cells. Like the contribution NK cells make to the maintenance of the cDC compartment, our data support a role for NK cell control of pDC IFN-a/b production for the promotion of antigen-specific CD8 T cell expansion. IFN-a/b can have both positive and negative effects on CD8 T cell responses [20, 21, [51] [52] [53] . For example, the optimal expansion of CD8 T cells in response to LCMV infection has been demonstrated to be dependent on the ability of the CD8 T cell compartment to receive IFN-a/b-mediated signals [53] . However, within the same viral system, IFN-a/b also drives the early attrition of both antigen-specific and bystander CD8 T cell populations [20, 21] . Here, we show that the delay in the expansion of antiviral effector CD8 T cells in the absence of an efficient NK cell response to MCMV occurs within the context of excessive production of IFN-a/b by pDCs. Moreover, we demonstrate that exogenous administration of IFN-a can override the ability of efficient NK cell activity to promote the accelerated expansion of functional antiviral CD8 T cells. The IFN-a-mediated effects on CD8 T cell expansion could be direct or a downstream consequence of the loss of cDCs. However, our results suggest a direct activity of IFN-a/b on CD8 T cells, as the administration of IFN-a leads to the complete disappearance of early antiviral CD8 T cells but only to an incomplete decrease in cDC numbers. Although not the focus of this study, our data also show that delayed CD8 T cell responses reach higher levels and are sustained for longer periods of time in the absence of efficient antiviral NK cell activity, which is consistent with previous observations [43] . This extended activation of the T cell compartment is required for MCMV control under these conditions and likely results from the poor ability of the innate immune system to control the virus early. One could hypothesize that this could also lead to chronic inflammation and long-term detrimental effects for the host, including the increased susceptibility of BALB/c mice to MCMV-induced T cell-dependent autoimmune diseases such as myocarditis [54] . Thus, efficient NK cell responses could be beneficial to the host not only by early direct antiviral effects but also by reducing the degree of antiviral T cell activation required for later control of viral replication and therefore the indirect costs this may bear for health. To our knowledge, this study is the first to demonstrate how naturally occurring, genetically determined differences in the interactions between a virus and its host can tilt the balance between the various functions of IFN-a/b towards conditions promoting the induction of early adaptive immunity rather than towards the development of a state of transient immunosuppression. Our data also confirm that NK cell antiviral functions prevent the generation of a chronic state of CD8 T cell activation later during the infection that may otherwise lead to immunopathology. A key question for future studies will be to determine how general this function of NK cells is for antiviral defense. Such mechanisms may also be in place during HIV infection where genetic [55] or functional [56] [57] [58] evidences suggest a role in the variation of NK cell functions for resistance or susceptibility to the development of AIDS, where detrimental effects of excessive levels of IFN-a/b [59] or of pro-inflammatory cytokines [60, 61] have been recently suggested, and where the initiation of antiviral CD8 T cell responses appears delayed [62] [63] [64] . This could also be the case for infections with highly pathogenic strains of influenza, as excessive production of pro-inflammatory cytokines have been shown to occur and proposed to play a crucial role in immunopathology [65, 66] , while NK cells can confer resistance through recognition of infected cells by the NKp46 receptor [67] . Very recently, researchers have demonstrated that human NK cells are able to recognize DCs infected with the influenza or Ebola viruses [68, 69] , while a correlation between unregulated IFN-a/b responses and a malfunction of the switch from innate to adaptive immunity has been reported during fatal SARS [70] , which further highlight the potential relevance of our observations to a variety of viral infections. Thus, it will be important to determine the impact of NK cell function on the modulation of innate and adaptive immunity and on the development of immunopathology in other models of viral infection. Moreover, our results imply that in individuals susceptible to such viral infections, pharmacological control of viral replication very early should benefit them not only by directly limiting viral cytopathogenicity, but also by establishing conditions better suited to the development of balanced, protective immunity, since this is the case in mice susceptible to MCMV infection when viral replication is controlled by drug treatment. A role for NK cells in the generation of antitumor CD8 T cell responses has been linked to their ability to increase inflammation via secreting IFN-c and promoting IL-12 production by DCs [71, 72] . Collectively, these studies and ours reveal a role for NK cells as mediators of an ''innate cytokine balance'' for the optimal generation of the immune response, whereby NK cells are required to produce cytokines to increase inflammation when it is intrinsically low (during tumor development) and to prevent excessive production of innate cytokines when it can be intrinsically high (in the context of pathogenic encounters). Of note is the apparent requirement, in both instances, for NK cells to engage in cognate receptor-mediated interactions to naturally develop immunoregulatory functions. For example, during the response to MCMV, the systemic activation of NK cells by IL-12 to produce IFN-c in a Ly49H À environment is not sufficient to promote their control of viral replication and their associated immunoregulatory functions. These cognate interactions could also enable NK cells to act as direct regulators of antiviral immunity, as already illustrated for the generation of antitumor CD8 T cell responses. During the response to MCMV, Ly49H expression by NK cells dramatically increases the efficiency of recognition and killing of infected cells. Ly49H may also enable NK cells to deliver IFNc in a proper place and time to help the priming of antiviral CD8 T cells in infected mice through a tripartite interaction with MCMV-infected m157-expressing DCs. The importance of these cognate interactions for the promotion of efficient NK cell functions may differ depending on the tissues and on the target cell types, as exemplified by the heightened requirement for Ly49H-dependent mechanisms for MCMV control in the lungs [73] . In conclusion, we demonstrate that efficient NK cell activity contributes to the optimal orchestration of innate and adaptive immunity during the course of a viral infection in vivo. The implications of these findings extend to the design of therapeutic strategies to fight human disease in two ways, as they emphasize i) that the benefits of innate cytokine production for immune cell recruitment and activation can be outweighed by detrimental effects that can directly reduce the capacity of the host to fight infection, as well as ii) the need to better understand and take into account the complex interactions of relatively rare cell types that occur in a physiological context for the purpose of exploiting the human immune system to promote health over disease. Mice. BALB/c mice (Charles River Laboratories, http://www.criver. com/) were purchased for use in these studies. C.B6-Klra8 Cmv1-r /UwaJ, referred to as Klra8, (Jackson Laboratory, http://www.jax.org/), B10.D2, and B10.D2-DAP12-deficient (DAP128) animals were bred in pathogen-free breeding facilities at the Centre d'Immunologie de Marseille-Luminy (Marseille, France). Experiments were conducted in accordance with institutional guidelines for animal care and use. Protocols have been approved by the French Provence ethical committee (number 04/2005) and the US Office of Laboratory Animal Welfare (assurance A5665-01). In vivo treatment protocols. Stocks of wild-type Smith Strain MCMV salivary gland extracts [14] and bacterial artificial chromosome (BAC)-derived wild-type, DN-SCP-MCMV [44] , Dm157-MCMV, and m157-revertant were prepared as previously described [73] . Infections were initiated on day 0 with the i.p. delivery of 5 3 10 3 PFU (for in vivo-derived virus) or 2 3 10 5 PFU (for in vitro-derived virus). These modest doses were chosen because they do not induce lymphopenia in infected animals who harbor leukocyte numbers equal to or greater than those of uninfected controls (not shown). In addition, these doses are likely to be closer to the physiologic doses reached upon natural exposure to the virus through contact with infected animals. For the in vivo arrest of DN-SCP-MCMV replication, 200 lg of doxycycline was delivered by i.p. injection 20 h postinfection followed immediately by the addition of 2 mg/ml doxycycline plus 5% sucrose in the drinking water. 50,000 units of recombinant mouse IFN-a (HyCult Biotech, http://www.hbt.nl/) was administered by i.p. injection at 30 and 48 h post-MCMV infection. This dose was titrated in ELISA and shown to correspond to 25 ng of the ELISA IFN-a standard. Thus, since the cytokine titers measured by ELISA in the serum of BALB/c mice at 36 h post-infection range between 5 and 10 ng/ml, and the volume of the lymph and blood of an adult mouse can roughly be estimated around 5 ml, the dose of rmIFN-a injected in Klra8 mice should be similar to the physiologic levels of the cytokine naturally induced in infected BALB/c mice. NK cells were depleted by delivery of 100 lg of purified anti-NK1.1 mAb (PK136) on days À1 pre-MCMV infection, and days 1 and 3 post-MCMV infection. Control mice were treated with 100 lg of mouse IgG (Jackson ImmunoResearch Laboratories, http://www. jacksonimmuno.com/). Uninfected mice were depleted of NK cells on the schedule of day 4 infected mice. To assess NK cell depletion, a combination of anti-DX5 and anti-TCRb staining was used in order to avoid the risk of underevaluating the numbers of remaining NK cells, as may have occurred due to a potential problem of epitope masking if the antibody used for immunophenotyping had been the same as the one used for depletion. NK cell depletions were greater than 98% ( Figure S5) . Isolation of lymphocytes. For the analysis of CD8 T cell populations, spleens were minced, passed through nylon mesh and washed in PBS, 5 mM EDTA, and 3% FCS (PBS/EDTA/FCS). For the analysis of DC populations, spleens were digested by collagenase (liberase CI; Boehringer Mannheim, http://www.roche.com/) and teased apart by repeating pipeting in PBS/EDTA/FCS. In both protocols, erythrocytes were osmotically lysed by ammonium chloride treatment. Thereafter, cell suspensions were kept in PBS/ EDTA/FCS unless specified otherwise. Total live splenocytes were counted by trypan blue exclusion using a hemocytometer. Total numbers of specific leukocyte subsets were calculated for each individual mouse as (percent of these cells in the live gate of total splenocytes) 3 (total live splenocyte numbers). CD8 T cell stimulation for IFN-c production. For intracellular IFNc detection, splenic lymphocytes were isolated and then incubated with IE-1 ( 168 YPHFMPTNL 176 ) or m164 peptide ( 257 AGPPRYSRI 265 ) (10 À7 M) pulsed P815.B7 cells at a 10:1 ratio for 6 h with brefeldin A (Sigma-Aldrich, http://www.sigmaaldrich.com/) added for the last 3 h of culture. Cells were then harvested and analyzed for CD8a, CD43, and intracellular IFN-c protein by three-color staining followed by flow cytometry. In vivo cytotoxicity assay. Antigen-specific CD8 T cell-mediated cytotoxicity was assayed as described in [74] . Briefly, splenocytes from naive mice were costained with PKH26 (Sigma-Aldrich) and either 1 lM, 100 nM, or 1 nM CFSE (Molecular Probes, http://probes. invitrogen.com/). Labeled cells were then pulsed with the indicated peptides, mixed in equal ratios, then transferred i.v. (5 3 10 6 total cells) into the indicated groups of mice. Lymphocytes were isolated from the spleens of recipient mice 4 h post-transfer and analyzed for PKH26 (all transferred cells) and level of CFSE expression (unpulsed or peptide-pulsed cells). Percent killing within the PKH26 þ gate was calculated by: 100 À ([(% peptide-pulsed in infected/% unpulsed in infected)/(% peptide-pulsed in uninfected/% unpulsed in uninfected)] 3 100). Quantification of viral titers and serum cytokine levels. Spleens were homogenized [14] , and viral titers were determined by plaque assay using mouse embryonic fibroblasts with centrifugal enhancement or NIH-3T3 cells [73] . Serum was collected at the indicated time points and cytokine levels were determined by ELISA for IFN-a (PBL Biomedical Laboratories, http://www.interferonsource.com/) and IL-12p70 (R&D Systems, http://www.rndsystems.com/) per the manufacturer's instructions. Abs and reagents. DX5-PE, CD43-PE (clone 1B11), CD11c-PE (clone HL3), CD8a-PerCp (clone 53-6.7), TCR-b-allophycocyanin (clone H57-597), IFN-c-allophycocyanin (clone XMG1.2), IL-12allophycocyanin (clone C15.6), and streptavidin-PE were purchased from BD Pharmingen (http://www.bdbiosciences.com/). CD8a-FITC (clone CT-CD8a) was purchased from Caltag (http://www.caltag.com/). Ki-67-FITC (clone MM1) was purchased from Novocastra (http://www. vision-bio.com/). Purified rat anti-mouse IFN-a (clone F18 and RMMA-1) and anti-IFN-b (clone RMMB-1) were purchased from TEBU-Bio (http://www.tebu-bio.com/). 120G8 mAb was provided by Schering-Plough (http://www.schering-plough.com/) and conjugated to Alexa Fluor-488 using a kit from Molecular Probes. Isotype controls for each mAb were purchased from the appropriate manufacturer. The following tetramers conjugated to PE were obtained through the NIH Tetramer Facility (Atlanta, Georgia, United States): H-2L(d)/IE-1( 168 YPHFMPTNL 176 ) and H-2D(d)/ m164( 257 AGPPRYSRI 265 ). The above-mentioned reagents were used for FACS analysis in this study. Flow cytometric analysis. Cells were first incubated with 2.4G2 mAb for 20 min. Cells were then stained with mAbs specific for cell surface markers or isotype controls for 30 min at 4 8C. Cells were then washed and fixed in 2% paraformaldehyde in PBS. Intracellular staining for Ki-67, IFN-c, and IL-12 was performed using the Cytofix/ Cytoperm kit (BD Pharmingen). Intracellular staining for IFN-a/b was performed as previously described [35] . Depending on the experiments, 2.5 3 10 5 to 2 3 10 7 events were collected on a FACSCalibur. The data were acquired and analyzed using CellQuest software (BD Biosciences). Isotype controls were used to set gates for the presented FACS analyses. Statistical analyses. Statistical analyses were performed in Microsoft Excel 5.0 (Microsoft Corporation, http://www.microsoft.com/) using Student's two-tailed t tests. Mean 6 standard deviation (SD) was calculated for each graph. If visually absent, error bars are too small to be depicted based on the scale of the y-axis.